TWI860159B - Establishing method for prediction model on probe mark defect, prediction system and prevention method thereof - Google Patents

Abstract

An establishing method for a prediction model on probe mark defect is adapted for solving the problem of the lack of a method for predicting probe mark defect in the conventional techniques. The establishing method includes a data collecting step, an importance analysis step, a factor re-filtering step and a model training step. In the importance analysis step, an importance analysis method is applied to define an important association data. In the factor re-filtering step, the important association data is transformed into a corresponding correlation coefficient matrix, which includes an association among each factors and defect and a correlation between any two of the factors, so that a predicting factor data is determined. In the model training step, the predicting factor data is inputted into a prediction model to be trained to establish a trained prediction model. This invention provides the effect of predicting the probability of occurrence of the probe mark defect.

Description

Method for establishing a prediction model for probe mark defects and its prediction system and prevention method

The present invention relates to a wafer process improvement, in particular to a method for predicting and preventing probe mark defects.

During the manufacturing process of semiconductor wafers, they need to go through multiple steps such as etching, deposition, wet process, and dry process. Each step will affect the quality and performance of the wafer. Therefore, in order to ensure the stability of the process and the consistency of quality, the wafer needs to be tested. When testing the wafer, the probe of the probe card is used to contact the specific detection position on the surface of the wafer to test the electrical parameters of the wafer, such as resistance, capacitance, current, etc.

When the probe contacts the surface of the wafer, the test probe will leave some marks on the wafer surface. These marks are called probe marks, which can be used to record the location of the test point. The probe mark should be formed at a specific detection position on the wafer surface, otherwise the deviation of the probe mark will cause unexpected changes in the electrical parameters of the wafer. Therefore, when the distance between the probe mark and its predetermined position deviates to a certain extent, the corresponding wafer will be considered defective, and this phenomenon of deteriorating the electrical properties of the wafer due to deviation is called probe mark damage. However, no research has yet proposed how to predict the occurrence of probe mark defects.

In view of this, there is still a need to improve the knowledge technology in the prediction of probe mark defects.

In order to solve the above problems, the purpose of the present invention is to provide a method for establishing a prediction model of probe mark defects, which can predict the defect probability of probe mark defects.

The second purpose of the present invention is to provide a method for establishing a prediction model for probe mark defects, which can ensure or improve the prediction accuracy of the corresponding prediction model during the training process and the application process after establishment.

Another object of the present invention is to provide a prediction system for predicting probe mark defects, which can predict the defect probability of probe mark defects.

Another object of the present invention is to provide a preventive method for reducing probe mark defects, which can reduce the risk of probe mark defects occurring during the detection process.

The quantifiers "one" or "a" used in the components and parts described throughout the present invention are only for the convenience of use and to provide a general meaning of the scope of the present invention; in the present invention, they should be interpreted as including one or at least one, and the single concept also includes the plural case, unless it is obvious that it means otherwise.

The terms "computer" or "system" described in the present invention may include at least one "processor" as a whole or individually. The processor refers to various data processing devices with specific functions and implemented by hardware or hardware and software to process and analyze information and/or generate corresponding control information, such as electronic controllers, servers, cloud platforms, virtual machines, desktop computers, laptops, tablet computers or smart phones, etc., which can be understood by those with ordinary knowledge in the technical field to which the present invention belongs. In addition, corresponding data receiving or transmitting units may be included to receive or transmit the required data. In addition, corresponding databases/storage units (especially non-transient memory units) may be included for reading and storing corresponding data. In particular, unless otherwise specifically excluded or contradictory, the processor may be a collection of multiple processors based on a distributed system architecture, used to contain/represent the process, mechanism and result of information stream processing between multiple processors.

The method for establishing a prediction model of a probe mark defect of the present invention is to execute the following steps through a computer, the steps comprising: a data collection step, collecting a plurality of different factors in each probe mark generation process, and forming a preliminary factor combination from the plurality of different factors collected each time, and associating each preliminary factor combination with information on whether a defect occurs to define a preliminary associated information; an importance analysis step, using an importance analysis method to screen the preliminary associated information to obtain information related to the occurrence of the defect; A combination of initially screened factors that are correlated with the defects is generated, and the combination of initially screened factors is correlated with information on whether a defect occurs to define an important correlation information; a factor re-screening step includes using a factor correlation analysis method to convert the important correlation information into a corresponding correlation coefficient matrix, wherein the correlation coefficient matrix includes a correlation between each factor and the defect and includes a correlation degree between any two factors; each factor corresponding to a correlation that is not less than a first screening threshold is selected to define a high correlation factor. sub-combination; when the corresponding correlation between any two factors is not less than a second screening threshold, the factor with the lower correlation with the defect is defined as a rejection factor; and the factor identical to the rejection factor is removed from the high correlation factor combination to obtain a corresponding re-screened factor combination, and the re-screened factor combination is associated with information on whether a defect occurs to define a prediction factor information; and a model training step, the prediction factor information is input into a prediction model to be trained , to establish a trained prediction model, which is established with a basic transformer architecture; the basic transformer architecture includes an encoder network and a decoder network; the encoder network sequentially includes a multi-head self-attention layer, a summing and normalization layer, a feedforward network layer and another summing and normalization layer; the decoder network sequentially includes a masked multi-head attention layer, a summing and normalization layer, a multi-head cross-attention layer, another summing and normalization layer, a feedforward network layer and another summing and normalization layer.

