TWI897769B - Method, device, apparatus and computer-readable storage medium for predicting wafer and surface nanomorphology thereof - Google Patents

Abstract

The present invention provides a method, apparatus, device, and computer-readable storage medium for predicting the nanotopography of a wafer and its surface, belonging to the field of semiconductor manufacturing technology. The method may include: detecting surface topography measurement data of the wafer during wafer processing; and predicting, based on the surface topography measurement data, the predicted surface topography value of the wafer corresponding to a set wavelength range after subsequent processing steps based on the surface topography measurement data based on a trained convolutional neural network (CNN) model.

Description

Method, device, apparatus and computer-readable storage medium for predicting wafer and surface nanomorphology thereof

Cross-references to related applications

This invention claims priority to Chinese Patent Application No. 202410956702.0 filed in China on July 17, 2024, the entire contents of which are incorporated herein by reference.

The present invention relates to the field of semiconductor manufacturing technology, and more particularly to a method, device, apparatus, and computer-readable storage medium for predicting the nanotopography of a wafer and its surface.

During the wafer manufacturing process, single-crystal silicon ingots are produced using the Czochralski method. These ingots then undergo a series of processes, including wire sawing, lapping, etching, grinding, and chemical mechanical polishing (CMP), ultimately resulting in single-crystal silicon wafers. For single-crystal silicon wafers, their surface morphology is a key quality parameter. The wafer's surface nanotopography (NT) is a key quality parameter within this morphology.

In related solutions, NT testing of wafers is typically performed after the final surface processing step, the CMP process. However, short-wavelength quality parameters such as nanotopography are largely determined by the processing steps prior to CMP. This means that during the implementation of these solutions, if problems arise in the pre-CMP process, they can only be discovered after the CMP process. Consequently, wafers with pre-CMP processing problems continue to undergo the full wafer manufacturing process, resulting in a waste of production resources.

In light of this, the present invention aims to provide a method, apparatus, device, and computer-readable storage medium for predicting wafer and surface nanotopography. These methods can predict wafer surface nanotopography parameters based on products produced after front-end processing in the wafer manufacturing process. This allows monitoring of product performance in the front-end processing flow based on the predicted nanotopography data, preventing defective wafers from the front-end processing flow from entering subsequent processing flows, thereby avoiding waste of production resources.

The technical solution of the present invention is implemented as follows:

In a first aspect, the present invention provides a method for predicting wafer surface nanotopography. The method comprises: detecting surface topography measurement data of the wafer during wafer processing; and predicting, based on the surface topography measurement data, predicted surface topography values corresponding to a predetermined wavelength range of the wafer after subsequent processing steps using a trained convolutional neural network (CNN) model.

In a second aspect, the present invention provides a device for predicting wafer surface nanotopography, comprising: a detection portion and a prediction portion; wherein the detection portion is configured to detect surface topography measurement data of the wafer during wafer processing; and the prediction portion is configured to predict, based on a trained convolutional neural network (CNN) model, predicted surface topography values corresponding to a set wavelength range of the wafer after subsequent processing based on the surface topography measurement data.

In a third aspect, the present invention provides a computing device comprising: a processor and a memory; the processor is configured to execute instructions stored in the memory to implement the wafer surface nanotopography prediction method described in the first aspect.

In a fourth aspect, the present invention provides a computer-readable storage medium storing at least one instruction. The at least one instruction is configured to be executed by a processor to implement the wafer surface nanotopography prediction method described in the first aspect.

In a fifth aspect, the present invention provides a wafer having a predicted nanotopography value of less than 5 nm at a 2 mm * 2 mm size and/or less than 10 nm at a 10 mm * 10 mm size. The wafer has a nanotopography value of less than 5 nm at a 2 mm * 2 mm size and less than 10 nm at a 10 mm * 10 mm size.

