A wafer defect root cause analysis method, system, device and medium based on process trajectory

Abstract

This invention relates to the field of semiconductor manufacturing data analysis technology, specifically to a method, system, device, and medium for wafer defect root cause analysis based on process trajectory. First, a standard process trajectory is generated through preprocessing. Second, heterogeneous processes are converted into a unified vector through substring kernel embedding. Then, a constrained recurrent neural network is used to achieve variable-length trajectory vectorization and defect density prediction. Finally, an additivity attribution score is calculated based on partial trajectory comparison, and the root cause is located by the cumulative curve jump point. Attribution evaluation is completed from consistency, accuracy, significance, and fit, improving process optimization and yield. It is applicable to all scenarios of front-end / back-end manufacturing, process integration, and yield ramp-up.

Description

Technical Field

[0001] This invention relates to the field of semiconductor manufacturing data analysis technology, and more specifically, to a method, system, device, and medium for wafer defect root cause analysis based on process trajectory. Background Technology

[0002] In semiconductor manufacturing, wafer defects are a core issue affecting chip yield, production efficiency, and manufacturing costs. Rapid, accurate, and interpretable location of upstream processes that generate defects is a critical technological requirement in wafer manufacturing. Semiconductor process flows have significant unique characteristics: they include hundreds of heterogeneous physical / chemical processes such as photolithography, etching, ion implantation, thin film deposition, and chemical mechanical polishing. The processing paths are lengthy, dynamic, and non-fixed due to rework, equipment waiting times, batch switching, and formulation adjustments. Furthermore, the effective sample size is small, there are many similar processes, and data is sparse during process integration and yield ramp-up stages.

[0003] Existing methods for defect root cause analysis suffer from significant technical limitations: classic Design of Experiments (DOE) methods are difficult to implement in scenarios involving multiple parameters and coupled processes; traditional linear regression, random forests, and gradient boosting models rely on fixed-dimensional vector inputs, making them unsuitable for variable-length process trajectories; sequence models such as Recurrent Neural Networks (RNNs) and Transformers are black-box structures, making interpretable attribution difficult; interpretable AI methods based on Shapley values ​​ignore the process timing structure and wafer-batch hierarchy, are insensitive to process deviations, and have weak attribution signals; existing process embedding methods mostly employ one-hot encoding, failing to utilize the correlation between similar equipment and similar formulations, resulting in extremely low feature learning efficiency in small-sample scenarios. In summary, the industry urgently needs a complete solution for wafer defect root cause analysis that can adapt to variable-length heterogeneous process trajectories, balance prediction accuracy and causal attribution interpretability, support quantitative evaluation of attribution effects, and is suitable for small-sample learning. Summary of the Invention

[0004] This invention addresses the challenges in semiconductor manufacturing, such as longer wafer processing trajectories, heterogeneous processes, limited sample sizes, the inability of traditional root cause analysis to perform time-series modeling, uninterpretable attributions, and a lack of quantitative evaluation mechanisms. It proposes a wafer defect root cause analysis method, system, equipment, and medium based on process trajectories. Through process embedding, process trajectory regression, partial trajectory causal attribution, and multi-dimensional attribution evaluation, it achieves accurate prediction of defect density, automatic location of defect root causes, and reliable quantification of attribution results, providing a practical data-driven solution for process debugging, yield improvement, and process optimization.

[0005] The specific implementation details of this invention are as follows: A method for analyzing the root causes of wafer defects based on process traces, specifically including the following steps: Step S1: Construct the process trajectory based on the full-process processing data collected from the wafer manufacturing execution system; Step S2: Construct a vocabulary based on the process trajectory, call the substring kernel function to obtain the kernel matrix, and map it to obtain the proc2vec vector; Step S3: Construct a process trajectory sequence based on the proc2vec vector, build a defect density prediction model, and obtain the defect density prediction value; Step S4: Calculate the attribution score using the Rubin causal model, generate the cumulative attribution curve, and identify the jump point to obtain the root cause process; Step S5: Calculate the overall reliability score based on the established evaluation system and issue an alert for unreliable results.

[0006] To better realize the present invention, step S1 further includes the following steps: Step S11: Collect full-process processing data from the wafer manufacturing execution system and collect corresponding measured defect density data from process-constrained yield detection points; Step S12: Cleaning process data and measured defect density data; Step S13: Based on the full process data after cleaning, the measured defect density data, the total number of processes the wafer has passed through, and the set timestamps, unify the processing flow of each wafer into a process trajectory with time-series markers.

