Method and system for cross-physical-field generation of simulation workflows for process defect traceability

Abstract

The application relates to the technical field of industrial manufacturing process defect analysis and simulation, and particularly discloses a method and system for generating a simulation workflow of process defect traceability across physical fields. Key parameter boundaries are extracted from multi-physical field historical data to construct a low-dimensional boundary anchor point set; a three-layer progressive logic is used to evaluate process risks in real time and identify defect modes; mechanism-driven enhanced time sequence data is generated based on the identification results, a causally constrained generative adversarial network is trained, and a simulation workflow with embedded visual causal chains is formed; finally, the workflow is used to guide process parameter optimization, and new data is fed back to update the knowledge base, so that continuous enhancement optimization from defect tracing, simulation generation to autonomous optimization is realized.

Description

Technical Field

[0001] This invention relates to the field of industrial manufacturing process defect analysis and simulation technology, specifically to a method and system for generating simulation workflows across physical fields to trace the causes of process defects. Background Technology

[0002] In the high-end manufacturing sector, the prevention and control of process defects are crucial in determining product quality and cost. Currently, most mainstream defect analysis methods rely on post-inspection data statistics from a single physical field or isolated simulations based on idealized assumptions.

[0003] At the defect tracing level, traditional methods struggle to systematically decouple the coupling effects of multiple physics fields. Existing technologies typically analyze each physics field's data independently, lacking effective cross-field correlation and causal inference methods. Industry reports indicate that isolated analysis prevents the accurate identification of the root cause of nearly 40% of complex defects, leading to recurring problems and wasting significant R&D resources on trial-and-error troubleshooting. At the simulation modeling level, the widely used deterministic simulation models based on first principles or finite element methods, while physically clear, are highly dependent on expert experience and computationally expensive. More importantly, these models are typically used to simulate design conditions, not specifically for reverse engineering defect evolution paths. When faced with defects caused by nonlinear coupling of multiple small parameter deviations, traditional simulation models often deviate significantly from the actual defect evolution process due to slight inaccuracies in initial and boundary conditions. In highly complex processes, their accuracy for defect prediction is generally below 65%, limiting their direct application in high-value decision-making. At the process optimization and prevention level, current practices often exhibit open-loop characteristics. Optimization schemes are often derived from expert experience or limited experimental designs. Whether the adjusted new process parameters truly break the causal chain of defect generation lacks an efficient and automated verification and feedback mechanism. This leads to conservative and inefficient exploration of process windows. Furthermore, optimization results are difficult to directly translate into reusable, structured knowledge, failing to address new defect risks introduced by new materials and equipment, and hindering the continuous improvement of process robustness. Summary of the Invention

[0004] The technical problem to be solved by this invention is to overcome the shortcomings of existing technologies, such as difficulty in tracing the causes of process defects, disconnect between simulation and real cause and effect, and lack of feedback in the optimization process. This invention provides a method and system for generating simulation workflows across physical fields to trace the causes of process defects. It extracts key physical field parameters from historical defect data and generates high-fidelity defect evolution simulation sequences through intelligent judgment and adversarial training, thereby realizing automatic identification of defect causes, path simulation and process optimization.

[0005] The specific technical solution of this application is as follows:

[0006] According to one aspect of this application, a method for generating a simulation workflow across physics fields to trace the causes of process defects is provided, comprising:

[0007] Starting from historically actual process defects, we collect multi-physics field sensing data corresponding to the moment of occurrence. Through least squares fitting and sensitivity analysis, we extract the critical physical field parameter combination that induces the defect, forming a low-dimensional boundary anchor point set oriented towards the cause of the defect.

[0008] The parameters from the low-dimensional boundary anchor point set are input into the improved AdaBoost.M2 classifier, and progressive judgment is performed. Specifically, it is determined whether the parameters approximate any defect anchor point boundary; if they approximate, it is determined whether other field anomalies are excited through physical field coupling; if they are excited, it is compared with the historical defect path library to determine whether the coupling response matches the known defect evolution pattern; if all three layers are triggered, a high-risk cross-field boundary combination and causal path map are output; if not all three layers are triggered, it is marked as a potential novel defect and manual review is triggered.

[0009] Based on high-risk cross-field boundary combinations and causal path maps, nonlinear stretching is performed along the defect evolution time axis in the physical parameter space to generate enhanced time-series data containing the complete defect evolution process; conservative neighborhood sampling is adopted for potential novel paths; all generated data retain causal labels and time-series markers between physical fields, forming a defect mechanism-driven training set.

[0010] Using the defect mechanism training set as input, the adversarial network training is initiated. The generator of the adversarial network outputs a simulation sequence. The decision of the adversarial network determines whether the sequence completely reproduces the causal chain of defect triggering, inter-field transmission and final manifestation. Through adversarial training between the generator and the decision until Nash equilibrium is reached, a simulation workflow with an embedded visual causal chain is output.

[0011] The simulation workflow is input into the intelligent process parameter tuning module, which generates process adjustment instructions based on the key control points of the causal chain. The adjusted new operating condition data is then used to update the low-dimensional boundary anchor point set. If the defect is eliminated, a new safe interval is recorded. If the defect morphology changes, a new causal path is added and the process returns to step two for re-judgment, thus forming a continuous self-reinforcing defect prevention and optimization.

[0012] As a further option of the method of the present invention, the critical physical field parameter combination extraction step includes:

[0013] Based on multi-physics sensing data from historical defect cases, the first-order Sobol exponent of each process parameter on the defect index is calculated through global sensitivity analysis, and key parameters whose contribution exceeds the preset threshold are selected.

[0014] In the key parameter space, through local fitting and gradient analysis, the critical point that causes the defect index to exceed the threshold is found, forming a low-dimensional boundary anchor point set. Metadata is attached to each anchor point, including the associated defect type, gradient direction, and causal path seed.

