WO2025255673A1

WO2025255673A1 - An automated ai-based wafer bonder for microleds

Info

Publication number: WO2025255673A1
Application number: PCT/CA2025/050820
Authority: WO
Inventors: Gholamreza Chaji; John Cronin; Michael D'andrea
Original assignee: Vuereal Inc
Current assignee: Vuereal Inc
Priority date: 2024-06-14
Filing date: 2025-06-13
Publication date: 2025-12-18
Anticipated expiration: 2026-12-14

Abstract

The present disclosure provides an automated wafer bonding of LED substrates system in which an automated wafer bonding inspection module is used to determine the quality of incoming processing material and an automated wafer bonding process module is used to process the processing material, an automated wafer bonding post-process module is used to inspect the resultant process material for acceptability, and an automated wafer bonding historical module is used to analyze the post process results of the most recent process, against all historical data to find enhanced process changes in the automated wafer bonding process module and an automated wafer bonding integration module to update the automated wafer bonding process module based upon the results of the automated wafer bonding historical module.

Description

AN AUTOMATED AI-BASED WAFER BONDER FOR MICROLEDS

BACKGROUND AND FIELD OF THE INVENTION

[1] The present disclosure is generally related to an automated wafer bonding process of a microLED chips system.

[2] In the wafer bonding process, numerous parameters can significantly impact the outcome. Residues or particulate matter lingering on the wafer surface pose a significant challenge, as they can obstruct the bonding process, resulting in weak bonds or defects within the bonded interface. Additionally, the roughness of the wafer surfaces must be meticulously controlled. Excessive roughness may hinder intimate contact between the wafers, whereas insufficient roughness can impede the mechanical interlocking required for certain bonding techniques. Moreover, issues stemming from uneven heating can arise due to suboptimal control of ramp-up and holding temperatures within the wafer bonder. This irregular heating may lead to non-uniform force application, creating weak spots along the bond interface. Furthermore, rapid or uneven cooling rates following the bonding process can induce thermal stress, potentially causing wafer cracking, exacerbated by inadequate sensor feedback. Additionally, excessive force during the release phase can inflict damage on the wafers, a consequence often attributable to sensor sensitivity issues or feedback system delays. Lastly, the presence of weak points within the bond interface poses a critical concern. These areas, characterized by insufficient adhesion, can manifest as micro-voids, incomplete fusion of bonding materials, or regions with diminished mechanical strength. Such weaknesses compromise the overall integrity and functionality of the microLED device. Accurately detecting these weak points is paramount for ensuring the production of high-quality and reliable microLED displays. Thus, there is a need in the prior art for an automated wafer bonding process of microLED chips system.

SUMMARY

[3] The present invention relates to an automated wafer bonding (WB) method for microLED chips system, comprising, an automated wafer bonding inspection module that measures at least one parameter from a first group of parameters and calculates an output of at least one parameter from a second group of parameters to help control the automated wafer bonding processing module, an automated wafer bonding processing module which inputs the at least one parameter from said second group to control the wafer bonding using at least one parameter from a third group of parameters, an automated wafer bonding post-processing module which measures at least one parameter from a fourth group of parameters that determines how well the parameters of said first and second group were running, an automated wafer bonding historical module which calculates a plurality of group parameters based upon parameters of the first group, second group, third group, and fourth group and an automated wafer bonding integration module to update the automated wafer bonding processing module based upon the results of the automated wafer bonding historical module

BRIEF DESCRIPTIONS OF THE DRAWINGS

[4] FIG. 1 : Illustrates an automated wafer bonder for microLEDs, according to an embodiment.

[5] FIG. 2: Illustrates a Base Module, according to an embodiment.

[6] FIG. 3 : Illustrates a WB Inspection Module, according to an embodiment.

[7] FIG. 4: Illustrates a WB Process Module, according to an embodiment.

[8] FIG. 5: Illustrates a WB Post-Process Module, according to an embodiment.

[9] FIG. 6: Illustrates a WB Historical Module, according to an embodiment.

[10] FIG. 7: Illustrates a WB Integration Module, according to an embodiment.

[11] FIG. 8: Illustrates a WB Inspection Database, according to an embodiment.

[12] FIG. 9: Illustrates a WB Process Database, according to an embodiment.

[13] FIG. 10: Illustrates a WB Post-Process Database, according to an embodiment.

DETAILED DESCRIPTION

[14] Embodiments of the present disclosure will be described more fully hereinafter with reference to the accompanying drawings in which like numerals represent like elements throughout the several figures, and in which example embodiments are shown. Embodiments of the claims may, however, be embodied in many different forms and should not be construed as limited to the embodiments set forth herein. The examples set forth herein are non-limiting examples and are merely examples among other possible examples.

[15] FIG. 1 illustrates a system for an automated wafer bonder for microLEDs. This system comprises a 3rd party wafer bonding Al network 102 which integrates pre-quality inspection, processing, post-processing, historical analysis, and continuous improvement in a wafer bonding system. The WB inspection module 110 measures essential parameters and calculates outputs to guide the subsequent wafer bonding process. The WB process module 112 acts as the core controller, utilizing inputs from the WB inspection module 110 to govern the wafer bonding process to ensure optimal heating profiles, bonding force application, cooling rate, release force, bond uniformity, bond alignment, etc. The WB post-process module 114 conducts inspections to assess the execution of parameters and identify potential deviations. The WB historical module 116 analyzes data collected across various stages, allowing for continuous learning and pattern recognition. The WB integration module 118 then employs historical insights to update and optimize the process, enabling adaptive improvements over time.

[16] Further, embodiments may include a communication interface 104, which may be a hardware or software component that enables communication between the 3rd party wafer bonding Al network 102 and the wafer bonding pre-process 128, wafer bonding process 134, and wafer bonding post-process 150. The communication interface 104 may include a set of protocols, rules, and standards that define how information is transmitted and received between the devices. The communication interface 104 may be a physical connector, wireless network, or software application and may include components such as drivers, software libraries, and firmware that may be used to control and manage the communication process. In some embodiments, the communication interface 104 may be compatible with USB, Bluetooth, or Wi-Fi. The communication interface 104 may communicate with a network. Examples of networks may include but are not limited to, the Internet, a cloud network, a Wireless Fidelity (Wi-Fi) network, a Wireless Local Area Network (WLAN), a Local Area Network (LAN), a telephone line (POTS), Long Term Evolution (LTE), and/or a Metropolitan Area Network (MAN).

[17] Further, embodiments may include a memory 106, which may include suitable logic, circuitry, and/or interfaces that may be configured to store a machine code and/or a computer program with at least one code section executable by a processor. Examples of implementation of the memory 106 may include, but are not limited to, Random Access Memory (RAM), Read Only Memory (ROM), Hard Disk Drive (HDD), and/or a Secure Digital (SD) card.

[18] Further, embodiments may include a base module 108, which initiates the WB inspection module 110, the WB process module 112, the WB post-process module 114, the WB historical module 116, and the WB integration module 118.

[19] Further, embodiments may include a WB inspection module 110, which begins by being initiated by the base module 108. The WB inspection module 110 connects to the wafer bonding pre-process 130. The WB inspection module 110 collects the data from the wafer bonding pre- process 130. The WB inspection module 110 performs the inspection algorithm. The WB inspection module 110 sends the determined process data from the inspection algorithm to the WB process module 112. The WB inspection module 110 stores the data in the WB inspection database 120. The WB inspection module 110 returns to the base module 108.

[20] Further, embodiments may include a WB process module 112, which begins by being initiated by the base module 108. The WB process module 112 receives the process data, such as the data on the microLED processed wafer, from the WB inspection module 110. The WB process module 112 connects to the wafer bonding process 140. The WB process module 112 collects the data from the wafer bonding process 140. The WB process module 112 performs the control algorithm. The WB process module 112 adjusts the control parameters of the wafer bonding process 140. The WB process module 112 executes the wafer bonding process. The WB process module 112 stores the data in the WB process database 122. The WB process module 112 returns to the base module 108.

[21] Further, embodiments may include a WB post-process module 114, which begins by being initiated by the base module 108. The WB post-process module 114 connects to the wafer bonding post-process 150. The WB post-process module 114 collects the post-process data. The WB postprocess module 114 extracts the data from the WB inspection database 120 and WB process database 122. The WB post-process module 114 performs the post-process algorithm. The WB post-process module 114 performs the post-process inspection. The WB post-process module 114 stores the data in the WB post-process database 124. The WB post-process module 114 returns to the base module 108. [22] Further, embodiments may include a WB historical module 116, which begins by being initiated by the base module 108. The WB historical module 116 connects to the wafer bonding pre-process 130, the wafer bonding process 140, and the wafer bonding post-process 150. The WB historical module 116 aggregates the data from the various types of processes and stores the data in the WB network database 126. The WB historical module 116 performs the historical machine learning algorithm on the historical data stored in the WB network database 126. The WB historical module 116 sends the process adjustments to the WB inspection module 110, the WB process module 112, and the WB post-process module 114. The WB historical module 116 returns to the base module 108.

[23] Further, embodiments may include a WB integration module 118, which begins by being initiated by the base module 108. The WB integration module 118 performs the integration machine learning algorithm. The WB integration module 118 connects to the wafer bonding pre- process Al module 136, the wafer bonding process Al module 146, and the wafer bonding postprocess Al module 156. The WB integration module 118 sends the process adjustments to the wafer bonding pre-process Al module 136, the wafer bonding process Al module 146, and the wafer bonding post-process Al module 156. The WB integration module 118 returns to the base module 108.

[24] Further, embodiments may include a WB inspection database 120, which provides an example of the results of the inspection algorithm performed in the WB inspection module 110. The WB inspection database 120 contains the wafer number or ID, the defect density before the cleaning process, the defect density after the cleaning process with the parameters optimized by the inspection algorithm, the delta or effectiveness of the cleaning process, etc. In some embodiments, the WB inspection database 120 may include other parameter data collected or calculated during the WB inspection module 110 process, such as surface contamination, surface roughness, wafer bow and warp, thermal expansion mismatch, chemical incompatibility, alignment accuracy, void formation, electrical issues, etc.

[25] Further, embodiments may include a WB process database 122, which provides an example of the results of the control algorithm performed in the WB process module 112. The WB process database 122 contains the wafer number or ID, the ramp-up time to 300 degrees Celsius in minutes, and the bonding strength in megapascals optimized through the use of Al in the control algorithm, etc. The WB process database 122 may contain the data for each wafer that is being processed by the wafer bonding process 140. In some embodiments, the WB process database 122 may store other parameter data collected or calculated from the WB process module 112, such as temperature application, bonding force application, cooling rate, release force, bond uniformity, bond alignment, bonding pressure, etc.

[26] Further, embodiments may include a WB post-process database 124, which provides an example of the results of the post-process algorithm performed in the WB post-process module 114. The WB post-process database 124 contains the wafer number or ID, the outputted annealing temperature of the post-process algorithm, the warpage of the wafer measured in microns, etc. The WB process database 122 may contain the data for each wafer that is being processed by the wafer bonding post-process 150. In some embodiments, the WB post-process database 124 may store other parameter data collected or calculated from the WB post-process module 114, such as bonding strength evaluation, void detection, surface roughness measurement, wafer warpage assessment, residual stress, defect density, etc.

[27] Further, embodiments may include a WB network database 126 may contain the historical data from the various processes performed by the wafer bonding pre-process 130, wafer bonding process 140, and wafer bonding post-process 150. The WB network database 126 may contain the data parameters collected during the inspection process, such as surface contamination, surface roughness, wafer bow and warp, thermal expansion mismatch, chemical incompatibility, alignment accuracy, void formation, electrical issues, etc., control parameters of the process, such as temperature application, bonding force application, cooling rate, release force, bond uniformity, bond alignment, bonding pressure, etc. The wafer bonding process 140 may include the bonding equipment, such as wafer bonders, vacuum chambers, or thermal bonding systems, etc., and data parameters collected during post-processing, such as bonding strength evaluation, void detection, surface roughness measurement, wafer warpage assessment, residual stress, defect density, etc.

[28] Further, embodiments may include a cloud 128, which may be a network of remote servers that provide on-demand computing resources and services over the Internet. The cloud 128 may consist of a collection of servers, storage devices, and networking equipment. Users may access the cloud 128 through a variety of devices, such as computers, smartphones, and tablets, using internet connectivity. In some embodiments, the architecture of the cloud 128 may be based on a distributed computing model, with multiple servers working together to provide services to users.

[29] Further, embodiments may include wafer bonding pre-process 130, which may be the processes, such as surface preparation, alignment, surface activation, etc., and equipment, such as plasma cleaners, wet bench, alignment systems, etc. utilized before the wafer bonding process 140, designed to prepare the microLED processed wafers for optimal performance during the wafer bonding process. The pre-process may involve collecting a plurality of parameter data including surface contamination, surface roughness, wafer bow and warp, thermal expansion mismatch, chemical incompatibility, alignment accuracy, void formation, electrical issues, etc. In some embodiments, the wafer bonding pre-process 130 may send and receive data from the 3rd party wafer bonding Al network 102, including parameter adjustments, control adjustments, material or workpiece data, decisions on material inspections, inputs to be used by the wafer bonding pre- process Al module 136, etc.

[30] Further, embodiments may include a communication interface 132, which may be a hardware or software component that enables communication between the wafer bonding pre- process 130 and the 3rd party wafer bonding Al network 102. In some embodiments, the wafer bonding pre-process 130 may communicate with the wafer bonding process 140, and wafer bonding post-process 150. The communication interface 132 may include a set of protocols, rules, and standards that define how information is transmitted and received between the devices. The communication interface 132 may be a physical connector, wireless network, or software application and may include components such as drivers, software libraries, and firmware that may be used to control and manage the communication process.

[31] In some embodiments, the communication interface 132 may be compatible with USB, Bluetooth, or Wi-Fi. The communication interface 132 may communicate with a network. Examples of networks may include but are not limited to, the Internet, a cloud network, a Wireless Fidelity (Wi-Fi) network, a Wireless Local Area Network (WLAN), a Local Area Network (LAN), a telephone line (POTS), Long Term Evolution (LTE), and/or a Metropolitan Area Network (MAN).