[0007] To better realize the present invention, step S2 further includes the following steps: Step S21: Extract string attributes from the process trajectory and concatenate them into process markers according to the set delimiters; Step S22: Construct a vocabulary based on process markers. Based on the size of the vocabulary, call the substring kernel function to calculate the similarity between any two process markers and obtain the kernel matrix. Step S23: Perform eigenvalue decomposition on the kernel matrix and map the process to a proc2vec vector.

[0008] To better realize the present invention, step S3 further includes the following steps: Step S31: Construct a process trajectory sequence from the proc2vec vectors according to the time series, and call the constrained recurrent neural network to construct the process trajectory vectorization module; Step S32: Convert the process trajectory into a trajectory vector of a set dimension, and use the trajectory vector as input to construct a defect density prediction model using a fully connected network with L1 regularization; Step S33: Construct an optimization objective function based on the set model learnable parameters, regularization strength, embedding dimension, and the actual defect density of the wafer; Step S34: Train the defect density prediction model based on the optimization objective function to obtain the defect density prediction value.

[0009] To better realize the present invention, step S4 further includes the following steps: Step S41: Call the Rubin causal model, compare the partial trajectory prediction results that include the k-th process with those that do not, and calculate the process attribution score; Step S42: Obtain the cumulative attribution curve based on the time-series cumulative process attribution scores; Step S43: Obtain the root cause process that causes high defect density based on the cumulative attribution curve.

[0010] To better realize the present invention, step S5 further includes the following steps: Step S51: Establish an attribution consistency evaluation system, calculate the variance of attribution scores for wafers of the same type and batch in the same process, and determine that the attribution is stable if the variance is less than the set variance threshold. Step S52: Establish a prediction accuracy evaluation system, calculate the determination coefficient between the predicted defect density and the measured value. If the determination coefficient is greater than the set determination coefficient threshold, the trajectory regression model prediction is deemed effective. Step S53: Establish a jump significance evaluation system, calculate the difference of the cumulative attribution curve point by point, and determine the points with a difference greater than 3 times the global standard deviation as significant root cause jump points; Step S54: Compare the root cause process with the results of manual annotation, and calculate the precision, recall and F1 score to measure the consistency between the attribution results and engineering experience. Step S55: Independently calculate the four indicators: attribution consistency, prediction accuracy, jump significance, and expert agreement. Combine them in an equally weighted manner to obtain a comprehensive credibility score, and output the reliability level judgment result according to the preset threshold.

[0011] Based on the above-mentioned wafer defect root cause analysis method based on process trajectory, in order to better realize the present invention, a wafer defect root cause analysis system based on process trajectory is further proposed to execute the above-mentioned wafer defect root cause analysis method based on process trajectory; including a data access and storage module, a process trajectory preprocessing module, a process embedding module, a process trajectory regression module, a process trajectory attribution module, an attribution evaluation module, and a visualization output module. The data access and storage module is used to connect to the MES system and the measurement data system, collect process parameters, time information, and defect density data in real time, and persistently store raw data, trajectory sequences, model parameters, attribution results, and evaluation reports, providing a unified data interface for subsequent modules. The process trajectory preprocessing module is used to perform deduplication, cleaning, missing value filling, timestamp alignment and outlier removal on the raw data, and convert multi-source heterogeneous data into a standard time-series process trajectory format to ensure that the subsequent modeling data is standardized and reliable. The process embedding module is used to generate process tags based on the high-level attributes of the process, calculate the similarity matrix using the substring kernel and complete the kernel embedding, and output the proc2vec vector of each process to achieve unified vectorization of heterogeneous processes. The process trajectory regression module is used to load the proc2vec sequence, generate route2vec trajectory vectors through a constrained recurrent neural network (CRNN), and train a defect density prediction model, supporting prediction of complete and partial trajectories. The process trajectory attribution module is used to calculate the trajectory prediction difference that includes / removes the current process for each process, obtain the attribution score, generate the cumulative attribution curve, and automatically identify the jump point and output the root cause process list. The attribution evaluation module is used to calculate attribution consistency, prediction accuracy, jump significance, and expert agreement based on preset rules, output a comprehensive credibility score, and issue an alarm for unreliable results. The visualization output module is used to display proc2vec clustering distribution, cumulative attribution curve, attribution score ranking, root cause process details and evaluation report in chart form, supporting engineers to quickly locate problems and perform process optimization.