[0015] As a further option of the method of the present invention, the improved AdaBoost.M2 classifier performs the first layer of judgment in the progressive judgment, which includes:

[0016] Calculate the Mahalanobis distance from the real-time parameter status to each boundary anchor point. If the minimum Mahalanobis distance is less than the preset threshold, it is determined to be approaching the boundary and the second layer of judgment is triggered.

[0017] The formula for calculating Mahalanobis distance is: ;in, It is the real-time parameter status. To the Anchor points Mahalanobis distance; It is a real-time monitored or input process parameter status vector. It is the first A boundary anchor vector, Key process parameters The historical covariance matrix, is the transpose of the vector, and -1 is the inverse of the historical covariance matrix.

[0018] As a further option of the method of the present invention, the second layer of judgment includes:

[0019] Based on a simplified physical field coupling model, predict the change vector of other physical field observations under the current parameter perturbation;

[0020] If the predicted change exceeds the normal fluctuation range, it is determined to trigger a cross-field anomaly and trigger the third layer of judgment.

[0021] The third-level judgment includes:

[0022] The current evolutionary segment is dynamically time-warped and matched with the patterns in the historical defect path library. If the matching degree is higher than the threshold, it is determined to be a known defect pattern, and the corresponding causal path map is output.

[0023] As a further option of the method of the present invention, the generation of enhanced time-series data containing the complete defect evolution process includes:

[0024] Based on the causal path graph, the time axis of physical parameters is nonlinearly stretched to achieve higher time resolution near key events.

[0025] Using the boundary anchor point as the endpoint, the starting point of the evolution path is determined in reverse, generating the parameter evolution trajectory from the starting point to the endpoint;

[0026] The trajectory is used as a time-varying boundary condition input to the forward physical simulation model to generate a multiphysics spatiotemporal evolution sequence, and random perturbations are introduced to enhance data diversity.

[0027] As a further option of the method of the present invention, conservative neighborhood sampling is used for potential novel defect paths, including:

[0028] Sampling is performed within a neighborhood where the potential defect parameter point is centered and its Mahalanobis distance is less than a preset radius;

[0029] Run short-term simulations for each sampling point to observe whether any anomalies are triggered, and save the short sequences along with the parameters as a preliminary exploratory dataset for novel defects.

[0030] As a further option of the method of the present invention, the adversarial network is a conditional generative adversarial network with causal constraints, and its loss function includes adversarial loss and causal consistency loss: ;in, It is the total loss during the training of the adversarial generative network. It is to combat losses. It is a loss of causal consistency. It is used to balance the weights between adversarial loss and causal loss.

[0031] As a further option of the method of the present invention, the training of the adversarial network adopts a phased strategy:

[0032] The first stage involves independently training a causal checker network to enable it to identify key events and times from sequences.

[0033] The second stage fixes the parameters of the causal checker and embeds them into the cGAN framework, guiding the generator to optimize both data distribution and causal logic during adversarial training.

[0034] As a further option of the method of the present invention, the intelligent process parameter tuning module uses Bayesian optimization method to find the parameter combination that minimizes the defect index in the simulation workflow based on the key control points in the causal chain.

[0035] The optimized new operating condition data is fed back to the system knowledge base: if the defect is eliminated, it is recorded as a safe process interval; if the defect morphology changes, the new evolution path is extracted and the historical defect path library is updated to achieve continuous self-enhancement of the system.

[0036] Another aspect of this application provides a system for generating a cross-physics simulation workflow for tracing the causes of process defects, the system comprising:

[0037] The data extraction and anchor point generation module is used to collect multi-physics field sensing data corresponding to the time of occurrence of real process defects in history. Through least squares fitting and sensitivity analysis, it extracts the critical physical field parameter combination that induces the defect and forms a low-dimensional boundary anchor point set oriented towards the cause of the defect.

[0038] The intelligent classification and pattern recognition module, including the improved AdaBoost.M2 classification engine, receives parameters from a low-dimensional boundary anchor point set and performs progressive judgments: first, it determines whether the parameters approximate any defect anchor point boundary; if they do, it determines whether physical field coupling has triggered anomalies in other fields; if so, it compares the historical defect path library to determine whether the coupling response matches known defect evolution patterns; if all three layers are triggered, it outputs high-risk cross-field boundary combinations and causal path maps; if not all three layers are triggered, it is marked as a potential novel defect and triggers manual review.

[0039] The enhanced time-series data generation module is used to perform nonlinear stretching along the defect evolution time axis in the physical parameter space based on high-risk cross-field boundary combinations and causal path maps, generating enhanced time-series data containing the complete defect evolution process; conservative neighborhood sampling is used for potential novel paths; all generated data retain causal labels and temporal markers between physical fields, forming a defect mechanism-driven training set;

[0040] The simulation workflow training and generation module includes a causal constraint adversarial generative network, whose input is a defect mechanism-driven training set. The adversarial generative network includes a generator and a decision unit. The generator is used to output a simulation sequence, and the decision unit is used to determine whether the sequence completely reproduces the causal chain of defect triggering, inter-field transmission and final manifestation. Through adversarial training until Nash equilibrium is reached, a simulation workflow with an embedded visual causal chain is output.

[0041] The process optimization and link update module is used to input the simulation workflow into the intelligent tuning sub-module of process parameters and generate process adjustment instructions based on the key control points of the causal chain. The adjusted new operating condition data is fed back to the data extraction and anchor point generation module to update the low-dimensional boundary anchor point set: if the defect is eliminated, the new safe interval is recorded; if the defect morphology changes, a new causal path is added and the intelligent classification and pattern recognition module is triggered to re-judge, forming a continuous self-reinforcing defect prevention and optimization.

[0042] The beneficial effects of this application are as follows:

[0043] This method significantly improves the accuracy and efficiency of tracing the origins of complex process defects. Through multi-physics coupled sensitivity backpropagation and three-layer progressive causal matching, the system can pinpoint key parameter combinations that induce defects from historical data, forming a low-dimensional boundary anchor set. This overcomes the bottleneck of traditional single-physics analysis, which fails to accurately determine the cause of nearly 40% of defects. Enhanced data generation and simulation training based on causal paths make the reproduction and prediction of defect evolution more accurate, potentially raising the defect prediction accuracy of traditional simulation models from below 65% to a higher level.

