Wafer cutting yield prediction method based on multi-source process parameters

Abstract

The invention discloses a wafer cutting yield prediction method based on multi-source process parameters, and aims to solve the problems of inaccurate prediction and poor model robustness caused by dependence on false correlation in the prior art. The method comprises the steps of collecting and preprocessing multi-source process data to form a historical data set; constructing a causal graph based on a causal discovery algorithm, and identifying key process parameters having a direct causal effect on the cutting yield to form a core causal feature set; constructing a neural network prediction model by taking the feature set as input, and introducing causal priori constraints in training; adversarial robust training is carried out by mixing an adversarial sample and an original sample, and a robust prediction model is obtained; and inputting the core causal characteristics of the wafer to be tested into the model, and outputting the predicted cutting yield. According to the technical scheme, causals and correlations can be effectively distinguished, model robustness, interpretability and prediction precision are improved, and accurate prediction and active management and control of the wafer cutting yield are achieved.

Description

Technical Field

[0001] This invention belongs to the field of semiconductor manufacturing technology, and specifically discloses a method for predicting wafer dicing yield based on multi-source process parameters. Background Technology

[0002] Wafer dicing is a critical process for separating individual chips from a wafer, and its yield directly impacts the cost and performance of the final product. Currently, the industry commonly uses machine learning-based methods to predict dicing yield by analyzing historical process parameters. However, these methods suffer from two main drawbacks: First, they confuse correlation with causation. Traditional models rely on statistical correlation for prediction, but complex interdependencies exist among process parameters. Many parameters have only "spurious correlations" with dicing yield rather than "true causation." When minor drifts occur in the production process or new disturbances are introduced, the predictive performance of correlation-based models drops sharply, exhibiting poor generalization ability. Second, they are sensitive to noise and disturbances. Sensor noise and material batch fluctuations are unavoidable in the production environment. Traditional models are highly sensitive to these minor disturbances, leading to unstable prediction results and failing to provide a reliable basis for refined process control. Therefore, there is an urgent need in this field for a wafer dicing yield prediction method that can reveal true causal relationships and is robust to production environment disturbances. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides a method for predicting wafer dicing yield based on multi-source process parameters, comprising the following specific steps: Step S1: Collect multi-source process data related to wafer dicing yield, and perform preprocessing and alignment to form a historical dataset; Step S2: Based on the causal discovery algorithm, learn the causal graph between variables from the historical data, and identify key process parameters that have a direct causal effect on the cutting yield, thus forming a core causal feature set; Step S3: Using the core causal feature set as input, construct a neural network prediction model, and introduce causal prior constraints based on the causal graph during the model training process; Step S4: During the training phase of the neural network prediction model, adversarial examples are generated for the core causal feature set, and the original samples and adversarial examples are mixed for adversarial robust training to obtain a robust prediction model. Step S5: Input the core causal features from the real-time process data of the wafer to be predicted into the robust prediction model, and output the predicted dicing yield of the wafer.

[0004] Preferably, in step S1, the multi-source process data includes structured parameters from the preceding manufacturing process, setting parameters from the cutting equipment, and temporal characteristics of vibration and acoustic emission signals from online sensors.

[0005] Preferably, the preprocessing and alignment in step S1 includes data cleaning, outlier handling, timestamp alignment, and feature standardization.

[0006] Preferably, in step S2, in the scenario where unobserved confounding variables are not considered, the causal discovery algorithm adopts the PC algorithm; the significance level of the PC algorithm is 0.05, and the maximum condition set size is 5; in the scenario where unobserved confounding variables are considered, the causal discovery algorithm adopts the FCI algorithm; identifying key process parameters includes locating the direct causal parent node of the cutting yield node, calculating its average treatment effect, and selecting parameters with an average treatment effect exceeding 0.1 as core causal features. Both algorithms can effectively handle the complex dependencies between high-dimensional process parameters, and the accuracy of causal graph construction can reach over 85%.

[0007] Preferably, in step S2, identifying key process parameters with a direct causal effect on cutting yield includes locating the direct causal parent node of the cutting yield node in the causal graph and calculating its average treatment effect. The average treatment effect calculation employs a latent outcome framework based on counterfactual reasoning, using the do operator for intervention calculation, and selecting parameters with an average treatment effect exceeding a preset threshold of 0.1 as core causal features. This threshold can be adjusted from 0.05 to 0.2 according to actual production requirements to ensure that the selected features have a significant causal influence.