[0012] Based on the above-mentioned wafer defect root cause analysis method based on process trajectory, in order to better realize the present invention, an electronic device is further proposed, including a memory and a processor; the memory stores a computer program; when the computer program is executed on the processor, the above-mentioned wafer defect root cause analysis method based on process trajectory is implemented.

[0013] Based on the above-mentioned wafer defect root cause analysis method based on process trajectory, in order to better realize the present invention, a computer-readable storage medium is further proposed, wherein computer instructions are stored on the computer-readable storage medium; when the computer instructions are executed on the above-mentioned electronic device, the above-mentioned wafer defect root cause analysis method based on process trajectory is realized.

[0014] The present invention has the following beneficial effects: (1) This invention uses kernel embedding to uniformly represent heterogeneous processes as proc2vec vectors, making full use of the similarity between equipment and recipe, and still maintaining high feature expression efficiency in small sample scenarios.

[0015] (2) The present invention uses a constrained recurrent neural network to construct route2vec, which can directly process variable-length, time-dependent process trajectories, breaking through the limitations of traditional fixed-dimensional models.

[0016] (3) The causal attribution based on partial trajectory comparison in this invention has strict additivity, the attribution process is transparent and interpretable, and there is no black box reasoning.

[0017] (4) The present invention has a built-in multi-dimensional attribution evaluation mechanism to realize automatic verification and reliable quantification of root cause results; verified on real front-end process data, the defect density prediction determination coefficient can reach 0.87, which can accurately identify the root cause of defects caused by abnormal waiting time, and can be directly used for semiconductor process debugging and yield improvement, with strong engineering applicability. Attached Figure Description

[0018] Figure 1 This is a schematic diagram illustrating the entire process of defect root cause analysis completed collaboratively by various modules provided by the present invention. Detailed Implementation

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments, and therefore should not be regarded as a limitation on the scope of protection. 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.

[0020] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "set up," "connected," and "linked" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0021] Example 1: This embodiment specifically includes the following steps: Step S1: Construct the process trajectory based on the full-process processing data collected from the wafer manufacturing execution system; Step S1 specifically includes the following steps: Step S11: Collect full-process processing data from the wafer manufacturing execution system and collect corresponding measured defect density data from process-constrained yield detection points; Step S12: Cleaning process data and measured defect density data; Step S13: Based on the full process data after cleaning, the measured defect density data, the total number of processes the wafer has passed through, and the set timestamps, unify the processing flow of each wafer into a process trajectory with time-series markers.

[0022] Step S2: Construct a vocabulary based on the process trajectory, call the substring kernel function to obtain the kernel matrix, and map it to obtain the proc2vec vector; Step S2 specifically includes the following steps: Step S21: Extract string attributes from the process trajectory and concatenate them into process markers according to the set delimiters; Step S22: Construct a vocabulary based on process markers. Based on the size of the vocabulary, call the substring kernel function to calculate the similarity between any two process markers and obtain the kernel matrix. Step S23: Perform eigenvalue decomposition on the kernel matrix and map the process to a proc2vec vector.

[0023] Step S3: Construct a process trajectory sequence based on the proc2vec vector, build a defect density prediction model, and obtain the defect density prediction value; Step S3 specifically includes the following steps: Step S31: Construct a process trajectory sequence from the proc2vec vectors according to the time series, and call the constrained recurrent neural network to construct the process trajectory vectorization module; Step S32: Convert the process trajectory into a trajectory vector of a set dimension, and use the trajectory vector as input to construct a defect density prediction model using a fully connected network with L1 regularization; Step S33: Construct an optimization objective function based on the set model learnable parameters, regularization strength, embedding dimension, and the actual defect density of the wafer; Step S34: Train the defect density prediction model based on the optimization objective function to obtain the defect density prediction value.

[0024] Step S4: Calculate the attribution score using the Rubin causal model, generate the cumulative attribution curve, and identify the jump point to obtain the root cause process; Step S4 specifically includes the following steps: Step S41: Call the Rubin causal model, compare the partial trajectory prediction results that include the k-th process with those that do not, and calculate the process attribution score; Step S42: Obtain the cumulative attribution curve based on the time-series cumulative process attribution scores; Step S43: Obtain the root cause process that causes high defect density based on the cumulative attribution curve.

