Wafer deformation recognition method and wafer deformation recognition system

Abstract

A wafer deformation recognition method includes acquiring measurement data of a wafer at a starting station after the wafer is loaded, establishing a plurality of machine learning models and a deep learning model, selecting at least one machine learning model from the plurality of machine learning models based on a plurality of confidence levels of the machine learning models for the wafer, integrating the at least one machine learning model and the deep learning model into an ensemble learning model, training the ensemble learning model, and inferring a deformation amount of the wafer to identify whether the wafer is good or bad by the ensemble learning model based on the measurement data of the wafer.

Description

Technical Field

[0001] This invention describes a wafer change recognition method and a wafer change recognition system, particularly a wafer change recognition method and a wafer change recognition system implemented using machine learning models and deep learning models. Prior Technology

[0002] In the semiconductor industry, wafer defects have always been a significant issue, as they affect yield, which in turn impacts production efficiency and product yield. Traditional wafer defect detection methods primarily rely on manual inspection, which is time-consuming, labor-intensive, and susceptible to human error. Therefore, the industry has been seeking automated, objective, and efficient wafer defect detection methods.

[0003] In recent years, machine learning technology has made significant progress in image recognition and has begun to be applied to wafer defect detection. However, most existing machine learning models are based on real defect images for judgment, while risk assessment of wafer shape still requires manual methods or numerical methods to identify each wafer. However, training personnel to perform manual wafer shape inspection is very time-consuming, and achieving an accuracy of over 90% has been virtually impossible in the past. In other words, traditionally, risk assessment and deformation prediction of wafer shape have relied mainly on manual visual inspection, which is time-consuming and manpower-intensive, and easily affected by subjective factors, making it difficult to achieve high accuracy. Furthermore, existing technologies mostly detect defects after they occur, lacking preventative mechanisms and failing to effectively avoid production stoppages and cost losses caused by wafer breakage and other problems.

[0004] Therefore, developing a machine learning-based wafer shape recognition method to predict and screen wafers that may cause production anomalies, and combining it with a statistical process control system to achieve an automated early warning and defense mechanism, is an important design issue. Summary of the Invention

[0005] This invention proposes a wafer deformation identification method. The wafer deformation identification method includes obtaining measurement data of the wafer at the starting site after wafer fabrication; establishing a plurality of machine learning models and a deep learning model; selecting at least one machine learning model from among the machine learning models based on a plurality of confidence levels of the wafer; integrating the at least one machine learning model with the deep learning model into an ensemble learning model; and inferring the wafer deformation based on the wafer measurement data using the trained ensemble learning model to identify the wafer's quality.

[0006] Another embodiment of the present invention provides a wafer deformation identification system. The wafer deformation identification system includes a data processing module, a memory, and a processor. The memory is coupled to the data processing module. The processor is coupled to the memory and is used to drive a plurality of machine learning models and deep learning models. The data processing module acquires measurement data of the wafer at the starting site after wafer fabrication and processes the measurement data. The memory caches the processed measurement data. The processor establishes the machine learning models and deep learning models. Based on a plurality of confidence levels of the wafer by the machine learning models, the processor selects at least one machine learning model from the machine learning models. The processor integrates the at least one machine learning model with the deep learning model into an ensemble learning model and trains the ensemble learning model. Based on the wafer's measurement data, the processor uses the trained ensemble learning model to infer the wafer's deformation to identify whether the wafer is of good or bad quality. Simple Explanation of the Diagram

