A method and system for rapid prediction of high-temperature annealing microstructure evolution of grain-oriented electrical steel based on cellular automaton and bidirectional consistent deep learning agent model, and a medium

Abstract

The application discloses a kind of based on cellular automaton and the rapid prediction and process reverse design method, system and medium of high temperature annealing organization evolution of oriented electrical steel of bidirectional consistency deep learning agent model, with initial state feature s and annealing process vector u as input, generate the virtual data set of "(s,u)→ODF" in batch through cellular automaton model in preset process space, and train bidirectional consistency deep learning agent model, realize the rapid prediction of high temperature annealing texture and the process of target texture guide reverse push.This application can realize rapid forward prediction on the basis of absorbing mechanism model credibility, and further realize executable annealing process path from target texture reverse calculation, so as to support stable, replicable high temperature annealing process design and quality consistency control.

Description

Technical Field

[0001] This invention relates to metallic materials and intelligent manufacturing technology, and more specifically, to a method, system, and medium for rapid prediction of microstructure evolution and reverse engineering of high-temperature annealing of oriented electrical steel based on cellular automata and bidirectional consistency deep learning agent model. Background Technology

[0002] Grain-oriented electrical steel (typically with a silicon content of approximately 3.0%–3.5%) is a key soft magnetic alloy material whose magnetic properties (such as magnetic induction and iron loss) are highly sensitive to texture and microstructure. High-temperature annealing (e.g., final high-temperature annealing / secondary recrystallization annealing) is a crucial step in its fabrication process. It primarily induces grain growth and orientation-selective evolution under high-temperature atmospheres to achieve targeted texture enhancement, exemplified by Goss orientation, and stabilizes the final microstructure and properties while controlling the degree of anomalous grain growth (AGG). Because the high-temperature annealing process involves grain boundary migration, inhibitor pinning / dissolution, orientation-dependent growth, and the dynamic evolution of multi-field coupling, its microstructure evolution is extremely sensitive to process variables such as temperature, holding time, heating profile, and atmosphere parameters. The process window is narrow, and batch-to-batch fluctuations are significant, directly impacting product consistency and yield.

[0003] Currently, the prediction of microstructure and optimization of high-temperature annealing mainly rely on the following technical routes, but all of them have obvious limitations:

[0004] (1) Empirical trial and error and offline verification method: This method relies on engineering experience and a large number of experiments to back-infer process adjustments through the characterization results after annealing. This method has a long cycle, high cost, and is difficult to cover the high-dimensional process space with multiple coupled variables. It can usually only optimize locally and has insufficient stability.

[0005] (2) Pure physical mechanism simulation methods (such as phase field, cellular automata CA, etc.): can describe the mechanisms of grain boundary migration and grain evolution, and have high physical reliability. However, when considering factors such as thickness difference, inhibitor evolution and orientation selectivity, parameter calibration is complicated and computational cost is high. It is difficult to quickly scan and evaluate large-scale process paths, and it is even more difficult to directly realize the inverse design of process parameters from the target texture.

[0006] (3) Pure data-driven machine learning methods: Although they have the ability to fit nonlinearity, their training depends on a large amount of high-quality labeled data (EBSD / texture data after annealing). In actual production, the high cost of high-temperature annealing and the scarcity of samples make it difficult to obtain data and insufficient coverage, resulting in poor model generalization ability, insufficient robustness across batches, and lack of verifiable reverse design capability, making it difficult to use for closed-loop process control.

[0007] In summary, high-temperature annealing microstructure prediction and control faces a core contradiction: achieving speed, accuracy, and reverse design simultaneously is difficult. On the one hand, there are numerous and strongly coupled process variables, making microstructure evolution highly sensitive to minute parameter perturbations. On the other hand, experimental data is scarce and characterization is lagging, making it difficult to support purely data-driven methods. Furthermore, while physical models are reliable, their computational and calibration costs are high, failing to meet the demands of industrial settings for rapid assessment, process window scanning, and target texture-guided control. Especially in scenarios where target texture (e.g., Goss enhancement) and microstructure risk (e.g., AGG runaway) need to be simultaneously constrained, existing methods lack an engineering-practical technical system capable of achieving rapid forward prediction based on the reliability of mechanistic models, and further enabling the reverse derivation of executable annealing process paths from target texture, thereby supporting stable and reproducible high-temperature annealing process design and quality consistency control. Summary of the Invention

[0008] To address the shortcomings of existing technologies, the present invention aims to provide a method, system, and medium for rapid prediction of microstructure evolution and reverse engineering of oriented electrical steel under high-temperature annealing based on cellular automata and bidirectional consistency deep learning surrogate models. This method enables rapid forward prediction based on the credibility of the absorption mechanism model and further enables reverse engineering of executable annealing process paths from the target texture, thereby supporting stable and reproducible high-temperature annealing process design and quality consistency control.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] The first aspect of this invention provides a method for rapid prediction of microstructure evolution and reverse engineering of high-temperature annealing of oriented electrical steel based on cellular automata and bidirectional consistency deep learning agent model;

[0011] Using initial state features s and annealing process vector u as input, a virtual dataset of “(s,u)→ODF” is generated in batches within a preset process space through a cellular automata model based on physical mechanisms. A bidirectional consistency deep learning surrogate model is then trained to achieve rapid prediction of high-temperature annealing texture and process reverse inference guided by the target texture.

[0012] Preferably, the method for rapid prediction of microstructure evolution and reverse engineering of high-temperature annealed grain-oriented electrical steel includes the following steps:

[0013] S1, Multi-source data acquisition and preprocessing;

[0014] S2, Dataset generation based on the cellular automata model;

[0015] S3, Construction and training of the bidirectional consistency deep learning agent model;

[0016] S4, Fast forward prediction and visualization of ODF slice images under high temperature annealing conditions;

[0017] S5, Target Texture-Oriented Process Reverse Design and Constraint Optimization;

[0018] S6, closed-loop verification and online incremental learning.