The present invention provides a method, apparatus, device, and computer-readable storage medium for predicting wafer and surface nanotopography. This method utilizes a trained CNN model and surface topography measurement data from a wafer after completing a current process to predict the wafer's predicted surface topography after completing a subsequent process. These predicted surface topography values not only enable monitoring of each process step to ensure the stability and consistency of the processing equipment used, but also monitor the products of each process step to ensure that qualified products advance to the next process step, thereby improving the microtopography quality of the wafer surface and preventing subsequent processes from producing products with a low probability of future qualification, thereby reducing processing costs. Furthermore, these predicted surface topography values can be used to determine whether there are any issues with the completed process step, assisting in process adjustments.

50: Prediction device

501: Detection section

502: Prediction Section

503: Training Section

504: Feedback Section

70: Computing equipment

710: Processor

720: Storage

Figure 1 is a flow chart of a method for predicting wafer surface nanotopography provided by the present invention.

Figure 2 is a schematic diagram of the CNN model architecture provided by the present invention.

Figure 3 is a schematic diagram of the transformation of the regression task into a classification task provided by the present invention.

Figure 4 is a schematic diagram of the loss error curve provided by the present invention.

Figure 5 is a schematic diagram of the composition of a device for predicting wafer surface nanotopography provided by the present invention.

Figure 6 is a schematic diagram of another device for predicting wafer surface nanotopography provided by the present invention.

Figure 7 is a schematic diagram of the structure of a computing device provided by the present invention.

The following will combine the drawings in this invention to clearly and completely describe the technical solutions in this invention.

In this approach, the wafer surface undergoing the CMP process is divided into analysis areas of a set size. The peak-to-valley values of the filtered measurement data at each sampling point within each analysis area are then ranked in ascending order. The value at the corresponding position in this ascending order is then selected according to a set percentile (e.g., 99.5%) as the wafer's nanotopography (NT) value.

From the perspective of the wafer surface, the filtered measurement data at each sampling point in each region will exhibit overall fluctuations in the measurement data at each sampling point. From a wave perspective, this fluctuation can be viewed as the result of the superposition of signals of different wavelengths. Different wavelengths correspond to the size of the area where the fluctuation occurs. Specifically, the smaller the signal wavelength, the smaller the area where the fluctuation occurs; the larger the signal wavelength, the larger the area where the fluctuation occurs. Among the fluctuations exhibited by measurement data, longer-wavelength signal data can include parameters such as bow and warp. These parameters are largely finalized after the wire cutting process and are minimally affected by subsequent processing steps. However, shorter-wavelength signal data, such as nanoscale topography, are affected by all front-end processing steps prior to the CMP process.

Furthermore, as semiconductor device line widths continue to shrink and device structures evolve toward multi-layer stacks, the requirements for the surface roughness of silicon wafers are becoming increasingly stringent, with stricter requirements placed on the roughness of even smaller areas. It is highly likely that surface topography parameters corresponding to signal data with even shorter wavelengths will emerge in the future.

Taking the NT value as an example, the NT value of a wafer obtained by the relevant solution is the NT value of the wafer after the CMP process. This data only represents the performance of the entire wafer processing process and cannot evaluate the performance of each processing step within the process. Moreover, as advanced processes place increasingly stringent requirements on the wafer surface microtopography, it is necessary to monitor each processing step to ensure the stability and consistency of the processing equipment used in the process. The products of each process step must also be monitored to ensure that qualified products enter the next process step, thereby improving the microtopography quality of the wafer surface.

Based on this, FIG1 illustrates a method for predicting wafer surface nanotopography provided by the present invention, which includes steps S101 to S102.

In step S101, during the wafer processing process, the surface topography measurement data of the wafer is detected.

In the present invention, the surface topography measurement data of the wafer is obtained after completing any process in the wafer processing.

In some examples, after completing any processing step (such as wire sawing), a single-point metrology solution is used to measure the surface height of a sample point on the surface of the wafer that has undergone wire sawing to obtain the current surface topography measurement data of the wafer.

In the above example, a single-shot measurement solution is one in which only one sampling point can be measured during a single measurement process. For example, a contact measurement solution can be used, such as using a probe to contact the surface of the wafer being measured and move horizontally across the wafer surface. This horizontal movement causes the probe to undergo longitudinal displacement due to height variations on the wafer surface. This longitudinal displacement is sensed by a displacement sensor, which converts the sensed longitudinal displacement value into height data for the wafer surface being measured, i.e., historical measurement data on the wafer surface height. Non-contact measurement solutions, such as capacitance measurement and laser focus measurement, can also be used.