[0008] Preferably, in step S3, the neural network prediction model is constructed using a deep feedforward neural network structure, comprising an input layer, three hidden layers, and an output layer. The number of neurons in the input layer is equal to the dimension of the core causal feature set. The number of neurons in the hidden layers are 128, 64, and 32, respectively, and the ReLU activation function is used. The output layer consists of a single neuron, using the Sigmoid activation function to output a predicted cut yield value between 0 and 1. The network optimizer is the Adam optimizer, with an initial learning rate set to 0.001 and a batch size set to 32.

[0009] Preferably, in step S3, the causal prior constraint is introduced using a structural constraint method. Between the input layer and the first hidden layer of the neural network, only the connection weights corresponding to the core causal feature set are retained, while the connection weights of non-core features are masked. Specifically, a binary mask matrix is ​​designed to force the weights corresponding to non-core features to zero, and these weights do not participate in gradient updates during training. This method ensures that the model learns only based on features with genuine causal relationships, effectively avoiding interference from spurious correlations.

[0010] Preferably, in step S3, the causal prior constraint is introduced by adding a causal regularization term to the loss function during model training. The causal regularization term is based on the intervention distribution defined by the causal graph, calculating the KL divergence between the model prediction and the ideal intervention result, with a regularization coefficient set to 0.01. The loss function consists of a weighted average of the mean squared error loss and the causal regularization term, with a weight ratio of 10:1, thus enhancing causal consistency while ensuring prediction accuracy.

[0011] Preferably, in step S4, the projective gradient descent method is used to generate adversarial examples for the core causal feature set. The perturbation generation step size is set to 0.01, the maximum number of iterations is 10, and the perturbation amplitude is limited to the feature value within a range of ±0.1. The generated perturbation is only applied to the data dimension corresponding to the core causal feature set, while the non-core feature dimensions remain unchanged, ensuring that adversarial training specifically improves the robustness of the model to perturbations of key process parameters.

[0012] Preferably, in step S4, the adversarial robust training mixes the original samples and adversarial samples in a 1:1 ratio to form an enhanced training set. During training, a dynamic sample weight adjustment strategy is employed, assigning higher weights to adversarial samples that are difficult to classify correctly, with the weight adjustment coefficient ranging from 0.5 to 2.0. Through this training method, the model maintains its predictive ability for normal samples while significantly improving its adaptability to fluctuations in process parameters, increasing prediction stability by more than 30%.

[0013] Preferably, step S5 is followed by step S6: calculating an uncertainty measure for predicting the cutting yield using Monte Carlo Dropout technology. Dropout is kept active during the prediction phase, and 50 forward propagations are performed to calculate the mean and variance of the prediction results. The uncertainty measure is defined as the prediction variance. When the variance exceeds a threshold of 0.05, the prediction result is considered unreliable and requires manual review or remeasurement.

[0014] Preferably, step S5 is followed by step S7: calculating the uncertainty measure of the predicted cutting yield using deep ensemble technology; training five neural network models with identical structures but different initializations using the deep ensemble technology, each model being trained on an independent subset of data. In the prediction phase, the outputs of all models are combined to calculate the prediction mean and variance, and the accuracy of the ensemble prediction is improved by 5% to 8% compared to a single model.

[0015] Preferably, an early warning signal is triggered when the predicted cutting yield and / or uncertainty measure exceed a preset threshold. The predicted cutting yield threshold is set according to product specifications, typically 85%; the uncertainty measure threshold is set to 0.05. The early warning signal is simultaneously issued through three methods: an audible and visual alarm, SMS notification, and a pop-up window on the industrial control system interface, ensuring that production personnel are promptly informed of any abnormalities.

[0016] Preferably, the method further includes establishing a dynamic relationship model between process parameters and dicing yield. Based on real-time process data acquired online, a sliding window method is used to extract time-series features, with the window size set to the data of the most recent 100 wafers and the step size being one wafer. A recurrent neural network is used to model the dynamic characteristics of process parameters changing over time, enabling early identification and warning of dicing process trends.

[0017] Preferably, the method further includes constructing an adaptive model update mechanism. When the prediction error of 10 consecutive wafers exceeds 5%, the model retraining process is automatically triggered. Retraining uses historical data from the most recent 1000 wafers to update the neural network weights while maintaining the original causal structure, ensuring that the model can adapt to the slow drift of the manufacturing process.

[0018] Preferably, the method is applied to a wafer manufacturing execution system, interacting with production equipment via the standard OPCUA protocol. The prediction results are displayed in real-time on the production line monitoring screen, updated within 30 seconds of each wafer being diced, providing timely decision support for process engineers.