[0019] Preferably, step S1 specifically includes:

[0020] Collect initial microstructure and texture data of the annealed plate or the plate to be annealed before entering high-temperature annealing;

[0021] The high-temperature annealing process parameters are expressed as the annealing process vector u;

[0022] The initial microstructure and texture data and the high-temperature annealing process parameters are standardized / normalized and missing value processing is performed to obtain normalized input. Simultaneously, the target output texture representation is unified into a fixed-size ODF slice image.

[0023] Preferably, the initial state feature s includes the initial grain size distribution, grain boundary angle distribution, and texture orientation information distribution;

[0024] The initial microstructure and texture data include electron backscatter diffraction data with a scan step size of 0.5–2 μm, from which initial grain size distribution, crystallographic orientation information, grain boundary type and orientation difference, and initial texture characteristics are extracted; and / or

[0025] The annealing process vector u includes at least a two-dimensional vector u=(T,t) consisting of the holding temperature T and the holding time t.

[0026] Preferably, the electron backscattering diffraction data is X-ray diffraction macroscopic texture spectrum or inhibitor-related characterization information; and / or

[0027] The annealing process vector u also includes heating rate HR, cooling rate CR, atmosphere parameter A, or segmented annealing curve parameters.

[0028] Preferably, the cellular automata model in step S2 is used to simulate grain boundary migration, grain growth / abnormal grain growth and texture evolution behavior during high-temperature annealing.

[0029] The cellular automaton model is run in batches within a preset process design space to generate ODF slice images from the initial state features s and the annealing process vector u to the annealed ODF slice images. "Virtual annotation" data .

[0030] Preferably, the cellular automata model integrates the following physical mechanisms:

[0031] Grain boundary mobility, expressed as ,in, Due to grain boundary orientation differences, For temperature, Pre-exponential factor, To activate energy, It is the gas constant;

[0032] The driving force and pinning effect determine the grain boundary migration rate, which is jointly determined by the driving force term and the inhibitory pinning term.

[0033] Orientation-selective growth introduces target orientation advantage through orientation-related migration rates or driving force weights.

[0034] Preferably, in step S2, measured electron backscatter diffraction data or X-ray diffraction texture data are used to calibrate the output of the cellular automata model to correct the deviation.

[0035] Preferably, step S3 includes the following steps:

[0036] S31, the forward prediction network F, uses the normalized initial state features. With annealing process vector As input, the output is the predicted ODF slice image. ;

[0037] S32, reverse-engineering network G, using normalized initial state features With target ODF slice image Alternatively, the target texture template can be used as input to output recommended process parameters. ;

[0038] S33, Loss Function and Training Strategy: During training, the forward fitting error, backward regression error, and bidirectional consistency error are minimized simultaneously to ensure that the backward design results can be validated by the forward model. Its form can be expressed as:

[0039] positive loss ;

[0040] Reverse loss ;

[0041] Bidirectional consistency loss .

[0042] Preferably, in step S31, the forward prediction network F employs a lightweight encoder-decoder structure to process the initial state features. With the annealing process vector After feature encoding and fusion, a 64×64 ODF slice image is generated through a two-dimensional decoding module; the output is connected to a Sigmoid function to limit the predicted values ​​to a numerical range consistent with the normalized label; and / or

[0043] In step S32, the reverse engineering network G adopts a lightweight structure of "image encoder + MLP regression head": for the target ODF slice image Convolutional encoding is performed to obtain the texture target features, which are then compared with the initial state features. The annealing process vector u is obtained by regression after fusion and mapped to the executable process boundary through inverse normalization; and / or

[0044] In step S33, the optimizer employs an optimization algorithm and saves model weights by early stopping the validation set; and / or superimposes process boundary constraints and curve smoothing constraints.

[0045] Preferably, the forward prediction network F includes an encoding module for feature fusion of the initial state features s and the annealing process vector u, and a decoding module for outputting the annealed ODF slice map.

[0046] The reverse design network G includes an image encoder for feature extraction of the target ODF slice image and a multilayer perceptron regression module for outputting annealing process vectors.

[0047] Preferably, step S4 specifically includes:

[0048] After the bidirectional consistency deep learning proxy model is trained, for any given initial state feature s and annealing process vector u, a normalization method consistent with step S1 is first applied to obtain... Then input the forward network F to obtain the predicted ODF slice image. ;

[0049] For the predicted ODF slice image Perform denormalization and output in a 64×64 layout to achieve rapid prediction and visualization of the texture after annealing.

[0050] Preferably, step S5 specifically includes:

[0051] For the target ODF slice image Or the texture template obtained by mapping the target index, will Inputting the reverse engineering network G yields the initial process values. ;

[0052] With the aforementioned initial process values As the initial solution, under the constraints of process upper / lower limits and heating / cooling rates, we further minimize... To obtain the optimal process that satisfies the target texture. .

[0053] Preferably, the heat preservation temperature T is [900℃, 1300℃], and the heat preservation time t is [10min, 3000min].

[0054] The resolution of the ODF slice image is 64×64 or 128×128;

[0055] The forward prediction network F and / or the reverse design network G include 2 to 6 fully connected layers, with 64 to 1024 hidden units per layer.

[0056] The second aspect of this invention provides a system for rapidly predicting the evolution of the microstructure of high-temperature annealing of grain-oriented electrical steel and for reverse engineering the process based on cellular automata and a bidirectional consistency deep learning surrogate model, as described in the first aspect of this invention. The system includes:

[0057] The data acquisition and preprocessing module is used to acquire and organize the initial state characteristics of the grain-oriented electrical steel before high-temperature annealing and the corresponding annealing process information, forming a consistent input for subsequent modeling. This module can analyze and extract features from data from electron backscatter diffraction (EBSD) and other related characterization methods, as well as production process records, generating an initial state feature vector representation *k* and an annealing process vector *k*. Simultaneously, this module performs denoising, missing value processing, outlier removal, unit unification, and standardization / normalization mapping on the data to obtain normalized data. Furthermore, the target output texture is uniformly converted into a fixed-resolution ODF slice image label format to ensure that data from different batches and sources can be aligned, trained, and compared within the same feature space.

