A contamination source identification method for wafer manufacturing based on AMC distribution between layers within FOUP

Abstract

The present invention discloses a method for identifying contamination sources in the wafer manufacturing process based on the distribution of AMCs between layers within a FOUP (formulated form of equipment), which relates to the technical field of contamination source identification. The method includes: analyzing airflow disturbances and contaminant diffusion characteristics, formulating an external detector deployment strategy, and collecting contaminant concentrations and associated equipment operating parameters in real time; establishing a spatiotemporal matrix of contaminant concentrations, and using wavelet transforms to extract the gradient characteristics of contaminant concentration changes. Combined with the operating parameters of associated equipment, preliminary location data of the contamination source is obtained; constructing a spatiotemporal gridded input matrix and designing a dual-branch deep learning model to generate a contamination source location probability map and obtain the location data of the contamination source; and evaluating the contribution weight of each associated equipment to contaminant diffusion, formulating and implementing a targeted remediation strategy. The present invention can accurately locate the location of contamination sources, improving the accuracy and response speed of contamination source identification.

Description

Technical Field

[0001] The present invention relates to the technical field of pollution source identification, and in particular to a method for identifying pollution sources in a wafer process based on the AMC distribution between layers within a FOUP. Background Art

[0002] During semiconductor manufacturing, wafers are exposed to complex, highly clean process environments and are extremely sensitive to airborne contaminant concentrations. Airborne molecular contaminants (AMCs), due to their small molecular weight, strong diffusivity, and difficulty in detection, have become a significant factor affecting wafer surface quality, device performance, and yield.

[0003] To ensure a clean environment during wafer transport and storage, front-opening unified pods (FOUPs) are widely used as wafer carriers and protectors. However, in actual production, AMC can infiltrate the interior of the FOUP due to factors such as decreased sealing, material volatilization, and residual process gases. AMC can diffuse unevenly within the internal space, causing contamination variations between wafer layers and potentially leading to batch-specific process defects.

[0004] Currently, monitoring AMC contamination within FOUPs (former optical fiber unpacking) primarily relies on fixed-position detectors for fixed-point data collection and analysis. This approach lacks the ability to dynamically model the spatial and temporal distribution of contaminants and locate contamination sources. Furthermore, traditional monitoring methods fail to fully integrate the structural characteristics of the FOUP's interior with the dynamics of airflow, resulting in low contamination source identification accuracy and slow response speeds. These methods are unable to meet the real-time and accurate contamination control requirements of high-precision wafer manufacturing processes.

[0005] Currently, no effective solutions have been proposed for the problems in related technologies. Summary of the Invention

[0006] In response to the problems in the related art, the present invention proposes a method for identifying contamination sources in the wafer process based on the AMC distribution between layers within the FOUP to overcome the above-mentioned technical problems existing in the existing related art.

[0007] To this end, the specific technical solutions adopted in the present invention are as follows:

[0008] A method for identifying contamination sources in the wafer process based on the AMC distribution between layers within a FOUP includes:

[0009] S1. Analyze airflow disturbances and contaminant diffusion characteristics based on the internal space of the transport box and the wafer level placement, formulate external detector deployment strategies, and collect contaminant concentrations and related equipment operating parameters in real time.

[0010] S2. Establish a spatiotemporal matrix of pollutant concentrations and use wavelet transform to extract the gradient characteristics of pollutant concentration changes. Combined with the operating parameters of related equipment, obtain preliminary location data of pollution sources;

[0011] S3. Construct a spatiotemporal gridded input matrix and design a dual-branch deep learning model to generate a pollution source location probability map and obtain the location data of the pollution source;

[0012] S4. Based on the location data of pollution sources, evaluate the contribution weight of each associated equipment to the spread of pollutants, and formulate and implement targeted governance strategies.

[0013] Optionally, based on the internal space of the transport box and the wafer level placement, analyze the airflow disturbance and contaminant diffusion characteristics, formulate an external detector deployment strategy, and collect pollutant concentrations and related equipment operating parameters in real time, including:

[0014] S11. Based on the interior space of the transport box, establish a parameterized geometric model, set fluid boundary conditions, and run computational fluid dynamics simulations to obtain baseline airflow field and pollutant diffusion characteristic data to build a digital twin model.

[0015] S12. Design the initial external detector deployment strategy based on wafer layer placement, collect initial pollutant concentrations, airflow velocities, and associated equipment operating parameters at each layer, and establish a spatiotemporal pollution distribution baseline.

[0016] S13. Combining the established pollution distribution baseline with the constructed digital twin model, the adversarial transfer learning algorithm is used to optimize the digital twin model;

[0017] S14. Predict the spatiotemporal distribution characteristics of pollutants using the optimized digital twin model, and update the initial external detector deployment strategy in combination with the multi-armed bandit algorithm to obtain the final external detector deployment strategy;

[0018] S15. Execute the final external detector deployment strategy and collect pollutant concentrations and associated equipment operating parameters in real time.