[0044] This method establishes a complete chain from defect analysis to process optimization, significantly improving process robustness and knowledge reuse capabilities. By embedding a visualized causal chain within a causal constraint adversarial generative network, the simulation workflow can not only reverse-engineer defects but also provide forward guidance for parameter tuning and offer automated verification feedback. The system can update the knowledge base with optimized safety process parameters, forming a continuous self-reinforcing mechanism. This effectively addresses the risks of novel defects introduced by new materials and equipment, reduces trial-and-error costs, and improves the efficiency of process window exploration. Attached Figure Description

[0045] Figure 1 A schematic diagram of the overall method for generating simulation workflows across physics fields to trace the causes of process defects;

[0046] Figure 2 S100 flowchart of a method for generating simulation workflows across physics fields to trace the causes of process defects;

[0047] Figure 3 S200 flowchart of a method for generating simulation workflows across physics fields to trace the causes of process defects;

[0048] Figure 4 S300 flowchart of a method for generating simulation workflows across physics fields to trace the causes of process defects;

[0049] Figure 5 S400 flowchart of a method for generating simulation workflows across physics fields to trace the causes of process defects;

[0050] Figure 6 The S500 flowchart illustrates a method for generating simulation workflows across physical fields to trace the causes of process defects. Detailed Implementation

[0051] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0052] The core of this invention lies in constructing an intelligent system that reverse-engineers the physical causes from historical defect data, generates high-fidelity simulation sequences, and forms a process optimization chain. Its theoretical foundation is built upon multi-physics coupling sensitivity inverse reasoning theory, progressive causal path matching theory, and adversarial generative network theory with causal constraints. By extracting low-dimensional key parameter boundaries from high-dimensional sensor data, using three-layer progressive logic to identify defect evolution patterns, and driving the generation of a simulation workflow with embedded physical causal chains, the system ultimately achieves autonomous optimization and enhancement of the process.

[0053] For those involved The process of a physical field, the moment when defects occur. The sensor data constitutes a high-dimensional observation vector. The goal is to find the low-dimensional key physical parameter vector that induces defects. Each parameter was quantified through global sensitivity analysis. Defect indicators The contribution of the first-order Sobol exponent. The calculation formula is: ;in, It is a parameter The first-order Sobol exponent, Except A vector consisting of all other process parameters besides It is a defect indicator. In the parameters Under these conditions, defect indicators Compared to all other parameters Conditional expectation, For parameters Find the variance. It is a defect indicator The total variance.

[0054] Filter out The parameters constitute a candidate set. Then, within the candidate parameter space, through fitting and gradient analysis, the parameters that make... Exceeding the defect threshold critical point set That is, the low-dimensional boundary anchor set .

[0055] Given real-time parameter status The three-level judgment logic is as follows:

[0056] First layer, boundary approximation: calculate the Mahalanobis distance to the nearest anchor point. : ;in, It is the real-time parameter status. To the Anchor points Mahalanobis distance; It is a real-time monitored or input process parameter status vector. It is the first A boundary anchor vector, Key process parameters The historical covariance matrix, is the transpose of the vector, and -1 is the inverse of the historical covariance matrix.

[0057] like If so, the next layer will be triggered.

[0058] The second layer, cross-field excitation: based on a simplified coupling model. Determine the disturbance Does it exceed the normal fluctuation range? If it does, the next layer is triggered. Among them, It is the vector of changes in other physical field observations predicted. It is a simplified physical field coupling model. It is a typical process parameter vector under normal and defect-free conditions.

[0059] The third layer, pattern matching: matching the current evolutionary segment. With historical defect path library The patterns in the data are dynamically time-warped for matching. If the matching degree is higher than the threshold, it is determined to be a known defect pattern.

[0060] Construct a Conditional Generative Adversarial Network (cGAN). Generator With noise and path tags Time series data is generated from the input. Discriminator Determine the authenticity and causal consistency of the data. The loss function includes adversarial loss. and causal loss : ;in, It is the total loss during the training of the adversarial generative network. It is to combat losses. It is a loss of causal consistency. It is used to balance the weights between adversarial loss and causal loss.

[0061] Forced generation of sequences to follow paths Defined causal timing rules.

[0062] Based on the above theory, the specific implementation methods of the present invention will be described in detail below.

[0063] Example 1:

[0064] Please see Figure 1 This illustrates a method for generating a simulation workflow across physics fields to trace the causes of process defects, provided by an embodiment of the present invention. The method includes:

[0065] S100: Based on historical defect multiphysics data, extract the critical physical field parameter combination that induces defects to form a low-dimensional boundary anchor point set.

[0066] S200: Input the parameters to be analyzed into the improved classifier, perform progressive three-layer judgment, and output high-risk cross-field boundary combinations and causal path maps, or mark potential new defects.

[0067] S300: Based on high-risk combinations and path maps, it generates enhanced time-series data containing the complete defect evolution process in the physical parameter space, forming a defect mechanism-driven training set.

[0068] S400: Using a defect mechanism training set as input, trains a causal-constrained adversarial generative network until it reaches equilibrium, and then outputs a simulation workflow with an embedded visual causal chain.

[0069] S500: Uses simulation workflow to guide intelligent optimization of process parameters and feeds back new operating condition data to S100 to update the boundary anchor set and causal path, forming a continuously self-reinforcing link.

[0070] The specific plan is as follows:

[0071] In a cross-physics field generation simulation workflow for tracing the causes of process defects, S100 condenses high-dimensional, redundant sensor information into low-dimensional, key physical parameter boundary conditions by mining historical defect data.