The method for establishing a prediction model of a probe mark defect of the present invention is to execute the following steps through a computer, the steps comprising: a data collection step, collecting a plurality of different factors in each probe mark generation process, and forming a preliminary factor combination from the plurality of different factors collected each time, and associating each preliminary factor combination with information on whether a defect occurs to define a preliminary associated information; an importance analysis step, using an importance analysis method to screen the preliminary associated information to obtain information related to the occurrence of the defect; A combination of factors is preliminarily screened, and the preliminarily screened combination of factors is associated with information on whether a defect occurs to define an important correlation information; a factor re-screening step includes using a factor correlation analysis method to convert the important correlation information into a corresponding correlation coefficient matrix, wherein the correlation coefficient matrix includes a correlation between each factor and the defect and a correlation between any two factors; each factor corresponding to the correlation being not less than a first screening threshold is selected to define a high correlation factor combination; between any two factors When the correlation between the two factors is not less than a second screening threshold, the factor with the lower correlation with the defect is defined as a rejection factor; and the factor identical to the rejection factor is removed from the high correlation factor combination to obtain a corresponding re-screened factor combination, and the re-screened factor combination is associated with information on whether a defect occurs to define a prediction factor information; and a model training step is performed, wherein the prediction factor information is input into a prediction model to be trained to establish a trained prediction model, and ... The prediction model is established with a sparse attention transformer architecture; the sparse attention transformer architecture includes an encoder network and a decoder network; the encoder network sequentially includes a multi-head self-attention layer, a summation and normalization layer, a feedforward network layer and another summation and normalization layer; the decoder network sequentially includes a sparse multi-head self-attention layer, a summation and normalization layer, a multi-head cross-attention layer, another summation and normalization layer, a feedforward network layer, another summation and normalization layer, a linear layer and a normalized exponential layer.

Accordingly, the method for establishing a prediction model of a probe mark defect of the present invention can ensure, through the combination of factors obtained in the importance analysis step and the factor re-screening step, that the combination of factors is highly correlated with the prediction target (i.e., the probability of a probe mark defect), thereby ensuring or improving the prediction accuracy of the corresponding prediction model during the training process and the application process after establishment; on the other hand, it can screen out factors that can be eliminated from factors with high correlation, thereby reducing the number of factors used to train/input the prediction model, thereby reducing the amount of calculation of the prediction model and improving the calculation efficiency; through the converter The attention mechanism in the architecture can focus more on the features that need to be paid attention to, and can achieve more accurate recognition/prediction. The configuration of each layer in the transformer architecture can improve the model training and computational efficiency. Or, through the interval selection mechanism in the sparse multi-head self-attention layer, on the one hand, the corresponding model can better capture the long-term dependency between factors/parameters, and can focus on calculating a few important positions with high attention scores, thereby reducing the amount of computation and improving computational efficiency. On the other hand, it can reduce the risk of overfitting of the corresponding model during training, and the corresponding model can be more easily generalized to new data, thereby improving the accuracy of the model.

Among them, the hyperparameters in the prediction model are configured as follows; in the initial hyperparameters, the batch size is 64, the training rounds are 50, and the learning rate is 0.0001; in the encoder network: the number of heads of the multi-head self-attention layer is 8, and the dimension of the attention key is 64; the dropout rate in the feedforward network layer is 0.1, and the output dimension of the fully connected layer is 64 and uses the ReLU activation function ; and in the decoder network: the number of heads of the sparse multi-head self-attention layer is 8, and the dimension of the attention key is 64; the number of heads of the multi-head cross-attention layer is 8, and the dimension of the attention key is 64; the dropout rate in the feedforward network layer is 0.1, the output dimension of the fully connected layer is 64 and uses the ReLU activation function; the output dimension of the linear layer is 2, and the ReLU activation function is used.

Among them, the importance analysis method is a random forest algorithm. In this way, it is possible to achieve the effect of obtaining highly correlated factors efficiently, accurately and stably.