[0058] The CA virtual experiment data generation module is used to run cellular automata (CA) mechanism models in batches within a preset process design space, constructing a "virtual labeled" training dataset to reduce reliance on large-scale measured texture data after annealing. Based on the input initial state features s and process vector u, this module simulates mechanisms such as grain boundary migration, grain growth / abnormal grain growth (AGG), inhibitor pinning and dissolution, and orientation-selective growth during high-temperature annealing, outputting an annealed ODF slice image Y, forming a data pair (s,u)→Y. The module supports process space sampling strategies (such as Latin hypercube sampling, mesh scanning, Bayesian DoE, etc.) and parallel computing scheduling to quickly cover high-dimensional process variable combinations; and optionally introduces a small amount of measured EBSD / XRD results to calibrate the CA output (e.g., correcting key texture peak intensity, peak position, and AGG risk statistics) to improve the physical fidelity of the virtual data and the consistency with real production line trends.

[0059] The bidirectional surrogate model training and inference module is used to build, train, and deploy a bidirectional consistency deep learning surrogate model to achieve fast forward prediction and verifiable reverse process inference of high-temperature annealing texture. This module includes a forward prediction network F and a reverse design network G: where F receives (s′, u′) and outputs a predicted ODF slice image. It employs a lightweight encoder-decoder structure to achieve fast inference; G receiver And return to the recommended output process It can employ an "image encoder + MLP regression head" structure and map the results to the executable process boundary. During the training phase, this module jointly minimizes the forward fitting error, backward regression error, and cycle consistency error, enabling the backward output to be quickly validated by the forward model, thereby ensuring the interpretability and feasibility of the backward recommended process. It also supports Adam / AdamW optimization, early stopping on the validation set, model weight management, and inference interface encapsulation, outputting millisecond-level texture prediction results and their visualization data that can be used for rapid on-site evaluation.

[0060] The target texture process inversion and constraint optimization module is used to reverse design the path from the target texture to the executable annealing process, and outputs the optimal process solution under multiple process constraints and quality risk constraints. This module uses the target ODF slice image. Alternatively, the texture template generated by index mapping can be used as input, and the initial process values ​​are first obtained by calling the inverse network G. , and then As the initial solution, under constraints such as upper / lower limits of the process, heating / cooling rate boundaries, and monotonicity / smoothness of the piecewise curve, the objective function is... Perform constrained optimization iterations to obtain the optimal process. Furthermore, this module can introduce multi-objective constraints for engineering implementation (such as maximizing Goss strength, texture uniformity, AGG runaway risk penalty, energy consumption / time cost constraints, etc.), and output executable process parameter tables or segmented annealing curves to achieve "target texture-oriented" process recommendation, window scanning, and risk-controlled optimization.

[0061] The closed-loop verification and online incremental learning module is used to apply the optimal process obtained through inversion. The module is implemented in experimental furnaces or production lines, and uses the measured results to perform closed-loop verification and continuous updates of the surrogate model and CA model, thereby improving cross-batch robustness and long-term availability. This module collects EBSD / XRD texture and microstructure indices of annealed sheet metal, compares the measured ODF / key indices with the model predictions for consistency and error assessment. When the error exceeds the threshold or a new operating condition distribution drift occurs, the module adds new samples to the data pool in the form of "real labeling" or "weak supervision constraints," triggering incremental fine-tuning of the bidirectional surrogate model (or updating the CA deviation calibration term). The updated model weights are then redistributed for the next round of inference and inversion. Through this closed-loop mechanism, continuous adaptation of the digital twin model, stable iteration of process recommendations, and dynamic enhancement of quality consistency control capabilities can be achieved.

[0062] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the program, when executed on a processor, implements the steps of the method for rapid prediction of microstructure evolution and reverse engineering of high-temperature annealing of oriented electrical steel based on cellular automata and bidirectional consistency deep learning surrogate model as described in the first aspect of the present invention.

[0063] This invention provides a method, system, and medium for rapid prediction of microstructure evolution and reverse engineering of grain-oriented electrical steel during high-temperature annealing, based on cellular automata and a bidirectional consistency deep learning surrogate model. It involves modeling and microstructure control of the high-temperature annealing process of grain-oriented electrical steel, oriented towards rolling process design and rapid optimization, and digital research and verification of the microstructure evolution mechanism. It also has the following beneficial effects:

[0064] (1) Reduce the reliance on EBSD data after large-scale experimental annealing to alleviate the problems of data scarcity and verification lag.

[0065] This invention constructs a training set primarily using CA virtual data, requiring only 3-12 sets of measured EBSD / XRD data to complete model calibration. Compared to pure data-driven methods that require hundreds of sets of measured labeled data, the data requirement is reduced by more than 90%, improving the deployability and scalability of this method in industrial scenarios.

[0066] (2) The complex forward modeling process of high-temperature annealing is compressed into a lightweight surrogate model to achieve rapid prediction of annealing texture.

[0067] Compared to directly calling CA or other physical models to perform a large number of parameter scans, this invention achieves millisecond-level inference through a proxy model, which can quickly evaluate and visualize a large number of candidate annealing processes within the production cycle, greatly improving process development efficiency.

[0068] (3) Achieve both forward prediction and verifiable reverse design within the same framework to support target texture-oriented control.

[0069] This invention constructs a closed-loop verifiable inversion mechanism through "forward network + reverse network + bidirectional consistency constraint", which can directly inversely deduce the executable annealing process from the target ODF (or target texture template) and quickly verify the inversion results through the forward model, thereby realizing process recommendation and control oriented towards target texture, breaking through the limitation of traditional methods that "can only perform forward modeling and are difficult to perform inversion".