[0072] Please refer to Figure 2 It shows a flowchart of an exemplary S100 step of this application, the contents of which include:

[0073] S110: In this embodiment, historically confirmed process defect cases are extracted from the manufacturing execution system and sensor network. For each defect case, the time when the defect was discovered is collected. All relevant physical field sensing data within the preceding and following time windows.

[0074] In this embodiment, the physical fields include temperature fields, pressure fields, flow velocity fields, stress fields, electromagnetic fields, etc. Data sources include various sensors distributed in different spatial locations.

[0075] In one possible implementation of this embodiment, data from all sensors are collected through a unified time-series data platform, using a network time protocol or hardware triggering to ensure strict timestamp alignment. Data preprocessing includes outlier removal, missing value imputation, and noise reduction filtering. The preprocessed multiphysics data constitutes a spatiotemporal data cube.

[0076] S120: In this embodiment, a complete set of adjustable process parameters corresponding to the defect cases is extracted from the process parameter library. The goal is to... From the parameters selected, those that have a significant impact on defect occurrence are identified. Key parameters.

[0077] In one possible implementation of this embodiment, the Sobol index method is used. Each process defect case is considered as a sample, with process parameters as input and a defect severity index as output. By calculating each parameter on all historical samples. First-order Sobol exponent Quantify its individual pair The contribution ratio of variance. The calculation formula is as described above.

[0078] Set contribution threshold To satisfy all The process parameters were incorporated into the preliminary key parameter set. At the same time, record the value range of each key parameter and its typical value distribution when the defect occurs.

[0079] S130: In this embodiment, in the initial key parameter set Composition In 3D space, the parameter points corresponding to historical defect cases constitute a point cloud. This step aims to find the critical boundary surrounding this defect point cloud.

[0080] In one possible implementation of this embodiment, support vector machines or logistic regression are used. A separating hypersurface is fitted in space to initially distinguish between defects and normal operating conditions. To obtain more physically meaningful boundary points, a detailed analysis is performed in the region near the separating hypersurface. For points near the boundary, defect indices are calculated through local linear fitting or by querying a high-fidelity physical simulation model. gradient at points near the boundary The gradient direction indicates The most sensitive direction of change. Search along the gradient direction to find the direction that makes... Just exceeding the defect threshold critical point .

[0081] In one possible implementation of this embodiment, a set of critical boundary points is obtained by approximating the defect point cloud from different directions. Each boundary point corresponds to a specific set of critical physical field parameters that can induce defects. All boundary points constitute a low-dimensional set of boundary anchor points. Simultaneously, typical multiphysics evolution sequence fragments that cause defects near each anchor point are extracted and used as the initial causal path seed corresponding to that anchor point, and stored in the path library seed area.

[0082] S140: In this embodiment, for each boundary anchor point Additional metadata, including the corresponding defect type, associated historical cases, and local gradient direction. And the associated causal path seed ID.

[0083] In one possible implementation of this embodiment, key parameters are based on all historical defect cases. The values ​​on are used to calculate key parameters. Empirical covariance matrix .

[0084] In a cross-physics field generation simulation workflow for tracing the causes of process defects, S200 builds an intelligent classification and judgment system with three-layer progressive logic based on the low-dimensional boundary anchor point set and path seeds generated by S100. This system is used to assess the risk of process status in real time or offline and identify defect evolution patterns.

[0085] Please refer to Figure 3 It shows a flowchart of an exemplary S200 step of this application, the contents of which include:

[0086] S210: In this embodiment, the AdaBoost.M2 algorithm is improved to handle classification tasks with logical dependencies, such as progressive judgments. The classifier's output is structured information containing the judgment results and the confidence scores of each layer.

[0087] In one possible implementation of this embodiment, a set of base classifiers is trained for each layer of decision logic, and then integrated by the AdaBoost.M2 framework.

[0088] The training objective of the first-layer base classifier is to determine the input parameters. The first layer determines whether the parameters approach any boundary anchor points. The second layer base classifier is trained to determine whether cross-physics anomalies are triggered given that the parameters approach the boundary. The third layer base classifier is trained to determine whether the evolution pattern matches a known defect path given that anomalies are triggered. The AdaBoost.M2 algorithm iteratively adjusts the sample weights and classifier weights to ultimately obtain a strong classifier for each layer.

[0089] S220: In this embodiment, when a new combination of process parameters... Upon entering the system, the first layer of strong classifier processes the data and calculates its approximation score.

[0090] In one possible implementation of this embodiment, the system performs parallel computing. to anchor set Mahalanobis distance of all anchor points The calculation formula is as described above.

[0091] If the approximation score is greater than the threshold and the minimum Mahalanobis distance is... If the condition is met, it is considered a close approximation, and the ID of the nearest anchor point and its distance value are recorded. Simultaneously, the associated path seed for that anchor point is retrieved from the metadata. If the condition is not met, it is considered safe.

[0092] S230: In this embodiment, when the first layer determines that the approximation is successful, the second layer is automatically triggered. The system calls the second-layer strong classifier, with the input being... And coupled response characteristics calculated in real time from a simplified physical model or queried from a cache.

[0093] In one possible implementation of this embodiment, the simplified physical model is a regression model trained on historical data, used to quickly predict changes in other indirectly controlled physical field observations given parameters. The system will... The predicted change is input together with the second-layer strong classifier.

[0094] If the output excitation confidence is greater than the threshold If this occurs, it is determined to be an excitation. Simultaneously, the system initiates high-frequency data acquisition or invokes high-precision simulation to obtain short-time data of actual or simulated multiphysics fields. This prepares for the third level of judgment.

[0095] S240: In this embodiment, when the second layer determines that an excitation has occurred, the third layer determination is triggered. The system will process short-sequence data. Alternatively, its feature vector can be input into the third-layer strong classifier.

[0096] In one possible implementation of this embodiment, the output of the third-layer strong classifier is the probability distribution over all known defect path patterns. If a certain known pattern exists... probability Greater than the threshold If the pattern is significantly higher than other patterns, it is determined to match a known pattern. The system retrieves data from the historical defect path database. Retrieve pattern The complete causal path map.