[32] Further, embodiments may include a memory 134 may include suitable logic, circuitry, and/or interfaces that may be configured to store a machine code and/or a computer program with at least one code section executable by a processor. Examples of implementation of the memory 134 may include, but are not limited to, Random Access Memory (RAM), Read Only Memory (ROM), Hard Disk Drive (HDD), and/or a Secure Digital (SD) card. [33] Further, embodiments may include a wafer bonding pre-process Al module 136 in which a predictive model may be performed to predict the final product of the wafer bonding process based upon the data collected from the wafer bonding pre-process 130, such as surface contamination, surface roughness, wafer bow and warp, thermal expansion mismatch, chemical incompatibility, alignment accuracy, void formation, electrical issues, etc. Predictive modeling involves the systematic analysis of historical data to identify patterns, correlations, and trends that can be used to build models capable of making accurate predictions about future outcomes In some embodiments, the wafer bonding pre-process Al module 136 may receive a predictive model from the WB integration module 118.

[34] In some embodiments, the wafer bonding pre-process Al module 136 may receive and send data to the WB inspection module 110. In some embodiments, the data received from the WB inspection module 110 may be inputted into the predictive model to determine if any parameters of the pre-process should be adjusted, such as surface contamination, surface roughness, wafer bow and warp, thermal expansion mismatch, chemical incompatibility, alignment accuracy, void formation, electrical issues, etc.

[35] Further, embodiments may include a wafer bonding pre-process database 138 which may include data parameters collected from the wafer bonding pre-process 130, such as surface contamination, surface roughness, wafer bow and warp, thermal expansion mismatch, chemical incompatibility, alignment accuracy, void formation, electrical issues, etc.

[36] In some embodiments, the data stored in the wafer bonding pre-process database 138 may be sent to the WB inspection module 110 and/or the WB historical module 116.

[37] Further, embodiments may include wafer bonding process 140, which may be systems or equipment for automated wafer bonding of microLED chips that bring the prepared surfaces of the LED wafer and the substrate wafer into intimate contact under controlled conditions of temperature, pressure, and atmosphere. The wafer bonding process 140 may involve collecting parameter data, such as temperature application, bonding force application, cooling rate, release force, bond uniformity, bond alignment, bonding pressure, etc. The wafer bonding process 140 may include the bonding equipment, such as wafer bonders, vacuum chambers, thermal bonding systems, etc.

[38] In some embodiments, the wafer bonding process 140 may be performed by the EVG 520 Wafer Bonder, SUSS MicroTec’s SB6e wafer bonder, etc. In some embodiments, the wafer bonding process 140 may send and receive data from the 3rd party wafer bonding Al network 102, including parameter adjustments, control adjustments, temperature application, bonding force application, cooling rate, release force, bond uniformity, bond alignment, bonding pressure, inputs to be used by the wafer bonding process Al module 146, etc.

[39] Further, embodiments may include a communication interface 142, which may be a hardware or software component that enables communication between the wafer bonding process 140 and the 3rd party wafer bonding Al network 102.

[40] In some embodiments, the wafer bonding process 140 may communicate with the wafer bonding pre-process 130, and wafer bonding post-process 150. The communication interface 142 may include a set of protocols, rules, and standards that define how information is transmitted and received between the devices. The communication interface 142 may be a physical connector, wireless network, or software application and may include components such as drivers, software libraries, and firmware that may be used to control and manage the communication process.

[41] In some embodiments, the communication interface 142 may be compatible with USB, Bluetooth, or Wi-Fi. The communication interface 142 may communicate with a network. Examples of networks may include but are not limited to, the Internet, a cloud network, a Wireless Fidelity (Wi-Fi) network, a Wireless Local Area Network (WLAN), a Local Area Network (LAN), a telephone line (POTS), Long Term Evolution (LTE), and/or a Metropolitan Area Network (MAN).

[42] Further, embodiments may include a memory 144 may include suitable logic, circuitry, and/or interfaces that may be configured to store a machine code and/or a computer program with at least one code section executable by a processor. Examples of implementation of the memory 144 may include, but are not limited to, Random Access Memory (RAM), Read Only Memory (ROM), Hard Disk Drive (HDD), and/or a Secure Digital (SD) card.

[43] Further, embodiments may include a wafer bonding process Al module 146 in which a predictive model may be performed to predict the final product of the wafer bonding process based upon the data collected from the wafer bonding process 140, such as temperature application, bonding force application, cooling rate, release force, bond uniformity, bond alignment, bonding pressure, etc. Predictive modeling involves the systematic analysis of historical data to identify patterns, correlations, and trends that can be used to build models capable of making accurate predictions about future outcomes.

[44] In some embodiments, the wafer bonding process Al module 146 may receive a predictive model from the WB integration module 118. In some embodiments, the wafer bonding process Al module 146 may receive and send data to the WB process module 112. In some embodiments, the data received from the WB process module 112 may be inputted into the predictive model to determine if any parameters of the wafer bonding process 140 should be adjusted, such as temperature application, bonding force application, cooling rate, release force, bond uniformity, bond alignment, bonding pressure, etc. [45] Further, embodiments may include a wafer bonding process database 148 which may include data parameters collected from the wafer bonding process 140, such as temperature application, bonding force application, cooling rate, release force, bond uniformity, bond alignment, bonding pressure, etc.

[46] In some embodiments, the data stored in the wafer bonding process database 148 may be sent to the WB process module 112 and/or the WB historical module 116.

[47] Further, embodiments may include wafer bonding post-process 150, which may be the processes and methods after the completion of the wafer bonding process 140, such as an annealing process, thin film removal, inspection and testing, etc. The post-process may involve parameters such as bonding strength evaluation, void detection, surface roughness measurement, wafer warpage assessment, residual stress, defect density, etc. The post-process may include using an annealing furnace, chemical etching systems, metrology tools, etc.

[48] In some embodiments, the wafer bonding post- process 150 may send and receive data from the 3rd party wafer bonding Al network 102, including parameter adjustments, control adjustments, bonding strength evaluation, void detection, surface roughness measurement, wafer warpage assessment, residual stress, defect density, inputs to be used by the wafer bonding post-process Al module 156, etc.

[49] Further, embodiments may include a communication interface 152, which may be a hardware or software component that enables communication between the wafer bonding postprocess 150 and the 3rd party wafer bonding Al network 102. In some embodiments, the wafer bonding post-process 150 may communicate with the wafer bonding pre-process 130, and wafer bonding process 140. The communication interface 152 may include a set of protocols, rules, and standards that define how information is transmitted and received between the devices. The communication interface 152 may be a physical connector, wireless network, or software application and may include components such as drivers, software libraries, and firmware that may be used to control and manage the communication process.

[50] In some embodiments, the communication interface 152 may be compatible with USB, Bluetooth, or Wi-Fi. The communication interface 152 may communicate with a network. Examples of networks may include but are not limited to, the Internet, a cloud network, a Wireless Fidelity (Wi-Fi) network, a Wireless Local Area Network (WLAN), a Local Area Network (LAN), a telephone line (POTS), Long Term Evolution (LTE), and/or a Metropolitan Area Network (MAN).

[51] Further, embodiments may include a memory 154 may include suitable logic, circuitry, and/or interfaces that may be configured to store a machine code and/or a computer program with at least one code section executable by a processor. Examples of implementation of the memory 154 may include, but are not limited to, Random Access Memory (RAM), Read Only Memory (ROM), Hard Disk Drive (HDD), and/or a Secure Digital (SD) card.

[52] Further, embodiments may include a wafer bonding post-process Al module 156 in which a predictive model may be performed to optimize the future final products based upon the data collected from the wafer bonding post-process 150, such as bonding strength evaluation, void detection, surface roughness measurement, wafer warpage assessment, residual stress, defect density, etc. Predictive modeling involves the systematic analysis of historical data to identify patterns, correlations, and trends that can be used to build models capable of making accurate predictions about future outcomes In some embodiments, the wafer bonding post-process Al module 156 may receive a predictive model from the WB integration module 118. [53] In some embodiments, the wafer bonding post-process Al module 156 may receive and send data to the WB post-process module 114.

[54] In some embodiments, the data received from the WB post-process module 114 may be inputted into the predictive model to determine if any parameters of the wafer bonding post-process 150 should be adjusted, such as temperature and pressure control, alignment accuracy, bonding force and time, surface preparation and cleanliness, adhesive material selection, post-bonding annealing or curing, etc.

[55] Further, embodiments may include a wafer bonding post-process database 158 which may include data parameters collected from the wafer bonding post-process 150, such as temperature and pressure control, alignment accuracy, bonding force and time, surface preparation and cleanliness, adhesive material selection, post-bonding annealing or curing, etc.

[56] In some embodiments, the data stored in the wafer bonding post-process database 158 may be sent to the WB post-process module 114 and/or the WB historical module 116.

[57] FIG. 2 illustrates the base module 108. The process begins with the base module 108 initiating, at step 200, the WB inspection module 110. For example, the WB inspection module 110 begins by being initiated by the base module 108. The WB inspection module 110 connects to the wafer bonding pre-process 130. The WB inspection module 110 connects with the wafer bonding pre-process 130, such as processes surface preparation, alignment, surface activation, etc., and equipment, such as plasma cleaners, wet bench, alignment systems, etc. In some embodiments, the WB inspection module 110 may transmit and receive data from the wafer bonding pre-process 130. In some embodiments, the wafer bonding pre-process 130 may measure surface contamination, surface roughness, wafer bow and warp, thermal expansion mismatch, chemical incompatibility, alignment accuracy, void formation, electrical issues, etc. [58] In some embodiments, the WB inspection module 110 may control or send inputs to control the wafer bonding pre-process 130. The WB inspection module 110 collects the data from the wafer bonding pre-process 130. The WB inspection module 110 collects the pre-processing data, such as surface contamination, surface roughness, wafer bow and warp, thermal expansion mismatch, chemical incompatibility, alignment accuracy, void formation, electrical issues, etc. The WB inspection module 110 performs the inspection algorithm. For example, the inspection algorithm may be an artificial intelligence or machine learning algorithm in which the data collected from the wafer bonding pre-process 130 is used to determine if a control parameter of the wafer bonding process should be adjusted or altered. The inspection algorithm may determine the void formation of the previously processed wafers and use this measurement to determine the parameters for the plasma cleaning for future wafers during the wafer bonding pre-process.

[59] In a microLED process, one typical contaminant that needs to be removed before wafer bonding is organic residues. These residues can come from various sources such as handling, processing, or even from the environment. Organic contaminants can interfere with the bonding process by preventing proper adhesion between the wafers or creating defects in the bonded interface. For example, the inspection algorithm may be a machine learning model that determines the optimal power and time settings for a plasma cleaning process based on the defect density. A training dataset may include defect density measurements before and after plasma cleaning, along with the corresponding power and time settings used for each cleaning process. The dataset may be used to train the inspection algorithm to learn the relationship between the power, time, and resulting defect density. Once the inspection algorithm is trained with the dataset, the desired defect density may be inputted for the post-plasma clean wafer, and the inspection algorithm may predict the optimal power and time settings to achieve the desired defect density. For example, if the

Y1 inputted desired defect density after the plasma cleaning is 0.20 defects/cm2 the inspection algorithm may recommend a power setting of 150 W for a 10-minute time period. The plasma cleaning process is then performed using the inspection algorithms' recommended power and time settings.

[60] In some embodiments, the defect density may be measured after the plasma cleaning to confirm the defect density matches the inputted desired defect density and validates the effectiveness of the inspection algorithm. For example, the inspection algorithm may measure the surface contamination of the wafer and modify wafer bonding pre-process parameters to minimize the void formation after the wafer bonding process, such as adjusting the plasma cleaning parameters including gas composition, gas flow rate, plasma power, treatment time, pressure, substrate temperature, plasma treatment mode, etc. The inspection algorithm utilizes the captured data, from the wafer bonding pre-process 130, to send data about the microLED processed wafer to the WB process module 112 to be used to adjust wafer bonding parameters used by the wafer bonding process 140. In some embodiments, the WB inspection module 110 may send the outputted data from the inspection algorithm to the WB process module 112 to optimize the parameters used in the process.

[61] In some embodiments, the inspection algorithm may be a machine learning regression model, such as a neural network or a decision tree, that is trained using historical data of defect densities and corresponding plasma clean power and time settings. The model learns the relationship between the input parameters, such as power and time, and the output, such as defect density based on the training data. Once trained, the model can be used to predict the optimal power and time settings that would result in a desired defect density. The model can be fine-tuned and updated over time as more data becomes available, allowing for continuous optimization of the plasma cleaning process.

[62] In some embodiments, the inspection algorithm may perform reinforcement learning in which the algorithm may interact with the plasma cleaning process by adjusting the power and time settings and observing the resulting defect density. The algorithm learns to optimize the power and time settings through trial and error, aiming to maximize a reward signal, for example achieving a target defect density. The algorithm may perform actions, such as adjusting power and time, in the environment, for example, the plasma cleaning process, and receive feedback, such as the defect density measurements, as rewards or penalties. Over time, the algorithm learns which combinations of power and time settings lead to the desired defect density and refines its decision-making strategy accordingly. Through continuous learning and exploration, the algorithm may adapt to changes in the process and environment, optimizing the plasma cleaning parameters in a dynamic and adaptive manner. The WB inspection module 110 sends the determined process data from the inspection algorithm to the WB process module 112. For example, the WB inspection module 110 may send the defect density to the WB process module 112 to optimize the wafer bonding process.

[63] In some embodiments, the WB inspection module 110 may send surface contamination, surface roughness, wafer bow and warp, thermal expansion mismatch, chemical incompatibility, alignment accuracy, void formation, electrical issues, etc. The WB inspection module 110 stores the data in the WB inspection database 120. The WB inspection module 110 stores the data outputted from the inspection algorithm in the WB inspection database 120, such as the wafer number or ID, the defect density before the cleaning process, the defect density after the cleaning process with the parameters optimized by the inspection algorithm, and the delta or effectiveness of the cleaning process, etc. [64] In some embodiments, the WB inspection database 120 may include other parameter data collected or calculated during the WB inspection module 110 process, such as surface contamination, surface roughness, wafer bow and warp, thermal expansion mismatch, chemical incompatibility, alignment accuracy, void formation, electrical issues, etc. The WB inspection module 110 returns to the base module 108. The base module 108 initiates, at step 202, the WB process module 112. For example, the WB process module 112 begins by being initiated by the base module 108. The WB process module 112 receives the pre-process data, such as the data on the microLED processed wafer, from the WB inspection module 110. The WB process module 112 receives the outputted data from the inspection algorithm performed in the WB inspection module 110, such as the defect density of the wafer, allowing the wafer bonding process 140 to determine the optimal control parameters to prevent any defects in the wafer bonding process.