[0019] Optionally, combining the established pollution distribution baseline with the constructed digital twin model, and optimizing the digital twin model using the adversarial transfer learning algorithm includes:

[0020] S131. Input the established pollution distribution baseline into the constructed digital twin model, simulate and obtain simulation data, and construct a data set for adversarial transfer learning by combining the initial pollutant concentration, airflow velocity, and equipment operating parameters;

[0021] S132. Based on the adversarial transfer learning algorithm, construct an adversarial transfer learning model, where the adversarial transfer learning model includes a convolutional encoder, a classifier, and a domain discriminator;

[0022] S133, using a convolutional encoder to extract spatiotemporal characteristics of pollutant concentrations in the data set to generate a feature vector;

[0023] S134. Use a domain discriminator to perform inter-domain difference analysis on the feature vectors, and combine a convolutional neural network to identify the distribution differences between the simulation data and the initial pollutant concentration, airflow velocity, and equipment operating parameters, and calculate the domain classification loss based on the identification results;

[0024] S135. Optimize the classifier and the domain discriminator based on the calculated domain classification loss and the adversarial training mechanism to obtain an optimized adversarial transfer learning model.

[0025] S136. Combine the optimized adversarial transfer learning model with the digital twin model to optimize the parameters of the digital twin model.

[0026] Optionally, the spatiotemporal distribution characteristics of pollutants are predicted using the optimized digital twin model, and the initial external detector deployment strategy is updated in combination with the multi-armed bandit algorithm to obtain the final external detector deployment strategy, including:

[0027] S141. Input the initial pollutant concentration, airflow velocity, and equipment operating parameters into the optimized digital twin model to predict the spatiotemporal distribution characteristics of the pollutant concentration in the interior space of the transport box;

[0028] S142. Based on the designed initial external detector deployment strategy, a nearest neighbor algorithm is used to calculate the spatiotemporal similarity between the external detector nodes, and a number of candidate external detector nodes that best match the current spatiotemporal distribution characteristics are identified;

[0029] S143. Calculate the average weight of each candidate external detector node based on the identified candidate external detector nodes and the preset pollutant prediction function, and evaluate the expected monitoring effect of each candidate external detector node based on the spatiotemporal distribution characteristics of the pollutant concentration;

[0030] S144. In combination with the multi-armed bandit algorithm, select the candidate external detector node with the best expected monitoring effect to update the initial external detector deployment strategy and obtain the final external detector deployment strategy.

[0031] Optionally, in combination with a multi-armed bandit algorithm, a candidate external detector node with the best expected monitoring effect is selected to update the initial external detector deployment strategy. The final external detector deployment strategy includes:

[0032] S1441. Map each candidate external detector node to a decision arm in a multi-armed bandit algorithm, and construct a reward function based on monitoring effectiveness;

[0033] S1442. Initialize the reward function of each decision arm based on the expected monitoring effect of each candidate external detector node and set the initial confidence interval;

[0034] S1443. Calculate the upper bound of each initial confidence interval based on the initial confidence interval of each decision arm, and select the candidate external detector node with the highest upper bound as the optimal external detector node through the exploration mechanism of the multi-armed bandit algorithm.

[0035] S1444. Generate external detector deployment recommendations for the next phase based on the optimal external detector node to update the external detector deployment strategy and obtain a final external detector deployment strategy.

[0036] Optionally, a spatiotemporal gridded input matrix is ​​constructed and a dual-branch deep learning model is designed to generate a pollution source location probability map. Obtaining the location data of the pollution source includes:

[0037] S31. Construct a spatiotemporal gridded input matrix based on the preliminary location data of pollution sources and the real-time collected pollutant concentrations;

[0038] S32. Design a dual-branch deep learning model based on convolutional neural networks and temporal attention mechanisms, and use the dual-branch deep learning model to extract the spatial and temporal characteristics of pollutants in the spatiotemporal gridded input matrix;

[0039] S33: Fuse the spatial and temporal features of pollutants, and inject the fusion results into the preset pollution source heat map through feature splicing and full connection layer compression to generate a pollution source location probability map;

[0040] S34. Based on the generated pollution source location probability map and Markov decision process, and combined with offline simulation training and greedy strategy, the location data of the pollution source is obtained.

[0041] Optionally, the spatial and temporal features of pollutants are fused, and the fusion results are injected into a preset pollution source heat map through feature splicing and full connection layer compression to generate a pollution source location probability map including:

[0042] S331. Align the extracted spatial and temporal features of the pollutants according to the newly deployed external detector nodes, and perform feature concatenation to generate a joint feature matrix.

[0043] S332. Input the generated joint feature matrix into the fully connected layer, compress it using nonlinear dimensionality reduction technology, and extract feature components using low-rank matrix decomposition technology to obtain a compressed feature representation;

[0044] S333: Perform weighted fusion of the compressed feature representation and the preset pollution source heat map, smooth the fusion result using Gaussian filtering, and output a pollution source location probability map.