[0097] If the probability of belonging to a novel pattern is the highest, or the probabilities of all known patterns are too low, it is determined to be a non-match with a known pattern, i.e., a potential novel defect. The system packages all relevant information of the current state and marks it as a case to be reviewed.

[0098] S250: In this embodiment, a structured output is generated based on the results of the three-layer judgment. If all three layers are triggered and a known pattern is matched, a high-risk status is output, including the high-risk parameter combination, the matched defect path pattern ID, the corresponding complete causal path graph, and the confidence level of each layer. Based on other triggering conditions, corresponding level of warning or alarm information is output.

[0099] In a cross-physics generation simulation workflow for tracing the causes of process defects, S300 uses the high-risk combinations and causal path maps identified by S200 as blueprints to synthesize a large amount of time-series data containing the complete evolution process of defects in the physical parameter space, providing rich, mechanism-driven training samples for subsequent training of high-fidelity simulators.

[0100] Please refer to Figure 4 It illustrates a flowchart of an exemplary S300 step of this application, the contents of which include:

[0101] S310: In this embodiment, the causal path graph This describes the sequence of key events and their temporal relationships from defect triggering to manifestation. The goal is to anchor static critical parameters. This expands to a parameter trajectory that dynamically changes over time and ultimately leads to defects. .

[0102] In one possible implementation of this embodiment, for the matched defect pattern and its anchor points First, determine the total time span of the evolution. In the parameter space, with Using the endpoint as the starting point, the safe starting point for determining the evolutionary path is determined by moving in reverse along the sensitive direction. .

[0103] Define a parameter evolution path from the starting point to the ending point. According to the map The event nodes are non-linearly stretched to achieve higher temporal resolution near critical events. This is achieved through a monotonically increasing time scaling function. This allows the time scaling function to have a larger slope near the event point.

[0104] S320: In this embodiment, the parameter evolution path is utilized. By combining forward physics simulation models or high-precision surrogate models, corresponding multiphysics spatiotemporal evolution sequences are generated. .

[0105] In one possible implementation of this embodiment, the forward physics simulation model can be a finite element analysis model or a computational fluid dynamics model. The simulation model is used as a time-varying boundary condition input, and the simulation is run to obtain multiphysics field data for the entire process.

[0106] To increase data diversity, in Reasonable random perturbations are introduced, such as superimposed smooth Gaussian process noise, or random fluctuations of minor factors are introduced into the simulation. By sampling different random seeds multiple times, a large number of different defect evolution sequences are generated from the anchor point and path map.

[0107] S330: In this embodiment, for cases marked as potential novel defects, a conservative exploration strategy is adopted because their evolution mode is unknown.

[0108] In one possible implementation of this embodiment, the parameter points are taken as an example of a potential novel defect. Centered on, its Mahalanobis distance is less than the smaller radius Sampling is performed within the neighborhood of the sampled point to obtain a series of parameter points. For each sampled point, a short simulation or analysis is run to observe whether it triggers anomalies. The short sequence is saved together with the corresponding sampling parameters as a preliminary exploratory dataset for novel defects.

[0109] S340: In this embodiment, rich tags are added to each generated time series data, including the defect pattern ID it follows, the time point and type of the key event, the start and end time of the sequence, the start and end anchor point IDs of the parameter evolution path used, and the physical quantity and location information corresponding to each data channel.

[0110] All labeled sequence data together constitute the defect mechanism-driven training set. Defect mechanism-driven training set The characteristic is that each sample is associated with a clear physical cause and evolutionary logic.

[0111] In a cross-physics field generation simulation workflow for tracing the causes of process defects, the S400 uses the mechanism-driven training set built by the S300 as fuel to train an adversarial generative network that can internalize the causal law of defect evolution. Its generator eventually becomes a plug-and-play simulation workflow that can output a visualized causal chain.

[0112] Please refer to Figure 5 It illustrates a flowchart of an exemplary S400 step of this application, the contents of which include:

[0113] S410: In this embodiment, a generator is designed. and discriminator The neural network structure. Generator With random noise vector and path tags As input, output multiphysics time series data Discriminator Input time series data and condition vector, output a probability scalar that represents the data as true and consistent with causal laws.

[0114] Loss functions include adversarial loss and loss of causal consistency The total loss is: ;

[0115] Adversarial loss can be implemented using Wasserstein GAN with Gradient Penalty (WGAN-GP) to improve training stability. Causal loss. Through a pre-trained causal checker network To achieve this, the adversarial generative network predicts the timing of key events from a sequence, with the loss defined as the error between the predicted timing and the time specified in the graph.

[0116] S420: In this embodiment, to efficiently and stably train the causal constraint adversarial generative network, a phased and modular training strategy is adopted to ensure that the generator simultaneously grasps the data distribution patterns and physical causal logic. First, the causal checker component is trained independently to enable it to have accurate event recognition capabilities; then, the parameters of the causal checker component are fixed, and the causal checker component is embedded into the adversarial training framework to guide the generator's learning direction.

[0117] In one possible implementation of this embodiment, the first stage uses a mechanism-driven training set constructed by S300 with precise event timestamp annotations to specifically train the causal checker network. Causality checker network The algorithm employs a structure combining one-dimensional temporal convolution with an attention mechanism. The input is a multiphysics sequence, and the output is the probability distribution of key events in the sequence and their temporal regression values. The training objective is to minimize the sum of the event classification cross-entropy loss and the temporal regression mean squared error loss. The second stage involves starting the generator. With discriminator Joint adversarial training is performed. During each generator parameter update, not only is the adversarial loss from the discriminator calculated, but also a causal checker with frozen parameters is generated from the input sequence. The causal consistency loss is calculated, and the two are weighted and backpropagated together to update the generator. The discriminator is trained to focus on distinguishing the authenticity of real sequences from generated sequences and evaluating their matching degree with condition labels.