The prediction factor information is composed of at least a factor of the minimum diameter of the needle tip, a factor of the correction position and the corresponding defect occurrence information. Thus, these factors are obtained through the importance analysis step and the factor re-screening step, and these factors are highly relevant and representative prediction factors, so that the prediction model can at least be trained and predicted by the data of the above factors, and can achieve the power of improving the computational efficiency and accuracy.

The prediction system for predicting probe mark defects of the present invention includes a prediction model established by the above-mentioned method for establishing a prediction model for probe mark defects. The prediction model is used to receive monitoring data to output a defect probability of whether a probe mark defect will occur; the type of each factor in the monitoring data is the same as the type of each factor in the prediction factor information.

Accordingly, the prediction system for predicting probe mark defects of the present invention can achieve the power of improving computational efficiency and accuracy by using the prediction model established by the above-mentioned method for establishing a prediction model of probe mark defects.

The prediction system for predicting probe mark defects of the present invention includes the above-mentioned prediction system. When the prediction system receives a monitoring data corresponding to a current workpiece and generates a corresponding defect probability, and in a situation where the defect probability exceeds a preset warning threshold, before executing the inspection procedure corresponding to other subsequent workpieces, the subsequent value of at least one factor in the monitoring data is adjusted to be different from the current value of the at least one factor in the monitoring data, so that the defect probability corresponding to the value of each factor in the subsequent monitoring data does not exceed the preset warning threshold.

Accordingly, the prediction system for predicting probe mark defects of the present invention can achieve the effect of reducing the probability of probe mark defects by applying the above prediction model and matching the relationship between the preset warning value and the defect probability, and then adjusting at least one factor in the subsequent monitoring data so that the corresponding defect probability does not exceed the preset warning threshold value.

S1: Data collection step

S1A: Data screening steps

S1A-1: Importance analysis steps

S1A-2: Factor re-screening step

S2: Model training steps

[Figure 1] Schematic diagram of the process of establishing a prediction model for probe mark defects of the present invention.

[Figure 2] A schematic diagram of a basic converter architecture.

[Figure 3] A schematic diagram of the configuration of a sparse attention switch architecture.

In order to make the above and other purposes, features and advantages of the present invention more clearly understood, the following specifically cites a preferred embodiment of the present invention and provides a detailed description in conjunction with the attached drawings.

Please refer to FIG. 1, which is a preferred embodiment of the method for establishing a prediction model of a probe mark defect of the present invention, which is to complete the training/establishment of the target model by executing a data collection step S1, an optional data screening step S1A and a model training step S2 through a computer.

In the data collection step S1, several different factors/parameters related to each probe mark generation process are collected, and a preliminary factor combination is formed from the several different factors collected each time, and each preliminary factor combination is associated with information on whether a defect occurs to define preliminary associated information. The content of the preliminary correlation information, for example, presents various factor combinations and whether defects occur at a certain time point; for example, at the first time point t1, the preliminary correlation information may include a corresponding parameter a1 of the first factor A, a corresponding parameter b1 of the second factor B, a corresponding parameter c1 of the third factor C, and a result r1 of whether a defect occurs; similarly, at the second time point t2, the preliminary correlation information may include a corresponding parameter a2 of the first factor A, a corresponding parameter b2 of the second factor B, a corresponding parameter c2 of the third factor C, and a result r2 of whether a defect occurs.

In detail, the information of the several different factors refers to the several different factors that generate probe marks on a workpiece (e.g., a wafer) when the workpiece is inspected using a probe card device, and the specific value or information of each factor can be obtained through a built-in system in the probe card device, and/or detected by additionally installing corresponding sensing elements, and/or obtained through subsequent related inspections/measurements. For example, in a specific embodiment of the present invention, the "defect" of concern refers to the "probe mark defect", and in particular, when the offset defined by the minimum distance between the corresponding probe mark and the edge of the probe card is not greater than a distance threshold during measurement, it is defined as the occurrence of the probe mark defect; preferably, the distance threshold is 5nm.

In a specific embodiment of the present invention, the several different factors or the preliminary factor combination include:

(1) Count by Lot: represents the number of workpiece batches tested by the probe card.

(2) Number of contacts: represents the number of contact points generated by the probe of the probe card on the workpiece when inspecting the workpiece.

(3) Needle Tip Min: The minimum diameter of all needle tip marks produced when the probe card's probe contacts the workpiece.

(4) Needle Tip Max: The maximum diameter of all needle tip marks produced when the probe card's probe contacts the workpiece.

(5) Needle Tip Ave: The average diameter of all needle tip marks produced when the probe card's probe contacts the workpiece.

(6) Probe height (Pb Height): The distance between the probe card and the workpiece before the probe card performs the inspection procedure.