[0045] Optionally, based on the generated pollution source location probability map and Markov decision process, and in combination with offline simulation training and greedy strategy, obtaining the location data of the pollution source includes:

[0046] S341. Based on the Markov decision process, the pollution source location probability map is used as the environmental state, the pollution source verification operation is used as the agent action, and the spatial and temporal characteristics of the pollutants are combined to establish the state transition rules;

[0047] S342. Construct a multi-objective reward function according to the established state transition rule;

[0048] S343. Generate a virtual pollution source scenario dataset using the optimized digital twin model, build a convolutional policy network based on the dataset, and perform offline simulation training using a proximal policy optimization algorithm in combination with a multi-objective reward function.

[0049] S344. Based on the action value output by the trained convolutional strategy network, the greedy strategy is used to select the optimal pollution source verification operation, and the pollution source location probability map is updated in combination with the feedback data from the external detector.

[0050] S345. Based on the updated pollution source location probability map, a verification path is generated, and the verification path is optimized in combination with the entropy value threshold to obtain the location data of the pollution source.

[0051] Optionally, the multi-objective reward function includes: operation cost penalty, successful positioning reward, information gain reward, probability density reward and distance penalty.

[0052] Optionally, based on the location data of the pollution source, the contribution weight of each associated device to the spread of pollutants is evaluated, and a targeted governance strategy is formulated and implemented, including:

[0053] S41. Construct a cause-effect diagram model by combining the location data of pollution sources, the real-time collected pollutant concentrations, and the operating parameters of related equipment;

[0054] S42. Based on the constructed causal graph model, use causal inference technology to evaluate the impact of each related equipment on the pollutant diffusion process, obtain the contribution weight, and formulate and implement targeted governance strategies based on the contribution weight.

[0055] The beneficial effects of the present invention are:

[0056] 1. The present invention realizes real-time monitoring and collection of environmental data by combining the flow of air in the transport box, the diffusion characteristics of pollutants, and the operating parameters of related equipment. By extracting and analyzing the concentration gradient characteristics of pollutants in these data, the location of the pollution source can be accurately located, thereby improving the accuracy and response speed of pollution source identification, and preventing the decline of product or environmental quality due to the failure to detect pollution in time.

[0057] 2. The present invention evaluates the contribution of each associated device to pollutant diffusion based on the pollution source location results, can accurately identify the pollution source, and formulate targeted governance strategies accordingly; combining the pollution source location information with the equipment operating status, the equipment operating parameters can be adjusted in real time, thereby reducing the risk of pollutant diffusion. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0059] Figure 1 This is a flow chart of a method for identifying contamination sources in a wafer process based on the AMC distribution between layers within a FOUP according to an embodiment of the present invention. DETAILED DESCRIPTION

[0060] To further illustrate each embodiment, the present invention provides drawings, which are part of the disclosure of the present invention. They are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. By referring to these contents, ordinary technicians in this field should be able to understand other possible implementation methods and the advantages of the present invention.

[0061] According to an embodiment of the present invention, a method for identifying contamination sources in a wafer process based on the AMC distribution between layers within a FOUP is provided.

[0062] The present invention will now be further described with reference to the accompanying drawings and specific embodiments. Figure 1 As shown, according to an embodiment of the present invention, a method for identifying contamination sources in a wafer process based on the AMC distribution between layers in a FOUP includes:

[0063] S1. Analyze airflow disturbances and contaminant diffusion characteristics based on the internal space of the transport box and the wafer level placement, formulate external detector deployment strategies, and collect contaminant concentrations and related equipment operating parameters in real time.

[0064] In this optional embodiment, based on the internal space of the transport box and the wafer level placement, the airflow disturbance and contaminant diffusion characteristics are analyzed, an external detector deployment strategy is formulated, and real-time collection of contaminant concentrations and associated equipment operating parameters includes:

[0065] S11. Based on the internal space of the transport box, a parameterized geometric model is established, fluid boundary conditions are set, and computational fluid dynamics simulation is run to obtain baseline airflow field and pollutant diffusion characteristic data to build a digital twin model.

[0066] It should be noted that based on the internal space of the transport box, a parameterized geometric model was established, fluid boundary conditions were set, and computational fluid dynamics simulations were run to obtain baseline airflow field and pollutant diffusion characteristic data to build a digital twin model. The following steps were taken:

[0067] First, a parametric geometric model is established based on the internal space of the transport box to clarify the spatial dimensions and connection relationships; secondly, key parameters such as the fluid inlet / outlet boundary conditions, initial pollutant concentration distribution, and material surface characteristics are set; then, computational fluid dynamics (CFD) simulation calculations are run to obtain the spatiotemporal distribution data of the velocity field, pressure field, and concentration field by solving the Navier-Stokes equations and pollutant transport equations; the simulation results are then gridded and feature extracted to establish a database containing key parameters such as airflow velocity and pollutant diffusion coefficient; finally, the simulation data is aligned with the real-time external detector monitoring data, and a digital twin model that can be updated in real time is constructed through dynamic calibration to provide a benchmark reference for subsequent pollution source identification.