[0118] S430: In this embodiment, the generator and discriminator are optimized alternately. During training, the discriminator loss, generator loss, causal consistency of the generated sequence, and physical plausibility are monitored. When the discriminator struggles to distinguish between true and false sequences, and the causal consistency of the generated sequence reaches a predetermined threshold, training is considered to have reached a practical equilibrium point, and training is stopped.

[0119] In one possible implementation of this embodiment, the system calculates and records the following core metrics after each training cycle:

[0120] The mean and distribution of the discriminator's output values ​​for real and generated samples;

[0121] The average causal consistency score obtained after the generated sequence is evaluated by the causality checker;

[0122] The pass rate of the generated sequence through a lightweight physical rule checker.

[0123] When the discriminator's accuracy in distinguishing between real and fake samples approaches 50%, the causal consistency score of the generated sequence remains stable above the predetermined threshold for several consecutive cycles, and the physical rule verification pass rate exceeds 98%, the system has reached a practical Nash equilibrium. At this point, the generator model weights should be saved immediately and training should be stopped to prevent overfitting or performance regression.

[0124] If the metrics fluctuate violently or continue to deteriorate during training, stabilization operations such as learning rate adjustment, gradient pruning, or reinitialization of some network layers will be automatically triggered.

[0125] S440: In this embodiment, the trained generator It is encapsulated as a callable simulation workflow module. The simulation workflow module takes the target defect mode and initial conditions as input and outputs simulated multiphysics time-series data.

[0126] In one possible implementation of this embodiment, the simulation workflow integrates visualization functionality. While generating the sequence, a causal checker or rule parser is invoked to analyze causal events in the sequence online, and these causal events are displayed synchronously with the changes in multiphysics data in the form of graphs, timeline markers, or animated highlights, forming an embedded visualized causal chain.

[0127] In a cross-physics simulation workflow for tracing the causes of process defects, the S500 transforms the simulation workflow generated by the S400 into actual productivity, which is used to guide the online optimization or offline design of process parameters. By feeding the optimization results back to the system knowledge base, a self-learning and continuously enhancing link is formed.

[0128] Please refer to Figure 6It illustrates a flowchart of an exemplary S500 step of this application, the contents of which include:

[0129] S510: Construction of the intelligent process parameter optimization module.

[0130] In this embodiment, the intelligent process parameter optimization module uses a simulated workflow as its core model and aims to eliminate or avoid specific defects to optimize process parameters. The intelligent process parameter optimization module receives high-risk alarms and causal path maps from S200 as input.

[0131] In one possible implementation of this embodiment, the module first analyzes the causal graph to identify key control points in the defective causal chain. These control points are typically early events or parameters with the strongest coupling that have a decisive impact on subsequent evolution.

[0132] S520: In this embodiment, the optimization module will adjust the current process parameters. Starting from this point, and using the simulation workflow as a constraint, we run optimization algorithms to find new parameters. This enables the simulation workflow to be in When running, defect metrics are minimized or fall below acceptable thresholds.

[0133] In one possible implementation of this embodiment, a Bayesian optimization method is employed. The simulation workflow is treated as a black-box function. Constructing an approximate black-box function for the proxy model The system intelligently selects evaluation points using a data acquisition function, finding the optimal solution within a limited number of simulations. The optimization process must simultaneously satisfy process constraints, constituting a constrained optimization problem. Upon convergence, a set of recommended process parameter adjustment instructions is output.

[0134] S530: In this embodiment, the optimized parameters It is applied to actual production or for verification in high-fidelity digital twins. During the verification operation, multi-physics sensor data is intensively acquired.

[0135] If the defect is successfully eliminated, the optimization is recorded as a success story. The system automatically extracts process parameters during the stable operation period. This will be used as a new safety point to update and expand the knowledge of the safety process zone in S100.

[0136] S540: In this embodiment, if the defect morphology changes after applying new parameters, it means that a new evolutionary path may have emerged. The system treats this run as a new defect evolution case, collects data, and executes a simplified S100 process to extract new critical features. The new evolutionary sequence is compared with the original map to form a new causal path or supplement the original map. This new path is input into the historical defect path library as a path to be verified.

[0137] The system can automatically schedule the verification of new paths. Once a new path is confirmed, it will be formally incorporated into the system's knowledge base. The classifier in S200 can be incrementally trained periodically using the updated path library, thereby gaining the ability to identify new defect patterns. S300 can also generate training data based on new paths, enhancing the simulation capabilities of the simulation workflow.

[0138] Through the S510 to S540 cycle, the entire system realizes a complete link from defect identification, cause analysis, simulation, optimization feedback to knowledge evolution, thereby achieving continuous self-reinforcement.

[0139] Example 2:

[0140] A system for cross-physics-based simulation workflow for tracing the causes of process defects includes:

[0141] The data extraction and anchor point generation module is used to collect multi-physics field sensing data corresponding to the time of occurrence of real process defects in history. Through least squares fitting and sensitivity analysis, it extracts the critical physical field parameter combination that induces the defect and forms a low-dimensional boundary anchor point set oriented towards the cause of the defect.

[0142] The intelligent classification and pattern recognition module, including the improved AdaBoost.M2 classification engine, receives parameters from a low-dimensional boundary anchor point set and performs progressive judgments: first, it determines whether the parameters approximate any defect anchor point boundary; if they do, it determines whether physical field coupling has triggered anomalies in other fields; if so, it compares the historical defect path library to determine whether the coupling response matches known defect evolution patterns; if all three layers are triggered, it outputs high-risk cross-field boundary combinations and causal path maps; if not all three layers are triggered, it is marked as a potential novel defect and triggers manual review.