[0025] Step S5: Calculate the overall reliability score based on the established evaluation system and issue an alert for unreliable results.

[0026] Step S5 specifically includes the following steps: Step S51: Establish an attribution consistency evaluation system, calculate the variance of attribution scores for wafers of the same type and batch in the same process, and determine that the attribution is stable if the variance Var is less than the set variance threshold; scoring rules: Var < 0.05: consistency score 25; 0.05 ≤ Var < 0.1: consistency score 15; Var ≥ 0.1: consistency score 0; Step S52: Establish a prediction accuracy evaluation system, calculate the determination coefficient between the predicted defect density and the measured value. If the determination coefficient R² is greater than the set determination coefficient threshold, the trajectory regression model is deemed to be effective. Based on the R² value, determine different levels of scoring rules: R² < 0.7: accuracy score 0; 0.7 ≤ R² < 0.85: accuracy score 15; R² ≥ 0.85: accuracy score 25. Step S53: Establish a jump significance evaluation system, calculate the difference point by point on the cumulative attribution curve, and determine the points where the difference γ is greater than 3 times the global standard deviation as significant root cause jump points; scoring rules: γ≥3: significance score 25; 2≤γ<3: significance score 15; γ<2: significance score 0; Step S54: Compare the root cause process with the manually labeled results, and calculate the precision, recall, and F1 score to measure the consistency between the attribution results and engineering experience. Scoring rules: F1 ≥ 0.85: consistency score 25; 0.7 ≤ F1 < 0.85: consistency score 15; F1 < 0.7: consistency score 0; Step S55: Independently calculate the four indicators: attribution consistency, prediction accuracy, jump significance, and expert agreement. Combine them using an equal-weighted approach to obtain a comprehensive reliability score, with equal weight for each of the four dimensions. Calculate the total score (Score). Output the reliability level judgment result according to the preset threshold. Judgment rules: Level A is reliable (Score ≥ 80), meaning the attribution result is highly reliable and can be directly used for process rectification; Level B is basically reliable (60 ≤ Score < 80), meaning the attribution result is for reference, and it is recommended to confirm it in conjunction with on-site verification; Level C is unreliable (Score < 60), meaning the attribution result is invalid and requires remodeling or data supplementation.

[0027] Working Principle: This embodiment achieves accurate prediction, root cause localization, and effect evaluation of wafer defects under variable-length heterogeneous process trajectories through process trajectory standardization, core embedding process vectorization, process trajectory regression prediction, causal attribution calculation, and multi-dimensional attribution evaluation. The method includes: preprocessing to generate standard process trajectories; converting heterogeneous processes into unified vectors through substring core embedding; using a constrained recurrent neural network to achieve variable-length trajectory vectorization and defect density prediction; calculating additivity attribution scores based on partial trajectory comparisons, and locating root causes by cumulative curve jump points; and completing attribution evaluation based on consistency, accuracy, significance, and fit. The system comprises seven modules: data, preprocessing, embedding, regression, attribution, evaluation, and visualization. This invention is adapted to variable-length semiconductor manufacturing trajectories, is efficient with small samples, provides interpretable attribution, and standardizes evaluation. In real wafer fab data, it achieves a prediction accuracy of 0.87, directly supporting process optimization and yield improvement, and is applicable to all scenarios of front-end / back-end manufacturing, process integration, and yield ramp-up.

[0028] Example 2: This embodiment is based on the above embodiment 1, such as... Figure 1 As shown, a specific embodiment will be used for illustration.

[0029] This embodiment consists of five steps: process trajectory data modeling, process vectorization, process trajectory regression modeling, process trajectory attribution calculation, and quantitative evaluation of attribution effect. Each step is executed sequentially to form a closed-loop analysis process.

[0030] Step S1: Modeling process trajectory data.

