Semiconductor gaseous molecular contaminant dynamic content detection system and method

Abstract

The application discloses a semiconductor gaseous molecular pollutant dynamic content detection system and method, and belongs to the technical field of artificial intelligence. A pollution source candidate feature tensor set is generated through a multi-modal time sequence synchronization mechanism. The feature tensor set is input into a pollution diffusion dynamic modeling network, and a space-time feature matrix containing a pollutant migration trend is output. A pollutant diffusion topological atlas is constructed based on a pollution concentration change rate, a weighted graph convolutional neural network is used to extract dynamic diffusion weights between nodes, and a migration path is corrected in combination with a boundary layer turbulence model. A pollutant source-sink prediction model is constructed, and a pollution concentration change prediction curve and a pollution source positioning result are output. The application has the technical advantages of accurate pollution tracing, real-time prediction response and fine pollution situation awareness, and is suitable for pollution control and dynamic scheduling scenes in a high-level clean environment.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a system and method for detecting the dynamic content of semiconductor gaseous molecular pollutants. Background Technology

[0002] In semiconductor manufacturing, gaseous molecular contaminants (such as ammonia, hydrogen sulfide, fluorides, and organic acids) have a significant impact on key processes such as photoresist, thin film deposition, and ion implantation. Even minute fluctuations in these contaminants can trigger wafer surface defects or interface contamination, leading to device failure. Therefore, the industry has placed higher demands on the monitoring accuracy, response speed, and traceability capabilities of gaseous contaminants.

[0003] However, existing pollutant detection systems generally suffer from the following technical bottlenecks: First, sampling devices rely on fixed locations and lack the ability to comprehensively model the dynamic diffusion process of pollutants in clean rooms; second, the coupling relationship between pollutant concentration monitoring and airflow field, heat exchange, and equipment movement status cannot be effectively extracted, making it difficult to predict pollutant migration trends; third, there is a lack of a real-time feedback system that integrates multi-source sensor data and deep learning mechanisms, resulting in the inability to provide timely early warning and intelligent intervention for sudden pollution events.

[0004] Therefore, there is an urgent need for a dynamic detection method that can integrate spatiotemporal perception of pollutants, airflow modeling, and intelligent prediction. This method can improve the identification, tracking, and scheduling response capabilities of trace pollutants in the semiconductor manufacturing environment by constructing a pollutant migration topology network, introducing a multi-scale feature decoupling mechanism, and a pollution response heatmap. Summary of the Invention

[0005] The purpose of this invention is to provide a system and method for detecting the dynamic content of semiconductor gaseous molecular pollutants, in order to overcome the shortcomings of the prior art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for detecting the dynamic content of semiconductor gaseous molecular contaminants, comprising:

[0007] Acquire multi-source heterogeneous sensing data from multiple areas within a cleanroom, and generate a candidate feature tensor set for pollution sources through a multimodal temporal synchronization mechanism;

[0008] Input the feature tensor set into the pollution diffusion dynamic modeling network, and output a spatiotemporal feature matrix containing the migration trend of pollutants;

[0009] A pollutant diffusion topology map is constructed based on the rate of change of pollution concentration. A weighted graph convolutional neural network is used to extract dynamic diffusion weights between nodes, and a boundary layer turbulence model is combined to correct the migration path.

[0010] A pollutant source-sink prediction model is constructed, and based on the reverse tracing and forward diffusion mechanism, the predicted curve of pollutant concentration change and the location results of potential pollution sources are output.

[0011] A loss function is constructed from the residuals of the pollution trend curve and the measured data, and a pollution response optimizer is trained to generate a high-resolution pollution prediction map.

[0012] Based on the pollution prediction map, a joint analysis of spectrum and information entropy is performed to extract pollution event mutation frequency, airflow disturbance response entropy, and migration path offset factor as dynamic quantitative indicators of pollution.

[0013] When any pollution quantification indicator exceeds the threshold, the system outputs pollution source coordinates, cleanliness level adjustment suggestions, and HVAC system optimization exhaust strategy.