[0019] Preferably, the method supports yield prediction for dicing multiple wafer types, distinguishing different types of wafers through product type coding. The model automatically loads the corresponding cause-effect graph and neural network parameters according to the product type, enabling simultaneous prediction of multiple products on a single production line, with a model switching response time of less than 2 seconds.

[0020] Compared with the prior art, the embodiments disclosed in this invention have the following beneficial effects: By employing a causal discovery algorithm to identify key process parameters with genuine causal effects on wafer dicing yield, the model effectively distinguishes between correlation and causal relationships, avoiding erroneous predictions based on spurious correlations. Adversarial robust training is used to enhance the model's robustness to production environment disturbances, significantly improving prediction stability. Causal prior constraints are introduced to ensure that the model learning process conforms to physical mechanisms, improving the model's interpretability and generalization ability. Combined with uncertainty measurement and early warning mechanisms, a reliable basis is provided for refined process control, ultimately achieving accurate prediction and proactive management of wafer dicing yield. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall technical solution architecture of the wafer dicing yield prediction method based on multi-source process parameters in this invention. Figure 2 This is a schematic diagram of the core principle framework of the present invention for identifying the core causal feature set based on the causal discovery algorithm and constructing a causal constraint neural network; Figure 3 This is a logical flowchart of the adversarial robust training and uncertainty-driven early warning mechanism in this invention. Detailed Implementation

[0022] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Currently, wafer dicing is a critical process for separating individual chips from a wafer, and its yield directly impacts the cost and performance of the final product. The industry commonly employs machine learning-based methods to predict dicing yield by analyzing historical process parameters. However, these methods suffer from two main drawbacks: First, they confuse correlation with causation. Traditional models rely on statistical correlation for prediction, but complex interdependencies exist among process parameters. Many parameters exhibit only "spurious correlations" with dicing yield rather than "true causation." When minor drifts occur in the production process or new disturbances are introduced, the predictive performance of correlation-based models drops sharply, exhibiting poor generalization ability. Second, they are sensitive to noise and disturbances. Sensor noise and batch material fluctuations are unavoidable in the production environment. Traditional models are highly sensitive to these minor disturbances, leading to unstable prediction results and failing to provide a reliable basis for refined process control. Therefore, there is an urgent need in this field for a wafer dicing yield prediction method that can reveal true causal relationships and is robust to production environment disturbances. To address the aforementioned technical issues, this invention proposes a causal discovery algorithm to identify key process parameters with genuine causal effects on wafer dicing yield, effectively distinguishing between correlation and causal relationships and avoiding erroneous predictions based on spurious correlations. Adversarial robust training is employed to enhance the model's robustness to production environment disturbances, significantly improving prediction stability. Causal prior constraints are introduced to ensure the model learning process conforms to physical mechanisms, improving the model's interpretability and generalization ability. Combining uncertainty measurement and early warning mechanisms provides a reliable basis for refined process control, ultimately achieving accurate prediction and proactive management of wafer dicing yield. This method is applied to a wafer dicing yield prediction method based on multi-source process parameters.

[0024] Reference Appendix Figure 1 The overall technical architecture of this invention includes a multi-source process data acquisition and preprocessing module, a causal feature recognition module, a causal constraint neural network construction module, an adversarial robust training module, a real-time prediction and uncertainty assessment module, and an early warning and adaptive update module. These modules work collaboratively to form a closed-loop cutting yield prediction and process optimization system.

[0025] Specifically, in step S1, multi-source process data related to wafer dicing yield is collected, preprocessed, and aligned to form a historical dataset. The multi-source process data includes structured parameters from preceding manufacturing processes, setting parameters from the dicing equipment, and temporal characteristics of vibration and acoustic emission signals from online sensors. The structured parameters from preceding manufacturing processes cover physical characteristics such as thin film deposition thickness, etching depth, and doping concentration. Data is collected at least three times per wafer for each critical process node, corresponding to process checkpoints after deposition, etching, and ion implantation. The setting parameters from the dicing equipment include process control parameters such as spindle speed range of 8000 to 30000 rpm, feed rate range of 0.1 to 5 mm / s, and dicing blade thickness of 100 to 500 micrometers. These parameters are set and recorded by the equipment controller before the dicing task begins. The vibration signal frequency range of the online sensor is 10 Hz to 10 kHz, and the acoustic emission signal frequency range is 50 kHz to 1 MHz. The sampling rate is no less than 100 kHz, ensuring that it can capture the micro-dynamic features during the cutting process, such as transient events like tool wear and material fracture.