In some cases, it's possible to obtain surface topography data for wafers after the current processing step using a single measurement process. This can be achieved using optical techniques such as Fizeau interferometry and differential interference spectroscopy. While these measurement methods offer high resolution and can obtain uniformly sampled data, they place high demands on the wafer surface, requiring a relatively smooth surface to ensure that incident light is reflected rather than scattered. This makes them unsuitable for wafer surfaces after all processing steps. For example, the surface of a wafer after wire sawing exhibits numerous line marks and high roughness, while the surface of a wafer after grinding lacks crystal planes. Therefore, in actual implementation, single-point measurement methods are more suitable for obtaining surface topography data.

In step S102, based on the trained convolutional neural network (CNN) model, the surface topography measurement data is used to predict the predicted surface topography values of the wafer corresponding to a set wavelength range after subsequent processing.

In the present invention, a subsequent process flow refers to a process flow following the process of obtaining wafer surface topography measurement data in step S101 during the wafer processing process. For example, if step S101 obtains wafer surface topography measurement data after the wire sawing process, the subsequent process flow described in step S102 may be a process following the wire sawing process during the wafer processing process, such as a lapping process, an etching process, a grinding process, or a CMP process.

In the present invention, the surface topography measurement data from all sampling points can be viewed as a three-dimensional fluctuation phenomenon across the entire wafer surface. This fluctuation phenomenon can be formed by the superposition of wave signals of different wavelengths. Based on this understanding, the surface topography value corresponding to a set wavelength range, such as the nanotopography (NT) value corresponding to the wavelength range of 22 microns to 20 millimeters, can be viewed as a statistical parameter for that wavelength range and can be obtained through filtering. In the present invention, the NT value is used as an example of a surface topography value corresponding to a set wavelength range. It is understood that surface topography values corresponding to other set wavelength ranges are equally applicable to the technical solution of the present invention and will not be elaborated here.

Notably, because filtering and convolution are similar computations, the process of acquiring NT values is highly similar to computer vision tasks in deep learning. Based on this understanding, the present invention employs a CNN model class suitable for computer vision tasks to address NT value prediction.

Once the CNN model is determined, the initialized CNN model can be trained using historical surface topography measurement data from wafers after the current process (e.g., wire sawing) and the NT values of these wafers after subsequent processes (e.g., CMP) as the dataset. This yields a trained CNN model capable of predicting the NT value after subsequent processes based on the surface topography measurement data from the current process.

In the present invention, the predicted surface topography value corresponding to the set wavelength range is still used as an example. After obtaining the predicted nanotopography (NT) value of the wafer in step S102, the currently completed processing flow can be evaluated.

In some examples, the operating status of the equipment executing the currently completed process can be determined based on the statistics of the wafer's predicted nanotopography values to ensure equipment stability and consistency and avoid machine errors. For example, the NT predicted values of all products processed daily by each piece of equipment can be monitored. If the NT predicted values show dispersion, excessive outliers, or an excessively large mean, this indicates that the equipment is unstable and requires downtime for maintenance.

In some examples, a comparison between the wafer's predicted nanotopography value and a preset evaluation index can be used to determine whether the wafer should proceed to subsequent processing steps. This allows for the selection of qualified products for subsequent processing, preventing the subsequent processing of products with a low likelihood of future qualification, thereby reducing processing costs.

In some examples, process parameters of the currently completed process flow are adjusted based on the wafer's predicted nanotopography values. For example, the predicted nanotopography values of the products of each process flow can be used to determine whether there are process issues within the completed process flow, thereby assisting in more targeted process adjustments. Furthermore, during process adjustments, these predicted nanotopography values can also be used as feedback to adjust the process parameters of the process flow.