[0014] Preferably, the pollution diffusion dynamic modeling network includes:

[0015] Multi-layer variable receptive field residual convolution is performed on the pollutant concentration tensor in the feature tensor set to extract multi-scale spatial concentration gradient features;

[0016] A flow-aware convolution module based on Bayesian weighting is introduced into the environmental wind speed vector tensor to extract the directional field features of pollutants along the dominant migration path of airflow;

[0017] By combining the frequency changes of boundary perturbation between the pollution source and the clean area, a self-attention time alignment mechanism is introduced to dynamically construct the temporal dependency map of pollution events;

[0018] The extracted multi-source spatial and temporal features are encoded and combined in a spatiotemporal cross-fusion layer to output a high-dimensional spatiotemporal feature matrix that includes pollutant diffusion trends, directional drift, and risk clustering areas.

[0019] Preferably, the construction of the pollutant diffusion topology map includes:

[0020] Based on the spatial gradient change rate of the pollution concentration time series tensor, a density threshold segmentation and region growth algorithm is used to extract the high gradient diffusion region of pollutants and generate an initial topological node set.

[0021] The node set is subjected to regional map construction based on Voronoi diagram partitioning to form spatial connection relationships between nodes;

[0022] The connection relationship is input into a graph convolutional neural network, and the dynamic diffusion weight matrix between nodes is extracted by joint training with concentration gradient difference and time-varying weights.

[0023] A boundary layer turbulence correction model is introduced, and the node transfer probability in the pollution diffusion path is adjusted based on the velocity shear tensor and the local Reynolds number to form a migration channel map.

[0024] Preferably, the construction of the pollutant source-sink prediction model includes:

[0025] On the pollutant diffusion topology map, a candidate set of pollutant source points is constructed based on the inverse gradient of node concentration, and high-confidence pollutant source nodes are screened by combining the maximum path entropy criterion.

[0026] Simultaneously, forward diffusion simulation is performed on the graph structure based on the rate of change of pollution concentration, and the transmission path of pollution trend is predicted using a time window sliding convolution model;

[0027] By performing path overlap analysis on the results of reverse tracing and forward diffusion, a pollution concentration change prediction curve is generated using the path consistency confidence function.

[0028] Based on the response delay and diffusion directionality of the source and end-point concentration evolution, the coordinate region of pollutant release is inverted, and the location results of potential pollution sources are output.

[0029] Preferably, the training of the pollution response optimizer includes:

[0030] Construct a residual time series between the fused pollution trend curve and the measured concentration value, and define a composite loss function that includes a prediction offset penalty term and a local gradient consistency term;

[0031] The residual sequence is input into a pollution response optimizer based on a variable weight attention mechanism to automatically adjust the spatial compensation factor within the prediction region.

[0032] In the optimizer training process, multi-scale regularization terms and reconstruction confidence constraints are introduced to improve the expression accuracy of the prediction map in boundary regions and low-concentration regions.

[0033] The final output is a high-resolution contamination prediction map covering key areas of the cleanroom, including continuous contaminant concentration fields, risk accumulation hotspots, and confidence level distribution information.

[0034] Preferably, the extraction of the dynamic quantitative indicators of pollution includes:

[0035] A fast Fourier transform is performed on the concentration change time series data in the pollution prediction graph to extract the main frequency component and calculate the high frequency energy ratio, and output the pollution event mutation frequency index.

[0036] Based on the local concentration response curve of the wind speed field disturbance area in the prediction map, a probability distribution model is constructed and the information entropy value is obtained to form the airflow disturbance response entropy index.

[0037] By comparing the changes in the main direction trajectory of the pollution diffusion path in multiple time slices, a path offset vector field is constructed, and its root mean square offset is calculated as the migration path offset factor.

[0038] Preferably, the multi-source heterogeneous sensing data includes the real-time concentration value of the target pollutant gas, environmental temperature and humidity parameters, ventilation airflow field status, and operating conditions of related process equipment.