[0143] The enhanced time-series data generation module is used to perform nonlinear stretching along the defect evolution time axis in the physical parameter space based on high-risk cross-field boundary combinations and causal path maps, generating enhanced time-series data containing the complete defect evolution process; conservative neighborhood sampling is used for potential novel paths; all generated data retain causal labels and temporal markers between physical fields, forming a defect mechanism-driven training set;

[0144] The simulation workflow training and generation module includes a causal constraint adversarial generative network, whose input is a defect mechanism-driven training set. The adversarial generative network includes a generator and a decision unit. The generator is used to output a simulation sequence, and the decision unit is used to determine whether the sequence completely reproduces the causal chain of defect triggering, inter-field transmission and final manifestation. Through adversarial training until Nash equilibrium is reached, a simulation workflow with an embedded visual causal chain is output.

[0145] The process optimization and link update module is used to input the simulation workflow into the intelligent tuning sub-module of process parameters and generate process adjustment instructions based on the key control points of the causal chain. The adjusted new operating condition data is fed back to the data extraction and anchor point generation module to update the low-dimensional boundary anchor point set: if the defect is eliminated, the new safe interval is recorded; if the defect morphology changes, a new causal path is added and the intelligent classification and pattern recognition module is triggered to re-judge, forming a continuous self-reinforcing defect prevention and optimization.

[0146] Those skilled in the art will understand that the embodiments of this application are provided as methods, systems, or computer program products. Therefore, this application takes the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application takes the form of a computer program product implemented on one or more computer storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer program code. The solutions in the embodiments of this application are implemented using various computer languages, exemplified by the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0147] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, are implemented by computer program instructions. These computer program instructions are provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart illustrations and / or block diagrams.

[0148] These computer program instructions are also stored in a computer read-memory that can direct a computer or other programmed data processing device to operate in a particular manner, such that the instructions stored in the computer read-memory produce an article of manufacture including instruction means that implement the functions specified in the flowchart or multiple flowcharts and / or block diagram blocks or multiple block diagrams.

[0149] These computer program instructions are also loaded onto a computer or other programming data processing device to cause a series of operational steps to be performed on the computer or other programming device to produce a computer-implemented process, such that the instructions, which execute on the computer or other programming device, provide steps for implementing the functions specified in the flowchart flow or multiple flows and / or the block diagram blocks or multiple blocks.

[0150] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0151] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for cross-physics field generation simulation workflow for tracing the causes of process defects, characterized in that, include: Starting from real-world process defects that occurred in history, we collect multi-physics field sensing data corresponding to the time of occurrence. Through least squares fitting and sensitivity analysis, we extract the critical physical field parameter combination that induces the defect, forming a low-dimensional boundary anchor point set oriented towards the cause of the defect. The parameters from the low-dimensional boundary anchor point set are input into the improved AdaBoost.M2 classifier, and progressive judgment is performed. Specifically, it is determined whether the parameters approximate any defect anchor point boundary; if they approximate, it is determined whether other field anomalies are excited through physical field coupling; if they are excited, it is compared with the historical defect path library to determine whether the coupling response matches the known defect evolution pattern; if all three layers are triggered, a high-risk cross-field boundary combination and causal path map are output; if not all three layers are triggered, it is marked as a potential novel defect and manual review is triggered. The improved AdaBoost.M2 classifier is as follows: For each layer of decision logic, a set of base classifiers is trained and then integrated using the AdaBoost.M2 framework; the training objective of the first layer of base classifiers is to determine the input parameters. Whether it approaches any boundary anchor point; the training goal of the second-layer base classifier is to determine whether a cross-physics anomaly is triggered given that the parameters approach the boundary; the training goal of the third-layer base classifier is to determine whether its evolution pattern matches a known defect path given that an anomaly is triggered; the AdaBoost.M2 algorithm iteratively adjusts the sample weights and classifier weights to ultimately obtain a strong classifier for each layer; Based on high-risk cross-field boundary combinations and causal path maps, nonlinear stretching is performed along the defect evolution time axis in the physical parameter space to generate enhanced time-series data containing the complete defect evolution process; conservative neighborhood sampling is adopted for potential novel paths; all generated data retain the causal labels and time-series markers between physical fields, forming a defect mechanism-driven training set; Using the defect mechanism training set as input, the adversarial network training is started. The generator of the adversarial network outputs the simulation sequence. The decision of the adversarial network determines whether the sequence completely reproduces the causal chain of defect triggering, inter-field transmission and final manifestation. Through adversarial training of the generator and decision until Nash equilibrium is reached, the simulation workflow with embedded visualization of the causal chain is output. The simulation workflow is input into the intelligent process parameter tuning module, which generates process adjustment instructions based on the key control points of the causal chain. The adjusted new operating condition data is then used to update the low-dimensional boundary anchor point set. If the defect is eliminated, a new safe interval is recorded. If the defect morphology changes, a new causal path is added and the progressive judgment is returned for re-judgment, forming a continuous self-reinforcing defect prevention and optimization.

2. The method for cross-physics field generation simulation workflow for tracing the causes of process defects according to claim 1, characterized in that, The critical physical field parameter combination extraction step includes: Based on multi-physics sensing data from historical defect cases, the first-order Sobol exponent of each process parameter on the defect index is calculated through global sensitivity analysis, and key parameters whose contribution exceeds the preset threshold are selected. In the key parameter space, through local fitting and gradient analysis, the critical point that causes the defect index to exceed the threshold is found, forming a low-dimensional boundary anchor point set. Metadata is attached to each anchor point, including the associated defect type, gradient direction, and causal path seed.

3. The method for cross-physics field generation simulation workflow for tracing the causes of process defects according to claim 1, characterized in that, The improved AdaBoost.M2 classifier performs the first layer of judgment in the progressive judgment process, which includes: Calculate the Mahalanobis distance from the real-time parameter status to each boundary anchor point. If the minimum Mahalanobis distance is less than the preset threshold, it is determined to be approaching the boundary and the second layer of judgment is triggered. The formula for calculating Mahalanobis distance is: ;in, It is the real-time parameter status. To the Anchor points Mahalanobis distance; It is a real-time monitored or input process parameter status vector. It is the first A boundary anchor vector, Key process parameters The historical covariance matrix, is the transpose of the vector, and -1 is the inverse of the historical covariance matrix.