(7) Probe Card Clean Cnt: This value indicates the degree of contamination on the probe card. The higher the value, the more serious the contamination.

(8) Probing Overdrive: The voltage value of the probe card device when measuring.

(9) Alignment End Position: Before the probe card moves vertically to form a probe mark on the corresponding workpiece surface, the initial position of the probe card moves horizontally to a preset measurement position.

In the optional data screening step S1A, an importance analysis step S1A-1 and an optional factor re-screening step S1A-2 are included. In the importance analysis step S1A-1, an importance analysis method is used to screen the preliminary correlation information to obtain a combination of preliminary screened factors that are highly correlated with the occurrence of the defect (which can be defined as a target characteristic), and the combination of preliminary screened factors is associated with information on whether a defect occurs to define an important correlation information. In particular, the importance analysis method can use any known method and is understandable to a person with ordinary knowledge in the field of the present invention; for example, a random forest algorithm can be applied as the importance analysis method. In detail, in the importance analysis step S1A-1, it is assumed that the factors initially defined to be relevant to the prediction results are i. After the data screening step S1A, j factors with higher correlation can be obtained from the i factors; i and j are both positive integers, and j is not greater than i. Therefore, through the importance analysis step S1A-1, factors with higher correlation to the results of interest can be screened out, and the factors with higher correlation can be used to learn the subsequent prediction model. In this way, on the one hand, the dimension of the overall training factors can be reduced, and the model training and calculation speed can be improved; on the other hand, the model is trained with factors with higher correlation, which helps to improve the accuracy of model prediction.

For example, in a specific implementation example of the present invention, based on the nine different factors obtained in the above data collection step S1 and the preliminary correlation information defined by the corresponding defect information/target characteristics, in the importance analysis step S1A-1, the preliminary correlation information is screened using the random forest algorithm in the importance analysis method, and five factors with significant effects on defect occurrence/target characteristics can be obtained to define the above important correlation information. In detail, the five factors in the important correlation information are probe height, minimum needle tip diameter, average needle tip diameter, maximum needle tip diameter and correction position, and the corresponding importance (sum is 1) is 0.3001, 0.2440, 0.2221, 0.1675, 0.0663 respectively.

In the optional factor re-screening step S1A-2, a factor correlation analysis method can be used to convert the preliminary correlation information or the important correlation information into a corresponding correlation coefficient matrix, which includes a correlation between each factor and the defect and a correlation between any two factors; select each factor corresponding to the correlation not less than a first screening threshold to obtain a high correlation factor combination that is highly correlated with the defect occurrence/target characteristic; at the same time, when the correlation between any two factors is not less than a second screening threshold, the two factors that are highly correlated with the defect are selected. The one with a lower correlation is defined as a rejection factor, that is, the rejection factor represents a factor that can be considered as a substantially repeated consideration among the factors for considering whether a defect occurs; then, the factor that is the same as the rejection factor is removed from the high correlation factor combination to obtain a corresponding re-screened factor combination, and the re-screened factor combination is associated with the information of whether a defect occurs to define a prediction factor information. In particular, the inter-factor correlation analysis method may use any known method and is understandable to those with ordinary knowledge in the field of the present invention; for example, the inter-factor correlation analysis method may be Pearson Correlation Coefficient, Spearman Rank Correlation Coefficient, Kendall Rank Correlation Coefficient, Point-biserial Correlation Coefficient or other bivariate regression analysis methods, and is not limited to the above-mentioned methods, to obtain the corresponding correlation coefficient matrix.

In a specific implementation example of the present invention, the five factors selected in the importance analysis step S1A-1 and the important correlation information defined by the corresponding defects are further screened in the factor screening step S1A-2, and the important correlation information is subjected to the correlation analysis method of the Pearson correlation coefficient, and the correlation coefficient matrix (Correlation Coefficient Matrix) of the important correlation information can be obtained as shown in the following table 1.

As shown in Table 1, the factors related to the probe defect mark are arranged from large to small according to the correlation: the minimum diameter of the needle tip (corresponding to the value of -0.04), the correction position (corresponding to the value of -0.034), the average diameter of the needle tip (corresponding to the value of -0.033), the maximum diameter of the needle tip (corresponding to the value of -0.021) and the height of the probe (-0.0015). Therefore, k factors with greater correlation can be selected from the above five factors for subsequent training of the correlation prediction model, where k is a positive integer and is not greater than j. In detail, based on the above correlation, the absolute values of the correlations of all factors can be added to obtain a sum of correlations, and then the influence of each factor on the result can be observed by dividing the correlation of each factor by the sum of the correlations.