[0039] The present invention also provides a semiconductor gaseous molecular contaminant dynamic content detection system, comprising:

[0040] Feature acquisition module: acquires multi-source heterogeneous sensing data from multiple areas within the cleanroom, and generates a candidate feature tensor set for pollution sources through a multimodal temporal synchronization mechanism;

[0041] Pollution modeling module: Inputs the feature tensor set into the pollution diffusion dynamic modeling network and outputs a spatiotemporal feature matrix containing the migration trend of pollutants;

[0042] Topology graph construction module: Constructs a pollutant diffusion topology graph based on the rate of change of pollution concentration, uses a weighted graph convolutional neural network to extract dynamic diffusion weights between nodes, and combines a boundary layer turbulence model to correct the migration path;

[0043] Source-sink prediction module: Constructs a pollutant source-sink prediction model, based on reverse tracing and forward diffusion mechanisms, and outputs pollutant concentration change prediction curves and potential pollution source location results;

[0044] Response optimization module: Constructs a loss function from the residuals of the pollution trend curve and the measured data, and trains the pollution response optimizer to generate a high-resolution pollution prediction map;

[0045] Pollution index analysis module: Based on the pollution prediction map, perform joint spectrum-information entropy analysis to extract pollution event mutation frequency, airflow disturbance response entropy, and migration path offset factor as dynamic quantitative indicators of pollution.

[0046] Early warning and decision-making module: When any pollution quantification indicator exceeds the threshold, it outputs pollution source coordinates, cleanliness level adjustment suggestions, and HVAC system optimization exhaust strategy.

[0047] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0048] 1. This invention overcomes the bottlenecks of existing technologies, such as reliance on a single data source, response lag, and insufficient path prediction accuracy, by constructing an integrated mechanism for multi-source heterogeneous sensing and pollution dynamic modeling. Employing a multimodal tensor fusion and temporal synchronization mechanism, it effectively integrates heterogeneous factors such as gas concentration, airflow disturbance, temperature and humidity environment, and equipment operating conditions, reconstructing the physical driving conditions of pollutant migration in real time, and significantly improving the comprehensiveness and dynamic analysis capabilities of pollution identification.

[0049] 2. This invention introduces a pollution evolution prediction path mechanism based on topological modeling and source-sink inference. Combining graph convolutional networks, turbulence correction models, and frequency domain entropy analysis methods, it achieves closed-loop modeling of pollution source initiation, diffusion paths, mutation risks, and dynamic response strategies. The technical effects are: significantly improved pollution path identification accuracy, shortened response time to mutation events, and the system's output pollution heatmap and early warning indicators possess stronger spatiotemporal continuity and decision-making guidance value, providing a forward-looking and adaptable intelligent means for cleanroom pollution control. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0051] Figure 1 This is a mind map of the method of the present invention.

[0052] Figure 2 This is a mind map of the system modules of the present invention. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] Example 1, please refer to Figure 1 As shown, the method for detecting the dynamic content of semiconductor gaseous molecular contaminants in this embodiment includes:

[0055] Acquire multi-source heterogeneous sensing data from multiple areas within a cleanroom, and generate a candidate feature tensor set for pollution sources through a multimodal temporal synchronization mechanism;

[0056] Input the feature tensor set into the pollution diffusion dynamic modeling network, and output a spatiotemporal feature matrix containing the migration trend of pollutants;

[0057] A pollutant diffusion topology map is constructed based on the rate of change of pollution concentration. A weighted graph convolutional neural network is used to extract dynamic diffusion weights between nodes, and a boundary layer turbulence model is combined to correct the migration path.

[0058] A pollutant source-sink prediction model is constructed, and based on the reverse tracing and forward diffusion mechanism, the predicted curve of pollutant concentration change and the location results of potential pollution sources are output.

[0059] A loss function is constructed from the residuals of the pollution trend curve and the measured data, and a pollution response optimizer is trained to generate a high-resolution pollution prediction map.

[0060] Based on the pollution prediction map, a joint analysis of spectrum and information entropy is performed to extract pollution event mutation frequency, airflow disturbance response entropy, and migration path offset factor as dynamic quantitative indicators of pollution.

[0061] When any pollution quantification indicator exceeds the threshold, the system outputs pollution source coordinates, cleanliness level adjustment suggestions, and HVAC system optimization exhaust strategy.

[0062] The multi-source heterogeneous sensing data includes the real-time concentration of the target pollutant gas, environmental temperature and humidity parameters, ventilation airflow field status, and operating conditions of related process equipment.

[0063] The system is deployed inside a semiconductor cleanroom and collects real-time concentration data of gaseous pollutants through a distributed high-precision sensor array, covering typical pollutant types such as ammonia, acidic gases, and volatile organic compounds. It also records the sampling timestamp and spatial location information simultaneously to form a high-resolution pollutant concentration tensor.