The technical solution shown in Figure 1 utilizes a trained CNN model and surface topography data from wafers after completing the current process to predict the wafer's surface topography after completing subsequent processes. This predicted topography not only monitors each process step to ensure the stability and consistency of the processing equipment used, but also monitors the products of each process step to ensure that qualified products advance to the next step. This improves the microscopic topography quality of the wafer surface and prevents subsequent processes from producing products with a low probability of future quality, thereby reducing processing costs. Furthermore, this data can be used to identify process issues within completed processes, assisting with process adjustments.

For the technical solution shown in FIG. 1 , in some examples, detecting the surface topography measurement data of the wafer includes: obtaining raw surface topography measurement data regarding the height of the wafer surface at each sampling point on the wafer surface using a single-point measurement solution after the wafer completes the current processing flow; and generating the surface topography measurement data based on the raw surface topography measurement data.

Regarding the above example, it should be noted that due to the large number of layers in a CNN model, as the network layer deepens, the design concept is usually to halve the feature size of the input image and double the data channels. However, for a single measurement solution on the wafer surface, there are usually not many sampling points. For example, taking a bare wafer with a radius of 150mm and being wire-cut, after removing the edge with an edge removal (EE) of 4mm, the surface height data is measured using the capacitance method to obtain the raw measurement data of the bare wafer surface. During the sampling measurement process, uniform sampling is performed using a polar coordinate system centered at the wafer center. This means that a radial direction is measured every 45°, with a sampling interval of 4mm in each radial direction. In this polar coordinate system, raw measurement data consisting of 8 x 37 sampling points is obtained. When this amount of raw measurement data is fed into the CNN model, it is attenuated after forward propagation through the first few layers of the CNN model, to the point where it is no longer useful in subsequent convolutional layers.

To address the aforementioned issue of small amounts of raw measurement data, the present invention performs height interpolation based on the raw measurement data, using the interpolated height interpolated data to expand the data volume of the raw measurement data, thereby ultimately obtaining surface topography measurement data of the wafer after completing the current processing flow.

Based on this, in some examples, forming the surface topography measurement data based on the original surface topography measurement data includes: converting all sampling points on the wafer surface from a polar coordinate system to a rectangular coordinate system; interpolating the original surface topography measurement data of all sampling points in the rectangular coordinate system using cubic spline interpolation to obtain interpolation points and height interpolation data at the interpolation points; and forming the surface topography measurement data from the original surface topography measurement data at all sampling points and the height interpolation data at all interpolation points.

Specifically, the original surface topography data sampled in a polar coordinate system appears uneven in a rectangular coordinate system. To address this unevenness, the present invention performs uniform sampling based on the locations of the sampling points and then uses cubic spline interpolation to reshape the final surface topography data from the aforementioned 8*37 shape to a 448*448 shape.

For the scheme shown in Figure 1, some possible implementations may also include a CNN model training process. This process may include: training an initialized convolutional neural network (CNN) model based on historical surface topography measurement data of wafers after completing the current process and surface topography values measured for the wafers after completing the subsequent process corresponding to a set wavelength range, to obtain a trained CNN model.

Regarding the aforementioned implementation method, the inspection process for the surface topography measurement data of historical wafers after the current processing flow, as well as its implementation method and examples, can be found in the aforementioned inspection process for surface topography measurement data and will not be elaborated upon here.

The CNN model described in the technical solution shown in Figure 1, as shown in Figure 2, includes not only convolutional layers connected to the input layer, maximum pooling layers, global average pooling layers, and fully connected layers connected to the output layer, but also multiple residual blocks and skip connections between the convolutional layers within each residual block, as indicated by the curved arrows in Figure 2. Each residual block includes multiple convolutional layers with the same number of convolution kernels, kernel size, and stride. In Figure 2, CNN model 2 is assumed to include two residual blocks, labeled R1 and R2. Each residual block includes four convolutional layers. For example, R1 includes convolutional layers R1-1, R1-2, R1-3, and R1-4, and R2 includes convolutional layers R2-1, R2-2, R2-3, and R2-4.