[0031] First, full-process data is collected from the wafer fab's Manufacturing Execution System (MES), including equipment ID, recipe ID, equipment type, lithography layer ID, path ID, and timestamp information for each process. Simultaneously, measured defect density data corresponding to process-limited yield (PLY) detection points are collected. The raw data is cleaned, removing samples with missing key attributes, abnormal timestamps, or defect densities significantly deviating from the reasonable range. Mildly missing data is then imputed using linear interpolation. The processing flow of each wafer is uniformly represented as a time-stamped process trajectory, formally described as:

[0032] Where L is the total number of processes that the wafer goes through, i.e. the process trace length; Let k be the set of high-level attributes for the k-th process. This is the timestamp corresponding to the k-th process step. This step transforms the unstructured, variable-length, and heterogeneous processing flow into a computable sequence of standard process trajectories.

[0033] Step S2: Process vectorization based on kernel embedding; For each process step, string attributes such as equipment ID, formula ID, equipment type, photolithography layer ID, and path ID are extracted from the pre-processed process trajectory and concatenated into a unique process identifier using a fixed delimiter.

[0034] Where ⊕ represents string concatenation. A vocabulary is constructed using all unique process markers, and the size of the vocabulary is denoted as Vd. A custom substring kernel function is used to calculate the similarity between any two process markers, resulting in a Vd×Vd dimensional kernel matrix K. Vd represents the vocabulary size. Eigenvalue decomposition is performed on the kernel matrix K, and the top D largest eigenvalues ​​are selected. Each process and its corresponding feature vector are mapped to a fixed D-dimensional vector, i.e., the proc2vec vector:

[0035] in Let K be the k-th largest eigenvalue of the kernel matrix K. Let be the i-th component of the corresponding feature vector, and D be the embedding dimension, preferably 64, 128, or 256. This step can effectively learn the correlation between similar equipment and similar formulas under small sample conditions, thereby improving the ability to express process features.

[0036] Step S3: Regression prediction based on process trajectory; The proc2vec vectors are arranged into a process trajectory sequence according to time sequence. A constrained recurrent neural network (C-RNN) is used to construct the process trajectory vectorization module route2vec, which updates part of the trajectory vectors time sequence by time sequence. The calculation formula is as follows: in, This is a partial trajectory vector (proc2vec vector) of the first k processes. This represents a partial trajectory vector for the first k-1 processes. For time mapping functions, the present invention preferably uses... After converting the arbitrary length process trajectory into a fixed-dimensional trajectory vector z, a defect density prediction model is constructed using a fully connected network with L1 regularization, with the trajectory vector as input: f(z) = MLP(z). The model optimizes by minimizing the regularized squared loss. Where N is the number of wafer samples, y(n) is the true defect density of the nth wafer, θ is the model learnable parameter, ν is the regularization strength (hyperparameter, preferably 0.001–0.01), and D is the embedding dimension (preferably 64 / 128 / 256).

[0037] After training, the model can output defect density predictions for the complete trajectory and any part of the trajectory.

[0038] Step S4: Causal attribution calculation based on process trajectory; Drawing upon the Rubin Causal Model (RCM), this study calculates the process attribution score by comparing partial trajectory predictions that include and exclude the k-th process. This represents the contribution of the k-th process to the defect, where... The attribution score for the k-th process is given. A positive value indicates that the process increases defect density, while a negative value indicates that it suppresses defect density. In this embodiment, the attribution score satisfies the additivity theorem. ,in This is the initial intercept term. The cumulative attribution curve is obtained by accumulating the attribution scores of each process in chronological order. The process corresponding to the position where a significant jump occurs on the curve is the root cause process causing high defect density.

[0039] Step S5: Quantitative evaluation method for attribution effect; To ensure the reliability, reproducibility, and applicability of root cause analysis results, this invention establishes a multi-dimensional attribution evaluation system: 1) Attribution consistency evaluation: Calculate the variance of attribution scores for wafers of the same type and batch at the same process. Variance < threshold indicates stable attribution; the smaller the variance, the more stable the attribution. 2) Prediction accuracy evaluation: Calculate the coefficient of determination R² between predicted defect density and measured values. R² > 0.85 indicates effective prediction; the higher the R², the more reliable the trajectory regression model. 3) Jump significance evaluation: Calculate the difference point by point on the cumulative attribution curve. Points with a difference greater than 3 times the global standard deviation are identified as significant root cause jump points. 4) Expert consistency evaluation: Compare the root causes output by the algorithm with the results manually labeled by domain experts, calculate precision, recall, and F1 score to measure the consistency between the attribution results and engineering experience. A comprehensive credibility score of Score=100 is used. A comprehensive score of 100 ≥ 80 is considered Grade A reliable, and the attribution results can be directly used for process rectification.