[0070] (4) The output results maintain the integrity of the texture information, which is convenient for coupling with performance evaluation or risk judgment.

[0071] This invention uses 64×64 ODF full-field slices as a unified output, which not only preserves the texture peaks, peak positions and distribution morphology information, but also facilitates the extraction of key indicators (such as Goss intensity, uniformity, AGG risk characteristics, etc.), providing a direct interface for subsequent linkage with magnetic performance models, quality criteria or process constraints. Attached Figure Description

[0072] Figure 1 This is a flowchart illustrating the rapid prediction of microstructure evolution and reverse engineering method for high-temperature annealing of grain-oriented electrical steel according to the present invention.

[0073] Figure 2 This is a schematic diagram of the CA "virtual experiment" module structure and DoE sampling in the rapid prediction and reverse design method for high-temperature annealing microstructure evolution of grain-oriented electrical steel in this invention.

[0074] Figure 3 This is a schematic diagram of the deep learning prediction model architecture used in Example 1 of the present invention, which is a method for rapid prediction of microstructure evolution and reverse design of high-temperature annealed oriented electrical steel.

[0075] Figure 4 This is a schematic diagram comparing the ODF diagram predicted by the deep learning model with the ODF effect obtained by actual CA simulation in Example 1 of the method for rapid prediction of microstructure evolution and reverse design of high-temperature annealing of oriented electrical steel in this invention. Detailed Implementation

[0076] To better understand the above-mentioned technical solutions of the present invention, the technical solutions of the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0077] This invention provides a method for rapid prediction of microstructure evolution and reverse engineering of high-temperature annealing of oriented electrical steel based on cellular automata and bidirectional consistency deep learning surrogate model;

[0078] Using the initial state features s and the annealing process vector u as input, a virtual dataset of “(s,u)→ODF(64×64)” is generated in batches within a preset process space through a cellular automata model, and a bidirectional consistency deep learning surrogate model is trained to achieve rapid prediction of high-temperature annealing texture and process reverse inference guided by the target texture.

[0079] Combination Figure 1 As shown, the method for rapid prediction of microstructure evolution and reverse engineering of high-temperature annealed grain structure of grain-oriented electrical steel according to the present invention includes the following steps:

[0080] S1, Multi-source data acquisition and preprocessing;

[0081] Collect initial microstructure and texture data of the annealed plate or the plate to be annealed before entering high-temperature annealing.

[0082] The high-temperature annealing process parameters are expressed as the annealing process vector u. In this invention, high-temperature annealing refers to the final high-temperature annealing / box annealing process (used for secondary recrystallization, impurity purification, and the formation of Goss orientation abnormal growth), and the general temperature range is 1100℃~1200℃.

[0083] The initial microstructure and texture data and high-temperature annealing process parameters were standardized / normalized and missing values ​​were processed to obtain normalized input. Meanwhile, the target output texture representation is unified into a fixed-size ODF slice image (pixel size of 64×64) as a supervision label for subsequent training.

[0084] ODF slice images are generated from raw orientation data acquired by EBSD and / or XRD. Specifically, the raw orientation data is first represented by Euler angles and orientation statistics are performed to obtain the orientation distribution function; then a preset ϕ2 section is selected, and the ODF intensity on this section is discretized and interpolated using a regular grid; finally, the discretized ODF intensity matrix is ​​converted into a two-dimensional image representation, which serves as the ODF slice image for model input or output.

[0085] Initial microstructure and texture data include electron backscatter diffraction (EBSD) data with a scan step size of 0.5–2 μm, from which initial grain size distribution, crystallographic orientation information, grain boundary type and orientation difference, and initial texture features are extracted. Electron backscatter diffraction data are X-ray diffraction (XRD) macroscopic texture spectra or inhibitor-related characterization information, used for subsequent calibration and validation.

[0086] Initial state features include commonly used material microstructure features such as initial grain size distribution, grain boundary angle distribution, and texture orientation information distribution.

[0087] The annealing process vector u includes at least a two-dimensional vector u=(T,t) consisting of the holding temperature T and the holding time t.

[0088] The annealing process vector u also includes heating rate HR, cooling rate CR, atmosphere parameters A (such as hydrogen-nitrogen ratio, nitrogen potential, etc.) or segmented annealing curve parameters.

[0089] S2, dataset generation based on cellular automata (CA) model, such as Figure 2 As shown;

[0090] A cellular automata model based on physical mechanisms was constructed to simulate grain boundary migration, grain growth / abnormal grain growth (AGG), and texture evolution behavior during high-temperature annealing.

[0091] Batch run cellular automata models within the predefined process design space (DoE) to generate ODF slice images from initial state features s and annealing process vector u to the annealed ODF slice images. "Virtual annotation" data Furthermore, a small amount of measured EBSD or XRD texture data can be introduced to calibrate the output of the CA model, thereby improving the physical fidelity and reliability of the generated data.

[0092] The cellular automata model integrates the following physical mechanisms:

[0093] Grain boundary mobility, expressed as ,in, Due to grain boundary orientation differences, For temperature, Pre-exponential factor, To activate energy, It is the gas constant;

[0094] The driving force and pinning effect determine the grain boundary migration rate, which is jointly determined by the driving force term and the inhibitory pinning term.

[0095] Orientation-selective growth introduces target orientation (e.g., Goss) advantage through orientation-related migration rates or driving force weights.

[0096] S3, Construction and training of a bidirectional consistency deep learning proxy model, such as Figure 3 As shown;

[0097] To simultaneously achieve forward prediction and reverse design of high-temperature annealing texture, this invention constructs a bidirectional surrogate model system consisting of a forward prediction network F and a reverse design network G, specifically including the following steps:

[0098] S31, the forward prediction network F, uses the normalized initial state features. With annealing process vector As input, the output is the predicted ODF slice image. Among them, the forward prediction network F adopts a lightweight encoder-decoder structure, which is used for the initial state features. With annealing process vector After feature encoding and fusion, a 64×64 ODF slice image is generated through a two-dimensional decoding module; the output is connected to a Sigmoid function to limit the predicted value to a numerical range consistent with the normalized label.