[0068] S12. Design the initial external detector deployment strategy based on wafer layer placement, collect the initial pollutant concentration, airflow velocity, and related equipment operating parameters at each layer, and establish a time-space-related pollution distribution baseline.

[0069] S13. Combine the established pollution distribution baseline with the constructed digital twin model and use the adversarial transfer learning algorithm to optimize the digital twin model.

[0070] In this optional embodiment, combining the established pollution distribution baseline with the constructed digital twin model, and optimizing the digital twin model using the adversarial transfer learning algorithm includes:

[0071] S131. Input the established pollution distribution baseline into the constructed digital twin model, simulate and obtain simulation data, and construct a data set for adversarial transfer learning by combining the initial pollutant concentration, airflow velocity, and equipment operating parameters;

[0072] S132. Based on the adversarial transfer learning algorithm, construct an adversarial transfer learning model, where the adversarial transfer learning model includes a convolutional encoder, a classifier, and a domain discriminator;

[0073] S133, using a convolutional encoder to extract spatiotemporal characteristics of pollutant concentrations in the data set to generate a feature vector;

[0074] S134. Use a domain discriminator to perform inter-domain difference analysis on the feature vectors, and combine a convolutional neural network to identify the distribution differences between the simulation data and the initial pollutant concentration, airflow velocity, and equipment operating parameters, and calculate the domain classification loss based on the identification results;

[0075] S135. Optimize the classifier and the domain discriminator based on the calculated domain classification loss and the adversarial training mechanism to obtain an optimized adversarial transfer learning model.

[0076] S136. Combine the optimized adversarial transfer learning model with the digital twin model to optimize the parameters of the digital twin model.

[0077] It should be noted that by inputting the pollution distribution baseline into the digital twin model and combining it with the adversarial transfer learning algorithm, the distribution differences between the simulated data and the actual data can be effectively overcome, and the accuracy of pollution source identification can be improved; the convolutional encoder is used to extract spatiotemporal features and generate feature vectors, and the domain discriminator is further used to analyze the differences between the data to ensure the model's adaptability to different data sources; the classifier and domain discriminator are optimized through adversarial training, which improves the model's sensitivity to pollutant concentration, airflow velocity and equipment operating parameters, enhances the model's generalization ability, and ensures real-time optimization and accurate prediction of the digital twin model in a dynamic environment; ultimately, the precise optimization of the digital twin model is achieved, making pollution source identification and control more efficient, reducing the cost of experiments and actual monitoring, and improving the stability of the production process and product quality.

[0078] S14. Use the optimized digital twin model to predict the spatiotemporal distribution characteristics of pollutants, and combine the multi-armed bandit algorithm to update the initial external detector deployment strategy to obtain the final external detector deployment strategy.

[0079] In this optional embodiment, the spatiotemporal distribution characteristics of pollutants are predicted using the optimized digital twin model, and the initial external detector deployment strategy is updated in combination with the multi-armed bandit algorithm to obtain the final external detector deployment strategy, including:

[0080] S141. Input the initial pollutant concentration, airflow velocity, and equipment operating parameters into the optimized digital twin model to predict the spatiotemporal distribution characteristics of the pollutant concentration in the interior space of the transport box;

[0081] S142. Based on the designed initial external detector deployment strategy, a nearest neighbor algorithm is used to calculate the spatiotemporal similarity between the external detector nodes, and a number of candidate external detector nodes that best match the current spatiotemporal distribution characteristics are identified;

[0082] S143. Calculate the average weight of each candidate external detector node based on the identified candidate external detector nodes and the preset pollutant prediction function, and evaluate the expected monitoring effect of each candidate external detector node based on the spatiotemporal distribution characteristics of the pollutant concentration;

[0083] S144. In combination with the multi-armed bandit algorithm, select the candidate external detector node with the best expected monitoring effect to update the initial external detector deployment strategy and obtain the final external detector deployment strategy.

[0084] It should be noted that by inputting the initial pollutant concentration, air flow velocity and equipment operating parameters into the optimized digital twin model, the temporal and spatial distribution of pollutant concentration inside the transport box can be accurately predicted, providing a reliable basis for the deployment of external detectors; the nearest neighbor algorithm is used to calculate the temporal and spatial similarity between external detector nodes, and the candidate nodes that best match the current distribution characteristics are effectively identified to ensure the scientific nature of the external detector layout; by calculating the average weight of the candidate nodes and evaluating their expected monitoring effects, the optimal node is selected in combination with the multi-armed bandit algorithm, and the external detector deployment strategy is dynamically updated, which significantly improves the representativeness and accuracy of the monitoring data, reduces the use of redundant external detectors, reduces deployment costs, and at the same time enhances the real-time and accuracy of pollution source identification, providing strong support for efficient and precise pollution control.