[65] In some embodiments, the WB process module 112 may receive surface contamination, surface roughness, wafer bow and warp, thermal expansion mismatch, chemical incompatibility, alignment accuracy, void formation, electrical issues, etc. The WB process module 112 connects to the wafer bonding process 140. The WB process module 112 connects with the wafer bonding process 140, such as systems or equipment for automated wafer bonding of microLED chips that bring the prepared surfaces of the LED wafer and the substrate wafer into intimate contact under controlled conditions of temperature, pressure, and atmosphere.

[66] In some embodiments, the WB process module 112 may transmit and receive data from the wafer bonding process 140. In some embodiments, the WB process module 112 may control or send inputs to control the wafer bonding process 140, such as controlling parameters of the wafer bonding process, for example, temperature application, bonding force application, cooling rate, release force, bond uniformity, bond alignment, bonding pressure, etc. The WB process module 112 collects the data from the wafer bonding process 140. The WB process module 112 collects the processing data, such as temperature application, bonding force application, cooling rate, release force, bond uniformity, bond alignment, bonding pressure, etc. The WB process module 112 performs the control algorithm. The control algorithm may use the output of the inspection algorithm performed in the WB inspection module 110 and the data collected from the wafer bonding process 140 as input. The control algorithm may clean and preprocess the collected data to handle missing values, and outliers, and ensure uniformity, and the dataset is divided into training and testing sets. The key features are identified that significantly impact the bonding strength during the wafer bonding process, such as temperature ramp-up times, temperature gradient, heating element placement, holding temperature, contact area variation, thermal resistance variation, stress, and strain, etc.

[67] The control algorithm utilizes the training dataset to train a linear regression model, which establishes a linear relationship between the selected features and the target variable, such as the bonding strength for the wafer bonding process. The trained model is then evaluated using the testing dataset to ensure its predictive accuracy.

[68] In some embodiments, metrics such as Mean Squared Error (MSE) or R-squared may be used to assess the performance of the linear regression model. The new data from the wafer bonding process 140 and the output of the inspection algorithm are inputted into the trained linear regression model, such as temperature ramp-up times, temperature gradient, heating element placement, holding temperature, contact area variation, thermal resistance variation, stress, and strain, etc. and the control algorithm outputs the predicted bonding strength value for the wafer bonding process. [69] In some embodiments, new data during the process may be continuously inputted into the control algorithm, allowing the algorithm to dynamically adapt to changing conditions and optimize the acceleration for optimal results. For example, unoptimized control of ramp-up and holding temperature in the wafer bonder may cause uneven heating. The optimal ramp-up rate for wafer bonding depends on various factors, including the materials being bonded, the bonding process used, and the desired properties of the bonded structure. Faster ramp-up rates may shorten the overall processing time, leading to higher throughput and productivity and rapid heating may accelerate bonding mechanisms such as surface activation or diffusion bonding, which may enhance bonding strength. Rapid temperature changes may induce thermal stresses, which may lead to a higher risk of defects such as delamination, voids, or cracks at the bonding interface and may result in non-uniform temperature distribution across the wafer surface, which may lead to variations in bonding strength and quality. Slower ramp-up rates may minimize thermal stresses and allow more time for bonding mechanisms to activate, which may reduce the risk of defects improve bonding quality, and may enable more uniform temperature distribution across the wafer surface, which may promote consistent bonding strength and quality. Slower ramp-up rates may prolong the overall processing time, potentially reducing throughput and productivity, and may increase the risk of contamination or outgassing at the bonding interface, affecting bonding quality.

[70] In some embodiments, the control algorithm may forecast the temperature profile required to achieve a ramp-up time of 15 minutes to reach 300°C by considering factors such as the initial temperature, heating capacity, wafer characteristics, and environmental conditions to predict the optimal temperature ramp-up trajectory. [71] In some embodiments, the control algorithm may continuously monitor temperature changes and compare them to the predicted temperature profile and may use sensor data and feedback mechanisms to assess the actual ramp-up time and temperature progression in real time.

[72] In some embodiments, the control algorithm may dynamically adjust heating parameters, such as power input, heating rate, and timing, to optimize the temperature ramp-up process and may leverage adaptive control system algorithms that respond to deviations from the predicted temperature profile and make proactive adjustments to ensure alignment with the target ramp-up time.

[73] In some embodiments, the control algorithm may learn from discrepancies between predicted and actual temperature profiles over time and refine its predictive models and control strategies by continuously adapting to changes in process conditions, equipment performance, and material properties to improve temperature ramp-up control and enhance bonding consistency and quality. In some embodiments, the control algorithm may integrate with wafer bonding equipment and process automation systems and may interface with temperature control units, heating elements, and other hardware components to implement precise temperature ramp-up profiles based on real-time predictions and feedback. The WB process module 112 adjusts the control parameters of the wafer bonding process 140. For example, the WB process module 112 adjusts the temperature application, such as the heating profile, of the wafer bonding process.

[74] In some embodiments, the WB process module 112 may adjust the bonding force application, cooling rate, release force, bond uniformity, bond alignment, bonding pressure, etc. The WB process module 112 executes the wafer bonding process. For example, the wafer bonding process may include the control parameter settings outputted by the control algorithm being used by the controller of the wafer bonding process 140 equipment to set the temperature ramp-up rate and holding temperature during the wafer bonding process. The WB process module 112 may monitor the wafer bonding process of the wafer bonding process 140 equipment and perform the control algorithm in real-time to continuously adjust the control parameters. The WB process module 112 stores, at step 414, the data in the WB process database 122. The WB process module 112 stores the data in the WB process database 122, such as the data collected from the wafer bonding process 140, the output of the inspection algorithm, and the output of the control algorithm. The WB process database 122 may contain the wafer number or ID, the ramp-up time to 300 degrees Celsius in minutes, and the bonding strength in megapascals optimized through the use of Al in the control algorithm, etc. The WB process database 122 may contain the data for each wafer that is being processed by the wafer bonding process 140. The WB process module 112 returns to the base module 108. The base module 108 initiates, at step 204, the WB post-process module 114. For example, the WB post-process module 114 begins by being initiated by the base module 108. The WB post-process module 114 connects to the wafer bonding post-process 150. The WB postprocess module 114 may connect to the wafer bonding post-process 150, such as the annealing process, thin film removal, inspection and testing, etc. The WB post-process module 114 collects the post-process data. The WB post-process module 114 may collect the parameter data, such as testing, etc. The post-process may involve parameters such as bonding strength evaluation, void detection, surface roughness measurement, wafer warpage assessment, residual stress, defect density, etc. The WB post-process module 114 extracts the data from the WB inspection database 120 and WB process database 122. The WB post -process module 114 extracts the data, such as material data, parameter data, etc. of the pre-process and the wafer bonding process. The WB postprocess module 114 performs the post-process algorithm. The WB post-process module 114 may perform a post-process algorithm that uses machine learning algorithms to analyze parameters related to defects, such as warpages. The post-process algorithm may focus on the process data outputted from the inspection algorithm performed in the WB inspection module 110 and the effectiveness of the parameters outputted by the control algorithm performed in the WB process module 112 during the wafer bonding process.

[75] The post-process algorithm may adjust control parameters of the current microLED processed wafers in the process to result in an improved wafer bonding process. For example, the post-process algorithm may optimize temperature and pressure control, alignment accuracy, bonding force and time, surface preparation and cleanliness, adhesive material selection, postbonding annealing or curing, etc. to minimize the warpage of the wafer. The post-bonding annealing or curing may address any residual stresses that may have been introduced during the bonding process and may provide stress relaxation, improved flatness, enhanced adhesion, etc.

[76] In some embodiments, the post-bonding treatment may streamline the process, and provide precise control and customization. In some embodiments, the post-process algorithm may minimize the warpage of the wafer by determining the annealing temperature, and annealing duration, providing uniform heating, and incorporating the material properties of the adhesives and intermediate layers used in the bonding process.

[77] In some embodiments, the post-process algorithm may gradually increase and decrease the temperature during the annealing process to minimize thermal shock and reduce the risk of inducing additional stress or warpage. [78] In some embodiments, the post-process algorithm may perform the annealing process in a controlled atmosphere, such as inert gas, which may prevent oxidation or other chemical reactions that might affect the microLEDs or the bonding materials.

[79] In some embodiments, the post-process algorithm may integrate the cumulative thermal budget and the interactions between different materials and processes used in microLED manufacturing.

[80] In some embodiments, the post-process algorithm may control the annealing temperature in the wafer bonding post-process 150 to optimize process parameters in real time to achieve desired material properties and minimize defects.

[81] In some embodiments, the post-process algorithm may use machine learning algorithms to analyze parameters related to warpage, such as annealing temperature, temperature ramp-up, temperature cool-down, atmosphere control, annealing temperature, etc. The post-process algorithm may focus on the process data outputted from the inspection algorithm performed in the WB inspection module 110 and the effectiveness of the parameters outputted by the control algorithm performed in the WB process module 112 during the wafer bonding process. The postprocess algorithm may adjust control parameters of the current microLED processed wafers in the process to result in an improved wafer bonding process. For example, the post-process inspection may result in a warpage identified in the wafers. The post-process algorithm may provide the data to the WB process module 112 to allow Al-controlled monitoring to result in minimized warpage by further adjusting the ramp-up temperature of the equipment during the wafer bonding process. The WB post-process module 114 performs the post-process inspection. The WB post-process module 114 may involve collecting parameter data such as bonding strength evaluation, void detection, surface roughness measurement, wafer warpage assessment, residual stress, defect density, etc. For example, the post-process may include the wafers undergoing annealing treatments at elevated temperatures to improve bond strength and relieve any residual stresses introduced during the bonding process.

[82] In some embodiments, thin films or sacrificial layers used during the bonding process may need to be removed which may be achieved through techniques like chemical etching or mechanical polishing. In some embodiments, inspection, and testing may be conducted to evaluate the quality of the bonded wafers, and parameters such as bond strength, interface uniformity, and defect density may be measured using techniques like scanning electron microscopy, atomic force microscopy, and optical profilometry. The WB post-process module 114 stores the data in the WB post-process database 124. The WB post-process module 114 stores the data in the WB postprocess database 124, such as the data collected from the wafer bonding post-process 150, the output of the post-process algorithm, etc. The WB post-process database 124 may contain the wafer number or ID, the outputted annealing temperature of the post-process algorithm, the warpage of the wafer measured in microns, etc. The WB process database 122 may contain the data for each wafer that is being processed by the wafer bonding post-process 150. The WB post-process module 114 returns to the base module 108. The base module 108 initiates, at step 206, the WB historical module 116. For example, the WB historical module 116 begins by being initiated by the base module 108. The WB historical module 116 connects to the wafer bonding pre-process 130, the wafer bonding process 140, and the wafer bonding post-process 150. The historical module 116 connects to the wafer bonding pre-process 130, such processes, such as surface preparation, alignment, surface activation, etc. and equipment, such as plasma cleaners, wet bench, alignment systems, etc., the wafer bonding process 140, such as systems or equipment for automated wafer bonding of microLED chips that brings the prepared surfaces of the LED wafer and the substrate wafer into intimate contact under controlled conditions of temperature, pressure, and atmosphere, and the wafer bonding post-process 150, such processes and methods after the completion of the wafer bonding process 140, such as an annealing process, thin film removal, inspection, and testing, etc. The WB historical module 116 aggregates the data from the various types of processes and stores the data in the WB network database 126. The WB network database 126 may contain the historical data from the various processes performed by the wafer bonding pre-process 130, wafer bonding process 140, and wafer bonding post-process 150. The WB network database 126 may contain the data parameters collected during the inspection process, such as surface contamination, surface roughness, wafer bow and warp, thermal expansion mismatch, chemical incompatibility, alignment accuracy, void formation, electrical issues, etc., control parameters of the process, such as temperature application, bonding force application, cooling rate, release force, bond uniformity, bond alignment, bonding pressure, etc., and data parameters collected during post-processing, such as bonding strength evaluation, void detection, surface roughness measurement, wafer warpage assessment, residual stress, defect density, etc.

[83] The post-process may include using an annealing furnace, chemical etching systems, metrology tools, etc. The WB historical module 116 performs the historical machine learning algorithm on the historical data stored in the WB network database 126. For example, the historical machine learning algorithm may use the historical data stored in the WB network database 126 to determine trends to understand how varying parameters affect the final quality and identify optimal parameter settings to create the best outcomes. The historical machine learning algorithm adjusts the parameters in real time for the WB inspection module 110, WB process module 112, and WB post-process module 114 to correct any defects. For example, the data may show defects in the wafers that occurred during the wafer bonding process. The historical machine learning algorithm may minimize the defects in the wafer by adjusting the power and time settings of the plasma cleaning process based on the defect density inputted in the WB inspection module 110. The historical machine learning algorithm may also adjust the ramp-up time to temperature occurring in the WB process module 112 to optimize the bonding strength and adjust the annealing temperature for the WB post-process module 114 to minimize warpage. In some embodiments, the historical machine learning algorithm may optimize the plasma pre-cleaning process to reduce void formation and enhance wafer bonding by analyzing historical data and identifying correlations between the pre-clean conditions and void formation rates.

[84] In some embodiments, the historical machine learning algorithm may be a predictive model, such as a regression model or neural network, to predict the optimal plasma pre-clean parameters based on the material type, wafer characteristics, and desired bonding outcomes.

[85] In some embodiments, the historical machine learning algorithm may be a support vector machine in which wafer surfaces are classified based on pre-clean effectiveness to predict whether a given set of pre-clean parameters will lead to successful bonding.

[86] In some embodiments, the historical machine learning algorithm may be a decision tree and random forest model in which the impact of various pre-clean parameters on void formation is analyzed to provide process adjustments.

[87] In some embodiments, the historical machine learning algorithm may perform a principal component analysis to identify the most critical pre-clean parameters that affect void formation from a large dataset. [88] In some embodiments, the historical machine learning algorithm may optimize temperature ramping profiles to enhance wafer bonding quality by using time-series analysis and clustering algorithms to identify successful temperature profiles and conditions that lead to high-quality bonding. The historical machine learning algorithm may predict the ideal temperature profile for a given set of wafer materials and bonding agents.