[0007] Figure 1 is a block diagram of an embodiment of the wafer deformation recognition system of the present invention. Figure 2 is a schematic diagram of the machine learning model training procedure in the wafer variation recognition system shown in Figure 1. Figure 3 is a schematic diagram of the deep learning model training procedure in the wafer deformation recognition system shown in Figure 1. Figure 4 is a flowchart of the artificial intelligence model built on simulation and statistics in the wafer variation recognition system of Figure 1. Figure 5 shows the process of applying an artificial intelligence model combined with a statistical process control unit to the production process in the wafer deformation recognition system shown in Figure 1. Figure 6 is a virtual contour map of a high-quality wafer in the wafer transformation identification system shown in Figure 1. Figure 7 is a virtual contour map of a degraded quality wafer in the wafer defect identification system shown in Figure 1. Figure 8 shows the simulation results of the inference accuracy and training data index in the wafer variation recognition system of Figure 1. Figure 9 shows the simulation results of the inferred loss function and training data index in the wafer variation recognition system of Figure 1. Implementation

[0008] 1 is a block diagram of an embodiment of the crystalline variable recognition system 100 of the present invention. The crystalline variation recognition system 100 is designed to predict the crystalline variation and identify wafers with a high risk of warping to prevent fragmentation or other defects during production. The circular variable identification system 100 can also integrate artificial intelligence models into the Statistical Process Control (SPC) system of wafer production to realize automated wafer quality monitoring and early warning to reduce machine downtime and maintenance costs due to wafer defects and improve production efficiency and product yield. The wafer variable identification system 100 contains a data processing module 10 , a memory 11 , and a processor 12 . Memory 11 is coupled to the data processing module 10 . Processor 12, coupled to memory 11, is used to drive a complex number of machine learning (Machine Learning) models and deep learning (Deep Learning) models.

[0009] In the circular variable recognition system 100 , the data processing module 10 may obtain and process the measurement data at the starting site after the wafer is cast. For example, the data processing module 10 may perform the data preprocessing procedure 10a as well as the cutting data procedure 10b . After obtaining the measurement data, the data processing module 10 first performs the data preprocessing procedure 10a to convert the raw data into a format available for model training. Since the measurement data may contain noisy or invalid data points, it is necessary to first perform screening and filtering to eliminate unnecessary information and focus on data points that are highly relevant to the wafer shape. It should be understood that the wafer may be positioned as a complex number of sampling points (such as point 49) containing a wafer center point and a complex number of annularly arranged points. The toroidally arranged points will be evenly distributed over different radii of the wafer, forming multiple concentric circles. Therefore, after the wafer is cast, the data processing module 10 can obtain the measurement data of each of those sampling points of the wafer at the starting site, containing its morphology, location and special parameters, which will be used to build the model and perform training. Moreover, the data processing module 10 can perform a cutting data procedure (Data Splitting) 10b for dividing the measurement data into different fields and distinguishing the collected wafer measurement data into training set 11a , validation set 11b , and test set 11c .

[0010] The data in training set 11a, validation set 11b, and test set 11c can be stored in memory 11. Training set 11a is the largest part of the dataset and is used to train the machine learning model. The model learns from the data in training set 11a and adjusts its internal parameters to recognize patterns and relationships in the data. Validation set 11b is used to evaluate the model's performance during training. During training, the model is periodically tested using validation set 11b to evaluate its generalization ability and adjust model parameters to prevent overfitting. Test set 11c can be used to evaluate the performance of the final model. The data in test set 11c has not been seen during training and validation, thus simulating the model's performance in real-world applications. In one embodiment, the splitting ratio of the dataset can be determined according to specific circumstances, for example, 70% of the data may be used for training set 11a, 20% for validation set 11b, and 10% for test set 11c. However, this embodiment is not limited to this.

[0011] Processor 12 can execute machine learning model training program 12a, deep learning model training program 12b, and ensemble learning model building program 12c. In short, processor 12 executes machine learning model training program 12a, building multiple models using different machine learning algorithms. Examples include support vector machine models, decision tree models, random forest models, and extreme gradient boosting (XGBoost) models. The processor can improve the accuracy of the models by adjusting the hyperparameters of each model, such as setting the range of hyperparameters, the magnitude of each change, and the step size. In other words, processor 12 can build multiple models using different machine learning algorithms, and these machine learning models are used for numerical analysis processing. Furthermore, processor 12 executes deep learning model training program 12b. For example, processor 12 can build and train a convolutional neural network (CNN) model, and the CNN model can be used for image analysis processing. Furthermore, in the ensemble learning model building procedure 12c, the processor 12 integrates the machine learning model trained in the machine learning model training procedure 12a (e.g., the top-ranked best model) with the deep learning model trained in the deep learning model training procedure 12b (e.g., a convolutional neural network model CNN) to build and train the ensemble learning model. Moreover, if both the machine learning model and the deep learning model within the wafer variation recognition system 100 have been trained and their inference accuracy has reached a certain level, the wafer variation recognition system 100 can deploy its inference mechanism and model to an artificial intelligence server. The artificial intelligence server will then be able to output accurate inference results.