It should be noted that the goal of computer vision tasks is typically image classification, object detection, instance segmentation, etc. Therefore, computer vision tasks are usually classification tasks, while NT value prediction tasks are regression tasks, so modifications need to be made to the CNN model.

In this invention, the output dimension of the last fully connected layer in the CNN model is modified accordingly and represented as FC-N, where FC is the abbreviation for the fully connected layer and N represents the number of output dimensions.

In some examples, N can be set to the number of intervals into which the predicted surface topography values are divided, so that the CNN model outputs the interval in which the predicted surface topography values are located; accordingly, the loss function used to train the CNN model is cross-entropy loss.

For example, the fully connected layer, which serves as the last layer in the CNN model, is changed to FC-50. This means that the predicted NT value is divided into 50 equal-width intervals, each of which corresponds to a class label. The final conclusion is that the predicted NT value belongs to a specific interval, that is, the regression task of predicting the NT value is treated as a classification task. Based on this change, the softmax classifier is still used to calculate the cross-entropy loss as the loss function during the training phase. As shown in Figure 3, the horizontal coordinate of the coordinate system represents the predicted value (Predicted Value), which is specifically a box within the coordinate system. The box represents the 50 intervals that the predicted NT value can belong to. The dotted line represents the ideal regression curve. It can be seen that the regression task of predicting NT values can be treated as a classification task. Furthermore, this modification enhances the robustness of the model and avoids the risk of overfitting, which in turn prevents unknown variations caused by the instability of the downstream processing.

In some examples, as shown in Figure 2, the fully connected layer serving as the last layer in the CNN model is replaced with an FC-1 layer to linearly output the predicted surface topography value; accordingly, the loss function used to train the CNN model is the mean squared error (MSE).

For the above example, you can select an appropriate value of N according to actual needs during implementation to achieve interval division with different precisions or linear output.

After determining the specific model and corresponding modifications, in some examples, training an initialized CNN model based on historical surface topography measurement data of wafers after completing the current process and surface topography values measured for the historical wafers after completing the subsequent process corresponding to a set wavelength range to obtain the trained CNN model includes: forming a dataset with the historical surface topography measurement data and the measured surface topography values of the historical wafers; dividing the dataset into a training set and a validation set; inputting the training set into the initialized CNN model to update the network parameters of the CNN model to obtain a preliminarily trained CNN model; and inputting the validation set into the preliminarily trained CNN model for validation to evaluate its performance.

It should be noted that, still using the NT value as an example, historical surface topography measurement data of wafers after the current process is used as input data, and historical NT values of wafers after subsequent processes are used as output data for training until the loss function is minimized. 90% of the input-output data set is selected as the training set, and the remaining 10% is used as the validation set. In conjunction with the aforementioned implementation, when the last fully connected layer in the CNN model is FC-50, the loss function is cross-entropy loss. When the output dimension of the last fully connected layer in the CNN model is set to 1, the loss function is MSE.

Based on the aforementioned technical solution, the present invention is described in a specific embodiment. In this embodiment, a bare wafer with a radius of 150mm after wire cutting is used as an example. After the edge of each bare wafer is removed according to an edge removal amount (EE) of 4mm, the surface height data of the bare wafer is sampled and measured using the capacitance method and interpolated to obtain the surface morphology measurement data of each bare wafer. During the sampling measurement process, uniform sampling is performed according to the polar coordinate system with the wafer center as the pole, that is, a radial direction is measured every 45°, and the sampling interval in each radial direction is 4mm. In this polar coordinate system, the original surface topography measurement data of 8*37 sampling points can be obtained, that is, a feature map size of 1*8*37 is obtained. Based on this feature map size, it is reshaped into a shape of 1*448*448 by interpolation to form a tensor of [-1,1,448,448]. The CNN model after the above modifications is trained. As shown in Figure 4, during the training process, the error loss of the training set (train loss) and the error loss of the validation set (Val The loss converges with the number of training epochs. Furthermore, for the trained models, namely the regression task model (with the output dimension of the fully connected layer set to 1) and the classification task model (with the output dimension of the fully connected layer set to the number of intervals divided by the predicted surface topography values), see Table 1. Using 16,023 wafers as an example for prediction, both the regression and classification task models achieved good prediction results thanks to the generalization of the model:

As shown in Table 1, using 16,023 wafers as an example, in the classification task mode, the precision and recall were 0.72 and 0.6, respectively. In the regression task mode, regardless of the aforementioned index values, the ^R² correlation between the predicted and actual NT values was 0.71. This demonstrates that the model provided by this invention is capable of highly accurate prediction of post-processing NT.