[0085] In this optional embodiment, in combination with the multi-armed bandit algorithm, a candidate external detector node with the best expected monitoring effect is selected to update the initial external detector deployment strategy. Obtaining the final external detector deployment strategy includes:

[0086] S1441. Map each candidate external detector node to a decision arm in a multi-armed bandit algorithm, and construct a reward function based on monitoring effectiveness;

[0087] S1442. Initialize the reward function of each decision arm based on the expected monitoring effect of each candidate external detector node and set the initial confidence interval;

[0088] S1443. Calculate the upper bound of each initial confidence interval based on the initial confidence interval of each decision arm, and select the candidate external detector node with the highest upper bound as the optimal external detector node through the exploration mechanism of the multi-armed bandit algorithm.

[0089] S1444. Generate external detector deployment recommendations for the next phase based on the optimal external detector node to update the external detector deployment strategy and obtain a final external detector deployment strategy.

[0090] It should be noted that by mapping candidate external detector nodes to decision arms of the multi-armed bandit algorithm and constructing a reward function based on monitoring efficiency, the expected monitoring effect of each external detector node can be effectively evaluated, and the dynamic optimization of the external detector deployment strategy can be achieved; initializing the reward function and confidence interval of each decision arm ensures the balance between exploration and utilization of the algorithm, and by calculating the upper limit of the confidence interval and selecting the optimal external detector node, the pertinence and efficiency of the external detector deployment are improved; based on the optimal node, deployment recommendations are generated, and the deployment strategy is continuously updated to ultimately obtain the optimal external detector deployment plan, which improves the coverage and accuracy of pollution monitoring, reduces the use of redundant external detectors, reduces deployment costs, enhances the real-time response capability of the system, and provides a reliable data foundation for accurate pollution source identification and efficient governance.

[0091] S15. Execute the final external detector deployment strategy and collect pollutant concentrations and associated equipment operating parameters in real time.

[0092] S2. Establish a spatiotemporal matrix of pollutant concentrations and use wavelet transform to extract the gradient characteristics of pollutant concentration changes. Combined with the operating parameters of related equipment, obtain preliminary location data of pollution sources.

[0093] It should be noted that the establishment of a spatiotemporal matrix of pollutant concentrations, the use of wavelet transform to extract the gradient characteristics of pollutant concentration changes, and the combination of operating parameters of related equipment to obtain preliminary location data of pollution sources include:

[0094] First, pollutant concentrations are collected and organized into a space-time matrix according to their temporal and spatial distribution, ensuring that the data at each moment and in each internal space can accurately reflect the changes in pollutants. Then, the space-time matrix is ​​processed using wavelet transform to extract the gradient characteristics of pollutant concentration changes and capture their fluctuation trends in time and space. Next, these gradient characteristics are combined with equipment operating parameters (such as airflow velocity, equipment status, etc.) to analyze the impact of different factors on pollutant diffusion. The preliminary location data of the pollution source is obtained through model fusion methods, providing a basis for subsequent pollution source positioning and control.

[0095] S3. Construct a spatiotemporal gridded input matrix and design a dual-branch deep learning model to generate a pollution source location probability map and obtain the location data of the pollution source.

[0096] In this optional embodiment, a spatiotemporal gridded input matrix is ​​constructed, and a dual-branch deep learning model is designed to generate a pollution source location probability map. Obtaining the location data of the pollution source includes:

[0097] S31. Based on the preliminary location data of pollution sources and the real-time collected pollutant concentrations, a spatiotemporal gridded input matrix is ​​constructed.

[0098] S32. Based on the convolutional neural network and temporal attention mechanism, a dual-branch deep learning model is designed, and the dual-branch deep learning model is used to extract the spatial and temporal characteristics of pollutants in the spatiotemporal gridded input matrix.

[0099] It should be noted that based on the convolutional neural network and the temporal attention mechanism, a dual-branch deep learning model is designed. The spatial and temporal characteristics of pollutants in the spatiotemporal gridded input matrix are extracted using the dual-branch deep learning model.

[0100] First, a spatiotemporal gridded input matrix is ​​constructed, in which the pollutant concentration and spatial information of each time step are integrated as input features. Next, a two-branch deep learning model is designed. The first branch uses a convolutional neural network to process spatial features and extract the spatial distribution characteristics of pollutants through multiple convolutional layers. The second branch uses a temporal attention mechanism to weight the pollutant concentration at each time step and learn the characteristics of important moments in the time dimension, thereby capturing the temporal variation patterns of pollutants. The output features of the two branches are then fused, combining spatial and temporal features to ultimately form a complete spatiotemporal feature representation for use in subsequent tasks (such as pollution source identification). Through this model, the spatial and temporal variation patterns of pollutants can be captured simultaneously, improving the accuracy and stability of pollution source positioning.

[0101] S33. The spatial and temporal features of pollutants are integrated, and the fusion results are injected into the preset pollution source heat map through feature splicing and full connection layer compression to generate a pollution source location probability map.