[0012] Figure 2 is a schematic diagram of the machine learning model training procedure 12a in the wafer variable recognition system 100. In Figure 2, the machine learning model training procedure 12a uses multiple machine learning models, including machine learning model 12-1a, machine learning model 12-2a, and machine learning model 12-Na. N is a positive integer greater than 2. Machine learning models 12-1a to 12-Na can correspond to different machine learning algorithms and support operations such as Support Vector Machine (SVM), Decision Tree, Random Forest, and Extreme Gradient Boosting (XGBoost) as mentioned above. In training multiple machine learning models 12-1a to 12-Na, the machine learning model training procedure 12a can automatically adjust model parameters (such as hyperparameters) to improve model accuracy. In Figure 2, the accuracy of each machine learning model corresponds to its confusion matrix. It should be understood that a confusion matrix is ​​a visualization tool used to evaluate the classification performance of a machine learning model. It displays the model's performance in predicting different categories, including the number of correct and incorrect predictions. A confusion matrix is ​​typically presented in tabular form, where rows represent the categories predicted by the model and columns represent the true categories of the data. In this embodiment, the predicted category corresponds to a wafer's predicted quality of good or bad. Similarly, the true category corresponds to a wafer's predicted quality of good or bad. Therefore, the confusion matrix in this embodiment can be represented as follows, see Confusion Matrix T1. True quality: Excellent True quality: Deteriorated Predicted quality: Excellent TP% (Truly Excellent) FP% (False Excellent, Miss) Predicted quality: deterioration FN% (False Alarm) TN% (True Degradation) Confusion matrix T1

[0013] Based on the confusion matrix T1, the accuracy of the machine learning model can be expressed as follows:

[0014] As mentioned above, accuracy is defined as the proportion of samples correctly predicted by the model out of the total number of samples. Therefore, in the machine learning model training procedure 12a, each machine learning model can use a confusion matrix to generate a corresponding accuracy. For example, machine learning model 12-1a corresponds to accuracy 12-11a. Machine learning model 12-2a corresponds to accuracy 12-12a. And so on, machine learning model 12-Na corresponds to accuracy 12-1Na. As mentioned earlier, processor 12 can sort the accuracy 12-11a to 12-1Na corresponding to machine learning models 12-1a to 12-Na, and select at least one better machine learning model to conform to the current wafer deformation identification algorithm, thereby increasing its identification reliability.

[0015] Figure 3 is a schematic diagram of the deep learning model training program 12b in the wafer deformation recognition system 100. The deep learning model training program 12b may include a measurement data to image conversion program 12-1b, a high-order framework construction program (such as a Keras Model) 12-2b, and measurement for accuracy 12-3b. In the measurement data to image conversion program 12-1b, this program converts the measurement data into images to facilitate the training of the deep learning model. For example, the measurement data to image conversion program 12-1b uses the Python programming language to draw a virtual contour map from the 49 measurement data of the wafer to represent the shape of the wafer surface. The purpose of the measurement data to image conversion program 12-1b is to convert the original numerical data into image data to facilitate training using a convolutional neural network (CNN) model. As mentioned earlier, CNN models are adept at processing image data and can effectively extract features from images, thereby improving the accuracy of the model. In Figure 3, the output of the data conversion graphics program 12-1b is a virtual contour plot, which will be used in conjunction with a high-order framework program (such as Keras Model) 12-2b to train a deep learning model.