[0040] The other parts of this embodiment are the same as those in Embodiment 1 above, so they will not be described again.

[0041] Example 3: Based on any one of Examples 1-2 above, this embodiment takes the complete execution on a certain wafer fab front-end (FEOL) process dataset as an example, including 787 wafers, hundreds of processes and corresponding defect density annotations.

[0042] Step S1: Process trajectory data acquisition and standardization processing; The entire process data of 787 wafers was collected from the wafer fab MES system's process trajectory data acquisition and preprocessing, including equipment ID, recipe ID, equipment type, lithography layer ID, path ID, and timestamp. PLY inspection point defect density measurement data was also collected simultaneously. The raw data was cleaned: wafer samples with more than 10% missing process information were removed, data with abnormal timestamp jumps were corrected, and a small number of missing attributes were filled in by interpolation using the average value of the same batch. The processing flow of each wafer was organized into a standard time-series process trajectory ξ, with trajectory length L distributed between 120 and 350, which is a typical variable-length process trajectory.

[0043] Step S2: Process embedding proc2vec generation and verification; High-level attributes of each process step are extracted from each process trajectory and concatenated into process markers according to fixed rules, resulting in 426 unique process markers. A vocabulary of size Vd=426 is constructed. A 426×426 kernel matrix K is calculated using a custom substring kernel, and the embedding dimension D=128 is set. Eigenvalue decomposition is performed on the kernel matrix to generate a 128-dimensional proc2vec vector. The proc2vec vector is visualized using tSNE dimensionality reduction. The results show that the processes are clearly clustered according to equipment type, with similar equipment being closer together. The correlation coefficient with defect density reaches 0.61, which is significantly better than one-hot encoding (0.52) and constant vector (0.27), proving that the embedding effect is effective.

[0044] Step S3: Training the route2vec process trajectory regression model; The proc2vec vector is input sequentially into a constrained recurrent neural network (C-RNN), with the time weighting function using... The model updates part of the trajectory vector step by step, converting the variable-length process trajectory into a 128-dimensional fixed route2vec vector; it constructs a fully connected network without hidden layers as the prediction model, sets the L1 regularization coefficient ν=0.005, and trains the model with the goal of minimizing the regularization squared loss; after the model is trained, it predicts the defect density of the test set wafers, and the coefficient of determination between the predicted value and the measured value is R²=0.87, which meets the requirements of industrial-grade high-precision prediction.

[0045] Step S4: Process trajectory attribution calculation and root cause localization; A high-defect-density wafer was selected from the test set, and the attribution score was calculated for each process step. The cumulative attribution curve is obtained by accumulating the values ​​over time. The difference is calculated point by point on the cumulative attribution curve, and two obvious jump points A and B are identified, with the difference amplitude exceeding three times the global standard deviation. Further querying the corresponding process information confirms that the jump points are caused by abnormal long waiting times in two processes, which are determined to be the root cause of the defect.

[0046] Step S5: Quantitative evaluation of attribution effect; The root cause analysis results were evaluated using multiple dimensions: Regarding attribution consistency, the variance of attribution scores for the same root cause process across the 30 wafers in the same batch was less than 0.02, indicating stable attribution with a score less than 0.05, resulting in a consistency score of 25; regarding prediction accuracy, the model… The accuracy score is 25, indicating reliable prediction. Regarding the significance of jumps, both points A and B meet the criteria for significant jumps, resulting in a jump significance score of 25. In terms of expert agreement, the precision rate is 92%, recall rate is 89%, and F1 score is 90.5%, showing a high degree of consistency with expert judgment. The F1 score is greater than 0.85, resulting in a score of 25. The overall evaluation result is reliable and can be directly used for process optimization.

[0047] Step S6: Root cause analysis and implementation of process improvements; The system outputs the root cause process number, corresponding equipment number, anomaly waiting time, and attribution reliability score. Based on the results, process engineers perform chamber inspections, adjust scheduling strategies, and optimize maintenance cycles for the relevant equipment. The system pushes the root cause workstation to the EAP equipment automation system via API, triggering equipment maintenance, recipe verification, and scheduling optimization. After implementation, the density of similar defects decreased by 37%, and the yield increased by 2.1 percentage points. This verifies the practical engineering value of the method of this invention.