[89] In some embodiments, the historical machine learning algorithm may use a control algorithm, such as a PID controller augmented with Al to adjust the heating profiles in real-time based on ongoing performance data and model predictions. In some embodiments, the historical machine learning algorithm may be a gradient-boosting machine to predict optimal temperature profiles based on past successful bonds and provide high accuracy by learning from errors in previous predictions.

[90] In some embodiments, the historical machine learning algorithm may be a recurrent neural network and long short-term memory network to predict the outcomes of bonding processes over time.

[91] In some embodiments, the historical machine learning algorithm may perform Bayesian optimization to efficiently search the space of temperature profiles to find the optimal settings by balancing the exploration of new parameters with the exploitation of known data. The WB historical module 116 sends the process adjustments to the WB inspection module 110, the WB process module 112, and the WB post-process module 114. The historical machine learning algorithm may minimize the defects occurring in the wafer by adjusting the plasma cleaning parameters of the wafer bonding pre-process based on the defect density determined in the WB inspection module 110. The historical machine learning algorithm may also adjust the temperature ramp-up time in the WB process module 112 to optimize bonding strength for the WB post-process module 114 to inspect. The adjustments are sent to the WB inspection module 110 and WB process module 112 to create a more consistent final product which is then verified by the WB post-process module 114. The WB historical module 116 returns to the base module 108. The base module 108 initiates, at step 208, the WB integration module 118. For example, the WB integration module 118 begins by being initiated by the base module 108. The WB integration module 118 performs the integration machine learning algorithm.

[92] The integration machine learning algorithm may utilize data across the different stages of the wafer bonding process to enhance the overall process efficiency and outcome quality. For example, the integration machine learning algorithm may utilize insights from void formation data and plasma pre-clean parameters to optimize post-anneal and curing processes for reduced warpage. The integration machine learning algorithm may perform deep learning, such as a convolutional neural network, that analyzes historical data on void formation rates and preclean conditions to identify patterns where specific types of voids or contamination levels correlate with increased warpage post-annealing. The integration machine learning algorithm may predict postprocess warpage based on the pre-clean data and adjust post-anneal and curing temperatures and times dynamically to target specific conditions that have historically led to less warpage which enables proactive adjustments to post-process parameters based on early-stage data to reduce the need for extensive post-process corrections and improved wafer quality. For example, the integration machine learning algorithm may leverage temperature profile data from the wafer bonding process to fine-tune post-anneal and curing parameters for optimal stress relief and minimal warpage. The integration machine learning algorithm may perform regression analysis and gradient boosting and collect and analyze data on successful temperature ramps and outcomes to identify which profiles lead to the best bonding quality with minimal internal stresses that may lead to warpage.

[93] The integration machine learning algorithm may use regression analysis and gradient boosting to create a predictive model that recommends post-anneal and curing settings based on the heating profiles used during wafer bonding and adjust post-process conditions in real-time based on the recommendations which ensure that the post-process annealing and curing are tailored to the conditions experienced during bonding to improve bond integrity and reduce warpage. For example, the integration machine learning algorithm may apply insights from post-process anneal and curing outcomes to optimize the plasma preclean step to reduce the void formation and enhance initial bonding quality. The integration machine learning algorithm may perform reinforcement learning which learns from the relationship between post-process outcomes, such as warpage, anneal success, etc., and pre-process conditions and may use feedback from post-process evaluations to suggest adjustments in the plasma preclean parameters. The integration machine learning algorithm may dynamically adjust preclean parameters such as plasma power, duration, and gas mix based on what has been learned about their impact on final wafer quality and creates a feedback loop where post-process performance directly informs pre-process setup which allows for continuous improvement of the preclean process based on comprehensive outcomes analysis and leads to better initial bonding quality and reduced correction requirements later in the process.

[94] In some embodiments, the integration machine learning algorithm may use the data stored in the WB network database 126 to perform a predictive modeling algorithm to make predictions or forecasts of the final product based on historical data and patterns from the previously created products. Predictive modeling involves the systematic analysis of historical data to identify patterns, correlations, and trends that can be used to build models capable of making accurate predictions about future outcomes. For example, the integration machine learning algorithm may be used for minimizing warpage across the wafer bonding pre-process 130, wafer bonding process 140, and wafer bonding post-process 150. The historical data in the WB network database 126 may include information on warpage and relevant operational parameters such as void formation, preclean parameters, temperature profile data, annealing and curing temperature and times, etc. The data is cleaned and preprocessed, addressing any missing values and ensuring that all variables are in a suitable format for modeling. This may involve normalization or scaling of numerical features. The dataset is split into training and testing sets. The training set is used to train the predictive model, and the testing set assesses its performance on unseen data. The model used may be linear regression, decision trees, ensemble methods, neural networks, etc. Then key features are identified that influence the warpage and engineer new features if necessary. For example, the interaction between pre-clean parameters and temperature profiles may have a significant impact on the warpage. The selected predictive model is trained using the training dataset. The model learns patterns and relationships between operational parameters and warpage. The model's performance is validated on a separate validation dataset and hyperparameters are fine-tuned to optimize its accuracy and generalization. The model is then evaluated and deployed by being sent to the wafer bonding pre-process 130, wafer bonding process 140, and wafer bonding post-process 150. The WB integration module 118 connects to the wafer bonding pre-process Al module 136, the wafer bonding process Al module 146, and the wafer bonding post-process Al module 156. The WB integration module 118 connects to the wafer bonding pre-process 130, such as processes, such as surface preparation, alignment, surface activation, etc. and equipment, such as plasma cleaners, wet bench, alignment systems, etc., the wafer bonding process 140, such as systems or equipment for automated wafer bonding of microLED chips that brings the prepared surfaces of the LED wafer and the substrate wafer into intimate contact under controlled conditions of temperature, pressure, and atmosphere, and the wafer bonding post-process 150, such processes and methods after the completion of the wafer bonding process 140, such as an annealing process, thin film removal, inspection, and testing, etc. The WB integration module 118 sends the process adjustments to the wafer bonding pre-process Al module 136, the wafer bonding process Al module 146, and the wafer bonding post-process Al module 156. For example, the WB integration module 118 may send the void formation, pre-clean parameters, etc. adjustments to each of the processes to minimize the warpage of the wafer during the wafer bonding process.

[95] In some embodiments, the WB integration module 118 may send the integration machine learning algorithm to the processes, allowing the systems to further enhance the optimization. The WB integration module 118 returns to the base module 108.

[96] FIG. 3 illustrates the WB inspection module 110. The process begins with the WB inspection module 110 being initiated, at step 300, by the base module 108. The WB inspection module 110 connects, at step 302, to the wafer bonding pre-process 130. The WB inspection module 110 connects with the wafer bonding pre-process 130, such as processes surface preparation, alignment, surface activation, etc., and equipment, such as plasma cleaners, wet bench, alignment systems, etc. In some embodiments, the WB inspection module 110 may transmit and receive data from the wafer bonding pre-process 130. In some embodiments, the wafer bonding pre-process 130 may measure surface contamination, surface roughness, wafer bow and warp, thermal expansion mismatch, chemical incompatibility, alignment accuracy, void formation, electrical issues, etc. [97] In some embodiments, the WB inspection module 110 may control or send inputs to control the wafer bonding pre-process 130. The WB inspection module 110 collects, at step 304, the data from the wafer bonding pre-process 130. The WB inspection module 110 collects the preprocessing data, such as surface contamination, surface roughness, wafer bow and warp, thermal expansion mismatch, chemical incompatibility, alignment accuracy, void formation, electrical issues, etc.

[98] Edge Al and TinyML Example 1: In some embodiments, the WB inspection module 110 may utilize a TinyML-enabled Edge Al model deployed on in-line imaging hardware to perform on-device, low-power classification of bond anomalies during image acquisition, enabling immediate wafer-level rejection or flagging without dependency on upstream inspection servers. For example, a shift operator may query, “Is this wafer showing early void patterns consistent with clamp misalignment?” while an embedded diagnostic agent may autonomously prompt, “Flag images for review if concentric artifact contrast exceeds the baseline threshold.” In response, the WB inspection module 110 may acquire real-time image frames from the pre-processing WB equipment 128, process them locally using a quantized convolutional model trained on defect archetypes stored in the WB inspection database 120, and perform immediate classification of features such as ring voids, edge delamination, or stress patterns. If a match is detected, the device may annotate the wafer ID, display an alert on the in-line UI, and log metadata for delayed upload to the WB network database 126. Internally, the TinyML model may leverage low-memory feature extractors and edge-optimized pooling layers to enable sub-second classification within the constraints of embedded image sensor processors. Outputs may include an on-device flag signal, a locally cached defect snapshot, and a priority inspection recommendation appended to the batch manifest. By embedding defect recognition logic directly at the point of image capture, the WB inspection module 110 may reduce inspection latency, mitigate defect propagation, and maintain yield continuity even under bandwidth or compute constraints. The WB inspection module 110 performs, at step 306, the inspection algorithm. For example, the inspection algorithm may be an artificial intelligence or machine learning algorithm in which the data collected from the wafer bonding pre-process 130 is used to determine if a control parameter of the wafer bonding process should be adjusted or altered.

[99] The inspection algorithm may determine the void formation of the previously processed wafers and use this measurement to determine the parameters for the plasma cleaning for future wafers during the wafer bonding pre-process. In a microLED process, one typical contaminant that needs to be removed before wafer bonding is organic residues. These residues can come from various sources such as handling, processing, or even from the environment. Organic contaminants can interfere with the bonding process by preventing proper adhesion between the wafers or creating defects in the bonded interface. For example, the inspection algorithm may be a machine learning model that determines the optimal power and time settings for a plasma cleaning process based on the defect density. A training dataset may include defect density measurements before and after plasma cleaning, along with the corresponding power and time settings used for each cleaning process. The dataset may be used to train the inspection algorithm to learn the relationship between the power, time, and resulting defect density. Once the inspection algorithm is trained with the dataset, the desired defect density may be inputted for the post-plasma clean wafer, and the inspection algorithm may predict the optimal power and time settings to achieve the desired defect density. For example, if the inputted desired defect density after the plasma cleaning is 0.20 defects/cm2 the inspection algorithm may recommend a power setting of 150 W for a 10-minute time period. The plasma cleaning process is then performed using the inspection algorithms' recommended power and time settings.

[100] In some embodiments, the defect density may be measured after the plasma cleaning to confirm the defect density matches the inputted desired defect density and validates the effectiveness of the inspection algorithm. For example, the inspection algorithm may measure the surface contamination of the wafer and modify wafer bonding pre-process parameters to minimize the void formation after the wafer bonding process, such as adjusting the plasma cleaning parameters including gas composition, gas flow rate, plasma power, treatment time, pressure, substrate temperature, plasma treatment mode, etc. The inspection algorithm utilizes the captured data, from the wafer bonding pre-process 130, to send data about the microLED processed wafer to the WB process module 112 to be used to adjust wafer bonding parameters used by the wafer bonding process 140. In some embodiments, the WB inspection module 110 may send the outputted data from the inspection algorithm to the WB process module 112 to optimize the parameters used in the process. In some embodiments, the inspection algorithm may be a machine learning regression model, such as a neural network or a decision tree, that is trained using historical data of defect densities and corresponding plasma clean power and time settings. The model learns the relationship between the input parameters, such as power and time, and the output, such as defect density based on the training data. Once trained, the model can be used to predict the optimal power and time settings that would result in a desired defect density. The model can be fine-tuned and updated over time as more data becomes available, allowing for continuous optimization of the plasma cleaning process.

[101] In some embodiments, the inspection algorithm may perform reinforcement learning in which the algorithm may interact with the plasma cleaning process by adjusting the power and time settings and observing the resulting defect density. The algorithm learns to optimize the power and time settings through trial and error, aiming to maximize a reward signal, for example achieving a target defect density. The algorithm may perform actions, such as adjusting power and time, in the environment, for example, the plasma cleaning process, and receive feedback, such as the defect density measurements, as rewards or penalties. Over time, the algorithm learns which combinations of power and time settings lead to the desired defect density and refines its decision-making strategy accordingly. Through continuous learning and exploration, the algorithm may adapt to changes in the process and environment, optimizing plasma cleaning parameters in a dynamic and adaptive manner.

[102] Al Agents Example 1 : In some embodiments, the WB inspection module 110 may include a dedicated Al agent configured to dynamically adjust inspection frequency and retest scheduling based on live defect trends and material stack metadata, thereby optimizing inspection resource allocation without compromising detection sensitivity. For example, a yield analyst may prompt, “Increase post-bond inspection for hybrid wafers showing oxide thinning,” while the agent may autonomously trigger, “Escalate inspection density if void probability exceeds baseline in three consecutive lots.” In execution, the Al agent may continuously analyze defect maps and bonding anomaly scores from the WB inspection database 120, correlate these with material properties and bonding pressure logs from the WB process database 122, and cross-reference inspection throughput constraints and past pattern recurrence rates in the WB network database 126. Upon detecting emerging high-risk trends, the agent may proactively adjust inspection sample density, increase reinspection quotas for selected wafers, or route flagged lots to an alternate imaging modality. Internally, the agent may use risk-weighted inspection cost modeling and a reward- optimized policy function trained on historical yield impact cases to determine when intensified inspection is warranted. Outputs may include an updated inspection schedule annotated with trigger logic and priority flags, as well as a traceable rationale signal passed to the WB integration module 118 to inform downstream control decisions. By embedding an adaptive Al agent in the WB inspection module 110, the system may improve defect coverage during bond quality transitions while minimizing unnecessary inspection overhead.