[0016] In the high-order framework construction procedure 12-2b, the Keras model can be used. The Keras model is a deep learning model used to convert wafer shape data into images and perform image recognition to achieve higher prediction accuracy. The Keras model is a type of neural network model that can employ a multi-layered structure, including an input layer, convolutional layers, maximum pooling layers, flattened layers, and dense layers for feature extraction and prediction. Furthermore, Keras model parameter tuning can include epochs (each time all training data has been reviewed is counted as one epoch), batch size (determining the amount of data the model reviews for each parameter update), and the learning rate, among other things. Keras model parameter tuning can be accomplished using automatic parameter tuning tools to achieve optimal prediction results. Similarly, the deep learning model training procedure 12b can also continuously calculate the accuracy 12-3b to ensure the high stability and reliability of the convolutional neural network (CNN) model.

[0017] In simple terms, in the circular deformation recognition system 100, after the data processing module 10 processes the measurement data, the memory 11 can cache the processed measurement data. The processor 12 establishes these machine learning models and deep learning models. The processor 12 can select at least one machine learning model from these machine learning models based on the multiple confidence levels of the wafer. The processor 12 can integrate at least one machine learning model with the deep learning model into an ensemble learning model and train the ensemble learning model. Finally, based on the wafer's measurement data, the processor 12 uses the trained ensemble learning model to infer the wafer's deformation to identify whether the wafer is of good or bad quality. The details of training and applying the machine learning and deep learning models in the circular deformation recognition system 100 will be explained below.

[0018] Figure 4 is a flowchart illustrating the construction of the artificial intelligence model based on simulation and statistics in the wafer variation recognition system 100. The flowchart in Figure 4 includes steps S401 to S413. Any reasonable technical modifications to steps S401 to S413 fall within the scope disclosed in the embodiments. Steps S401 to S413 are explained below. Step S401: Measurement data acquisition; Step S402: Classify data using simulation / statistical methods; Step S403: Establish multiple models; Step S404: Validate multiple models; Step S405: Calculate the confidence level of the model; Step S406: Select the best model; Step S407: Merge into a new model; Step S408: Calculate the confidence level of the new model; Step S409: Mislabeled data; Step S410: Retrain a new model; Step S411: Validate the new model; Step S412: Generate the coefficient of determination R², and determine if R² is greater than the goodness-of-fit threshold. If so, Execute step S413; if not, return to step S410; Step S413: Deploy new models on AI servers.

[0019] In step S401, the data processing module 10 acquires measurement data of the wafer at the starting site after wafer fabrication and processes the measurement data. As mentioned above, the measurement data may include wafer warpage, curvature, and total thickness, etc. The measurement data can be combined with the location of these sampling points on the wafer and multiple key parameters, and assigned corresponding multiple weights to form a virtual contour map. In step S402, the data can be classified using simulation / statistical methods. Classifying data using simulation / statistical methods refers to classifying the measurement data into different groups, such as good and bad wafers, using simulation or statistical methods. Simulating how deformation occurs can help understand the distribution of data and identify key factors affecting the results. For example, when training a machine learning model, simulated data can be used to first label which wafers are normal and which wafers have saddle-shaped deformation. Then, the actual measured data is input into the model, allowing the model to learn to judge whether the wafer is normal. In step S403, processor 12 can build multiple models, such as six machine learning models and a deep learning model like CNN (a total of seven models). Next, in step S404, processor 12 can verify the multiple machine learning models and one deep learning model. In steps S405 and S406, processor 12 can obtain the confidence levels of these six machine learning models for the wafer. It should be understood that confidence level refers to the degree of certainty a machine model has about its predictions. Accuracy, as mentioned above, refers to the correctness of a machine model's predictions. The correlation between confidence and accuracy is that a prediction with high confidence usually also has high accuracy. This is because if a model is very certain about its predictions, it is more likely to be correct. However, confidence and accuracy are not always perfectly aligned. A model may have high confidence in a prediction but actually be wrong. This could be due to insufficient training data or inherent flaws in the model itself. In evaluating the performance of machine models, embodiments may also consider both confidence and accuracy simultaneously. This is because a good model should have both high confidence and high accuracy. Therefore, in one embodiment, in step S406, the processor 12 can sort the machine learning models from high confidence to low confidence based on these confidence levels, and select at least one machine learning model with a confidence level higher than a threshold (e.g., selecting 2 out of 6 machine learning models). In another embodiment, in step S406, the processor 12 can sort the machine learning models simultaneously based on both confidence and accuracy, and select at least one machine learning model with both high confidence and high accuracy. Any reasonable technical modifications fall within the scope disclosed in these embodiments.