In detail, in the process of selecting k factors, corresponding to the aforementioned selection process of the combination of highly correlated factors, a first screening threshold may be used (for example, it may be defined as not more than 100% divided by the number of current factors, but is not limited thereto), and it may be defined that when the correlation ratio of a factor is not less than the first screening threshold, the factor is selected as information for training in the subsequent prediction model. It can be understood that the higher the correlation ratio corresponding to the factor, the higher the degree of dependence of the factor on the target characteristic. For example, taking Table 1 as an example, the total correlation is 0.1295, the correlation of the minimum tip diameter accounts for 30.89%, the correlation of the correction position accounts for 26.25%, the correlation of the average tip diameter accounts for 25.48%, the correlation of the maximum tip diameter accounts for 16.22%, and the correlation of the probe height accounts for 1.16%; at this time, the first screening threshold can be defined as any reasonable value according to actual needs, such as 5%, and the corresponding high correlation factor combination is composed of factors such as the minimum tip diameter, the correction position, the average tip diameter and the maximum tip diameter.

Next, corresponding to the aforementioned selection of the shaving factors, the correlation between any two factors can also be observed from Table 1, and a second screening threshold can be set. When the correlation between any two factors is not less than the second screening threshold, the factor with the higher correlation between the two factors and the target characteristic can be selected as the representative factor, and the factor with the lower correlation can be selected as the elimination factor. In addition, the factors that are the same as the elimination factor in the aforementioned high correlation factor combination are removed to simplify the amount of information (the number of factor types) input into the prediction model. It can be understood that the higher the correlation between the two factors, the higher the degree to which one of the two factors can reflect the other. For example, taking Table 1 as an example, it can be observed that the correlation between any two of the minimum tip diameter, the average tip diameter, and the maximum tip diameter is at least 0.77 (the corresponding second screening threshold can be set to be no greater than 0.77, for example), and the correlation ratio between the factor of the minimum tip diameter and the defect is the highest among these factors, so that the factor of the minimum tip diameter corresponds to the aforementioned representative factor, and the two factors of the average tip diameter and the maximum tip diameter correspond to the aforementioned shaving factor.

Finally, the factors identical to the rejection factor [average needle tip diameter, maximum needle tip diameter] are removed from the high correlation factor combination [minimum needle tip diameter, correction position, average needle tip diameter, maximum needle tip diameter] to define the re-screened factor [minimum needle tip diameter, correction position]. In other words, through the above-mentioned relevant screening rules of correlation and correlation degree, when training the corresponding prediction model, it is possible to consider using the factor of the minimum diameter of the needle tip to represent the two factors corresponding to the average diameter of the needle tip and the maximum diameter of the needle tip; that is, the factor re-screening step S1A-2 simultaneously considers the first screening threshold (the correlation ratio affecting the occurrence of defects) and the second screening threshold (the correlation degree between factors) to obtain the corresponding prediction factor information (including the two factors of the minimum diameter of the needle tip and the correction position and the corresponding defect occurrence information), and use the prediction factor information to train the corresponding prediction model.

According to the above-mentioned data collection step S1, the optional importance analysis step S1A-1 in the data screening step S1A and the optional factor re-screening step S1A-2, several different training data combinations one to three can be obtained, which are:

(1) Training data combination 1: The training data used to train the prediction model (corresponding to the aforementioned preliminary correlation information) is only the factors obtained through the data collection step S1 and the corresponding defect occurrence information. The number of types of the factors is i.

(2) Training data combination 2: The training data used to train the prediction model (corresponding to the aforementioned important correlation information) is the factors and corresponding defect occurrence information obtained through the data collection step S1 and the importance analysis step S1A-1. The number of types of the factors is j.

(3) Training data combination 3: The training data used to train the prediction model (corresponding to the aforementioned prediction factor information) is the factors and corresponding defect occurrence information obtained through the data collection step S1, the importance analysis step S1A-1 and the factor re-screening step S1A-2. The number of types of the factors is k.

It should be noted that the training data combinations also include situations where j factors are exactly the same as i factors, or situations where k factors are exactly the same as j factors. The data used to train the subsequent prediction model may include the aforementioned preliminary correlation information (corresponding to the data collection step S1), preferably the aforementioned importance correlation information (corresponding to the importance analysis step S1A-1), and more preferably the aforementioned prediction factor information (corresponding to the factor re-screening step S1A-2).

In the model training step S2, the present invention proposes a prediction model to be trained, which is established using a machine learning network architecture, especially a transformer architecture, and the aforementioned prediction factor information is input into the prediction model to be trained to establish a trained prediction model. For example, the transformer architecture can be a basic transformer architecture or a sparse attention transformer architecture.