[0048] The other parts of this embodiment are the same as any one of the above embodiments 1-2, so they will not be described again.

[0049] Example 4: Based on any one of Embodiments 1-3 above, this embodiment proposes a wafer defect root cause analysis system based on process trajectory, used to execute the above-mentioned wafer defect root cause analysis method based on process trajectory; including a data access and storage module, a process trajectory preprocessing module, a process embedding module, a process trajectory regression module, a process trajectory attribution module, an attribution evaluation module, and a visualization output module; The data access and storage module is used to connect to the MES system and the measurement data system, collect process parameters, time information, and defect density data in real time, and persistently store raw data, trajectory sequences, model parameters, attribution results, and evaluation reports, providing a unified data interface for subsequent modules. The process trajectory preprocessing module is used to perform deduplication, cleaning, missing value filling, timestamp alignment and outlier removal on the raw data, and convert multi-source heterogeneous data into a standard time-series process trajectory format to ensure that the subsequent modeling data is standardized and reliable. The process embedding module is used to generate process tags based on the high-level attributes of the process, calculate the similarity matrix using the substring kernel and complete the kernel embedding, and output the proc2vec vector of each process to achieve unified vectorization of heterogeneous processes. The process trajectory regression module is used to load the proc2vec sequence, generate route2vec trajectory vectors through a constrained recurrent neural network (CRNN), and train a defect density prediction model, supporting prediction of complete and partial trajectories. The process trajectory attribution module is used to calculate the trajectory prediction difference that includes / removes the current process for each process, obtain the attribution score, generate the cumulative attribution curve, and automatically identify the jump point and output the root cause process list. The attribution evaluation module is used to calculate attribution consistency, prediction accuracy, jump significance, and expert agreement based on preset rules, output a comprehensive credibility score, and issue an alarm for unreliable results. The visualization output module is used to display proc2vec clustering distribution, cumulative attribution curve, attribution score ranking, root cause process details and evaluation report in chart form, supporting engineers to quickly locate problems and perform process optimization.

[0050] This embodiment also proposes an electronic device, including a memory and a processor; the memory stores a computer program; when the computer program is executed on the processor, it implements the above-described wafer defect root cause analysis method based on process trajectory.

[0051] This embodiment also proposes a computer-readable storage medium storing computer instructions; when the computer instructions are executed on the aforementioned electronic device, the aforementioned wafer defect root cause analysis method based on process trajectory is implemented.

[0052] The other parts of this embodiment are the same as any one of the embodiments 1-3 above, so they will not be described again.

[0053] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications or equivalent changes made to the above embodiments based on the technical essence of the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for analyzing the root causes of wafer defects based on process traces, characterized in that, Specifically, the following steps are included: Step S1: Construct the process trajectory based on the full-process processing data collected from the wafer manufacturing execution system; Step S2: Construct a vocabulary based on the process trajectory, call the substring kernel function to obtain the kernel matrix, and map it to obtain the proc2vec vector; Step S3: Construct a process trajectory sequence based on the proc2vec vector, build a defect density prediction model, and obtain the defect density prediction value; Step S4: Calculate the attribution score using the Rubin causal model, generate the cumulative attribution curve, and identify the jump point to obtain the root cause process; Step S5: Calculate the overall reliability score based on the established evaluation system and issue an alert for unreliable results.

2. The wafer defect root cause analysis method based on process trajectory according to claim 1, characterized in that, Step S1 specifically includes the following steps: Step S11: Collect full-process processing data from the wafer manufacturing execution system and collect corresponding measured defect density data from process-constrained yield detection points; Step S12: Cleaning process data and measured defect density data; Step S13: Based on the full process data after cleaning, the measured defect density data, the total number of processes the wafer has passed through, and the set timestamps, unify the processing flow of each wafer into a process trajectory with time-series markers.

3. The wafer defect root cause analysis method based on process trajectory according to claim 1, characterized in that, Step S2 specifically includes the following steps: Step S21: Extract string attributes from the process trajectory and concatenate them into process markers according to the set delimiters; Step S22: Construct a vocabulary based on process markers. Based on the size of the vocabulary, call the substring kernel function to calculate the similarity between any two process markers and obtain the kernel matrix. Step S23: Perform eigenvalue decomposition on the kernel matrix and map the process to a proc2vec vector.