[0020] In step S406, processor 12 selects at least one machine learning model (e.g., two) and combines it with a deep learning model such as a CNN (a total of three models). Next, in step S407, processor 12 integrates the at least one machine learning model with the deep learning model to establish an ensemble learning model. Next, in step S408, processor 12 calculates the confidence score of the new model (ensemble learning model). In step S409, processor 12 labels misclassified data. For example, processor 12 uses the ensemble learning model to generate results for misclassified data and labels the misclassified data. In step S410, processor 12 retrains the ensemble learning model using the new dataset and the labeled misclassified data to update the model's parameters and structure. In step S411, processor 12 validates the ensemble learning model. In step S412, processor 12 validates the trained ensemble learning model to generate the coefficient of determination (R-squared) R². The coefficient of determination R², also known as the determination coefficient, is a comprehensive indicator reflecting the model's good fit. The coefficient of determination R² ranges from 0 to 1. A higher R² value indicates a better model fit and stronger explanatory power. Conversely, a lower R² value indicates a lower model fit and weaker explanatory power. It should be understood that the R² value can be used to evaluate the predictive ability of an ensemble learning model. For example, if the R² value is greater than the goodness-of-fit threshold, the ensemble learning model has good predictive ability and can be deployed to the AI ​​server according to step S413. Conversely, if the R² value is less than or equal to the goodness-of-fit threshold, the ensemble learning model has poor predictive ability, and the process returns to step S410 to continue training the ensemble learning model.

[0021] In the wafer deformation recognition system 100, the trained ensemble learning model can infer the wafer deformation to identify whether the wafer is of good or bad quality. In one embodiment, the trained ensemble learning model can infer the wafer deformation based on wafer measurement data to generate a probability corresponding to the wafer. For example, the probability of the wafer predicting good quality is 90%. The wafer deformation recognition system 100 can identify whether the wafer is of good or bad quality based on the probability corresponding to the wafer. As mentioned above, accuracy is defined as the proportion of samples correctly predicted by the model to the total number of samples. Therefore, if the trained ensemble learning model has high accuracy, the result of the trained ensemble learning model predicting whether the wafer is of good or bad quality will be very similar to the actual result of the wafer's good or bad quality, thus having high reference value. Furthermore, the method by which the wafer deformation recognition system 100 identifies whether the wafer is of good or bad quality can be any reasonable decision strategy. For example, assuming that the probability corresponding to the wafer is positive, if the probability corresponding to the wafer is greater than or equal to a threshold value, the prediction of the wafer as good or bad quality can be identified. If the probability corresponding to the wafer is less than the threshold value, the wafer can be identified as having degraded quality.

[0022] Figure 5 illustrates the process of applying an artificial intelligence model combined with a statistical process control (SPC) terminal to the production stage in the wafer variation recognition system 100. The process in Figure 5 includes steps S501 to S511. Any reasonable technical modifications to steps S501 to S511 fall within the scope disclosed in this embodiment. Steps S501 to S511 are explained below. Step S501: Wafer casting; Step S502: Measurement sites; Step S503: Upload measurement data to the AI ​​server; Step S504: New model calculations and inferences; Step S505: After calculation, the data, along with the original data, is returned to the standard process control (SPC) terminal. Step S506: Is there an alarm signal at the statistical process control terminal? If yes, proceed to step S507; if no, proceed to step S510. Step S507: Site engineers confirmed; Step S508: Is this a true result? If yes, proceed to step S509; if no, proceed to step S511. Step S509: Abnormal wafers were separated. Step S510: Automatic continuation of the process. Step S511: Continue with the process.