[0026] In step S1 above, preprocessing and alignment include data cleaning, outlier handling, timestamp alignment, and feature standardization. Data cleaning employs outlier detection based on the 3σ principle, removing or correcting data points exceeding three standard deviations above or below the mean. The mean and standard deviation are dynamically calculated based on historical data from the most recent 500 wafers within a sliding window to accommodate slow process drift. Timestamp alignment uses interpolation to unify process parameters with different sampling frequencies to the same time base, achieving millisecond-level interpolation accuracy. For low-frequency data such as structured parameters, zero-order hold interpolation is used; for high-frequency vibration and acoustic emission signals, cubic spline interpolation is used to preserve the dynamic characteristics of the signal. Feature standardization uses a max-min normalization method to map all process parameters to a numerical range of 0 to 1, eliminating the impact of dimensional differences on model training. The maximum and minimum values ​​used for normalization are also dynamically updated based on a sliding window with a window size of 1000 wafers.

[0027] Reference Appendix Figure 2In step S2, a causal graph between variables is learned from the historical data based on a causal discovery algorithm, and key process parameters with direct causal effects on dicing yield are identified, forming a core causal feature set. The causal discovery algorithm is either the PC algorithm or the FCI algorithm. The PC algorithm constructs a causal framework through conditional independence testing and then determines the causal direction through orientation rules, with a significance level set to 0.05 and a maximum condition set size of 5. The FCI algorithm further processes potential confounding factors based on the PC algorithm, and can identify some directed acyclic graphs. It is suitable for industrial scenarios with unobserved confounding variables, such as unmonitored environmental temperature and humidity or wafer batch differences. Both algorithms can effectively handle complex dependencies between high-dimensional process parameters, and the accuracy of causal graph construction can reach over 85%. Specifically, identifying key process parameters with direct causal effects on dicing yield in step S2 includes locating the direct causal parent node of the dicing yield node in the causal graph and calculating its average treatment effect. The average treatment effect calculation adopts a potential outcome framework based on counterfactual reasoning, and uses the do operator for intervention calculation, selecting parameters with an average treatment effect exceeding a preset threshold of 0.1 as core causal features. This threshold can be adjusted from 0.05 to 0.2 according to actual production requirements to ensure that the selected features have a significant causal influence. For example, if the causal graph shows that "spindle speed" is the direct parent node of "cutting yield" and its average processing effect is 0.15, then "spindle speed" is included in the core causal feature set; while "thickness of thin film deposition" is highly correlated with "cutting yield", but is only indirectly related in the causal graph, so it is excluded.

[0028] Step S3 involves constructing a neural network prediction model using the core causal feature set as input, and introducing causal prior constraints based on the causal graph during model training. The neural network prediction model employs a deep feedforward neural network structure, comprising an input layer, three hidden layers, and an output layer. The number of neurons in the input layer equals the dimension of the core causal feature set. The number of neurons in the hidden layers are 128, 64, and 32, respectively, with the ReLU activation function. The output layer consists of a single neuron, using the Sigmoid activation function to output a predicted cut yield value between 0 and 1. The network optimizer uses the Adam optimizer, with an initial learning rate set to 0.001 and a batch size set to 32.

[0029] Introducing causal prior constraints using a structural constraint method involves retaining only the connection weights corresponding to core causal features between the input layer and the first hidden layer of the neural network, while masking the connection weights for non-core features. Specifically, a binary mask matrix is ​​designed to force the weights corresponding to non-core features to zero, preventing them from participating in gradient updates during training. This method ensures that the model learns only based on features with genuine causal relationships, effectively avoiding interference from spurious correlations. Alternatively, causal prior constraints can be introduced by adding a causal regularization term to the loss function during model training. The causal regularization term is based on the intervention distribution defined by the causal graph, calculating the KL divergence between the model's prediction and the ideal intervention result, with a regularization coefficient set to 0.01. The loss function consists of a weighted average squared error loss and a causal regularization term, with a weight ratio of 10:1, enhancing causal consistency while maintaining prediction accuracy.