4. The wafer defect root cause analysis method based on process trajectory according to claim 1, characterized in that, Step S3 specifically includes the following steps: Step S31: Construct a process trajectory sequence from the proc2vec vectors according to the time series, and call the constrained recurrent neural network to construct the process trajectory vectorization module; Step S32: Convert the process trajectory into a trajectory vector of a set dimension, and use the trajectory vector as input to construct a defect density prediction model using a fully connected network with L1 regularization; Step S33: Construct an optimization objective function based on the set model learnable parameters, regularization strength, embedding dimension, and the actual defect density of the wafer; Step S34: Train the defect density prediction model based on the optimization objective function to obtain the defect density prediction value.

5. The wafer defect root cause analysis method based on process trajectory according to claim 1, characterized in that, Step S4 specifically includes the following steps: Step S41: Call the Rubin causal model, compare the partial trajectory prediction results that include the k-th process with those that do not, and calculate the process attribution score; Step S42: Obtain the cumulative attribution curve based on the time-series cumulative process attribution scores; Step S43: Obtain the root cause process that causes high defect density based on the cumulative attribution curve.

6. The wafer defect root cause analysis method based on process trajectory according to claim 1, characterized in that, Step S5 specifically includes the following steps: Step S51: Establish an attribution consistency evaluation system, calculate the variance of attribution scores for wafers of the same type and batch in the same process, and determine that the attribution is stable if the variance is less than the set variance threshold. Step S52: Establish a prediction accuracy evaluation system, calculate the determination coefficient between the predicted defect density and the measured value. If the determination coefficient is greater than the set determination coefficient threshold, the trajectory regression model prediction is deemed effective. Step S53: Establish a jump significance evaluation system, calculate the difference of the cumulative attribution curve point by point, and determine the points with a difference greater than 3 times the global standard deviation as significant root cause jump points; Step S54: Compare the root cause process with the results of manual annotation and calculate the expert agreement. Step S55: Based on attribution consistency, prediction accuracy, jump significance, and expert agreement, a comprehensive reliability score is obtained by fusion in an equally weighted manner, and the reliability level judgment result is output according to the set threshold.

7. A wafer defect root cause analysis system based on process trajectory, used to execute the wafer defect root cause analysis method based on process trajectory as described in claim 1; characterized in that, It includes a data access and storage module, a process trajectory preprocessing module, a process embedding module, a process trajectory regression module, a process trajectory attribution module, an attribution evaluation module, and a visualization output module; The data access and storage module is used to connect to the MES system and the measurement data system, collect process parameters, time information, and defect density data in real time, and persistently store raw data, trajectory sequences, model parameters, attribution results, and evaluation reports, providing a unified data interface for subsequent modules. The process trajectory preprocessing module is used to perform deduplication, cleaning, missing value filling, timestamp alignment and outlier removal on the raw data, and convert multi-source heterogeneous data into a standard time-series process trajectory format to ensure that the subsequent modeling data is standardized and reliable. The process embedding module is used to generate process tags based on the high-level attributes of the process, calculate the similarity matrix using the substring kernel and complete the kernel embedding, and output the proc2vec vector of each process to achieve unified vectorization of heterogeneous processes. The process trajectory regression module is used to load the proc2vec sequence, generate route2vec trajectory vectors through a constrained recurrent neural network (CRNN), and train a defect density prediction model, supporting prediction of complete and partial trajectories. The process trajectory attribution module is used to calculate the trajectory prediction difference that includes / removes the current process for each process, obtain the attribution score, generate the cumulative attribution curve, and automatically identify the jump point and output the root cause process list. The attribution evaluation module is used to calculate attribution consistency, prediction accuracy, jump significance, and expert agreement based on preset rules, output a comprehensive credibility score, and issue an alarm for unreliable results. The visualization output module is used to display proc2vec clustering distribution, cumulative attribution curve, attribution score ranking, root cause process details and evaluation report in chart form, supporting engineers to quickly locate problems and perform process optimization.

8. An electronic device, characterized in that, It includes a memory and a processor; the memory stores a computer program; when the computer program is executed on the processor, it implements the wafer defect root cause analysis method based on process trajectory as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions; when the computer instructions are executed on the electronic device as described in claim 8, they implement the wafer defect root cause analysis method based on process trajectory as described in any one of claims 1-6.

Images