[0099] S32, reverse-engineering network G, using normalized initial state features With target ODF slice image Alternatively, the target texture template can be used as input to output recommended process parameters. The reverse design network G employs a lightweight structure of "image encoder + MLP regression head": it processes the target ODF slice image... Convolutional encoding is performed to obtain the texture target features, which are then compared with the initial state features. After fusion, the annealing process vector u is obtained by regression and then mapped to the executable process boundary through inverse normalization.

[0100] S33, Loss Function and Training Strategy: During training, the forward fitting error, backward regression error, and bidirectional consistency error are minimized simultaneously to ensure that the backward design results can be validated by the forward model. Its form can be expressed as:

[0101] positive loss ;

[0102] Reverse loss ;

[0103] Bidirectional consistency loss .

[0104] The optimizer uses Adam / AdamW and saves model weights through early stopping on the validation set. Preferably, process boundary constraints and curve smoothing constraints can be superimposed to ensure that the output process is executable.

[0105] The specific definition of the loss function:

[0106]

[0107]

[0108]

[0109] The overall loss function can be expressed as:

[0110] L=α*L_forward+β*L_backward+γ*L_cycle

[0111] Where α, β, and γ are weighting coefficients.

[0112] The forward prediction network F includes an encoding module for feature fusion of the initial state features s and the annealing process vector u, and a decoding module for outputting the ODF slice map after annealing.

[0113] The reverse design network G includes an image encoder for feature extraction of the target ODF slice map and a multilayer perceptron regression module for outputting annealing process vectors.

[0114] S4, Fast forward prediction and visualization of ODF slice images under high temperature annealing conditions;

[0115] After the bidirectional consistency deep learning surrogate model is trained, for any given initial state feature s and annealing process vector u, the normalization method consistent with step S1 is first used to obtain... Then input the forward network F to obtain the predicted ODF slice image. .

[0116] For the predicted ODF slice image Performing denormalization and outputting in a 64×64 layout enables rapid prediction and visualization of the texture after annealing, providing an efficient tool for process window scanning and risk assessment.

[0117] S5, Target Texture-Oriented Process Reverse Design and Constraint Optimization;

[0118] Target ODF slice image Or the texture template obtained by mapping the target index, will Input the reverse engineering network G to obtain the initial process values. .

[0119] Initial process value As the initial solution, under the constraints of process upper / lower limits and heating / cooling rates, we further minimize... To obtain the optimal process that satisfies the target texture. .

[0120] Step S5 enables rapid reverse engineering and control from the “target texture” to the “executable process”.

[0121] S6, closed-loop verification and online incremental learning;

[0122] The optimal process obtained in step S5 The process was performed in an experimental furnace or production line, and the annealed sheet was subjected to electron backscatter diffraction / X-ray diffraction tests to obtain the measured texture results and microstructure parameters.

[0123] The measured results are compared with the predictions of the bidirectional consistency deep learning agent model. When the error exceeds the threshold, new data pairs are added to the training set and the bidirectional consistency deep learning agent model is incrementally fine-tuned to achieve continuous optimization and adaptive updating of the digital twin model.

[0124] Through the above steps, this invention can quickly obtain ODF slice image prediction results under different cold rolling paths without repeating large-scale CA model calculations or a large number of trial high-temperature annealing experiments, providing process engineers with an efficient tool for high-temperature annealing process design, texture control and performance stability improvement.

[0125] The dataset generation method for the cellular automata model in this invention specifically includes:

[0126] The CA model characterizes grain boundary migration, grain growth / abnormal grain growth (AGG), and texture evolution processes, and can introduce mechanistic terms such as orientation difference and temperature-related grain boundary mobility and inhibitor pinning effect. Preferably, the output deviation of the CA model is calibrated using a small amount of measured EBSD / XRD data to improve the physical fidelity of the virtual data.

[0127] The key to the structure and training method of the bidirectional consistency deep learning agent model in this invention lies in:

[0128] Forward prediction network F: Implementation The end-to-end prediction outputs a 64×64 ODF slice image;

[0129] Reverse design network G: Implementation The end-to-end inversion outputs executable annealing process parameters that satisfy boundary constraints;

[0130] And through a two-way consistency mechanism, the reverse output can be verified by the forward model.

[0131] The loss function and constraint design of the bidirectional consistency deep learning proxy model in this invention are as follows:

[0132] Simultaneously minimize the positive fitting error, the back regression error, and the consistency error, so that and Constraints such as process boundaries, upper limits for heating / cooling rates, and curve smoothing are superimposed to ensure that the inversion process is executable.

[0133] This invention also provides a system for rapidly predicting the microstructure evolution of high-temperature annealed grain-oriented electrical steel and for reverse engineering the process of implementing the method of rapid prediction and process reverse engineering of high-temperature annealed grain-oriented electrical steel of this invention, comprising:

[0134] The data acquisition and preprocessing module is used to acquire and organize the initial state characteristics of the grain-oriented electrical steel before high-temperature annealing and the corresponding annealing process information, forming a consistent input for subsequent modeling. This module can analyze and extract features from data from electron backscatter diffraction (EBSD) and other related characterization methods, as well as production process records, generating an initial state feature vector representation *k* and an annealing process vector *k*. Simultaneously, this module performs denoising, missing value processing, outlier removal, unit unification, and standardization / normalization mapping on the data to obtain normalized data. Furthermore, the target output texture is uniformly converted into a fixed-resolution ODF slice image label format to ensure that data from different batches and sources can be aligned, trained, and compared within the same feature space.