[103] Multimodal Al Example 2: In some embodiments, the WB inspection module 110 may utilize a multimodal Al model to improve defect classification accuracy by fusing post-bond image data, force and temperature profiles from the bonding process, and technician-entered annotations. For example, a process engineer may query, “What combination of signals explains the peripheral voids seen on Lot T5?” while a quality control agent may request, “Identify anomalies that correlate with clamp edge flare across multiple tools.” In response, the WB inspection module 110 may retrieve visual inspection outputs from the pre-processing WB equipment 128, time-aligned bonding conditions from the WB process database 122, and operator-tagged feedback from the WB inspection database 120. The multimodal model may synchronize image features with time-series thermal and mechanical telemetry and contextual tags using a shared embedding space, allowing it to detect complex interaction patterns, such as clamp-induced edge compression artifacts amplified by platen heat gradient. Internally, the model may employ late fusion of encoded visual, numerical, and linguistic signals and compare them to failure archetypes from the WB network database 126 to assign a confidence-ranked defect cause classification. Outputs may include a fused analysis summary linking defect shape and distribution to pressure non-uniformity, and a literature reference indicating similar behavior observed in plasma-activated oxide bonding under edge- constrained heating. By embedding multimodal Al into the WB inspection module 110, the system may transform disparate inspection and process records into unified, high-confidence diagnostic insights for yield-impacting bond anomalies.

[104] Generative Video Al Example 1. In some embodiments, the WB inspection module 110 may employ a generative video Al model to simulate coating deformation and void nucleation during the bonding process based on early-cycle clamp behavior and initial temperature profiles, enabling predictive visualization of defect formation before full bonding completion. For example, a process engineer may request, “Simulate void pattern development if platen tilt exceeds 2 degrees Celsius during ramp-up,” while a tooling specialist may prompt, “Show probable interface damage if clamp overshoot occurs within the first 5 seconds.” In response, the WB inspection module 110 may gather early bonding cycle telemetry — such as platen force distribution, clamp angle, and initial heat ramp rate — from the pre-processing WB equipment 128, retrieve baseline inspection images and surface metadata from the WB inspection database 120, and use a generative video model trained on bond formation sequences stored in the WB network database 126 to synthesize predictive frame-by-frame outcomes. Internally, the system may apply spatiotemporal generation layers conditioned on tool-specific motion signatures and thermal inputs to forecast deformation artifacts — such as radial void rings or center-out delamination — based on current or hypothetical bonding dynamics. The output may include a synthetic video illustrating how interface disruption is likely to progress under the specified setup, along with a literature-sourced annotation linking early clamp instability to void onset in oxide wafer bonding. By embedding generative video forecasting into the WB inspection module 110, the system may offer engineers an anticipatory tool for evaluating bonding stability and preemptively refining tool alignment and ramp profiles.

[105] Synthetic Data and AI-Enabled Digital Twins Example 1 : In some embodiments, the WB inspection module 110 may employ synthetic data generation to augment training of defect classification models by creating labeled bond failure images under underrepresented material stacks and clamp configurations. For example, a machine learning engineer may prompt, “Generate synthetic edge void cases for low-temperature bonds on hybrid oxide stacks,” while a yield optimization agent may request, “Balance the training dataset with more samples showing delamination in high-pressure, short-dwell profiles.” In response, the WB inspection module 110 may analyze inspection image distributions stored in the WB inspection database 120 and identify sparsely represented bonding conditions based on process metadata from the pre-processing WB equipment 128. A generative model trained on image-process correlation pairs archived in the WB network database 126 may then synthesize high-fidelity defect image samples labeled with simulated bonding context — including ramp rate, temperature profile, and clamp geometry. Internally, the model may use conditional generation techniques (e.g., cGAN or diffusion -based synthesis) to produce realistic defect imagery while ensuring structural consistency with known failure signatures. Outputs may include an expanded training set containing synthetic defect images with associated metadata tags, and a performance report showing improved model recall on minority defect classes, accompanied by a citation on the efficacy of synthetic augmentation for industrial image classification. By embedding synthetic image generation within the WB inspection module 110, the system may improve classifier robustness, reduce bias, and accelerate adaptation to new bonding materials and tool variants.

[106] The WB inspection module 110 sends, at step 308, the determined process data from the inspection algorithm to the WB process module 112. For example, the WB inspection module 110 may send the defect density to the WB process module 112 to optimize the wafer bonding process. In some embodiments, the WB inspection module 110 may send surface contamination, surface roughness, wafer bow and warp, thermal expansion mismatch, chemical incompatibility, alignment accuracy, void formation, electrical issues, etc. The WB inspection module 110 stores, at step 310, the data in the WB inspection database 120. The WB inspection module 110 stores the data outputted from the inspection algorithm in the WB inspection database 120, such as the wafer number or ID, the defect density before the cleaning process, the defect density after the cleaning process with the parameters optimized by the inspection algorithm, and the delta or effectiveness of the cleaning process, etc.

[107] In some embodiments, the WB inspection database 120 may include other parameter data collected or calculated during the WB inspection module 110 process, such as surface contamination, surface roughness, wafer bow and warp, thermal expansion mismatch, chemical incompatibility, alignment accuracy, void formation, electrical issues, etc. The WB inspection module 110 returns, at step 312, to the base module 108. 306, inspection algorithm serves as a pivotal component in optimizing the plasma cleaning process during the wafer bonding pre- process, particularly in the context of microLED production. This process is crucial for ensuring the wafers are free from organic residues and other contaminants that can adversely affect the bonding quality and, ultimately, the device performance. The inspection algorithm, leveraging principles of artificial intelligence (Al) and machine learning (ML), takes on the task of analyzing data collected from the wafer bonding pre-process to make informed decisions about the control parameters of the wafer bonding process. The foundation of this inspection algorithm at step 306 involves collecting historical data on the plasma cleaning process. This data includes crucial variables such as the power and time settings of the plasma cleaner and the resulting defect densities observed on the wafers. By training a Decision Tree Regressor — a type of Al algorithm — on this dataset, the model learns to predict the outcome (defect density) based on the input parameters (power and time settings). The decision tree operates by creating a hierarchical structure that splits the data into branches based on those input parameters, aiming to group the data in such a way that the variance in defect density across groups is minimized.

[108] This process is iterative and continues until the model adequately represents the relationship between the cleaning process settings and the defect outcomes. Once trained, the inspection algorithm at step 306 is capable of predicting the defect density for new sets of power and time settings, or conversely, determining the optimal settings required to achieve a target defect density. This prediction or optimization process is crucial for adjusting the plasma cleaning parameters in a way that minimizes defects, thereby enhancing the bonding quality of the wafers. The real utility of the inspection algorithm is realized when these predictions are used to adjust the plasma cleaning parameters in real-world manufacturing processes. By implementing the recommended settings, the wafer processing team can potentially reduce defect density, improving yield and device quality. Moreover, the inspection algorithm's role does not end with a single prediction. It is involved in a continuous loop of feedback and learning, where the actual outcomes (measured defect densities after applying the recommended settings) are fed back into the system. This post- process data allows the algorithm to refine its predictions over time, adjusting to changes in the manufacturing environment or materials. This aspect of continuous learning is critical for maintaining the accuracy and relevance of the inspection algorithm over time. In summary, the "306, inspection algorithm" embodies a dynamic Al-driven approach to optimizing the wafer plasma cleaning process. Through the application of machine learning techniques, specifically the use of a Decision Tree Regressor, the algorithm analyzes historical process data to recommend adjustments to the cleaning parameters. This proactive approach to process optimization not only aims to reduce defect densities but also enhances the overall efficiency and output quality of the wafer bonding process, showcasing the potential of Al and ML technologies in advancing manufacturing processes in the semiconductor industry.

[109] FIG. 4 illustrates the WB process module 112. The process begins with the WB process module 112 being initiated, at step 400, by the base module 108. The WB process module 112 receives, at step 402, the pre-process data, such as the data on the microLED processed wafer, from the WB inspection module 110. The WB process module 112 receives the outputted data from the inspection algorithm performed in the WB inspection module 110, such as the defect density of the wafer, allowing the wafer bonding process 140 to determine the optimal control parameters to prevent any defects in the wafer bonding process. In some embodiments, the WB process module 112 may receive surface contamination, surface roughness, wafer bow and warp, thermal expansion mismatch, chemical incompatibility, alignment accuracy, void formation, electrical issues, etc. The WB process module 112 connects, at step 404, to the wafer bonding process 140. The WB process module 112 connects with the wafer bonding process 140, such as systems or equipment for automated wafer bonding of microLED chips that bring the prepared surfaces of the LED wafer and the substrate wafer into intimate contact under controlled conditions of temperature, pressure, and atmosphere. In some embodiments, the WB process module 112 may transmit and receive data from the wafer bonding process 140. In some embodiments, the WB process module 112 may control or send inputs to control the wafer bonding process 140, such as controlling parameters of the wafer bonding process, for example, temperature application, bonding force application, cooling rate, release force, bond uniformity, bond alignment, bonding pressure, etc. The WB process module 112 collects, at step 406, the data from the wafer bonding process 140. The WB process module 112 collects the processing data, such as temperature application, bonding force application, cooling rate, release force, bond uniformity, bond alignment, bonding pressure, etc. The WB process module 112 performs, at step 408, the control algorithm. The control algorithm may use the output of the inspection algorithm performed in the WB inspection module 110 and the data collected from the wafer bonding process 140 as input. The control algorithm may clean and preprocess the collected data to handle missing values, and outliers, and ensure uniformity, and the dataset is divided into training and testing sets. The key features are identified that significantly impact the bonding strength during the wafer bonding process, such as temperature ramp-up times, temperature gradient, heating element placement, holding temperature, contact area variation, thermal resistance variation, stress, and strain, etc. The control algorithm utilizes the training dataset to train a linear regression model, which establishes a linear relationship between the selected features and the target variable, such as the bonding strength for the wafer bonding process. The trained model is then evaluated using the testing dataset to ensure its predictive accuracy.

[HO] In some embodiments, metrics such as Mean Squared Error (MSE) or R-squared may be used to assess the performance of the linear regression model. The new data from the wafer bonding process 140 and the output of the inspection algorithm are inputted into the trained linear regression model, such as temperature ramp-up times, temperature gradient, heating element placement, holding temperature, contact area variation, thermal resistance variation, stress, and strain, etc. and the control algorithm outputs the predicted bonding strength value for the wafer bonding process.

[Hl] In some embodiments, new data during the process may be continuously inputted into the control algorithm, allowing the algorithm to dynamically adapt to changing conditions and optimize the acceleration for optimal results. For example, unoptimized control of ramp-up and holding temperature in the wafer bonder may cause uneven heating. The optimal ramp-up rate for wafer bonding depends on various factors, including the materials being bonded, the bonding process used, and the desired properties of the bonded structure. Faster ramp-up rates may shorten the overall processing time, leading to higher throughput and productivity and rapid heating may accelerate bonding mechanisms such as surface activation or diffusion bonding, which may enhance bonding strength. Rapid temperature changes may induce thermal stresses, which may lead to a higher risk of defects such as delamination, voids, or cracks at the bonding interface and may result in non-uniform temperature distribution across the wafer surface, which may lead to variations in bonding strength and quality. Slower ramp-up rates may minimize thermal stresses and allow more time for bonding mechanisms to activate, which may reduce the risk of defects improve bonding quality, and may enable more uniform temperature distribution across the wafer surface, which may promote consistent bonding strength and quality. Slower ramp-up rates may prolong the overall processing time, potentially reducing throughput and productivity, and may increase the risk of contamination or outgassing at the bonding interface, affecting bonding quality.

[112] In some embodiments, the control algorithm may forecast the temperature profile required to achieve a ramp-up time of 15 minutes to reach 300°C by considering factors such as the initial temperature, heating capacity, wafer characteristics, and environmental conditions to predict the optimal temperature ramp-up trajectory.

[113] In some embodiments, the control algorithm may continuously monitor temperature changes and compare them to the predicted temperature profile and may use sensor data and feedback mechanisms to assess the actual ramp-up time and temperature progression in real time. In some embodiments, the control algorithm may dynamically adjust heating parameters, such as power input, heating rate, and timing, to optimize the temperature ramp-up process and may leverage adaptive control system algorithms that respond to deviations from the predicted temperature profile and make proactive adjustments to ensure alignment with the target ramp-up time.

[114] In some embodiments, the control algorithm may learn from discrepancies between predicted and actual temperature profiles over time and refine its predictive models and control strategies by continuously adapting to changes in process conditions, equipment performance, and material properties to improve temperature ramp-up control and enhance bonding consistency and quality.

[115] In some embodiments, the control algorithm may integrate with wafer bonding equipment and process automation systems and may interface with temperature control units, heating elements, and other hardware components to implement precise temperature ramp-up profiles based on real-time predictions and feedback.

[116] Synthetic Data and AI-Generated Digital Twins Example 2: In some embodiments, the WB process module 112 may incorporate an Al-generated digital twin framework to simulate bonding dynamics under variable stack materials, clamp geometries, and thermal profiles, enabling predictive validation of new tool recipes before deployment. For example, a process engineer may prompt, “Simulate void formation risk for bonded Si-Si stacks using asymmetric clamp profiles,” while a tooling developer may request, “Predict bond uniformity if dwell time is reduced to 8 seconds for high-modulus substrates.” In response, the WB process module 112 may extract historical bond performance data from the WB process database 122, including pressure ramp rates, thermal gradients, and post-bond inspection results, and apply a generative model trained on bonding trajectories stored in the WB network database 126. Internally, the model may condition synthetic bonding simulations on physical parameters — such as layer modulus mismatch or ramp timing — and output digital twin predictions of spatial stress distributions, void propagation likelihood, or interface decohesion under each scenario. Outputs may include a synthetic wafer batch summary with predicted defect probability heatmaps and a confidence annotation, along with a citation referencing literature on delamination thresholds in thermocompression oxide bonds. By embedding digital twin simulations into the WB process module 112, the system may enable proactive recipe tuning and early risk detection without requiring physical wafer bonding, accelerating qualification cycles and protecting yield integrity. The WB process module 112 adjusts, at step 410, the control parameters of the wafer bonding process 140. For example, the WB process module 112 adjusts the temperature application, such as the heating profile, of the wafer bonding process. In some embodiments, the WB process module 112 may adjust the bonding force application, cooling rate, release force, bond uniformity, bond alignment, bonding pressure, etc.