Based on the wafer surface nanotopography prediction method provided by the aforementioned technical solution, the predicted nanotopography value of the wafer is less than 5nm at a 2mm*2mm size and/or less than 10nm at a 10mm*10mm size. The wafer nanotopography value is less than 5nm at a 2mm*2mm size and less than 10nm at a 10mm*10mm size.

It should be noted that the process for obtaining the predicted nanotopography values precedes obtaining the nanotopography values of the wafer. For example, the predicted nanotopography values can be obtained after the wire sawing process, and the nanotopography values within the wafer can be obtained after the CMP process. Alternatively, the predicted nanotopography values can be obtained after the CMP process, and the nanotopography values within the wafer can be obtained when the wafer flows into the back-end semiconductor manufacturing process. This is not discussed in detail in this invention.

Based on the same inventive concept as the aforementioned technical solution, FIG. 5 illustrates a wafer surface nanotopography prediction device 50 provided by the present invention. The prediction device 50 comprises a detection portion 501 and a prediction portion 502. The detection portion 501 is configured to detect surface topography measurement data of the wafer during wafer processing. The prediction portion 502 is configured to predict the wafer's surface topography value corresponding to a set wavelength range after subsequent processing based on the surface topography measurement data using a trained convolutional neural network (CNN) model.

In some examples, the output dimension N of the last fully connected layer in the CNN model is at least 1.

In some examples, when the output dimension of the fully connected layer is 1, the predicted surface topography value is output linearly; accordingly, the loss function used for the CNN model is the minimum mean square error (MSE); when the output dimension of the fully connected layer is greater than 1, the output dimension of the fully connected layer represents the number of intervals into which the predicted surface topography value is divided; accordingly, the loss function used for training the CNN model is the cross-entropy loss.

In some examples, the detection portion 501 is configured to: after the wafer completes the current processing flow, obtain raw surface topography measurement data regarding the height of the currently processed wafer surface at each sampling point on the wafer surface using a single-point measurement solution; and generate the surface topography measurement data based on the raw surface topography measurement data.

In some examples, the detection portion 501 is configured to: convert all sampling points on the wafer surface from a polar coordinate system to a rectangular coordinate system; interpolate the original surface topography measurement data of all sampling points in the rectangular coordinate system using cubic spline interpolation to obtain interpolation points and height interpolation data at the interpolation points; and form the surface topography measurement data from the original surface topography measurement data at all sampling points and the height interpolation data at all interpolation points.

In some examples, referring to FIG. 6 , the wafer surface nanotopography prediction device 50 further includes a training portion 503 configured to train an initialized CNN model based on historical surface topography measurement data of wafers after completing a current process flow and surface topography values corresponding to a set wavelength range measured after completing a subsequent process flow, thereby obtaining the trained CNN model.

In some examples, the training portion 503 is configured to: form a dataset from the historical wafer surface topography measurement data and the measured surface topography values of the historical wafer; divide the dataset into a training set and a validation set; input the training set into the initialized CNN model to update the network parameters of the CNN model to obtain a preliminarily trained CNN model; and input the validation set into the preliminarily trained CNN model for validation to evaluate performance.

In some examples, referring to FIG. 6 , the wafer surface nanotopography prediction device 50 further includes a feedback portion 504 configured to: determine the operating status of equipment executing a currently completed process flow based on statistics of the wafer's predicted surface topography values; or, determine whether to continue the wafer with a subsequent process flow based on a comparison of the wafer's predicted surface topography values with a preset evaluation index; or, adjust process parameters of the currently completed process flow based on the wafer's predicted surface topography values.