[0023] In step S501, wafer feeding refers to the initial step of sending wafers into the semiconductor manufacturing plant's production line. The wafers are placed at an initial station, ready for subsequent production processes. In step S502, the measurement station refers to the initial station set up on the wafer production line, used to measure various parameters of the wafer, such as wafer thickness, warpage, and curvature. These parameters are used for subsequent analysis and judgment. In step S503, after obtaining the wafer measurement data at the measurement station, this data is uploaded to the AI ​​server. The server uses a trained ensemble learning model to analyze and judge this data. In step S504, the AI ​​server uses the latest ensemble learning model to perform calculations and inferences on the uploaded measurement data to determine whether the wafer meets production standards. In step S505, after completing the calculations and inferences, the AI ​​server sends the results, along with the original measurement data, back to the Statistical Process Control (SPC). The SPC performs further analysis and judgment based on this data to decide whether to issue an alarm.

[0024] In step S506, the Statistical Process Control (SPC) is checked for alarms. SPC is a tool that uses statistical methods to monitor and improve the production process. It helps identify and analyze anomalies in the production process and allows for timely adjustments to ensure consistent and stable product quality. In wafer manufacturing, SPC can monitor various key parameters, such as wafer thickness, warpage, curvature, and other factors that may affect wafer quality. When the SPC issues an alarm, it indicates a potential or possible anomaly in the wafer. Engineers or technicians must be alerted to take necessary measures, such as adjusting production parameters, inspecting equipment, or rejecting defective wafers, to prevent the problem from escalating. Therefore, in step S507, when the SPC issues an alarm, the site engineer needs to verify the alarm's authenticity and take necessary measures to resolve the issue. When the Statistical Process Control (SPC) does not issue an alarm, it indicates that the wafer production process is normal. Therefore, according to step S510, the process can automatically continue, sending the wafer to the next production station to continue subsequent production steps. However, when the SPC issues an alarm, and the engineer confirms the alarm's authenticity, according to step S509, the abnormal wafer needs to be separated and sent to a specific processing station for additional inspection, repair, or scrapping to prevent these wafers from continuing to undergo subsequent production steps, affecting product quality and yield. However, if the engineer confirms that the alarm is a false alarm, it indicates that the wafer production process is normal. Therefore, according to step S511, the process can continue.

[0025] Figure 6 shows the virtual contour map VM1 of a high-quality wafer in the wafer deformation recognition system 100. Figure 7 shows the virtual contour map VM2 of a low-quality wafer in the wafer deformation recognition system 100. In this embodiment, the virtual contour map is a visualization tool used to represent the shape of the wafer surface, similar to contour lines on a map. It is drawn using the Python programming language based on measurement data from multiple points on the wafer. The virtual contour map can show the undulations of the wafer surface and whether there is a saddle-shaped deformation. The virtual contour map converts the original numerical data into image data to facilitate training using a convolutional neural network model. In Figure 6, the contour lines of the virtual contour map VM1 of the high-quality wafer are relatively sparse and the color distribution is uniform, indicating that the surface of the high-quality wafer is flat and there are no obvious undulations or saddle-shaped deformations. However, in Figure 7, the virtual contour map VM2 of the degraded quality wafer shows relatively dense contour lines and uneven color distribution, indicating that the surface of the degraded wafer has obvious undulations and saddle-shaped deformation. In the circular deformation recognition system 100, a deep learning model (such as CNN) is particularly well-suited for image analysis processing. Therefore, virtual contour maps are highly suitable for using deep learning models to infer whether a wafer is deformed and its potential quality.