[117] Edge Al and TinyML Example 2: In some embodiments, the WB inspection module 110 may employ a generative video Al model to simulate coating deformation and void nucleation during the bonding process based on early-cycle clamp behavior and initial temperature profiles, enabling predictive visualization of defect formation before full bonding completion. For example, a process engineer may request, “Simulate void pattern development if platen tilt exceeds 2 degrees Celsius during ramp-up,” while a tooling specialist may prompt, “Show probable interface damage if clamp overshoot occurs within the first 5 seconds.” In response, the WB inspection module 110 may gather early bonding cycle telemetry — such as platen force distribution, clamp angle, and initial heat ramp rate — from the pre-processing WB equipment 128, retrieve baseline inspection images and surface metadata from the WB inspection database 120, and use a generative video model trained on bond formation sequences stored in the WB network database 126 to synthesize predictive frame-by-frame outcomes. Internally, the system may apply spatiotemporal generation layers conditioned on tool-specific motion signatures and thermal inputs to forecast deformation artifacts — such as radial void rings or center-out delamination — based on current or hypothetical bonding dynamics. The output may include a synthetic video illustrating how interface disruption is likely to progress under the specified setup, along with a literature-sourced annotation linking early clamp instability to void onset in oxide wafer bonding. By embedding generative video forecasting into the WB inspection module 110, the system may offer engineers an anticipatory tool for evaluating bonding stability and preemptively refining tool alignment and ramp profiles.

[118] The WB process module 112 executes, at step 412, the wafer bonding process. For example, the wafer bonding process may include the control parameter settings outputted by the control algorithm being used by the controller of the wafer bonding process 140 equipment to set the temperature ramp-up rate and holding temperature during the wafer bonding process. The WB process module 112 may monitor the wafer bonding process of the wafer bonding process 140 equipment and perform the control algorithm in real-time to continuously adjust the control parameters. The WB process module 112 stores, at step 414, the data in the WB process database 122. The WB process module 112 stores the data in the WB process database 122, such as the data collected from the wafer bonding process 140, the output of the inspection algorithm, and the output of the control algorithm. The WB process database 122 may contain the wafer number or ID, the ramp-up time to 300 degrees Celsius in minutes, and the bonding strength in megapascals optimized through the use of Al in the control algorithm, etc. The WB process database 122 may contain the data for each wafer that is being processed by the wafer bonding process 140. The WB process module 112 returns, at step 416, to the base module 108.

[119] The WB process module 112 performs a crucial operation at step 408 by carrying out the control algorithm, a complex procedure that has a big impact on the wafer bonding process's outcome. This control algorithm is skilled at using both data taken directly from the wafer bonding process and the deep insights offered by the output of the inspection algorithm from the WB inspection module 110. The first steps entail painstaking data preprocessing and cleaning, which is an essential step in guaranteeing the consistency and integrity of the dataset by resolving problems like outliers and missing values. Then, using this clean dataset, sets for training and testing are created, opening the door for a methodical examination of important factors that affect bonding strength, such as temperature ramp-up times, thermal gradients, heating element placement, variations in contact area, and thermal resistance. The core of the control system is the use of linear regression, a straightforward but effective machine learning technique. The curated dataset is used to train the model in order to find a linear relationship between the many carefully chosen characteristics and the final objective of reaching the ideal bonding strength. Instead of relying solely on intuition, the testing dataset is used to thoroughly assess the trained model's performance. Metrics like as Mean Squared Error (MSE) or R-squared are utilized to measure the model's predicted accuracy. When the model is judged to be reliable, it is tested using fresh data inputs that represent continuous fluctuations in the wafer bonding process, such as shifts in stress levels and temperature. The control method can forecast the final bonding strength using linear regression, which makes it a useful prediction tool for optimizing the wafer bonding procedure in the quest for excellence. This algorithm is unique in that it can dynamically adapt to changing conditions in the industrial environment by absorbing continuous data streams and recalibrating forecasts and suggestions in real-time. Beyond simple prediction, the control algorithm is a proactive approach to wafer bonding process optimization. The quality and consistency of the bonding, for example, can be greatly impacted by knowing and managing the temperature ramp- up rates. The algorithm makes recommendations for changes to reduce the likelihood of faults like delamination or to guarantee even heating throughout the wafer.

[120] The algorithm's iterative learning approach enables it to gradually improve its control methods and predictive models by utilizing feedback mechanisms and real-time data to continuously improve the accuracy and efficiency of the wafer bonding process. The control algorithm at step 408 is essentially a symbiosis of practical application and data-driven insights, allowing the WB process module 112 to negotiate the wafer bonding process's complexity with previously unheard-of precision. The method assists in the careful adjustment of control parameters by utilizing the intelligent application of linear regression and continuous learning, highlighting the transformational potential of basic Al algorithms in manufacturing excellence.

[121] FIG. 5 illustrates the WB post-process module 114. The process begins with the WB postprocess module 114 being initiated, at step 500, by the base module 108. The WB post-process module 114 connects, at step 502, to the wafer bonding post-process 150. The WB post-process module 114 may connect to the wafer bonding post-process 150, such as the annealing process, thin film removal, inspection and testing, etc. The WB post-process module 114 collects, at step 504, the post-process data. The WB post-process module 114 may collect the parameter data, such as testing, etc. The post-process may involve parameters such as bonding strength evaluation, void detection, surface roughness measurement, wafer warpage assessment, residual stress, defect density, etc. The WB post-process module 114 extracts, at step 506, the data from the WB inspection database 120 and WB process database 122. The WB post-process module 114 extracts the data, such as material data, parameter data, etc., of the pre-process and the wafer bonding process. [122] Multimodal Al Example 2: In some embodiments, the WB post-process module 114 may incorporate a multimodal Al model to correlate delamination risk with post-bond surface imagery, cure cycle telemetry, and fixture-level metadata in order to identify failure modes linked to toolspecific or material-specific conditions. For example, a reliability engineer may ask, “What tool and temperature signatures align with the film separation seen on Lot G12?” while a system planner may prompt, “Identify recurring patterns in delamination failures for bonded Si-Si stacks under long dwell profiles.” In response, the WB post-process module 114 may retrieve surface image maps captured via the post-processing WB equipment 148, align those with thermal ramp logs and chamber pressure curves from the WB post-processing database 124, and integrate relevant tool configuration metadata from the WB network database 126. The multimodal model may synchronize visual, thermal, and parametric features to detect latent relationships — such as tool- induced overpressure combined with ramp-down asymmetry causing stress accumulation at the bond interface. Internally, the system may apply cross-modal embedding alignment to fuse image feature maps with time-series sensor trends and perform similarity analysis against historical defect templates. Outputs may include a ranked risk explanation associating specific fixture IDs and cure conditions with delamination clusters, and a literature-supported recommendation noting that asymmetric ramp-down has been linked to interfacial decohesion in thermocompression oxide bonds. By embedding multimodal Al in the WB post-process module 114, the system may more effectively trace root causes of latent failures that emerge after bonding, improving both yield diagnostics and fixture qualification workflows.

[123] Generative Video Al Example 2. In some embodiments, the WB post-process module 114 may integrate a generative video Al model to simulate time-evolved visualizations of post-bond defect emergence — such as void expansion or delamination propagation — under varying thermal and mechanical conditions. For example, a materials engineer may prompt, “Visualize how void growth might evolve if ramp-down is shortened by 30 seconds,” while a process integrator may request, “Simulate delamination spread in a bonded stack with non-uniform pressure zones.” In response, the WB post-process module 114 may retrieve image sequences and stress distribution maps from the post-processing WB equipment 148, pair these with thermal and pressure cycle data from the WB post-processing database 124, and use a generative video model trained on historical bond degradation trajectories from the WB network database 126 to synthesize forward -projected visualizations of defect evolution. Internally, the model may apply temporally conditioned diffusion or GAN architectures to produce frame sequences reflecting predicted deformation or delamination propagation paths based on the specified process modifications. Outputs may include a video forecast showing how minor asymmetries in pressure or thermal ramp could lead to bond layer separation over time, accompanied by a reference to literature demonstrating that rapid temperature transitions exacerbate interfacial stress in thermocompression oxide bonding. By embedding generative video modeling into the WB post-process module 114, the system may offer proactive visualization tools for risk assessment and parameter tuning, enabling virtual evaluation of bonding failure modes without sacrificing wafers. The WB post-process module 114 performs, at step 508, the post-process algorithm. The WB post-process module 114 may perform a postprocess algorithm that uses machine learning algorithms to analyze parameters related to defects, such as warpages. The post-process algorithm may focus on the process data outputted from the inspection algorithm performed in the WB inspection module 110 and the effectiveness of the parameters outputted by the control algorithm performed in the WB process module 112 during the wafer bonding process. The post-process algorithm may adjust control parameters of the current microLED processed wafers in the process to result in an improved wafer bonding process. For example, the post-process algorithm may optimize temperature and pressure control, alignment accuracy, bonding force and time, surface preparation and cleanliness, adhesive material selection, post-bonding annealing or curing, etc. to minimize the warpage of the wafer. The post-bonding annealing or curing may address any residual stresses that may have been introduced during the bonding process and may provide stress relaxation, improved flatness, enhanced adhesion, etc.

[124] In some embodiments, the post-bonding treatment may streamline the process, and provide precise control and customization. In some embodiments, the post-process algorithm may minimize the warpage of the wafer by determining the annealing temperature, and annealing duration, providing uniform heating, and incorporating the material properties of the adhesives and intermediate layers used in the bonding process.

[125] In some embodiments, the post-process algorithm may gradually increase and decrease the temperature during the annealing process to minimize thermal shock and reduce the risk of inducing additional stress or warpage. In some embodiments, the post-process algorithm may perform the annealing process in a controlled atmosphere, such as inert gas, which may prevent oxidation or other chemical reactions that might affect the microLEDs or the bonding materials.

[126] In some embodiments, the post-process algorithm may integrate the cumulative thermal budget and the interactions between different materials and processes used in microLED manufacturing.

[127] In some embodiments, the post-process algorithm may control the annealing temperature in the wafer bonding post-process 150 to optimize process parameters in real time to achieve desired material properties and minimize defects. [128] In some embodiments, the post-process algorithm may use machine learning algorithms to analyze parameters related to warpage, such as annealing temperature, temperature ramp-up, temperature cool-down, atmosphere control, annealing temperature, etc. The post-process algorithm may focus on the process data outputted from the inspection algorithm performed in the WB inspection module 110 and the effectiveness of the parameters outputted by the control algorithm performed in the WB process module 112 during the wafer bonding process. The postprocess algorithm may adjust control parameters of the current microLED processed wafers in the process to result in an improved wafer bonding process. For example, the post-process inspection may result in a warpage identified in the wafers. The post-process algorithm may provide the data to the WB process module 112 to allow Al-controlled monitoring to result in minimized warpage by further adjusting the ramp-up temperature of the equipment during the wafer bonding process. The WB post-process module 114 performs, at step 510, the post-process inspection. The WB postprocess module 114 may involve collecting parameter data such as bonding strength evaluation, void detection, surface roughness measurement, wafer warpage assessment, residual stress, defect density, etc. For example, the post-process may include the wafers undergoing annealing treatments at elevated temperatures to improve bond strength and relieve any residual stresses introduced during the bonding process. In some embodiments, thin films or sacrificial layers used during the bonding process may need to be removed which may be achieved through techniques like chemical etching or mechanical polishing.

[129] In some embodiments, inspection, and testing may be conducted to evaluate the quality of the bonded wafers, and parameters such as bond strength, interface uniformity, and defect density may be measured using techniques like scanning electron microscopy, atomic force microscopy, and optical profilometry. The WB post-process module 114 stores, at step 512, the data in the WB post-process database 124. The WB post-process module 114 stores the data in the WB postprocess database 124, such as the data collected from the wafer bonding post-process 150, the output of the post-process algorithm, etc. The WB post-process database 124 may contain the wafer number or ID, the outputted annealing temperature of the post-process algorithm, the warpage of the wafer measured in microns, etc. The WB process database 122 may contain the data for each wafer that is being processed by the wafer bonding post-process 150. The WB post-process module 114 returns, at step 514, to the base module 108.

[130] The Random Forest algorithm is a viable Al technique that could be used in step 508, where the WB post-process module 114 is tasked with optimizing the wafer bonding process to avoid defects like warpages. Within the machine learning field, this approach is well-known for its excellent accuracy and robustness while handling complicated datasets. In order for Random Forest to function, it builds several decision trees during the training stage and outputs the mean prediction (regression) or mode of the classes (classification) for each tree in response to an input. Because of its ensemble method, it is quite effective in handling various data-related issues, such as overfitting, which is a prevalent problem with decision tree algorithms. Step 508 would require Random Forest to follow a few crucial steps. First, the process data produced by the inspection algorithm in the WB inspection module 110 and the efficacy of the parameters produced by the control algorithm in the WB process module 112. would be entered. The parameters pertaining to temperature and pressure management, alignment precision, bonding force and duration, and any modifications made during the surface preparation and cleaning stage, in addition to adhesive material selection and post-bonding annealing or curing parameters, may all be included in this data ensemble. Then, in order to provide a clean, consistent dataset that is ready for analysis, the Random Forest method would preprocess this input data to manage any abnormalities, such as missing values or outliers. The system finds patterns and connections between the different process parameters and the occurrence of warpage in the wafers by examining this dataset. This analysis is not simple; it includes a wide range of factors from the whole wafer bonding process, such as the rates at which the temperature ramps up and cools down, the control of the environment during the annealing process, and the total thermal budget over the course of the manufacturing process. Following the training phase, the Random Forest algorithm uses its learning to predict the possibility of warpage under different scenarios or recommend the best parameters to reduce warpage. For instance, it can suggest particular annealing temperatures and times that offer consistent heating while accounting for the adhesives' and the intermediate layers' material characteristics during the bonding process. It might also recommend annealing at gradually varying temperatures to avoid thermal shock and lower the chance of causing more stress or warpage. The program may also dynamically adjust to changes in processes occurring in real time. For example, the Random Forest algorithm can examine the data, pinpoint the causes of warpage found during the post-process inspection, and suggest changes to the ramp-up temperature or other parameters for subsequent wafers. This flexibility guarantees ongoing progress in the wafer bonding procedure, utilizing Al-managed observation to minimize warpage and increase bonding quality. The Random Forest method works well in this situation because it can handle high-dimensional input and produce precise, trustworthy predictions. Because of this capabilities, it can provide insightful analysis on how to best optimize the wafer bonding process, which makes it a crucial instrument for producing microLED production results that meet high standards. By putting such an algorithm into practice, the WB post-process module 114 gives the wafer bonding process the opportunity to make data-driven decisions that increase process reliability, lower errors, and improve overall device performance. [131] FIG. 6 illustrates the WB historical module 116. The process begins with the WB historical module 116 being initiated, at step 600, by the base module 108. The WB historical module 116 connects, at step 602, to the wafer bonding pre-process 130, the wafer bonding process 140, and the wafer bonding post-process 150. The historical module 116 connects to the wafer bonding pre- process 130, such processes, such as surface preparation, alignment, surface activation, etc. and equipment, such as plasma cleaners, wet bench, alignment systems, etc., the wafer bonding process 140, such as systems or equipment for automated wafer bonding of microLED chips that brings the prepared surfaces of the LED wafer and the substrate wafer into intimate contact under controlled conditions of temperature, pressure, and atmosphere, and the wafer bonding postprocess 150, such processes and methods after the completion of the wafer bonding process 140, such as an annealing process, thin film removal, inspection and testing, etc. The WB historical module 116 aggregates, at step 604, the data from the various types of processes and stores the data in the WB network database 126. The WB network database 126 may contain the historical data from the various processes performed by the wafer bonding pre-process 130, wafer bonding process 140, and wafer bonding post-process 150. The WB network database 126 may contain the data parameters collected during the inspection process, such as surface contamination, surface roughness, wafer bow and warp, thermal expansion mismatch, chemical incompatibility, alignment accuracy, void formation, electrical issues, etc., control parameters of the process, such as temperature application, bonding force application, cooling rate, release force, bond uniformity, bond alignment, bonding pressure, etc., and data parameters collected during post-processing, such as bonding strength evaluation, void detection, surface roughness measurement, wafer warpage assessment, residual stress, defect density, etc. The post-process may include using an annealing furnace, chemical etching systems, metrology tools, etc. The WB historical module 116 performs, at step 606, the historical machine learning algorithm on the historical data stored in the WB network database 126. For example, the historical machine learning algorithm may use the historical data stored in the WB network database 126 to determine trends to understand how varying parameters affect the final quality and identify optimal parameter settings to create the best outcomes. The historical machine learning algorithm adjusts the parameters in real time for the WB inspection module 110, WB process module 112, and WB post-process module 114 to correct any defects. For example, the data may show defects in the wafers that occurred during the wafer bonding process. The historical machine learning algorithm may minimize the defects in the wafer by adjusting the power and time settings of the plasma cleaning process based on the defect density inputted in the WB inspection module 110. The historical machine learning algorithm may also adjust the ramp-up time to temperature occurring in the WB process module 112 to optimize the bonding strength and adjust the annealing temperature for the WB post-process module 114 to minimize warpage.