[0102] In this optional embodiment, the spatial and temporal features of pollutants are fused, and the fusion results are injected into a preset pollution source heat map through feature splicing and full connection layer compression to generate a pollution source location probability map. The map includes:

[0103] S331. Align the extracted spatial and temporal features of the pollutants according to the newly deployed external detector nodes, and perform feature concatenation to generate a joint feature matrix.

[0104] S332. Input the generated joint feature matrix into the fully connected layer, compress it using nonlinear dimensionality reduction technology, and extract feature components using low-rank matrix decomposition technology to obtain a compressed feature representation;

[0105] S333: Perform weighted fusion of the compressed feature representation and the preset pollution source heat map, smooth the fusion result using Gaussian filtering, and output a pollution source location probability map.

[0106] It should be noted that by aligning and concatenating the spatial and temporal features of pollutants extracted by the newly deployed external detector nodes, a joint feature matrix can be generated, thereby achieving effective fusion of multi-source data; by inputting the joint feature matrix into the fully connected layer and compressing it using nonlinear dimensionality reduction technology, the data dimension can be effectively reduced, computational efficiency can be improved, and key feature information can be retained at the same time; low-rank matrix decomposition technology further extracts feature components, optimizes feature representation, and ensures the accuracy and precision of pollution source positioning; finally, the compressed feature representation is weightedly fused with the pollution source heat map, and smoothed using Gaussian filtering to reduce noise interference and output an accurate pollution source positioning probability map, providing efficient and reliable support for real-time monitoring and control of pollution sources.

[0107] S34. Based on the generated pollution source location probability map and Markov decision process, and combined with offline simulation training and greedy strategy, the location data of the pollution source is obtained.

[0108] In this optional embodiment, based on the generated pollution source location probability map and Markov decision process, and in combination with offline simulation training and greedy strategy, obtaining the location data of the pollution source includes:

[0109] S341. Based on the Markov decision process, the pollution source location probability map is used as the environmental state, the pollution source verification operation is used as the intelligent agent action, and the spatial and temporal characteristics of the pollutants are combined to establish the state transition rules.

[0110] S342. Construct a multi-objective reward function based on the established state transition rules.

[0111] In this optional embodiment, the multi-objective reward function includes: operation cost penalty, successful positioning reward, information gain reward, probability density reward and distance penalty.

[0112] S343. Use the optimized digital twin model to generate a virtual pollution source scene dataset, build a convolutional policy network based on the dataset, and combine it with a multi-objective reward function to perform offline simulation training through the proximal policy optimization algorithm.

[0113] S344. Based on the action value output by the trained convolutional strategy network, the greedy strategy is used to select the optimal pollution source verification operation, and the pollution source location probability map is updated in combination with the feedback data from the external detector.

[0114] S345. Based on the updated pollution source location probability map, a verification path is generated, and the verification path is optimized in combination with the entropy value threshold to obtain the location data of the pollution source.

[0115] It should be noted that by combining the Markov decision process with the digital twin model, the pollution source location problem can be transformed into an intelligent decision-making task, thereby improving the accuracy and efficiency of pollution source location; by constructing a multi-objective reward function, the model can comprehensively consider multiple factors such as operating cost, positioning success rate, information gain, probability density and verification distance, optimize the pollution source verification strategy, and avoid the limitations of single-objective optimization; the optimized digital twin model is used to generate a virtual data set, and a convolutional policy network is constructed based on it, and offline training is performed through the proximal policy optimization algorithm to ensure the robustness and stability of the model in various environments; after training, the greedy strategy is used to select the optimal operation, and the pollution source location probability map is dynamically updated in combination with the external detector feedback data, realizing the accurate generation of pollution source verification paths, and further optimizing the verification paths through the entropy value threshold, and finally obtaining accurate pollution source location data; the efficiency and accuracy of pollution source location are improved, and it can adapt to the changing environment in real time and optimize the decision path, significantly reducing the cost of pollution source identification.

[0116] S4. Based on the location data of pollution sources, evaluate the contribution weight of each associated equipment to the spread of pollutants, and formulate and implement targeted governance strategies.

[0117] In this optional embodiment, based on the location data of the pollution source, the contribution weight of each associated device to the spread of pollutants is evaluated, and a targeted control strategy is formulated and implemented, including:

[0118] S41. Construct a cause-effect diagram model by combining the location data of pollution sources, the real-time collected pollutant concentrations, and the operating parameters of related equipment;

[0119] S42. Based on the constructed causal graph model, use causal inference technology to evaluate the impact of each related equipment on the pollutant diffusion process, obtain the contribution weight, and formulate and implement targeted governance strategies based on the contribution weight.

[0120] It should be noted that by combining pollution source location data, real-time pollutant concentrations and equipment operating parameters to construct a causal graph model, the causal relationship between pollution sources and equipment can be clarified, revealing the source and propagation path of pollutant diffusion; using causal inference technology to evaluate the impact of each related equipment on pollutant diffusion, it is possible to accurately identify which equipment has a significant contribution to pollutant diffusion, thereby obtaining the contribution weight of each device; this information can help formulate more accurate and effective targeted governance strategies, optimize equipment operation or adjust processes, reduce the risk of pollutant diffusion, and improve governance efficiency; it can achieve precise and intelligent control of pollution sources, improve the accuracy and response speed of pollution management, and reduce the cost of pollution control in the production process.