[0064] Meanwhile, the system collects temperature and relative humidity information of the factory environment in real time as important background parameters for changes in pollutant concentration and diffusion rate. The temperature and humidity data sampling frequency is no less than 1Hz, covering the multi-functional areas of the clean room, to ensure the accurate representation of the pollution behavior modeling by temperature and humidity disturbances.

[0065] The ventilation airflow field status is obtained by integrating wind speed and direction sensors with CFD model results to construct a three-dimensional ventilation airflow field vector diagram, which can reflect the guiding effect of aerodynamic conditions on the propagation path of pollutants in real time. The wind speed range covers 0.05–1.2 m / s, and the spatial granularity is no greater than 0.5 m.

[0066] In addition, the system connects to the process scheduling platform to collect operating status information of key equipment related to pollutant release, including the operating status, opening and closing sequence and exhaust pressure data of lithography machines, etching machines and cleaning equipment. Combined with time stamps and spatial partitions, a potential correlation mapping between pollution sources and process events is established.

[0067] The collaborative acquisition and unified temporal-spatial modeling of multi-source heterogeneous sensing data provide a rich and computable data foundation for the full-factor coupled analysis of pollution migration mechanisms, pollution trend prediction, and source tracing, significantly improving the scenario adaptability and data accuracy of pollution dynamic modeling.

[0068] The pollution diffusion dynamic modeling network includes:

[0069] Multi-layer variable receptive field residual convolution is performed on the pollutant concentration tensor in the feature tensor set to extract multi-scale spatial concentration gradient features;

[0070] The variable receptive field residual convolutional structure employs a dynamic kernel size adjustment mechanism. By inserting dilated convolutions and scale-dilated layers into deep convolutional channels, the model can capture both local gradient changes and overall migration trends of pollutants at different spatial scales. Simultaneously, residual connections preserve the original pollutant concentration boundary information, improving the spatial resolution of the pollution distribution field.

[0071] A flow-aware convolution module based on Bayesian weighting is introduced into the environmental wind speed vector tensor to extract the directional field features of pollutants along the dominant migration path of airflow;

[0072] The flow direction sensing module dynamically adjusts the response weights of the convolution kernel in each direction using a Bayesian confidence update method, and derives the probability distribution of the main migration direction through the prior similarity between the historical airflow trajectory and the current velocity vector. The direction field tensor output by this module can effectively represent the actual diffusion trend of pollutants driven by unstable airflow.

[0073] By combining the frequency changes of boundary perturbation between the pollution source and the clean area, a self-attention time alignment mechanism is introduced to dynamically construct the temporal dependency map of pollution events;

[0074] The self-attention time alignment mechanism calculates the correlation between pollution intensity changes and boundary perturbation spectra in different time slices, adaptively adjusts the time window length and weight configuration, thereby enhancing the model's sensitivity to early signals of sudden pollution events and constructing a time-dependent graph with dynamic backtracking capabilities.

[0075] The extracted multi-source spatial and temporal features are encoded and combined in a spatiotemporal cross-fusion layer to output a high-dimensional spatiotemporal feature matrix that includes pollutant diffusion trends, directional drift, and risk clustering areas.

[0076] The spatiotemporal cross-fusion layer adopts a parallel design of spatial and temporal paths, which perform feature reconstruction and joint encoding operations respectively. It also integrates multi-dimensional heterogeneous information such as pollution concentration, wind speed direction and disturbance response through a channel attention mechanism. The final output spatiotemporal feature matrix has good pollution migration coherence and spatial risk focusing ability, providing an accurate feature foundation for subsequent topology modeling and trend prediction.

[0077] The construction of the pollutant diffusion topology map includes:

[0078] Based on the spatial gradient change rate of the pollution concentration time series tensor, a density threshold segmentation and region growth algorithm is used to extract the high gradient diffusion region of pollutants and generate an initial topological node set.