In detail, as shown in Figure 2, the basic converter architecture has an encoder network (Encoder) and a decoder network (Decoder). The encoder network sequentially includes a multi-head self-attention layer (Multi-head Self-attention Layer), an addition and normalization layer (Add & Norm Layer), a feedforward network layer (Feedforward Network) and another addition and normalization layer. The decoder network sequentially includes a masked multi-head attention layer (Masked Multi-head Attention Layer), an addition and normalization layer, a multi-head cross-attention layer (Multi-head Cross-attention Layer), another addition and normalization layer, a feedforward network layer, another addition and normalization layer, a linear layer and a normalized exponential layer.

Specifically, as shown in FIG. 3, the sparse attention transformer architecture can be regarded as a transformation of the basic transformer architecture; wherein the sparse attention transformer architecture replaces the masked multi-head attention layer in the basic transformer architecture with a sparse multi-head self-attention layer.

It should be noted that the basic converter architecture and the sparse attention converter architecture are both technical features in the field that are commonly known and can be understood and implemented based on existing technologies, so the details of the data flow and processing results between the layers in the corresponding architectures will not be repeated; in addition, the prediction model of the present invention can also be established using various other machine learning network architectures, and is not limited to the examples given in the present invention.

In a specific implementation example of the present invention, in the model training step S2, the prediction model is trained with the prediction factor information, and the prediction model is established with the basic converter architecture; wherein the prediction factor information can be used as training data for subsequent models, and corresponds to the combination of the training data mentioned above three. In detail, the prediction factor information (as training data) can be split into a training data set and a validation data set to complete the training of the prediction model. Among them, the prediction accuracy of the prediction models established based on the important correlation information (five factors and defect occurrence information, corresponding only to the result of the importance analysis step S1A-1) and the prediction factor information (two factors and defect occurrence information, including the results of the importance analysis step S1A-1 and the factor re-screening step S1A-2) is compared. The prediction accuracy using the important correlation information (corresponding to the importance analysis step S1A-1) is 92.67%, and the prediction accuracy using the prediction factor information (corresponding to steps S1A-1 and S1A-2) is 93.27%, which can prove that the preliminary collected data can be used through the importance analysis Step S1A-1 and the factor re-screening step S1A-2 are used to confirm whether the target characteristic (e.g., defect occurrence) to be predicted is the same as the type and quantity of the factors initially selected and the type and quantity of the factors obtained by the final screening. It can be verified/ensured that the selected factors are highly correlated with the target characteristic. Alternatively, when the type of factors obtained by the final screening is the result of a reduction in the type of factors initially selected, the factors with a lower correlation with the target characteristic can be removed from the selected factors, thereby reducing the dimension of the input information/training data (reducing the type of factors) and improving the prediction accuracy of the corresponding prediction model during the training process and in the application process after establishment. Among them, the five factors in the important related information are the aforementioned probe height, minimum needle tip diameter, average needle tip diameter, maximum needle tip diameter and correction position; the two factors in the prediction factor information are the aforementioned minimum needle tip diameter and correction position.

In particular, with respect to the prediction accuracy of the prediction model established by the prediction factor information, based on the same prediction model architecture, the present invention further uses different combinations of factors to establish corresponding prediction models, and compares the prediction accuracy of the prediction models established by various factor combinations as shown in Table 2 below. The results in Table 2 verify that the factors selected by the above-mentioned importance analysis step S1A-1 and factor re-screening step S1A-2 can indeed achieve the effect of having the highest prediction accuracy in establishing the corresponding prediction model.

In particular, based on the two factors of the minimum needle tip diameter and the correction position, the prediction model of the present invention is further compared using the basic converter architecture and the sparse attention converter architecture respectively; wherein the prediction accuracy of the prediction model established using the basic converter architecture is 93.27% as mentioned above, and the prediction accuracy of the prediction model established using the sparse attention converter architecture is 93.98%. In other words, the prediction model of the present invention can use the sparse attention converter architecture to replace the basic converter architecture to improve the prediction accuracy of the established model.

Preferably, the following hyperparameter combination is applied in the training model established using the sparse attention transformer architecture.

1. In the initial hyperparameters:

(1) The batch size (batch_size) is 64 (the default is 32).

(2) The number of training rounds (epochs) is 50.

(3) The learning rate is 0.0001 (the default setting is 0.0001).

2. Hyperparameter configuration in the encoder network:

(1) In the multi-head self-attention layer: the number of heads is 8, and the dimension of the attention key (key_dim) is 64.

(2) In the feedforward network layer: the dropout rate is set to 0.1; the output dimension of the fully connected layer (Dense) is 64, and the ReLU activation function is used.

3. Hyperparameter configuration in the decoder network:

(1) In the sparse attention layer: the number of heads is 8 and the dimension of the attention key is 64.

(2) In the multi-head cross-attention layer: the number of heads is 8 and the dimension of the attention key is 64.