[0121] In summary, the above-mentioned technical solutions of the present invention combine the flow of air within the transport box, the diffusion characteristics of pollutants, and the operating parameters of related equipment to achieve real-time monitoring and collection of environmental data. By extracting and analyzing the pollutant concentration gradient characteristics in this data, the location of the pollution source can be accurately located, thereby improving the accuracy and response speed of pollution source identification and preventing the decline in product or environmental quality caused by the failure to detect pollution in a timely manner. By evaluating the contribution of each associated equipment to the diffusion of pollutants based on the pollution source location results, the pollution source can be accurately identified and targeted treatment strategies can be formulated accordingly. Combining pollution source location information with the operating status of the equipment can adjust the equipment operating parameters in real time, thereby reducing the risk of pollutant diffusion.

[0122] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for identifying contamination sources in wafer manufacturing based on the AMC distribution between layers within a FOUP, characterized by: The method includes: S1. Analyze airflow disturbances and contaminant diffusion characteristics based on the internal space of the transport box and the wafer level placement, formulate external detector deployment strategies, and collect contaminant concentrations and related equipment operating parameters in real time. S2. Establish a spatiotemporal matrix of pollutant concentrations and use wavelet transform to extract the gradient characteristics of pollutant concentration changes. Combined with the operating parameters of related equipment, obtain preliminary location data of pollution sources; S3. Construct a spatiotemporal gridded input matrix and design a dual-branch deep learning model to generate a pollution source location probability map and obtain the location data of the pollution source; S4. Based on the location data of pollution sources, evaluate the contribution weight of each associated device to the spread of pollutants, and formulate and implement targeted governance strategies; Said S1 comprises: S11. Based on the interior space of the transport box, establish a parameterized geometric model, set fluid boundary conditions, and run computational fluid dynamics simulations to obtain baseline airflow field and pollutant diffusion characteristic data to build a digital twin model. S12. Design the initial external detector deployment strategy based on wafer layer placement, collect initial pollutant concentrations, airflow velocities, and associated equipment operating parameters at each layer, and establish a spatiotemporal pollution distribution baseline. S13. Combining the established pollution distribution baseline with the constructed digital twin model, the adversarial transfer learning algorithm is used to optimize the digital twin model; S14. Predict the spatiotemporal distribution characteristics of pollutants using the optimized digital twin model, and update the initial external detector deployment strategy in combination with the multi-armed bandit algorithm to obtain the final external detector deployment strategy; S15. Execute the final external detector deployment strategy and collect pollutant concentrations and associated equipment operating parameters in real time.

2. The method for identifying contamination sources in wafer manufacturing process based on AMC distribution between layers in a FOUP according to claim 1, characterized in that: Combining the established pollution distribution baseline with the constructed digital twin model and optimizing the digital twin model using the adversarial transfer learning algorithm includes: S131. Input the established pollution distribution baseline into the constructed digital twin model, simulate and obtain simulation data, and construct a data set for adversarial transfer learning by combining the initial pollutant concentration, airflow velocity, and equipment operating parameters; S132. Construct an adversarial transfer learning model based on an adversarial transfer learning algorithm, wherein the adversarial transfer learning model includes a convolutional encoder, a classifier, and a domain discriminator; S133, using a convolutional encoder to extract spatiotemporal characteristics of pollutant concentrations in the data set to generate a feature vector; S134. Use a domain discriminator to perform inter-domain difference analysis on the feature vectors, and combine a convolutional neural network to identify the distribution differences between the simulation data and the initial pollutant concentration, airflow velocity, and equipment operating parameters, and calculate the domain classification loss based on the identification results; S135. Optimize the classifier and the domain discriminator based on the calculated domain classification loss and the adversarial training mechanism to obtain an optimized adversarial transfer learning model. S136. Combine the optimized adversarial transfer learning model with the digital twin model to optimize the parameters of the digital twin model.

3. The method for identifying contamination sources in wafer manufacturing process based on AMC distribution between layers in a FOUP according to claim 2, characterized in that: The method of using the optimized digital twin model to predict the spatiotemporal distribution characteristics of pollutants and combining the multi-armed bandit algorithm to update the initial external detector deployment strategy to obtain the final external detector deployment strategy includes: S141. Input the initial pollutant concentration, airflow velocity, and equipment operating parameters into the optimized digital twin model to predict the spatiotemporal distribution characteristics of the pollutant concentration in the interior space of the transport box; S142. Based on the designed initial external detector deployment strategy, a nearest neighbor algorithm is used to calculate the spatiotemporal similarity between the external detector nodes, and a number of candidate external detector nodes that best match the current spatiotemporal distribution characteristics are identified; S143. Calculate the average weight of each candidate external detector node based on the identified candidate external detector nodes and the preset pollutant prediction function, and evaluate the expected monitoring effect of each candidate external detector node based on the spatiotemporal distribution characteristics of the pollutant concentration; S144. In combination with the multi-armed bandit algorithm, select the candidate external detector node with the best expected monitoring effect to update the initial external detector deployment strategy and obtain the final external detector deployment strategy.