[0026] Figure 8 shows the simulation results of the inference accuracy and training data index in the wafer shape recognition system 100. In Figure 8, the X-axis represents the training data index, and the Y-axis represents accuracy. As mentioned earlier, accuracy is defined as the proportion of samples correctly predicted by the model out of the total number of samples. Therefore, accuracy is a value between 0 and 1. When the accuracy is 0, it means that the model did not correctly predict any samples at all, and all predictions were incorrect. When the accuracy is 1, it means that the model correctly predicted all samples, and all predictions were correct. The higher the quality of the training data, the higher the accuracy of the model. This is because the higher the quality of the training data, the better the model can learn the patterns and rules in the data, and thus make more accurate predictions. In Figure 8, the average accuracy of the model's inference reaches over 90%. This indicates that the inference method of the wafer shape recognition system 100 is superior to the traditional method that can only be performed manually by trainers for wafer shape detection. Furthermore, to increase the stability of model accuracy, the wafer variation recognition system 100, when training the model, not only increases the amount of training data but also emphasizes the quality and distribution of the training data. In the embodiment, the wafer variation recognition system 100 achieves an accuracy of approximately 95.1% for the Support Vector Machine (SVM) model, approximately 94.3% for the Random Forest model, and approximately 93.8% for the Logistic Regression (LR) model.

[0027] Figure 9 shows the simulation results of the inference loss function and training data index in the wafer variation recognition system 100. In Figure 9, the X-axis represents the training data index, and the Y-axis represents the loss function. It should be understood that the loss function is an important metric in the deep learning model training process, used to measure the difference between the model's predicted values ​​and the actual values. The smaller the loss function value, the closer the model's prediction is to the true value, and the better the model's performance. The loss function can be reduced by adjusting model parameters, such as the learning rate, batch size, and number of training iterations. In the wafer variation recognition system 100, the loss function can be used to evaluate the performance of the deep learning model and serve as a reference metric for model training. In Figure 9, the loss function converges as the amount of training data increases. This indicates that the model gradually reduces the prediction error during training. Furthermore, in Figure 9, the convergence speed of the loss function is relatively fast, approaching convergence around 250 training data points. Therefore, in addition to its excellent accuracy, the wafer deformation recognition system 100 also has a fast convergence speed when training the model.

[0028] In summary, the embodiments provide a wafer deformation identification system and method. Through the establishment of a machine learning model, it can be used to predict and determine whether the wafer shape is a saddle shape that leads to breakage. For example, in the Metal-Organic Chemical Vapor Deposition (MOCVD) process, these wafers are prone to deformation due to high temperatures and material properties, leading to breakage and affecting machine operation. Therefore, the wafer deformation identification system uses shape measurement tools to collect multi-point shape data of the wafer and converts it into contour maps. Then, machine learning and deep learning models are used to train and test this data. The wafer deformation identification system first identifies the key shape features leading to breakage through simulation and statistical analysis, and uses this knowledge to train the machine learning model. During model training, multiple models are used, and the best model is selected based on its accuracy. Then, these best models are combined into an integrated model to predict new wafer shapes. The prediction results of this integrated model are fed back to the statistical process control system of wafer production for defensive measures and to trigger alarms. In summary, the wafer defect identification system and method are used to predict and screen wafers that may cause production anomalies. Combined with the statistical process control system, they enable automated early warning and defense mechanisms to reduce machine downtime and maintenance costs caused by wafer defects, and improve production efficiency and product yield. The above description is only a preferred embodiment of the present invention. All equivalent changes and modifications made in accordance with the claims of the present invention shall be covered by the present invention.