(3) In the feedforward network layer: the dropout rate is set to 0.1; the output dimension of the fully connected layer is 64, and the ReLU activation function is used.

(4) In the linear layer: the output dimension is 2, and the ReLU activation function is used.

According to the above content, the present invention proposes a method for establishing a prediction model of a probe mark defect, including the data collection step S1 (only step S1), preferably the additional importance analysis step S1A-1 (including steps S1 and S1A-1), and more preferably the additional factor re-screening step S1A-2 (including steps S1, S1A-1, S1A-2), the various factors obtained and the corresponding information on whether the probe mark defect occurs are used as a training data (that is, the prediction factor information is the same as the training data), and the training data (the prediction factor information) is input into a prediction model to be trained to complete the training/establishment of the prediction model.

Next, through the prediction model established by the above method, the present invention proposes a prediction system having the prediction model, and by continuously receiving/monitoring the current monitoring data through the prediction model, a defect probability of whether a probe mark defect will be generated can be obtained/predicted; wherein the type of each factor in the monitoring data is the same as the type of each factor in the training data (the prediction factor information).

In particular, based on the above-mentioned prediction system, the present invention proposes a method for preventing probe mark defects, comprising: in a situation where the prediction system receives/monitors a monitoring data corresponding to a current workpiece and generates a corresponding defect probability, comparing whether the defect probability exceeds a preset warning threshold; and in a situation where the defect probability exceeds the preset warning threshold, before the execution of the subsequent inspection procedures corresponding to other workpieces, adjusting the subsequent value of at least one factor in the monitoring data to be different from the current value of the at least one factor in the monitoring data, so that the defect probability corresponding to the value of each factor in the subsequent monitoring data will not exceed the preset warning threshold, so as to avoid the continuous use of values with a high probability of generating defects, thereby reducing the risk of generating defects and achieving the effect of preventing the occurrence of defects. Among them, the preset warning threshold can be set according to the actual processing requirements; the method of adjusting the value of at least one factor in the monitoring data has a corresponding adjustment method according to the characteristics of each factor.

Optionally, a data combination formed by different values of each factor in the training data/monitoring data can be pre-entered to pre-establish a predicted probability comparison data, which includes a defect probability corresponding to various different data combinations for whether a probe mark defect will occur. In this way, the predicted probability comparison data can be used as a reference for adjusting the values of each factor in the monitoring data, so that the adjusted defect probability of each factor is lower than the preset warning threshold.

Preferably, in a situation where the defect probability exceeds the preset warning threshold, the adjustment method for adjusting the value of at least one factor in the monitoring data can be performed by a computer according to pre-established rules to perform corresponding adjustments, thereby achieving automated adjustment.

In summary, the method for establishing a prediction model for probe mark defects of the present invention can ensure or improve the prediction accuracy of the corresponding prediction model during the training process and the application process after establishment through the combination of factors obtained by the importance analysis step and the factor re-screening step. In addition, the accuracy of the established prediction model can be improved by using the training model established by the sparse attention converter architecture and the corresponding hyperparameters. The prediction system of the present invention, through the prediction model established by the above method, can generate a corresponding defect probability when receiving the corresponding factor to monitor the risk of defects in the current state. The prevention method of the present invention can adjust the corresponding factor in real time by comparing whether the defect probability exceeds the preset warning threshold to reduce the risk of defects occurring during the process of the probe card detecting wafer characteristics.

Although the present invention has been disclosed using the above preferred embodiments, they are not intended to limit the present invention. Any person skilled in the art may make various changes and modifications to the above embodiments within the spirit and scope of the present invention, and the changes and modifications are still within the technical scope protected by the present invention. Therefore, the protection scope of the present invention shall include all changes within the meaning and equivalent scope recorded in the attached patent application scope.

S1: Data collection step

S1A: Data screening steps

S1A-1: Importance analysis steps

S1A-2: Factor re-screening step

S2: Model training steps

Claims (7)