[0030] Reference Appendix Figure 3 In step S4, during the training phase of the neural network prediction model, adversarial examples are generated for the core causal feature set, and the original samples and adversarial examples are mixed for adversarial robust training to obtain a robust prediction model. The projection gradient descent method is used to generate adversarial examples for the core causal feature set. The perturbation generation step size is set to 0.01, the maximum number of iterations is 10, and the perturbation amplitude is limited to the feature value within a range of ±0.1. The generated perturbation is only applied to the data dimensions corresponding to the core causal feature set; non-core feature dimensions retain their original values, ensuring that adversarial training specifically improves the model's robustness to perturbations of key process parameters. In adversarial robust training, the original samples and adversarial examples are mixed in a 1:1 ratio to form an enhanced training set. During training, a dynamic sample weight adjustment strategy is used, assigning higher weights to adversarial examples that are difficult to classify correctly, with weight adjustment coefficients ranging from 0.5 to 2.0. Through this training method, the model maintains its ability to predict normal samples while significantly improving its adaptability to fluctuations in process parameters, increasing prediction stability by more than 30%.

[0031] Step S5 involves inputting the core causal features from the real-time process data of the wafer to be predicted into the robust prediction model, outputting the predicted dicing yield of the wafer. Following step S5, the model further utilizes Monte Carlo Dropout technology to calculate an uncertainty metric for the predicted dicing yield. Dropout is kept active during the prediction phase, and 50 forward propagations are performed to calculate the mean and variance of the prediction results. The uncertainty metric is defined as the prediction variance; when the variance exceeds a threshold of 0.05, the prediction result is considered unreliable and requires manual review or remeasurement. Furthermore, following step S5, the model utilizes deep ensemble technology to calculate the uncertainty metric for the predicted dicing yield. Five neural network models with identical structures but different initializations are trained using deep ensemble technology, each trained on an independent subset of data. During the prediction phase, the outputs of all models are combined to calculate the prediction mean and variance. The accuracy of the ensemble prediction is 5% to 8% higher than that of a single model.

[0032] An early warning signal is triggered when the predicted cutting yield and / or uncertainty measure exceed a preset threshold. The predicted cutting yield threshold is set according to product specifications, typically 95%; the uncertainty measure threshold is set to 0.05. The early warning signal is simultaneously issued through three methods: an audible and visual alarm, SMS notification, and a pop-up window on the industrial control system interface, ensuring that production personnel are promptly informed of any abnormalities.

[0033] The method also includes establishing a dynamic relationship model between process parameters and dicing yield. Based on real-time process data acquired online, a sliding window method is used to extract time-series features, with the window size set to the data of the most recent 100 wafers and a step size of one wafer. A recurrent neural network is used to model the dynamic characteristics of process parameters changing over time, enabling early identification and warning of dicing process trends. The method also includes constructing an adaptive model update mechanism. When the prediction error of 10 consecutive wafers exceeds 5%, the model retraining process is automatically triggered. Retraining uses historical data from the most recent 1000 wafers, updating the neural network weights while maintaining the original causal structure, ensuring the model can adapt to the slow drift of the production process.

[0034] The method is applied to a wafer manufacturing execution system (WMS) and interacts with production equipment via the standard OPCUA protocol. Prediction results are displayed in real-time on a large monitoring screen on the production line, updated within 30 seconds of each wafer being diced, providing timely decision support for process engineers. The method supports yield prediction for multiple wafer types, distinguishing different wafer types through product type coding. The model automatically loads the corresponding cause-effect graph and neural network parameters based on the product type, enabling simultaneous prediction of multiple products on a single production line, with a model switching response time of less than 2 seconds.

[0035] To illustrate the effectiveness of this invention, an application example is provided: the method described in this invention is deployed on a 12-inch wafer dicing production line. This production line simultaneously produces logic chips and power devices. The system first collects historical data from the past six months, containing complete process records for 15,000 wafers. Step S1 involves preprocessing the data to form a standardized historical dataset. Step S2 uses the FCI algorithm to construct a causal graph, identifying the core features with a direct causal effect on the yield of logic chip dicing as: spindle speed, feed rate, dicing blade thickness, and peak frequency of the acoustic emission signal; for power devices, the core features are: spindle speed, thin film deposition thickness, and RMS value of the vibration signal. Step S3 constructs causal constraint neural networks for both types of wafers and introduces causal regularization terms for training. Step S4 uses projective gradient descent to generate adversarial examples for adversarial robust training. The resulting robust prediction model has an average absolute error of 2.1% on the test set, significantly lower than the 4.8% of traditional correlation models. During the subsequent three months of online operation, the system successfully issued warnings three times for declining cutting yield due to tool wear, with an average lead time of five wafers, thus gaining valuable time for process adjustments. Meanwhile, the model's prediction variance consistently remained below 0.03, demonstrating excellent stability.