[0079] By calculating the rate of change of local spatial gradient of pollution concentration in continuous time frames, regions with drastic fluctuations in pollutant concentration are identified. High-risk node regions are then screened using a set density threshold. Subsequently, a region growing algorithm is used to merge adjacent high-gradient points into clustered regions, ultimately extracting the initial topological node set that constitutes the main path of pollution propagation. This node set represents the possible sources, relay points, and risk convergence regions of pollution diffusion.

[0080] The node set is subjected to regional map construction based on Voronoi diagram partitioning to form spatial connection relationships between nodes;

[0081] The Voronoi diagram partitioning mechanism constructs polygonal regions based on the Euclidean distances between topological nodes in three-dimensional space. Each node controls its nearest neighbor, thus forming the spatial distribution and adjacency relationships between nodes. This structure preserves the geographical continuity of pollution diffusion and is suitable for high-resolution clean space modeling.

[0082] The connection relationship is input into a graph convolutional neural network, and the dynamic diffusion weight matrix between nodes is extracted by joint training with concentration gradient difference and time-varying weights.

[0083] The graph convolutional neural network (GCN) receives the spatial connectivity graph between nodes and the time series pollution concentration difference as input features. By constructing a diffusion weight propagation model, it jointly trains the concentration gradient between node pairs and their temporal evolution trend. The output diffusion weight matrix can express the migration intensity and directionality of pollution in space.

[0084] A boundary layer turbulence correction model is introduced, and the node transfer probability in the pollution diffusion path is adjusted based on the velocity shear tensor and the local Reynolds number to form a migration channel map.

[0085] The turbulence-corrected model calculates the velocity shear tensor and local Reynolds number at each connecting edge based on the microscale airflow variation characteristics around the pollution source, and adjusts the transfer probability weights between nodes according to the mass transfer-diffusion coupling model to generate a migration channel map that can dynamically evolve over time. This channel map not only reflects the main path of pollution diffusion, but also accommodates the offset propagation trajectory driven by local disturbances, enhancing the model's adaptability to unsteady pollution behavior.

[0086] The construction of the pollutant source-sink prediction model includes:

[0087] On the pollutant diffusion topology map, a candidate set of pollutant source points is constructed based on the inverse gradient of node concentration, and high-confidence pollutant source nodes are screened by combining the maximum path entropy criterion.

[0088] The system first calculates the reverse concentration gradient values ​​of each node in the topology map and then traces the nodes whose concentration gradually increases along the path, forming a cluster of candidate pollution sources. Subsequently, the path entropy criterion is introduced, using path uncertainty as the criterion. By maximizing the information entropy value of the candidate path set, the system selects the core source points with the most stable initial direction of pollution diffusion and the most concentrated paths, thereby improving the confidence accuracy of source tracing.

[0089] Simultaneously, forward diffusion simulation is performed on the graph structure based on the rate of change of pollution concentration, and the transmission path of pollution trend is predicted using a time window sliding convolution model;

[0090] The temporal difference sequence of pollution concentration is mapped to each node in the topology. Its evolution trend is extracted by convolution operation with sliding time window. Combined with the diffusion weight between nodes, forward simulation is performed to generate a spectrum of possible transport paths and velocity changes of pollutants, which is used to predict potential downstream risk areas.

[0091] By performing path overlap analysis on the results of reverse tracing and forward diffusion, a pollution concentration change prediction curve is generated using the path consistency confidence function.

[0092] By matching the spatial overlap rate and temporal phase difference between the reverse and forward paths, a path consistency confidence function is constructed, outputting a pollution trend prediction curve after path fusion. This prediction curve can accurately reflect the concentration diffusion rhythm and impact range after pollution source initiation, supporting the generation of subsequent dynamic response strategies.

[0093] Based on the response delay and diffusion directionality of source and end-point concentration evolution, the coordinate region where pollutants may be released is retrieved, and the location results of potential pollution sources are output.

[0094] The system constructs an inversion model based on the concentration change time delay between source nodes and terminal nodes, the path geometry angle, and the coupling degree of the main airflow direction. It then performs location inversion by combining the pollution evolution characteristics of multiple time slices, and finally outputs the pollutant release coordinate area with high confidence, providing a location basis for pollution control and clean scheduling.