[132] In some embodiments, the historical machine learning algorithm may optimize the plasma pre-cleaning process to reduce void formation and enhance wafer bonding by analyzing historical data and identifying correlations between the pre-clean conditions and void formation rates.

[133] In some embodiments, the historical machine learning algorithm may be a predictive model, such as a regression model or neural network, to predict the optimal plasma pre-clean parameters based on the material type, wafer characteristics, and desired bonding outcomes.

[134] In some embodiments, the historical machine learning algorithm may be a support vector machine in which wafer surfaces are classified based on pre-clean effectiveness to predict whether a given set of pre-clean parameters will lead to successful bonding. [135] In some embodiments, the historical machine learning algorithm may be a decision tree and random forest model in which the impact of various pre-clean parameters on void formation is analyzed to provide process adjustments.

[136] In some embodiments, the historical machine learning algorithm may perform a principal component analysis to identify the most critical pre-clean parameters that affect void formation from a large dataset.

[137] In some embodiments, the historical machine learning algorithm may optimize temperature ramping profiles to enhance wafer bonding quality by using time-series analysis and clustering algorithms to identify successful temperature profiles and conditions that lead to high-quality bonding. The historical machine learning algorithm may predict the ideal temperature profile for a given set of wafer materials and bonding agents.

[138] In some embodiments, the historical machine learning algorithm may use a control algorithm, such as a PID controller augmented with Al to adjust the heating profiles in real-time based on ongoing performance data and model predictions.

[139] In some embodiments, the historical machine learning algorithm may be a gradientboosting machine to predict optimal temperature profiles based on past successful bonds and provide high accuracy by learning from errors in previous predictions.

[140] In some embodiments, the historical machine learning algorithm may be a recurrent neural network and long short-term memory network to predict the outcomes of bonding processes over time. [141] In some embodiments, the historical machine learning algorithm may perform Bayesian optimization to efficiently search the space of temperature profiles to find the optimal settings by balancing the exploration of new parameters with the exploitation of known data.

[142] In some embodiments, the historical machine learning algorithm may control postannealing and curing temperatures and times to reduce warpage by utilizing regression analysis or neural networks to establish relationships between anneal or curing conditions and warpage outcomes. For example, the historical machine learning algorithm may implement an optimization framework, such as genetic algorithms, simulated annealing, etc. to determine the best anneal or curing parameters for minimizing warpage while maintaining or enhancing bond quality. In some embodiments, the historical machine learning algorithm may perform genetic programming for evolving expressions or algorithms that best describe the relationship between annealing or curing conditions and warpage, enabling dynamic adaptation of process parameters. In some embodiments, the historical machine learning algorithm may perform deep reinforcement learning to develop policies that dynamically adjust annealing and curing parameters in response to realtime feedback on warpage, learning optimal strategies over time. In some embodiments, the historical machine learning algorithm may perform gaussian process regression to model the uncertainty in the relationship between process conditions and warpage outcomes, allowing for more robust optimization under uncertainty.

[143] LLM Example 1: In some embodiments, the WB historical module 116 may employ a large language model (LLM) to automatically parse and synthesize unstructured process logs, technician notes, and support documentation stored within the WB network database 126, thereby enabling higher-level correlation analysis between subjective observations and objective bond outcomes. For example, a process engineer may ask, “What procedural notes are linked to bond strength outliers last quarter?” while an integration specialist may prompt, “Summarize thermal expansion mismatch mitigation tactics used on Line 4 in the past six months.”

[144] In response, the WB historical module 116 may retrieve unstructured records — including maintenance annotations, operator logs, and post-process comments — associated with wafer bonding process 140 and wafer bonding post-process 150, tokenize and embed the entries, and map them against structured defect patterns and outcome data using attention-weighted correlation scoring. Internally, the LLM may classify language segments into procedural domains (e.g., ramp control, cleaning variability, fixture tensioning), extract references to actionable process changes, and rank associations based on co-occurrence with defect tags such as “void formation” or “non- uniform bond strength.” Output may include a natural language summary of operational behaviors most frequently co-incident with critical defects, and a citation recommending parameter constraints consistent with published wafer bonding guidelines. By leveraging LLMs to unify narrative data with time-series process outcomes, the WB historical module 116 may unlock causal insight from otherwise siloed documentation, driving more informed adjustments and accelerating corrective feedback across the wafer bonding workflow. The WB historical module 116 sends, at step 608, the process adjustments to the WB inspection module 110, the WB process module 112, and the WB post-process module 114. The historical machine learning algorithm may minimize the defects occurring in the wafer by adjusting the plasma cleaning parameters of the wafer bonding pre-process based on the defect density determined in the WB inspection module 110. The historical machine learning algorithm may also adjust the temperature ramp-up time in the WB process module 112 to optimize bonding strength for the WB post-process module 114 to inspect. The adjustments are sent to the WB inspection module 110 and WB process module 112 to create a more consistent final product which is then verified by the WB post-process module 114. The

WB historical module 116 returns, at step 610, to the base module 108.

[145] FIG. 7 illustrates the WB integration module 118. The process begins with the WB integration module 118 being initiated, at step 700, by the base module 108. The WB integration module 118 performs, at step 702, the integration machine learning algorithm. The integration machine learning algorithm may utilize data across the different stages of the wafer bonding process to enhance the overall process efficiency and outcome quality. For example, the integration machine learning algorithm may utilize insights from void formation data and plasma pre-clean parameters to optimize post-anneal and curing processes for reduced warpage. The integration machine learning algorithm may perform deep learning, such as a convolutional neural network, that analyzes historical data on void formation rates and preclean conditions to identify patterns where specific types of voids or contamination levels correlate with increased warpage postannealing. The integration machine learning algorithm may predict post-process warpage based on the pre-clean data and adjust post-anneal and curing temperatures and times dynamically to target specific conditions that have historically led to less warpage which enables proactive adjustments to post-process parameters based on early-stage data to reduce the need for extensive post-process corrections and improved wafer quality. For example, the integration machine learning algorithm may leverage temperature profile data from the wafer bonding process to fine-tune post-anneal and curing parameters for optimal stress relief and minimal warpage.

[146] The integration machine learning algorithm may perform regression analysis and gradient boosting and collect and analyze data on successful temperature ramps and outcomes to identify which profiles lead to the best bonding quality with minimal internal stresses that may lead to warpage. The integration machine learning algorithm may use regression analysis and gradient boosting to create a predictive model that recommends post-anneal and curing settings based on the heating profiles used during wafer bonding and adjust post-process conditions in real-time based on the recommendations which ensure that the post-process annealing and curing are tailored to the conditions experienced during bonding to improve bond integrity and reduce warpage. For example, the integration machine learning algorithm may apply insights from post-process anneal and curing outcomes to optimize the plasma preclean step to reduce the void formation and enhance initial bonding quality. The integration machine learning algorithm may perform reinforcement learning which learns from the relationship between post-process outcomes, such as warpage, anneal success, etc., and pre-process conditions and may use feedback from post-process evaluations to suggest adjustments in the plasma preclean parameters. The integration machine learning algorithm may dynamically adjust preclean parameters such as plasma power, duration, and gas mix based on what has been learned about their impact on final wafer quality and creates a feedback loop where post-process performance directly informs pre-process setup which allows for continuous improvement of the preclean process based on comprehensive outcomes analysis and leads to better initial bonding quality and reduced correction requirements later in the process.

[147] In some embodiments, the integration machine learning algorithm may use the data stored in the WB network database 126 to perform a predictive modeling algorithm to make predictions or forecasts of the final product based on historical data and patterns from the previously created products. Predictive modeling involves the systematic analysis of historical data to identify patterns, correlations, and trends that can be used to build models capable of making accurate predictions about future outcomes. For example, the integration machine learning algorithm may be used for minimizing warpage across the wafer bonding pre-process 130, wafer bonding process 140, and wafer bonding post-process 150. The historical data in the WB network database 126 may include information on warpage and relevant operational parameters such as void formation, preclean parameters, temperature profile data, annealing and curing temperature and times, etc. The data is cleaned and preprocessed, addressing any missing values and ensuring that all variables are in a suitable format for modeling. This may involve normalization or scaling of numerical features. The dataset is split into training and testing sets. The training set is used to train the predictive model, and the testing set assesses its performance on unseen data. The model used may be linear regression, decision trees, ensemble methods, neural networks, etc. Then key features are identified that influence the warpage and engineer new features if necessary. For example, the interaction between pre-clean parameters and temperature profiles may have a significant impact on the warpage. The selected predictive model is trained using the training dataset. The model learns patterns and relationships between operational parameters and warpage. The model's performance is validated on a separate validation dataset and hyperparameters are fine-tuned to optimize its accuracy and generalization. The model is then evaluated and deployed by being sent to the wafer bonding pre-process 130, wafer bonding process 140, and wafer bonding post-process 150.

[148] Al Agents Example 2: In some embodiments, the WB integration module 118 may incorporate a persistent Al agent configured to autonomously coordinate real-time wafer bonding parameter adjustments across the WB inspection module 110, WB process module 112, and WB post-process module 114 in response to evolving defect and material condition trends. For example, an integration engineer may issue a directive such as, “Initiate tool-level force tuning if void rate increases on bonded pairs with low oxide thickness,” while the agent may independently trigger, “Reduce ramp rate on Tool 3 if thermal delta exceeds 5°C across platen zones over three lots.” In operation, the Al agent may continuously monitor image-based defect tags from the WB inspection database 120, temperature and force logs from the WB process database 122, and delamination or warpage alerts from the WB post-processing database 124. Upon detecting a pattern consistent with latent yield degradation, the agent may invoke the integration machine learning algorithm at step 702, generate updated platen force, ramp rate, or dwell time parameters, and dispatch them through the WB integration module 118. Internally, the agent may prioritize triggers using reinforcement-learned policies and defect trajectory forecasts grounded in patterns from the WB network database 126. Outputs may include an immediate parameter update with embedded rationale and a shift-level report that logs detected correlations and decision logic. By embedding an autonomous Al agent into the WB integration module 118, the system may deliver proactive, cross-module optimization in response to complex real-time process signals, improving bonding uniformity and defect suppression without operator delay. The WB integration module 118 connects, at step 704, to the wafer bonding pre-process Al module 136, the wafer bonding process Al module 146, and the wafer bonding post-process Al module 156. The WB integration module 118 connects to the wafer bonding pre-process 130, such as processes, such as surface preparation, alignment, surface activation, etc. and equipment, such as plasma cleaners, wet bench, alignment systems, etc., the wafer bonding process 140, such as systems or equipment for automated wafer bonding of microLED chips that brings the prepared surfaces of the LED wafer and the substrate wafer into intimate contact under controlled conditions of temperature, pressure, and atmosphere, and the wafer bonding post-process 150, such processes and methods after the completion of the wafer bonding process 140, such as an annealing process, thin film removal, inspection, and testing, etc.