[0036] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A method for predicting wafer dicing yield based on multi-source process parameters, characterized in that: Includes the following steps: Step S1: Collect multi-source process data related to wafer dicing yield, and perform preprocessing and alignment to form a historical dataset; Step S2: Based on the causal discovery algorithm, learn the causal graph between variables from the historical dataset and identify key process parameters that have a direct causal effect on the cutting yield, thus forming a core causal feature set; Step S3: Using the core causal feature set as input, construct a neural network prediction model, and introduce causal prior constraints based on the causal graph during the model training process; Step S4: During the training phase of the neural network prediction model, adversarial examples are generated for the core causal feature set, and the original samples and adversarial examples are mixed for adversarial robust training to obtain a robust prediction model. Step S5: Input the core causal features from the real-time process data of the wafer to be predicted into the robust prediction model, and output the predicted dicing yield of the wafer.

2. The wafer dicing yield prediction method based on multi-source process parameters according to claim 1, characterized in that: The multi-source process data includes structured parameters from preceding manufacturing processes, setting parameters from cutting equipment, and temporal characteristics of vibration and acoustic emission signals from online sensors.

3. The wafer dicing yield prediction method based on multi-source process parameters according to claim 1, characterized in that: In scenarios where unobserved confounding variables are not considered, the causal discovery algorithm employs the PC algorithm; the significance level of the PC algorithm is 0.05, and the maximum condition set size is 5; in scenarios where unobserved confounding variables are considered, the causal discovery algorithm employs the FCI algorithm; key process parameters are identified, including the direct causal parent nodes of the cutting yield node, and their average treatment effect is calculated, and parameters with an average treatment effect exceeding 0.1 are selected as core causal features.

4. The wafer dicing yield prediction method based on multi-source process parameters according to claim 1, characterized in that: The neural network prediction model is a deep feedforward neural network, which includes an input layer, three hidden layers, and an output layer. The number of neurons in the input layer is equal to the dimension of the core causal feature set. The number of neurons in the hidden layers are 128, 64, and 32, respectively, and the activation function is ReLU. The output layer uses the Sigmoid activation function to output a predicted cutting yield value between 0 and 1. The optimizer is Adam, with an initial learning rate of 0.001 and a batch size of 32.

5. The wafer dicing yield prediction method based on multi-source process parameters according to claim 1, characterized in that: The causal prior constraint is implemented through the structural constraint method. A binary mask array is set between the input layer and the first hidden layer to force the connection weights corresponding to non-core causal features to zero, and these weights do not participate in gradient updates during training.

6. The wafer dicing yield prediction method based on multi-source process parameters according to claim 1, characterized in that: The causal prior constraint is achieved by adding a causal regularization term to the loss function; the causal regularization term is based on the intervention distribution defined by the causal graph, calculates the KL divergence between the model prediction and the ideal intervention result, and has a regularization coefficient of 0.01; the loss function consists of the mean squared error loss and the causal regularization term with a weight ratio of 10:

1.

7. The wafer dicing yield prediction method based on multi-source process parameters according to claim 1, characterized in that: The adversarial examples are generated using the projective gradient descent method, with a perturbation generation step size of 0.01, a maximum number of iterations of 10, and the perturbation amplitude is limited to the feature value within a range of ±0.

1. The perturbation is applied only to the data dimension corresponding to the core causal feature set. The adversarial robust training mixes the original samples and adversarial samples in a 1:1 ratio and adopts a dynamic sample weight adjustment strategy with a weight adjustment coefficient ranging from 0.5 to 2.

0.

8. The wafer dicing yield prediction method based on multi-source process parameters according to claim 1, characterized in that: It also includes step S6, which uses Monte Carlo Dropout technology to calculate the uncertainty measure of the predicted cutting yield; the Monte Carlo Dropout remains active during the prediction phase and performs 50 forward propagations, using the prediction variance as the uncertainty measure.

9. The wafer dicing yield prediction method based on multi-source process parameters according to claim 8, characterized in that: It also includes step S7, which uses deep integration technology to calculate the uncertainty measure of the predicted cutting yield; and trains five neural network models with the same structure but different initializations using the deep integration technology, and outputs the predicted mean and variance in combination.

10. The wafer dicing yield prediction method based on multi-source process parameters according to claim 9, characterized in that: It also includes step S8, which establishes a dynamic relationship model between process parameters and cutting yield; and uses a recurrent neural network to model the dynamic characteristics of process parameters changing over time, so as to achieve early identification and warning of cutting process trends.

Images