[0095] The training of the pollution response optimizer includes:

[0096] Construct a residual time series between the fused pollution trend curve and the measured concentration value, and define a composite loss function that includes a prediction offset penalty term and a local gradient consistency term;

[0097] The system first calculates the difference between the predicted pollutant concentration curve and the measured concentration at the corresponding time point to form a residual time series. Based on this series, a composite loss function is designed, which includes an absolute offset penalty term between the predicted and measured values, as well as a structural constraint term for the concentration gradient difference between adjacent nodes. This balances global fitting accuracy with local spatial continuity, improving the model's generalization ability and boundary fitting performance.

[0098] The residual sequence is input into a pollution response optimizer based on a variable weight attention mechanism to automatically adjust the spatial compensation factor within the prediction region.

[0099] The pollution response optimizer adopts a multi-channel attention network architecture, dynamically allocates training resources according to the error distribution of different regions, guides variable attention weights through residual strength, automatically amplifies the attention to high error or pollution mutation regions, and weakens the adjustment intensity of prediction stable regions, thereby achieving fine adjustment of the prediction map.

[0100] In the optimizer training process, multi-scale regularization terms and reconstruction confidence constraints are introduced to improve the expression accuracy of the prediction map in boundary regions and low-concentration regions.

[0101] By introducing multi-scale spatial convolution paths to extract details of concentration changes at the boundaries, and adding reconstruction confidence constraints to suppress prediction uncertainty fluctuations, the pollution map maintains high confidence and structural stability in low concentration ranges and at abrupt boundary changes, avoiding false boundaries and inflated estimates.

[0102] The final output is a high-resolution contamination prediction map covering key areas of the cleanroom, including continuous contaminant concentration fields, risk accumulation hot zones, and confidence level distribution information.

[0103] The predicted map is a reconstruction result of the pollution concentration tensor output by the optimizer. It has complete spatial continuity and traceability of temporal evolution. At the same time, it outputs risk clustering heat map and pixel-level confidence heat map corresponding to the predicted map, providing intelligent decision-making basis for pollution zoning assessment, emergency response and dynamic cleanliness control.

[0104] The extraction of the dynamic quantitative indicators of pollution includes:

[0105] A fast Fourier transform is performed on the concentration change time series data in the pollution prediction graph to extract the main frequency component and calculate the high frequency energy ratio, and output the pollution event mutation frequency index.

[0106] The system performs a one-dimensional Fast Fourier Transform (FFT) on the concentration change sequences of each spatial node in the pollution prediction map to analyze their spectral structure. It identifies abrupt changes in concentration evolution by analyzing the dominant frequency energy distribution characteristics, and further calculates the proportion of high-frequency energy as a dynamic abrupt change frequency indicator for pollution events. This indicator can accurately capture the suddenness and intensity of pollution releases, improving the timeliness of risk identification.

[0107] Based on the local concentration response curve of the wind speed field disturbance area in the prediction map, a probability distribution model is constructed and the information entropy value is obtained to form the airflow disturbance response entropy index.

[0108] In areas with significant wind speed field disturbances, the system collects corresponding concentration response curves, establishes a probability density function model based on the degree of concentration fluctuation with airflow changes, and calculates its information entropy value. Higher information entropy indicates a more unstable response, characterizing the complexity of pollution disturbances driven by airflow in that area. This airflow disturbance response entropy index can be used to calibrate the degree of disturbance to pollution source diffusion, helping to identify turbulence-sensitive areas and anomalous propagation zones.

[0109] By comparing the changes in the main direction trajectory of the pollution diffusion path in multiple time slices, a path offset vector field is constructed, and its root mean square offset is calculated as the migration path offset factor.

[0110] Based on continuous time-series prediction maps, the system extracts the direction vectors of the main pollution migration paths for each time slice and constructs a path offset vector field through spatial registration and trajectory overlay. By calculating the root mean square offset value of this vector field across the entire diffusion domain, a migration path offset factor is formed, which measures the degree of diffusion instability and offset trend of pollutants under dynamic disturbances, providing a geometric evolution basis for abrupt change early warning.