It should be noted that the specific implementation of the functions configured in each "part" of the above-mentioned device can be found in the implementation methods and examples of the corresponding steps in the aforementioned wafer surface nanotopography prediction method, and will not be elaborated here.

Please refer to Figure 7, which shows a block diagram of a computing device provided by an exemplary embodiment of the present invention. In some examples, computing device 70 can be at least one of a smartphone, a smartwatch, a desktop computer, a laptop computer, a virtual reality terminal, an augmented reality terminal, a wireless terminal, and a portable laptop computer. Computing device 70 has a communication function and can access a wired network or a wireless network. Computing device 70 can generally refer to one of multiple terminals. As one skilled in the art will appreciate, the number of terminals can be greater or less. In some examples, computing device 70 can receive data based on the wired network or wireless network to which it is connected. It is understood that the computing device 70 is responsible for the calculation and processing work of the technical solution of the present invention, and the present invention is not limited to this.

As shown in FIG7 , the computing device of the present invention may include one or more of the following components: a processor 710 and a memory 720.

Preferably, processor 710 utilizes various interfaces and circuits to connect various components within the computing device. It executes instructions, programs, code sets, or instruction sets stored in memory 720, and calls data stored in memory 720 to perform various functions of the computing device and process data. Preferably, processor 710 can be implemented using at least one of the following hardware formats: digital signal processing (DSP), field-programmable gate array (FPGA), and programmable logic array (PLA). Processor 710 may integrate one or a combination of a central processing unit (CPU), a graphics processing unit (GPU), a neural network processing unit (NPU), and a baseband chip. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing content displayed on the touch screen; the NPU implements artificial intelligence (AI) functions; and the baseband chip handles wireless communications. It is understood that the baseband chip can be implemented independently of the processor 710.

The memory 720 may include random access memory (RAM) or read-only memory (ROM). Preferably, the memory 720 includes a non-transitory computer-readable storage medium. The memory 720 may be used to store instructions, programs, codes, code sets, or instruction sets. The memory 720 may include a program storage area and a data storage area. The program storage area may store instructions for implementing an operating system, instructions for at least one function (such as a touch function, a sound playback function, an image playback function, etc.), instructions for implementing the above method embodiments, etc. The data storage area may store data created based on the use of the computing device, etc.

Furthermore, those skilled in the art will appreciate that the structure of the computing device depicted in the above figures does not limit the computing device. The computing device may include more or fewer components than shown, or may combine certain components or have a different component layout. For example, the computing device may also include a display, a camera, a microphone, a speaker, an RF circuit, an input unit, sensors (such as an accelerometer, an angular velocity sensor, a light sensor, etc.), an audio circuit, a WiFi module, a power supply, a Bluetooth module, and other components, which will not be further described here.

The present invention also provides a computer-readable storage medium storing at least one instruction. The at least one instruction is configured to be executed by a processor to implement the wafer surface nanotopography prediction method described in the above embodiments.

The present invention also provides a computer program product comprising computer instructions stored in a computer-readable storage medium. A processor of a computing device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computing device to implement the wafer surface nanotopography prediction method described in each of the above embodiments.

The present invention also provides a wafer having a predicted nanotopography value of less than 5 nm at a 2 mm x 2 mm size and/or less than 10 nm at a 10 mm x 10 mm size, obtained using the wafer surface nanotopography prediction method described in each of the above embodiments. The wafer also has a nanotopography value of less than 5 nm at a 2 mm x 2 mm size and less than 10 nm at a 10 mm x 10 mm size.

Those skilled in the art will appreciate that the functions described in one or more of the above examples can be implemented using hardware, software, solid-state devices, or any combination thereof. When implemented using software, these functions can be stored in a computer-readable medium or transmitted as one or more instructions or codes on a computer-readable medium. Computer-readable media include computer storage media and communication media, where communication media includes any medium that facilitates the transfer of computer programs from one location to another. Storage media can be any available medium that can be accessed by a general-purpose or special-purpose computer.

The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications or substitutions that a person of ordinary skill in the art familiar with the present invention could easily conceive within the technical scope disclosed herein should be covered by the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be based on the scope of protection of the patent application.