[149] LLM Example 2 In some embodiments, the WB integration module 118 may incorporate a large language model (LLM) to generate human-readable rationales for machine-learned bond parameter adjustments by translating data correlations and decision traces into operator-facing explanations. For example, a line engineer may prompt, “Why were clamp force and platen temperature modified for Lot 682?” while a process trainer may request, “Summarize recent parameter changes and their rationale for operator onboarding.” In response, the WB integration module 118 may retrieve the control recommendations produced in step 702, correlate them with structured outcomes from the WB process database 122 and inspection trends from the WB inspection module 110, and use the LLM to generate plain-language narratives explaining how those inputs led to the selected optimization path. Internally, the model may align parameter deltas with time-series bonding trends, extract causal markers from historical data in the WB network database 126, and compile these into decision traces grounded in prior success rates and defect avoidance logic. Outputs may include a per-lot rationale card detailing the specific sensor patterns and historical precedents that triggered a parameter shift, as well as a contextual training summary citing literature-supported relationships — such as the effect of platen overshoot on void suppression in thermocompression bonds. By embedding an LLM into the WB integration module 118, the system may improve transparency and onboarding efficiency while reducing operator hesitation in the face of Al-driven process changes. The WB integration module 118 sends, at step 706, the process adjustments to the wafer bonding pre-process Al module 136, the wafer bonding process Al module 146, and the wafer bonding post-process Al module 156. For example, the WB integration module 118 may send the void formation, pre-clean parameters, etc. adjustments to each of the processes to minimize the warpage of the wafer during the wafer bonding process. In some embodiments, the WB integration module 118 may send the integration machine learning algorithm to the processes, allowing the systems to further enhance the optimization. The WB integration module 118 returns, at step 708, to the base module 108. [150] In order to increase process efficiency and product quality, the WB integration module 118 performs an integration machine learning algorithm at step 702 that synthesizes data from different phases of the wafer bonding process. Gradient Boosting is a suitable illustration of a straightforward but effective Al technique for this task. A versatile machine learning methodology, gradient boosting generalizes conventional boosting methods by allowing optimization of an arbitrary differentiable loss function, while still building models in a stage-wise manner. Gradient Boosting has the potential to optimize and inform subsequent stages of the wafer bonding process by utilizing a wide range of data points, including plasma pre-clean parameters and void formation metrics collected during the early stages of the process. Here, the objective would be to reduce warpage, a frequent flaw that has a big influence on the quality of the finished product. First, a training phase would be conducted via the Gradient Boosting-based integration machine learning method. It would examine past data on void formation rates, pre-clean conditions, and their relationship to post-annealing warpage results during this phase. The algorithm could then make educated guesses about how to modify post-anneal and curing temperatures and durations for new wafers, trying to replicate conditions that have historically resulted in less warpage, by identifying patterns in how particular types of voids or contamination levels lead to increased warpage. Through iteratively fixing mistakes from earlier models, this predictive model refines its predictions over time. Every new tree that is constructed fixes the mistakes in the previously assembled collection of trees. This procedure is carried out repeatedly until either a predetermined number of trees are planted or no appreciable progress can be made. In practical terms, this means that the algorithm may modify post-process parameters for the wafer bonding process dynamically based on previous insights and real-time data. This would allow for proactive adjustments that would lessen the need for significant post-process repairs. Furthermore, Gradient Boosting can be customized to forecast ideal temperature profiles for the post-annealing and curing phases in accordance with the particular circumstances encountered throughout the bonding procedure. This involves making adjustments for things like the heating profiles that are applied during the bonding of wafers, which are essential for maintaining bond integrity and minimizing warpage. The program can offer post-anneal and curing settings that result in the best bonding quality with the least amount of internal strains by examining successful temperature ramps and their results. The Gradient Boosting-based integrated machine learning technique, in its simplest form, is a comprehensive method for wafer bonding process optimization. It uses gradient boosting and regression analysis to anticipate post-process conditions and dynamically modify them in real-time based on collected data. This minimizes warpage and improves the overall quality of the microLED made wafers by making sure that every step of the wafer bonding process — from plasma pre-clean to post-annealing and curing — is precisely calibrated to produce the best results. The datasets would need to be extensive, spanning multiple production stages, in order to apply Gradient Boosting to the optimization of the wafer bonding process, specifically focusing at eliminating defects like warpage and boosting overall process efficiency. These are some instances of pertinent datasets. (1) Dataset of pre-cleaned parameters. The plasma power, duration, gas mix, pressure, and other details of the plasma pre-cleaning procedure are all included in detail in this dataset. Additionally, it would monitor the types of pollutants eliminated, the levels of contamination before and after cleaning, and any differences in the cleaning procedure between wafer batches. (2) Dataset for Wafer Bonding Process. Temperature ramp-up durations, applied pressure, bonding time, temperature gradient, heating element positioning, and ambient conditions (e.g., atmospheric composition) are examples of critical factors that would be included throughout the wafer bonding phase. Additionally important are data on heat resistance variance, contact area variation, and alignment precision. (3) Annealing and Curing Dataset Post-process. Details on post-bonding treatments, such as temperature ramp-up and cool-down rates, time intervals, and inert gas presence or annealing or curing temperatures. The results of these processes in terms of bonding strength, warpage, and residual stresses may also be included in this dataset. (4) The dataset of defects and outcomes, thorough documentation of all flaws found, including delamination, warpage, void formation, and any mechanical or electrical malfunctions. Measurements of bonding strength, interface uniformity, and any post-process quality evaluations like as atomic force microscope (AFM) scans or scanning electron microscopy (SEM) pictures would also be included in this collection. (5) Material Properties Dataset: Details on the materials that are utilized in the bonding process, including as the wafer and substrate materials, the adhesive materials, and any layers that are intermediate. Relevant characteristics include modulus of elasticity, thermal expansion coefficients, and response to various pressures and temperatures. The operational parameters dataset, comprising calibration data, maintenance records, and any alterations or adjustments made to the process parameters over time, comprises information about the overall operational parameters of the equipment utilized in each stage of the process. Dataset on Environmental Conditions (7). Records of the temperature, humidity, and any variations in these factors that could have an impact on the process results during each stage of the wafer bonding process. Before these datasets were divided into training and testing sets for the Gradient Boosting model, they would be preprocessed to address missing values, outliers, and guarantee homogeneity. The main characteristics found in these datasets would have a big influence on how well the model learned the intricate connections between process variables and results. Through iterative learning from this data, gradient boosting will increase its predictive accuracy over time and offer useful insights to improve the wafer bonding process for lower faults and more efficiency. [151] FIG. 8 illustrates the WB inspection database 120. The WB inspection database 120 provides an example of the results of the inspection algorithm performed in the WB inspection module 110. The WB inspection database 120 contains the wafer number or ID, the defect density before the cleaning process, the defect density after the cleaning process with the parameters optimized by the inspection algorithm, the delta or effectiveness of the cleaning process, etc. In some embodiments, the WB inspection database 120 may include other parameter data collected or calculated during the WB inspection module 110 process, such as surface contamination, surface roughness, wafer bow and warp, thermal expansion mismatch, chemical incompatibility, alignment accuracy, void formation, electrical issues, etc.

[152] FIG. 9 illustrates the WB process database 122. The WB process database 122 provides an example of the results of the control algorithm performed in the WB process module 112. The WB process database 122 contains the wafer number or ID, the ramp-up time to 300 degrees Celsius in minutes, and the bonding strength in megapascals optimized through the use of Al in the control algorithm, etc. The WB process database 122 may contain the data for each wafer that is being processed by the wafer bonding process 140. In some embodiments, the WB process database 122 may store other parameter data collected or calculated from the WB process module 112, such as temperature application, bonding force application, cooling rate, release force, bond uniformity, bond alignment, bonding pressure, etc.

[153] FIG. 10 illustrates the WB post-process database 124. The WB post-process database 124 provides an example of the results of the post-process algorithm performed in the WB post-process module 114. The WB post-process database 124 contains the wafer number or ID, the outputted annealing temperature of the post-process algorithm, the warpage of the wafer measured in microns, etc. The WB process database 122 may contain the data for each wafer that is being processed by the wafer bonding post-process 150. In some embodiments, the WB post-process database 124 may store other parameter data collected or calculated from the WB post-process module 114, such as bonding strength evaluation, void detection, surface roughness measurement, wafer warpage assessment, residual stress, defect density, etc.

[154] The functions performed in the processes and methods may be implemented in differing order. Furthermore, the outlined steps and operations are only provided as examples, and some of the steps and operations may be optional, combined into fewer steps and operations, or expanded into additional steps and operations without detracting from the essence of the disclosed embodiments.

Claims

CLAIMS What is claimed is:

1. An automated wafer bonding (WB) method for microLED chips system, comprising: an automated wafer bonding inspection module that measures at least one parameter from a first group of parameters and calculates an output of at least one parameter from a second group of parameters to help control the automated wafer bonding processing module; an automated wafer bonding processing module which inputs the at least one parameter from said second group to control the wafer bonding using at least one parameter from a third group of parameters; an automated wafer bonding post-processing module which measures at least one parameter from a fourth group of parameters that determines how well the parameters of said first and second group were running; an automated wafer bonding historical module which calculates a plurality of group parameters based upon parameters of the first group, second group, third group, and fourth group; and an automated wafer bonding integration module to update the automated wafer bonding processing module based upon the results of the automated wafer bonding historical module.

2. The method of claim 1, wherein the WB process module acts as the core controller, utilizing inputs from the WB inspection module to govern the wafer bonding process to ensure optimal heating profiles, bonding force application, cooling rate, release force, bond uniformity and bond alignment.

3. The method of claim 1, wherein the WB post-process module conducts inspections to assess the execution of parameters and identify potential deviations, and the WB historical module analyzes data collected across various stages, allowing for continuous learning and pattern recognition and subsequently a WB integration module then employs historical insights to update and optimize the process, enabling adaptive improvements over time.

4. The method of claim 1, wherein a communication interface, which is a hardware or a software component that enables communication between a WB Al network and a wafer bonding pre- process, the wafer bonding process, and a wafer bonding post-process.

5. The method of claim 4, wherein the communication interface comprises a set of protocols, rules, and standards that define how information is transmitted and received between devices is a physical connector, wireless network, or software application and include components such as drivers, software libraries, and firmware that are used to control and manage the communication process.

6. The method of claim 4, wherein the communication interface is compatible with a USB, a

Bluetooth, or a Wi-Fi and the communication interface communicates with a network.

7. The method of claim 1, wherein there is a memory that comprises suitable logic, circuitry, and/or interfaces that are configured to store a machine code and/or a computer program with at least one code section executable by a processor.

8. The method of claim 1, wherein there is a base module, which initiates the WB inspection module, the WB process module, the WB post-process module, the WB historical module, and the WB integration module.

9. The method of claim 8, wherein the WB inspection module begins by being initiated by the base module wherein the WB inspection module connects to the WB pre-process wherein the WB inspection module collects data from the WB pre-process.

10. The method of claim 8, wherein the WB inspection module performs an inspection algorithm, and the WB inspection module sends the determined process data from the inspection algorithm to the WB process module wherein the WB inspection module stores the data in the SC inspection database and SC inspection module returns to the base module.

11. The method of claim 8, wherein the WB process module begins by being initiated by the base module wherein the WB process module receives the process data, such as the data on the microLED processed wafer, from the WB inspection module and the WB process module connects to the WB process.

12. The method of claim 11, wherein the WB process module collects the data from the WB process and the WB process module performs a control algorithm and the WB process module adjusts the control parameters of the WB process and executes the WB process wherein WB process module stores the data in the WB process database and the WB process module returns to the base module.

13. The method of claim 8, wherein the WB post-process module begins by being initiated by the base module and the WB post-process module connects to the WB post-process wherein the WB post-process module collects the post-process data.

14. The method of claim 13, wherein the WB post-process module extracts the data from the WB inspection database and WB process database and performs the post-process algorithm subsequently the WB post-process module performs the post-process inspection wherein the WB post-process module stores the data in the WB post-process database and the WB post-process module returns to the base module.

15. The method of claim 8, wherein the WB historical module begins by being initiated by the base module and the WB historical module connects to the WB pre-process, the WB process, and the WB post-process wherein the WB historical module aggregates data from the various types of processes and stores the data in the WB network database.

16. The method of claim 15, wherein the WB historical module performs a historical machine learning algorithm on the historical data stored in the WB network database and sends the process adjustments to the WB inspection module, the WB process module, and the WB post-process module and subsequently the WB historical module returns to the base module.

17. The method of claim 8, wherein the WB integration module begins by being initiated by the base module and the WB integration module performs the integration machine learning algorithm and connects to the WB pre-process Al module, the WB process Al module, and the WB postprocess Al module.

18. The method of claim 17, wherein the WB integration module sends the process adjustments to the WB pre-process Al module , the WB process Al module , and the WB post-process Al module and the WB integration module returns to the base module.

19. The method of claim 10, wherein there is a WB inspection database which provides an example of the results of the inspection algorithm performed in the WB inspection module wherein the WB inspection database contains the wafer number or ID, a defect density before the cleaning process, a defect density after a cleaning process with the parameters optimized by the inspection algorithm, the delta or effectiveness of the cleaning process.

20. The method of claim 19, wherein the WB inspection database includes other parameter data collected or calculated during the WB inspection module’s process, such as surface contamination, surface roughness, wafer bow and warp, thermal expansion mismatch, chemical incompatibility, alignment accuracy, void formation and electrical issues.

21. The method of claim 12, wherein a WB process database contains the wafer number or ID, the ramp-up time to 300 degrees Celsius in minutes, and the bonding strength in megapascals optimized through the use of Al in the control algorithm, the data for each wafer that is being processed by the wafer bonding process. In some embodiments, the WB process database stores other parameter data collected or calculated from the WB process module, such as temperature application, bonding force application, cooling rate, release force, bond uniformity, bond alignment, bonding pressure.

22. The method of claim 21, wherein the WB process database stores other parameter data collected or calculated from the WB process module , such as temperature application, bonding force application, cooling rate, release force, bond uniformity, bond alignment and bonding pressure.

23. The method of claim 14, wherein a WB post-process database contains the wafer number or ID, the output annealing temperature of the post-process algorithm, the warpage of the wafer measured in microns, data for each wafer that is being processed by the wafer bonding postprocess. In some embodiments, the WB post-process database may store other parameter data collected or calculated from the WB post-process module, such as bonding strength evaluation, void detection, surface roughness measurement, wafer warpage assessment, residual stress, defect density, etc. .

24. The method of claim 23, wherein, the WB post-process database stores other parameter data collected or calculated from the WB post-process module, such as bonding strength evaluation, void detection, surface roughness measurement, wafer warpage assessment, residual stress and defect density.

25. The method of claim 16, wherein include a WB network database stores the historical data from the various processes performed by the wafer bonding pre-process, wafer bonding process, and wafer bonding post-process, the data parameters collected during the inspection process, such as surface contamination, surface roughness, wafer bow and warp, thermal expansion mismatch, chemical incompatibility, alignment accuracy, void formation, electrical issues, control parameters of the process, such as temperature application, bonding force application, cooling rate, release force, bond uniformity, bond alignment, bonding pressure.

26. The method of claim 25, wherein the wafer bonding process includes the bonding equipment, such as wafer bonders, vacuum chambers, or thermal bonding systems, and data parameters collected during post-processing, such as bonding strength evaluation, void detection, surface roughness measurement, wafer warpage assessment, residual stress and defect density.