[0111] Example 2, please refer to Figure 2 As shown, the semiconductor gaseous molecular contaminant dynamic content detection system of this embodiment includes:

[0112] Feature acquisition module: acquires multi-source heterogeneous sensing data from multiple areas within the cleanroom, and generates a candidate feature tensor set for pollution sources through a multimodal temporal synchronization mechanism;

[0113] Pollution modeling module: Inputs the feature tensor set into the pollution diffusion dynamic modeling network and outputs a spatiotemporal feature matrix containing the migration trend of pollutants;

[0114] Topology graph construction module: Constructs a pollutant diffusion topology graph based on the rate of change of pollution concentration, uses a weighted graph convolutional neural network to extract dynamic diffusion weights between nodes, and combines a boundary layer turbulence model to correct the migration path;

[0115] Source-sink prediction module: Constructs a pollutant source-sink prediction model, based on reverse tracing and forward diffusion mechanisms, and outputs pollutant concentration change prediction curves and potential pollution source location results;

[0116] Response optimization module: Constructs a loss function from the residuals of the pollution trend curve and the measured data, and trains the pollution response optimizer to generate a high-resolution pollution prediction map;

[0117] Pollution index analysis module: Based on the pollution prediction map, perform joint spectrum-information entropy analysis to extract pollution event mutation frequency, airflow disturbance response entropy, and migration path offset factor as dynamic quantitative indicators of pollution.

[0118] Early warning and decision-making module: When any pollution quantification indicator exceeds the threshold, it outputs pollution source coordinates, cleanliness level adjustment suggestions, and HVAC system optimization exhaust strategy.

[0119] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for detecting the dynamic content of gaseous molecular contaminants in semiconductors, characterized in that: include: Acquire multi-source heterogeneous sensing data from multiple areas within a cleanroom, and generate a candidate feature tensor set for pollution sources through a multimodal temporal synchronization mechanism; Input the feature tensor set into the pollution diffusion dynamic modeling network, and output a spatiotemporal feature matrix containing the migration trend of pollutants; A pollutant diffusion topology map is constructed based on the rate of change of pollution concentration. A weighted graph convolutional neural network is used to extract dynamic diffusion weights between nodes, and a boundary layer turbulence model is combined to correct the migration path. A pollutant source-sink prediction model is constructed, and based on the reverse tracing and forward diffusion mechanism, the predicted curve of pollutant concentration change and the location results of potential pollution sources are output. A loss function is constructed from the residuals of the pollution trend curve and the measured data, and a pollution response optimizer is trained to generate a high-resolution pollution prediction map. Based on the pollution prediction map, a joint analysis of spectrum and information entropy is performed to extract pollution event mutation frequency, airflow disturbance response entropy, and migration path offset factor as dynamic quantitative indicators of pollution. When any pollution quantification indicator exceeds the threshold, the system outputs pollution source coordinates, cleanliness level adjustment suggestions, and HVAC system optimization exhaust strategy.

2. The method for detecting the dynamic content of semiconductor gaseous molecular contaminants according to claim 1, characterized in that: The pollution diffusion dynamic modeling network includes: Multi-layer variable receptive field residual convolution is performed on the pollutant concentration tensor in the feature tensor set to extract multi-scale spatial concentration gradient features; A flow-aware convolution module based on Bayesian weighting is introduced into the environmental wind speed vector tensor to extract the directional field features of pollutants along the dominant migration path of airflow; By combining the frequency changes of boundary perturbation between the pollution source and the clean area, a self-attention time alignment mechanism is introduced to dynamically construct the temporal dependency map of pollution events; The extracted multi-source spatial and temporal features are encoded and combined in a spatiotemporal cross-fusion layer to output a high-dimensional spatiotemporal feature matrix that includes pollutant diffusion trends, directional drift, and risk clustering areas.

3. The method for detecting the dynamic content of semiconductor gaseous molecular contaminants according to claim 1, characterized in that: The construction of the pollutant diffusion topology map includes: Based on the spatial gradient change rate of the pollution concentration time series tensor, a density threshold segmentation and region growth algorithm is used to extract the high gradient diffusion region of pollutants and generate an initial topological node set. The node set is subjected to regional map construction based on Voronoi diagram partitioning to form spatial connection relationships between nodes; The connection relationship is input into a graph convolutional neural network, and the dynamic diffusion weight matrix between nodes is extracted by joint training with concentration gradient difference and time-varying weights. A boundary layer turbulence correction model is introduced, and the node transfer probability in the pollution diffusion path is adjusted based on the velocity shear tensor and the local Reynolds number to form a migration channel map.