A method for establishing a prediction model of a probe mark defect is performed by a computer to execute the following steps, the steps comprising: a data collection step, collecting a plurality of different factors in each probe mark generation process, and forming a preliminary factor combination from the plurality of different factors collected each time, and associating each preliminary factor combination with information on whether a defect occurs to define a preliminary associated information; an importance analysis step, using an importance analysis method to screen the preliminary associated information to obtain experience related to the occurrence of the defect; A preliminary screening factor combination is performed, and the preliminary screening factor combination is associated with information on whether a defect occurs to define an important correlation information; a factor re-screening step includes using a factor correlation analysis method to convert the important correlation information into a corresponding correlation coefficient matrix, wherein the correlation coefficient matrix includes a correlation between each factor and the defect and a correlation between any two factors; each factor corresponding to the correlation being not less than a first screening threshold is selected to define a high correlation factor combination; between any two factors, When the correlation between the factors is not less than a second screening threshold, the factor with the lower correlation with the defect is defined as a rejection factor; and the factor identical to the rejection factor is removed from the high correlation factor combination to obtain a corresponding re-screened factor combination, and the re-screened factor combination is associated with information on whether a defect occurs to define a prediction factor information; and a model training step is performed to input the prediction factor information into a prediction model to be trained to establish a trained prediction factor. The prediction model is established by a basic transformer architecture; the basic transformer architecture includes an encoder network and a decoder network; the encoder network includes a multi-head self-attention layer, a summing and normalization layer, a feedforward network layer and another summing and normalization layer in sequence; the decoder network includes a masked multi-head attention layer, a summing and normalization layer, a multi-head cross-attention layer, another summing and normalization layer, a feedforward network layer, a summing and normalization layer, a linear layer and a normalized exponential layer in sequence. A method for establishing a prediction model of a probe mark defect is performed by a computer to execute the following steps, the steps comprising: a data collection step, collecting a plurality of different factors in each probe mark generation process, and forming a preliminary factor combination from the plurality of different factors collected each time, and associating each preliminary factor combination with information on whether a defect occurs to define preliminary associated information; an importance analysis step, using an importance analysis method to screen the preliminary associated information to obtain preliminary associated information related to the occurrence of the defect. A step of screening factor combinations, and correlating the preliminarily screened factor combinations with information on whether defects occur, to define important correlation information; a step of factor re-screening, including using a factor correlation analysis method to convert the important correlation information into a corresponding correlation coefficient matrix, wherein the correlation coefficient matrix includes a correlation between each factor and the defect and a correlation between any two factors; selecting each factor corresponding to a correlation not less than a first screening threshold to define a high correlation factor combination; between any two factors When the corresponding correlation is not less than a second screening threshold, the factor with the lower correlation with the defect is defined as a rejection factor; and the factor identical to the rejection factor is removed from the high correlation factor combination to obtain a corresponding re-screened factor combination, and the re-screened factor combination is associated with information on whether a defect occurs to define a prediction factor information; and a model training step is performed to input the prediction factor information into a prediction model to be trained to establish a trained prediction model, the prediction factor information The test model is established with a sparse attention transformer architecture; The sparse attention transformer architecture includes an encoder network and a decoder network; the encoder network sequentially includes a multi-head self-attention layer, a summation and normalization layer, a feedforward network layer and another summation and normalization layer; the decoder network sequentially includes a sparse multi-head self-attention layer, a summation and normalization layer, a multi-head cross-attention layer, another summation and normalization layer, a feedforward network layer, a summation and normalization layer, a linear layer and a normalized exponential layer. As in claim 2, a method for establishing a prediction model of a probe mark defect, wherein the hyperparameters in the prediction model are configured as follows; in the initial hyperparameters: the batch size is 64, the training rounds are 50, and the learning rate is 0.0001; in the encoder network: the number of heads of the multi-head self-attention layer is 8, and the dimension of the attention key is 64; the discard rate in the feedforward network layer is 0. 1, the output dimension of the fully connected layer is 64 and uses the ReLU activation function; and in the decoder network: the number of heads of the sparse multi-head self-attention layer is 8, and the dimension of the attention key is 64; the discard rate in the feedforward network layer is 0.1, the output dimension of the fully connected layer is 64 and uses the ReLU activation function; the output dimension of the linear layer is 2, using the ReLU activation function. A method for establishing a prediction model for probe mark defects as in any one of claims 1 to 3, wherein the importance analysis method is a random forest algorithm. A method for establishing a prediction model for a probe mark defect as claimed in any one of claims 1 to 3, wherein the prediction factor information is composed of at least a factor of the minimum diameter of the needle tip, a factor of the correction position, and corresponding defect occurrence information. A prediction system for predicting probe mark defects, comprising a prediction model established by the method for establishing a prediction model for probe mark defects of any one of claim items 1 to 5, the prediction model is used to receive monitoring data to output a defect probability of whether a probe mark defect will occur; the type of each factor in the monitoring data is the same as the type of each factor in the prediction factor information. A preventive method for reducing probe mark defects, comprising a prediction system as claimed in claim 6, wherein the prediction system receives monitoring data corresponding to a current workpiece and generates a corresponding defect probability, and in a case where the defect probability exceeds a preset warning threshold, before executing the inspection procedure corresponding to other subsequent workpieces, the subsequent value of at least one factor in the monitoring data is adjusted to be different from the current value of the at least one factor in the monitoring data, so that the defect probability corresponding to the value of each factor in the subsequent monitoring data does not exceed the preset warning threshold.

Images