[0135] The CA virtual experiment data generation module is used to run cellular automata (CA) mechanism models in batches within a preset process design space, constructing a "virtual labeled" training dataset to reduce reliance on large-scale measured texture data after annealing. Based on the input initial state features s and process vector u, this module simulates mechanisms such as grain boundary migration, grain growth / abnormal grain growth (AGG), inhibitor pinning and dissolution, and orientation-selective growth during high-temperature annealing, outputting an annealed ODF slice image Y, forming a data pair (s,u)→Y. The module supports process space sampling strategies (such as Latin hypercube sampling, mesh scanning, Bayesian DoE, etc.) and parallel computing scheduling to quickly cover high-dimensional process variable combinations; and optionally introduces a small amount of measured EBSD / XRD results to calibrate the CA output (e.g., correcting key texture peak intensity, peak position, and AGG risk statistics) to improve the physical fidelity of the virtual data and the consistency with real production line trends.

[0136] The bidirectional surrogate model training and inference module is used to build, train, and deploy a bidirectional consistency deep learning surrogate model to achieve fast forward prediction and verifiable reverse process inference of high-temperature annealing texture. This module includes a forward prediction network F and a reverse design network G: where F receives (s′, u′) and outputs a predicted ODF slice image. It employs a lightweight encoder-decoder structure to achieve fast inference; G receiver And return to the recommended output process It can employ an "image encoder + MLP regression head" structure and map the results to the executable process boundary. During the training phase, this module jointly minimizes the forward fitting error, backward regression error, and cycle consistency error, enabling the backward output to be quickly validated by the forward model, thereby ensuring the interpretability and feasibility of the backward recommended process. It also supports Adam / AdamW optimization, early stopping on the validation set, model weight management, and inference interface encapsulation, outputting millisecond-level texture prediction results and their visualization data that can be used for rapid on-site evaluation.

[0137] The target texture process inversion and constraint optimization module is used to reverse design the path from the target texture to the executable annealing process, and outputs the optimal process solution under multiple process constraints and quality risk constraints. This module uses the target ODF slice image. Alternatively, the texture template generated by index mapping can be used as input, and the initial process values ​​are first obtained by calling the inverse network G. , and then As the initial solution, under constraints such as upper / lower limits of the process, heating / cooling rate boundaries, and monotonicity / smoothness of the piecewise curve, the objective function is... Perform constrained optimization iterations to obtain the optimal process. Furthermore, this module can introduce multi-objective constraints for engineering implementation (such as maximizing Goss strength, texture uniformity, AGG runaway risk penalty, energy consumption / time cost constraints, etc.), and output executable process parameter tables or segmented annealing curves to achieve "target texture-oriented" process recommendation, window scanning, and risk-controlled optimization.

[0138] The closed-loop verification and online incremental learning module is used to apply the optimal process obtained through inversion. The module is implemented in experimental furnaces or production lines, and uses the measured results to perform closed-loop verification and continuous updates of the surrogate model and CA model, thereby improving cross-batch robustness and long-term availability. This module collects EBSD / XRD texture and microstructure indices of annealed sheet metal, compares the measured ODF / key indices with the model predictions for consistency and error assessment. When the error exceeds the threshold or a new operating condition distribution drift occurs, the module adds new samples to the data pool in the form of "real labeling" or "weak supervision constraints," triggering incremental fine-tuning of the bidirectional surrogate model (or updating the CA deviation calibration term). The updated model weights are then redistributed for the next round of inference and inversion. Through this closed-loop mechanism, continuous adaptation of the digital twin model, stable iteration of process recommendations, and dynamic enhancement of quality consistency control capabilities can be achieved.

[0139] Example 1

[0140] In this embodiment 1, oriented electrical steel plates before annealing (before entering the high-temperature annealing process) of the same heat number were selected, and EBSD data (step size 1μm) were collected in the thickness direction. A CA model was constructed from the EBSD.

[0141] The design space for the high-temperature annealing process is set as follows:

[0142] Temperature range: T=1000℃~1250℃

[0143] Insulation time range: t=30min~3000min

[0144] Only with Constructing a two-dimensional process vector The min-max normalization mapping to [0,1] is used to obtain... The ODF slice images were uniformly set to 64×64, and the intensity was normalized using min-max to obtain the labels. .

[0145] Within the aforementioned design space, 500 sets of process points (T) were generated using Latin hypercube sampling. i ,t i In the CA model, the high-temperature annealing grain boundary migration and grain growth / texture evolution processes are simulated one by one. The ODF slice images after annealing are output as "ground values" and directly constitute the training set. No actual calibration is introduced.

[0146] The bidirectional consistency deep learning agent model employs a simplified dual-network structure:

[0147] Forward prediction network F: will (The initial ODF image was encoded using a lightweight CNN) and After fusion (2D MLP encoding), a 64×64 image is generated by 2 layers of upsampling convolution decoding, and the output is connected to a Sigmoid.

[0148] Reverse engineering network G: Input ,in The target ODF slice image (64×64) is encoded using a lightweight CNN and then... Fusion, using MLP regression output And mapped to through denormalization .

[0149] Training configuration: 500 datasets are partitioned into training / validation / test sets in a 7:1.5:1.5 ratio; the loss function used is... (All are MSE), optimizer Adam, learning rate 3×10-4 Batch size 16, maximum number of rounds 200, early stopping on the validation set.

[0150] After training, this Example 1 can achieve: given Forward fast prediction of ODF slice images; and given a target Direct reverse output process and use The inversion results are quickly verified.

[0151] Example 2

[0152] In this embodiment 2, two heat numbers of annealed plates were selected, with an EBSD step size of 0.8 μm. When constructing the CA model, an inhibitor equivalent characterization parameter was further introduced to enhance the model's ability to express AGG sensitivity.

[0153] The high-temperature annealing process vector adopts the preferred extended form:

[0154]

[0155] in, For heating rate, This refers to the cooling rate.

[0156] The design space is set as follows:

[0157] T=980℃~1280℃, t=20min~360min

[0158] HR=1℃ / min~15℃ / min, CR=1℃ / min~20℃ / min

[0159] All continuous variables are min-max normalized to [0,1].