Figure 1 is a flow chart, so symbols are omitted for simplicity of description.

Claims (9)

A method for predicting wafer surface nanotopography comprises: detecting surface topography measurement data of the wafer during wafer processing; predicting, based on the surface topography measurement data, predicted surface topography values of the wafer corresponding to a set wavelength range after subsequent processing, using a trained convolutional neural network (CNN) model; wherein the output dimension N of the last fully connected layer in the CNN model is at least 1; wherein, when the output dimension of the fully connected layer is 1, the predicted surface topography values are output linearly; and accordingly, the loss function of the CNN model is the minimum mean square error (MSE); When the output dimension of the fully connected layer is greater than 1, the output dimension of the fully connected layer represents the number of intervals into which the predicted surface morphology value is divided; accordingly, the loss function of the CNN model is cross entropy loss. The prediction method as described in claim 1, wherein the detecting the surface morphology measurement data of the wafer comprises: obtaining, for each sampling point on the surface of the wafer, raw surface morphology measurement data regarding the height of the wafer surface at each sampling point by a single-point measurement scheme after the wafer completes the current processing flow; and forming the surface morphology measurement data based on the raw surface morphology measurement data. The prediction method as described in claim 2, wherein the surface topography measurement data is formed based on the original surface topography measurement data, comprising: converting all sampling points on the wafer surface from a polar coordinate system to a rectangular coordinate system; interpolating the original surface topography measurement data of all sampling points in the rectangular coordinate system by cubic spline interpolation to obtain interpolation points and height interpolation data at the interpolation points; and forming the surface topography measurement data from the original surface topography measurement data at all sampling points and the height interpolation data at all interpolation points. The prediction method as described in claim 1 further includes: training an initialized CNN model based on the surface morphology measurement data of the historical wafer after completing the current processing flow and the measured surface morphology values corresponding to the set wavelength range of the historical wafer after completing the subsequent processing flow to obtain the trained CNN model. The prediction method as described in claim 1, wherein an initialized CNN model is trained based on the surface morphology measurement data of the historical wafer after completing the current processing flow and the measured surface morphology values of the historical wafer after completing the subsequent processing flow corresponding to the set wavelength range to obtain the trained CNN model, including: forming a dataset with the surface morphology measurement data of the historical wafer and the measured surface morphology values of the historical wafer; dividing the dataset into a training set and a verification set; inputting the training set into the initialized CNN model to update the network parameters of the CNN model to obtain a preliminarily trained CNN model; and inputting the verification set into the preliminarily trained CNN model for verification to evaluate performance. A device for predicting wafer surface nanotopography comprises: a detection portion and a prediction portion; wherein the detection portion is configured to detect surface topography measurement data of the wafer during wafer processing; the prediction portion is configured to predict, based on the surface topography measurement data, predicted surface topography values of the wafer after subsequent processing, corresponding to a set wavelength range, based on a trained convolutional neural network (CNN) model; wherein the output dimension N of the last fully connected layer in the CNN model is at least 1; wherein, when the output dimension of the fully connected layer is 1, the predicted surface topography values are output linearly; accordingly, the loss function of the CNN model is the minimum mean square error (MSE); When the output dimension of the fully connected layer is greater than 1, the output dimension of the fully connected layer represents the number of intervals into which the predicted surface morphology value is divided; accordingly, the loss function of the CNN model is cross entropy loss. A computing device comprising: a processor and a memory; the processor is used to execute instructions stored in the memory to implement the method for predicting wafer surface nanomorphology as described in any one of claims 1 to 5. A computer-readable storage medium stores at least one instruction, wherein the at least one instruction is used to be executed by a processor to implement the method for predicting wafer surface nanomorphology as described in any one of claims 1 to 5. A wafer having a predicted nanotopography value of less than 5 nm at a size of 2 mm*2 mm and/or a nanotopography value of less than 10 nm at a size of 10 mm*10 mm obtained by the wafer surface nanotopography prediction method as described in any one of claims 1 to 5.