4. The method for identifying contamination sources in wafer manufacturing process based on AMC distribution between layers in a FOUP according to claim 3, characterized in that: The method of combining the multi-armed bandit algorithm to select a candidate external detector node with the best expected monitoring effect to update the initial external detector deployment strategy and obtain the final external detector deployment strategy includes: S1441. Map each candidate external detector node to a decision arm in a multi-armed bandit algorithm, and construct a reward function based on monitoring effectiveness; S1442. Initialize the reward function of each decision arm based on the expected monitoring effect of each candidate external detector node and set the initial confidence interval; S1443. Calculate the upper bound of each initial confidence interval based on the initial confidence interval of each decision arm, and select the candidate external detector node with the highest upper bound as the optimal external detector node through the exploration mechanism of the multi-armed bandit algorithm. S1444. Generate external detector deployment recommendations for the next phase based on the optimal external detector node to update the external detector deployment strategy and obtain a final external detector deployment strategy.

5. The method for identifying contamination sources in wafer manufacturing process based on AMC distribution between layers in a FOUP according to claim 1, characterized in that: The construction of a spatiotemporal gridded input matrix and the design of a dual-branch deep learning model to generate a pollution source location probability map and obtain the location data of the pollution source include: S31. Construct a spatiotemporal gridded input matrix based on the preliminary location data of pollution sources and the real-time collected pollutant concentrations; S32. Design a dual-branch deep learning model based on convolutional neural networks and temporal attention mechanisms, and use the dual-branch deep learning model to extract the spatial and temporal characteristics of pollutants in the spatiotemporal gridded input matrix; S33: Fuse the spatial and temporal features of pollutants, and inject the fusion results into the preset pollution source heat map through feature splicing and full connection layer compression to generate a pollution source location probability map; S34. Based on the generated pollution source location probability map and Markov decision process, and combined with offline simulation training and greedy strategy, the location data of the pollution source is obtained.

6. The method for identifying contamination sources in wafer manufacturing process based on AMC distribution between layers in a FOUP according to claim 5, characterized in that: The method of fusing the spatial and temporal features of pollutants and injecting the fusion results into the preset pollution source heat map through feature splicing and full connection layer compression to generate the pollution source location probability map includes: S331. Align the extracted spatial and temporal features of the pollutants according to the newly deployed external detector nodes, and perform feature concatenation to generate a joint feature matrix. S332. Input the generated joint feature matrix into the fully connected layer, compress it using nonlinear dimensionality reduction technology, and extract feature components using low-rank matrix decomposition technology to obtain a compressed feature representation; S333: Perform weighted fusion of the compressed feature representation and the preset pollution source heat map, smooth the fusion result using Gaussian filtering, and output a pollution source location probability map.

7. The method for identifying contamination sources in wafer manufacturing process based on AMC distribution between layers in a FOUP according to claim 6, characterized in that: The pollution source location data is obtained based on the generated pollution source location probability map and Markov decision process, combined with offline simulation training and greedy strategy, including: S341. Based on the Markov decision process, the pollution source location probability map is used as the environmental state, the pollution source verification operation is used as the agent action, and the spatial and temporal characteristics of the pollutants are combined to establish the state transition rules; S342. Construct a multi-objective reward function according to the established state transition rule; S343. Generate a virtual pollution source scenario dataset using the optimized digital twin model, build a convolutional policy network based on the dataset, and perform offline simulation training using a proximal policy optimization algorithm in combination with a multi-objective reward function. S344. Based on the action value output by the trained convolutional strategy network, the greedy strategy is used to select the optimal pollution source verification operation, and the pollution source location probability map is updated in combination with the feedback data from the external detector. S345. Based on the updated pollution source location probability map, a verification path is generated, and the verification path is optimized in combination with the entropy value threshold to obtain the location data of the pollution source.

8. The method for identifying contamination sources in wafer manufacturing process based on AMC distribution between layers in a FOUP according to claim 7, characterized in that: The multi-objective reward function includes: operation cost penalty, successful positioning reward, information gain reward, probability density reward and distance penalty.

9. The method for identifying contamination sources in wafer manufacturing process based on AMC distribution between layers in a FOUP according to claim 1, characterized in that: Based on the location data of the pollution source, the contribution weight of each associated device to the spread of pollutants is evaluated, and a targeted governance strategy is formulated and implemented, including: S41. Construct a causal graph model by combining the location data of pollution sources, real-time pollutant concentrations and operating parameters of related equipment. Based on the constructed causal graph model, use causal inference technology to evaluate the impact of each related equipment on the pollutant diffusion process, obtain contribution weights, and formulate based on the contribution weights.

Images