[0160] Within the design space, 1200 sets of process points were generated, and the CA model was run to obtain virtual labeled ODF slice images. To improve the fidelity of the virtual data, this embodiment 2 additionally collected a small amount of measured data for calibration: 6 sets of annealing processes were selected for each heat number (12 sets in total), and the XRD macrotexture indices after annealing were obtained (optionally, 2-3 sets of EBSD were added). These measured data were used to correct the deviations of the key texture peak intensity / peak position and AGG risk statistics output by the CA model, so that the trend of the virtual data was consistent with the actual measurement, forming a "calibrated virtual dataset".

[0161] The bidirectional consistency deep learning agent model is trained using bidirectional consistency.

[0162] The forward prediction network F and the reverse design network G have the same structure as in Example 1, but are adapted to 5D process input and the hidden layers are appropriately widened (e.g., MLP width 256).

[0163] The loss function used is:

[0164]

[0165] in, constraint and ; For process boundaries and rate constraints (to ensure that the output is executable).

[0166] Reverse engineering employs a two-stage process: "initial value inversion + constraint refinement".

[0167] 1) First by Output initial process values ;

[0168] 2) with As the initial solution, minimize under process boundary and rate constraints. After a few iterations of refinement, we obtained .

[0169] Training configuration: 1200 groups divided in an 8:1:1 ratio; AdamW, learning rate 2×10⁻⁶. -4 Batch size 12, maximum number of rounds 250, early stop.

[0170] The results of Example 2 demonstrate that, while maintaining rapid inference, reverse engineering can output the process. It has higher executability and a more stable target texture achievement rate, making it the preferred implementation method for engineering projects.

[0171] Example 3

[0172] This embodiment 3 is designed for scenarios where data is limited or where it is desirable to further simplify the model structure: instead of training the reverse network G separately, only a forward proxy model F is trained, and then the differentiability of F is used to achieve reverse design (alternative inversion path).

[0173] In this embodiment 3, EBSD of the sheet material before annealing was acquired (step size 2μm), and the initial state s was simplified to an initial ODF slice image (64×64), without introducing additional inhibitor parameters (covering the "non-preferred simplified input" scheme). The output remains the same. =45° ODF slice image (64×64).

[0174] The process vector is parameterized using segmented annealing curves (an alternative that covers "more complex control modes"):

[0175] Assume the annealing curve is a 3-segment broken line with the following parameters:

[0176]

[0177] in, For segmented target temperatures, Define the segment duration and impose constraints. ≤ ≤ and ≥0.

[0178] Design space example:

[0179] =950℃~1280℃, =10min~180min

[0180] Total duration: + + ≤360min

[0181] The parameters are normalized to [0,1], and the temperature monotonic constraint is added to the training or inversion process as a penalty term.

[0182] Data generation: 800 piecewise curves are sampled in the above space, and the CA model is run to output ODF slice images as virtual supervision; at the same time, only 10 sets of measured XRD macrotexture indices are collected for weak supervision verification (not used in strong supervision training, but only for screening / verification and threshold correction), covering the "weak supervision alternative data" scheme.

[0183] The forward model F adopts a simpler structure:

[0184] Process MLP encoding (6-dimensional input, hidden layers 128-256-256, ReLU);

[0185] Initial ODFCNN encoding (3 convolutional layers);

[0186] After fusion, convolutional decoding generates a 64×64 ODF, which is then connected to a Sigmoid function.

[0187] The loss function is based on MSE, and it is preferable to add a temperature monotonicity / curve smoothing regularization term to ensure that the output has a stable response to changes in process parameters.

[0188] Reverse engineering relies solely on the differentiability of F: given the objective ODFY*, minimize it under constraints.

[0189]

[0190] Using Adam Iterative updates yield process curves that satisfy the constraints. Since this inversion process does not require training G, the structure is simpler, and the inversion results can be directly verified by F; furthermore, XRD weakly supervised data can be used to perform final screening of key texture indicators, thereby improving engineering reliability.

[0191] Training configuration: 800 groups divided in a 7:1.5:1.5 ratio; Adam dataset, learning rate 5×10⁻⁶. -4 Batch size 16, maximum number of rounds 200, early stop.

[0192] This embodiment 3 demonstrates that even without constructing a reverse network, forward prediction and executable reverse design can still be achieved within a single agent model framework, covering the alternative implementation path of this invention.

[0193] Those skilled in the art should recognize that the above embodiments are merely illustrative of the present invention and are not intended to limit the present invention. Any variations or modifications to the above embodiments that are within the spirit and essence of the present invention will fall within the scope of the claims of the present invention.