4. The method for detecting the dynamic content of semiconductor gaseous molecular contaminants according to claim 1, characterized in that: The construction of the pollutant source-sink prediction model includes: On the pollutant diffusion topology map, a candidate set of pollutant source points is constructed based on the inverse gradient of node concentration, and high-confidence pollutant source nodes are screened by combining the maximum path entropy criterion. Simultaneously, forward diffusion simulation is performed on the graph structure based on the rate of change of pollution concentration, and the transmission path of pollution trend is predicted using a time window sliding convolution model; By performing path overlap analysis on the results of reverse tracing and forward diffusion, a pollution concentration change prediction curve is generated using the path consistency confidence function. Based on the response delay and diffusion directionality of the source and end-point concentration evolution, the coordinate region of pollutant release is inverted, and the location results of potential pollution sources are output.

5. The method for detecting the dynamic content of semiconductor gaseous molecular contaminants according to claim 1, characterized in that: The training of the pollution response optimizer includes: Construct a residual time series between the fused pollution trend curve and the measured concentration value, and define a composite loss function that includes a prediction offset penalty term and a local gradient consistency term; The residual sequence is input into a pollution response optimizer based on a variable weight attention mechanism to automatically adjust the spatial compensation factor within the prediction region. In the optimizer training process, multi-scale regularization terms and reconstruction confidence constraints are introduced to improve the expression accuracy of the prediction map in boundary regions and low-concentration regions. The final output is a high-resolution contamination prediction map covering key areas of the cleanroom, including continuous contaminant concentration fields, risk accumulation hotspots, and confidence level distribution information.

6. The method for detecting the dynamic content of semiconductor gaseous molecular contaminants according to claim 1, characterized in that: The extraction of the dynamic quantitative indicators of pollution includes: A fast Fourier transform is performed on the concentration change time series data in the pollution prediction graph to extract the main frequency component and calculate the high frequency energy ratio, and output the pollution event mutation frequency index. Based on the local concentration response curve of the wind speed field disturbance area in the prediction map, a probability distribution model is constructed and the information entropy value is obtained to form the airflow disturbance response entropy index. By comparing the changes in the main direction trajectory of the pollution diffusion path in multiple time slices, a path offset vector field is constructed, and its root mean square offset is calculated as the migration path offset factor.

7. The method for detecting the dynamic content of semiconductor gaseous molecular contaminants according to claim 1, characterized in that: The multi-source heterogeneous sensing data includes the real-time concentration of the target pollutant gas, environmental temperature and humidity parameters, ventilation airflow field status, and operating conditions of related process equipment.

8. A semiconductor gaseous molecular contaminant dynamic content detection system, used to implement the semiconductor gaseous molecular contaminant dynamic content detection method according to any one of claims 1-7, characterized in that: include: Feature acquisition module: acquires multi-source heterogeneous sensing data from multiple areas within the cleanroom, and generates a candidate feature tensor set for pollution sources through a multimodal temporal synchronization mechanism; Pollution modeling module: Inputs the feature tensor set into the pollution diffusion dynamic modeling network and outputs a spatiotemporal feature matrix containing the migration trend of pollutants; Topology graph construction module: Constructs a pollutant diffusion topology graph based on the rate of change of pollution concentration, uses a weighted graph convolutional neural network to extract dynamic diffusion weights between nodes, and combines a boundary layer turbulence model to correct the migration path; Source-sink prediction module: Constructs a pollutant source-sink prediction model, based on reverse tracing and forward diffusion mechanisms, and outputs pollutant concentration change prediction curves and potential pollution source location results; Response optimization module: Constructs a loss function from the residuals of the pollution trend curve and the measured data, and trains the pollution response optimizer to generate a high-resolution pollution prediction map; Pollution index analysis module: Based on the pollution prediction map, perform joint spectrum-information entropy analysis to extract pollution event mutation frequency, airflow disturbance response entropy, and migration path offset factor as dynamic quantitative indicators of pollution. Early warning and decision-making module: When any pollution quantification indicator exceeds the threshold, it outputs pollution source coordinates, cleanliness level adjustment suggestions, and HVAC system optimization exhaust strategy.

Images