4. The method for cross-physics field generation simulation workflow for tracing the causes of process defects according to claim 3, characterized in that, The second layer of judgment includes: Based on a simplified physical field coupling model, predict the change vector of other physical field observations under the current parameter perturbation; If the predicted change exceeds the normal fluctuation range, it is determined to trigger a cross-field anomaly and trigger the third layer of judgment. The third-level judgment includes: The current evolutionary segment is dynamically time-warped and matched with the patterns in the historical defect path library. If the matching degree is higher than the threshold, it is determined to be a known defect pattern, and the corresponding causal path map is output.

5. The method for cross-physics field generation simulation workflow for tracing the causes of process defects according to claim 1, characterized in that, The generation of enhanced time-series data containing the complete defect evolution process includes: Based on the causal path graph, the time axis of physical parameters is nonlinearly stretched to achieve higher time resolution near key events. Using the boundary anchor point as the endpoint, the starting point of the evolution path is determined in reverse, generating the parameter evolution trajectory from the starting point to the endpoint; The trajectory is used as a time-varying boundary condition input to the forward physical simulation model to generate a multiphysics spatiotemporal evolution sequence, and random perturbations are introduced to enhance data diversity.

6. The method for cross-physics field generation simulation workflow for tracing the causes of process defects according to claim 5, characterized in that, Conservative neighborhood sampling is used for potential novel defect paths, including: Sampling is performed within a neighborhood where the potential defect parameter point is centered and its Mahalanobis distance is less than a preset radius; Run short-term simulations for each sampling point to observe whether any anomalies are triggered, and save the short sequences along with the parameters as a preliminary exploratory dataset for novel defects.

7. The method for cross-physics field generation simulation workflow for tracing the causes of process defects according to claim 1, characterized in that, The adversarial network is a conditional generative adversarial network with causal constraints, and its loss function includes adversarial loss and causal consistency loss: ;in, It is the total loss during the training of the adversarial generative network. It is to combat losses. It is a loss of causal consistency. It is used to balance the weights between adversarial loss and causal loss.

8. The method for cross-physics field generation simulation workflow for tracing the causes of process defects according to claim 7, characterized in that, The training of the adversarial network adopts a phased strategy: The first stage involves independently training a causal checker network to enable it to identify key events and times from sequences. The second stage fixes the parameters of the causal checker and embeds them into the cGAN framework, guiding the generator to optimize both data distribution and causal logic during adversarial training.

9. The method for cross-physics field generation simulation workflow for tracing the causes of process defects according to claim 1, characterized in that, The intelligent process parameter tuning module uses Bayesian optimization to find the parameter combination that minimizes the defect index in the simulation workflow, based on the key control points in the causal chain. The optimized new operating condition data is fed back to the system knowledge base: if the defect is eliminated, it is recorded as a safe process interval; If the defect morphology changes, a new evolution path is extracted and the historical defect path library is updated to achieve continuous self-enhancement of the system.

10. A system for generating a cross-physics simulation workflow for tracing the causes of process defects according to any one of claims 1-9, characterized in that, The system includes: The data extraction and anchor point generation module is used to collect multi-physics field sensing data corresponding to the time of occurrence of real process defects in history. Through least squares fitting and sensitivity analysis, it extracts the critical physical field parameter combination that induces the defect and forms a low-dimensional boundary anchor point set oriented towards the cause of the defect. The intelligent classification and pattern recognition module, including the improved AdaBoost.M2 classification engine, receives parameters from a low-dimensional boundary anchor point set and performs progressive judgments: first, it determines whether the parameters approximate any defect anchor point boundary; if they do, it determines whether physical field coupling has triggered anomalies in other fields; if so, it compares the historical defect path library to determine whether the coupling response matches known defect evolution patterns; if all three layers are triggered, it outputs high-risk cross-field boundary combinations and causal path maps; if not all three layers are triggered, it is marked as a potential novel defect and triggers manual review. The improved AdaBoost.M2 classifier is as follows: For each layer of decision logic, a set of base classifiers is trained and then integrated using the AdaBoost.M2 framework; the training objective of the first layer of base classifiers is to determine the input parameters. Whether it approaches any boundary anchor point; the training goal of the second-layer base classifier is to determine whether a cross-physics anomaly is triggered given that the parameters approach the boundary; the training goal of the third-layer base classifier is to determine whether its evolution pattern matches a known defect path given that an anomaly is triggered; the AdaBoost.M2 algorithm iteratively adjusts the sample weights and classifier weights to ultimately obtain a strong classifier for each layer; The enhanced time-series data generation module is used to perform nonlinear stretching along the defect evolution time axis in the physical parameter space based on high-risk cross-field boundary combinations and causal path maps, generating enhanced time-series data containing the complete defect evolution process; conservative neighborhood sampling is used for potential novel paths; all generated data retain causal labels and temporal markers between physical fields, forming a defect mechanism-driven training set; The simulation workflow training and generation module includes a causal constraint adversarial generative network, whose input is a defect mechanism-driven training set. The adversarial generative network includes a generator and a decision unit. The generator is used to output a simulation sequence, and the decision unit is used to determine whether the sequence completely reproduces the causal chain of defect triggering, inter-field transmission and final manifestation. Through adversarial training until Nash equilibrium is reached, a simulation workflow with an embedded visual causal chain is output. The process optimization and link update module is used to input the simulation workflow into the intelligent tuning sub-module of process parameters and generate process adjustment instructions based on the key control points of the causal chain. The adjusted new operating condition data is fed back to the data extraction and anchor point generation module to update the low-dimensional boundary anchor point set: if the defect is eliminated, the new safe interval is recorded; if the defect morphology changes, a new causal path is added and the intelligent classification and pattern recognition module is triggered to re-judge, forming a continuous self-reinforcing defect prevention and optimization.

Images