[0029] 100: Wafer Variation Recognition System 10: Data Processing Module 10a: Data Preprocessing Procedures 10b: Data Cutting Program 11: Memory 11a: Training set 11b: Validation set 11c: Test Set 12a: Machine Learning Model Training Procedure 12b: Deep Learning Model Training Procedure 12c: Ensemble Learning Model Building Program 12-1a to 12-Na: Machine Learning Models 12-11a to 12-1Na: Precision 12-1b: Measurement Data to Graphical Conversion Program 12-2b: Building a High-Order Framework Program 12-3b: Accuracy S401 to S413: Steps S501 to S511: Steps VM1: Virtual contour plot of high-quality wafers VM2: Virtual contour plot of degraded quality wafers

Claims

1. A wafer deformation identification method, comprising: acquiring measurement data of a wafer at a starting site after wafer fabrication; establishing a plurality of machine learning models and a deep learning model; acquiring a plurality of confidence levels of the machine learning models for the wafer; sorting the machine learning models from high confidence to low confidence based on the confidence levels; selecting at least one machine learning model from the machine learning models with a confidence level higher than a threshold value; integrating the selected at least one machine learning model with the deep learning model into an ensemble learning model; training the ensemble learning model; and inferring a deformation of the wafer based on the measurement data of the wafer, thereby identifying whether the wafer is of good or bad quality.

2. The method as described in claim 1, wherein obtaining measurement data of the wafer at the starting site after wafer fabrication comprises: positioning the wafer as a plurality of sampling points, the sampling points including a wafer center point and a plurality of annularly arranged points; and obtaining measurement data of each of the sampling points of the wafer at the starting site after wafer fabrication.

3. The method as described in claim 2, wherein the measurement data includes a warpage, a curvature and a total thickness of the wafer, and the measurement data, in conjunction with the positions of the sampling points on the wafer and a plurality of key parameters, is assigned corresponding weights to form a virtual contour map.

4. The method as described in claim 1, wherein the establishment of the machine learning models is to establish multiple models using different machine learning algorithms, the machine learning models being used for numerical analysis processing, the deep learning model being a convolutional neural network (CNN) model, and the CNN model being used for image analysis processing.

5. The method as described in claim 1, further comprising: validating the trained ensemble learning model to generate a coefficient of determination (R-squared); and determining, based on the coefficient of determination, whether the trained ensemble learning model should continue training; wherein the coefficient of determination is related to a goodness-of-fit index of the ensemble learning model, and the coefficient of determination is between 0 and 1.

6. The method as described in claim 5, wherein determining whether to continue training the trained ensemble learning model based on the determination coefficient comprises: if the determination coefficient is greater than a goodness-of-fit threshold, deploying the trained ensemble learning model on an artificial intelligence server; and if the determination coefficient is less than or equal to the goodness-of-fit threshold, continuing to train the trained ensemble learning model.

7. The method as described in claim 1, wherein the trained ensemble learning model infers the deformation of the wafer based on the measurement data of the wafer to identify whether the wafer is superior or inferior comprises: inferring the deformation of the wafer based on the measurement data of the wafer to generate a probability corresponding to the wafer; and identifying whether the wafer is superior or inferior based on the probability corresponding to the wafer.

8. The method as described in claim 7, wherein identifying whether a wafer is good or bad based on the probability corresponding to the wafer includes: if the probability corresponding to the wafer is greater than or equal to a threshold value, identifying the wafer as good quality; and if the probability corresponding to the wafer is less than the threshold value, identifying the wafer as bad quality.

9. A wafer variation recognition system, comprising: a data processing module; a memory coupled to the data processing module; and a processor coupled to the memory for driving a plurality of machine learning models and a deep learning model; wherein the data processing module acquires measurement data of a wafer at a starting site after wafer fabrication and processes the measurement data; the memory caches the processed measurement data; the processor establishes the machine learning models and the deep learning model; the processor acquires a plurality of confidence levels of the machine learning models for the wafer and sorts the machine learning models from high confidence to low confidence based on the confidence levels; the processor selects at least one machine learning model from the machine learning models with a confidence level higher than a threshold value; and the processor integrates the selected at least one machine learning model with the deep learning model into an ensemble learning model. The processor trains the ensemble learning model and uses the trained ensemble learning model to infer a shape variable of the wafer based on the measurement data of the wafer, so as to identify whether the wafer is of good or bad quality.

Images