Claims

1. A method for rapid prediction of microstructure evolution and reverse engineering of high-temperature annealing of grain-oriented electrical steel based on cellular automata and bidirectional consistency deep learning surrogate model, characterized in that: Using initial state features s and annealing process vector u as input, the initial state features s include initial grain size distribution, grain boundary angle distribution, and texture orientation information distribution; the annealing process vector u includes at least a two-dimensional vector u=(T,t) composed of holding temperature T and holding time t; a virtual dataset of "(s,u)→ODF" is generated in batches within a preset process space through a cellular automata model based on physical mechanisms, and a bidirectional consistency deep learning surrogate model is trained to achieve rapid prediction of high-temperature annealing texture and process reverse inference guided by target texture; The method for rapid prediction of microstructure evolution and reverse engineering of high-temperature annealing of grain-oriented electrical steel includes the following steps: S1, Multi-source data acquisition and preprocessing; The initial microstructure and texture data of the annealed plate or the plate to be annealed are collected before entering the high-temperature annealing process; the initial microstructure and texture data include electron backscatter diffraction data, and the initial grain size distribution, crystallographic orientation information, grain boundary type and orientation difference, and initial texture characteristics are extracted from them; The high-temperature annealing process parameters are expressed as the annealing process vector u; The initial state feature s and the high-temperature annealing process parameters are standardized / normalized and missing value processing is performed to obtain the normalized input. Simultaneously, the target output texture representation is unified into a fixed-size ODF slice image; S2, Dataset generation based on the cellular automata model; The cellular automata model integrates the following physical mechanisms: Grain boundary mobility, expressed as ,in, Due to grain boundary orientation differences, For temperature, Pre-exponential factor, To activate energy, It is the gas constant; The driving force and pinning effect determine the grain boundary migration rate, which is jointly determined by the driving force term and the inhibitory pinning term. Orientation-selective growth introduces target orientation advantage through orientation-related migration rates or driving force weights; S3, Construction and training of the bidirectional consistency deep learning agent model; S31, the forward prediction network F, employs a lightweight encoder-decoder structure, using normalized initial state features. With annealing process vector As input, the output is the predicted ODF slice image. ; S32, reverse-engineering network G, using normalized initial state features With target ODF slice image Alternatively, the target texture template can be used as input to output recommended process parameters. ; S33, loss function and training strategy: during training, the forward fitting error, backward regression error and bidirectional consistency error are minimized simultaneously to ensure that the backward design results can be validated by the forward model. S4, Fast forward prediction and visualization of ODF slice images under high temperature annealing conditions; After the bidirectional consistency deep learning proxy model is trained, for any given initial state feature s and annealing process vector u, a normalization method consistent with step S1 is first applied to obtain... Then input the forward network F to obtain the predicted ODF slice image. ; For the predicted ODF slice image Perform denormalization and output in 64×64 layout to achieve rapid prediction and visualization of texture after annealing; S5, Target Texture-Oriented Process Reverse Design and Constraint Optimization; S6, closed-loop verification and online incremental learning.

2. The method for rapid prediction of microstructure evolution and reverse engineering of high-temperature annealing of grain-oriented electrical steel based on cellular automata and bidirectional consistency deep learning surrogate model as described in claim 1, characterized in that: The scanning step size of the electron backscatter diffraction data is 0.5~2μm.

3. The method for rapid prediction of microstructure evolution and reverse engineering of high-temperature annealing of oriented electrical steel based on cellular automata and bidirectional consistency deep learning surrogate model as described in claim 2, characterized in that: The annealing process vector u also includes heating rate HR, cooling rate CR, atmosphere parameter A, or segmented annealing curve parameters.

4. The method for rapid prediction of microstructure evolution and reverse engineering of high-temperature annealing of grain-oriented electrical steel based on cellular automata and bidirectional consistency deep learning surrogate model as described in claim 1, characterized in that: The cellular automata model in step S2 is used to simulate grain boundary migration, grain growth / abnormal grain growth and texture evolution during high-temperature annealing. The cellular automaton model is run in batches within a preset process design space to generate ODF slice images from the initial state features s and the annealing process vector u to the annealed ODF slice images. "Virtual annotation" data .

5. The method for rapid prediction of microstructure evolution and reverse engineering of high-temperature annealing of grain-oriented electrical steel based on cellular automata and bidirectional consistency deep learning surrogate model as described in claim 4, characterized in that: In step S2, measured electron backscatter diffraction data or X-ray diffraction texture data are introduced to calibrate the output of the cellular automata model.

6. The method for rapid prediction of microstructure evolution and reverse engineering of high-temperature annealing of oriented electrical steel based on cellular automata and bidirectional consistency deep learning surrogate model as described in claim 1, is characterized in that... In step S33, the loss function can be expressed in the following form: positive loss ; Reverse loss ; Bidirectional consistency loss .

7. The method for rapid prediction of microstructure evolution and reverse engineering of high-temperature annealing of grain-oriented electrical steel based on cellular automata and bidirectional consistency deep learning surrogate model as described in claim 6, characterized in that: In step S31, the forward prediction network F employs a lightweight encoder-decoder structure to process the initial state features. With the annealing process vector After feature encoding and fusion, a 64×64 ODF slice image is generated through a two-dimensional decoding module; the output is connected to a Sigmoid to limit the predicted value to a numerical range consistent with the normalized label. and / or In step S32, the reverse engineering network G adopts a lightweight structure of "image encoder + MLP regression head": for the target ODF slice image Convolutional encoding is performed to obtain the texture target features, which are then compared with the initial state features. The annealing process vector u is obtained by regression after fusion and then mapped to the executable process boundary through inverse normalization. and / or In step S33, the optimizer employs an optimization algorithm and saves model weights by early stopping the validation set; and / or superimposes process boundary constraints and curve smoothing constraints.

8. The method for rapid prediction of microstructure evolution and reverse engineering of high-temperature annealing of grain-oriented electrical steel based on cellular automata and bidirectional consistency deep learning surrogate model as described in claim 6, characterized in that: The forward prediction network F includes an encoding module for feature fusion of initial state features s and annealing process vector u, and a decoding module for outputting the ODF slice map after annealing. The reverse design network G includes an image encoder for feature extraction of the target ODF slice image and a multilayer perceptron regression module for outputting annealing process vectors.

9. The method for rapid prediction of microstructure evolution and reverse engineering of high-temperature annealing of oriented electrical steel based on cellular automata and bidirectional consistency deep learning surrogate model as described in claim 7, is characterized in that... Step S5 specifically includes: For the target ODF slice image Or the texture template obtained by mapping the target index, will Inputting the reverse engineering network G yields the initial process values. ; With the aforementioned initial process values As the initial solution, under the constraints of process upper / lower limits and heating / cooling rates, we further minimize... To obtain the optimal process that satisfies the target texture. .

10. The method for rapid prediction of microstructure evolution and reverse engineering of high-temperature annealing of oriented electrical steel based on cellular automata and bidirectional consistency deep learning surrogate model as described in claim 6, characterized in that: The heat preservation temperature T is [900℃, 1300℃], and the heat preservation time t is [10min, 3000min]; The resolution of the ODF slice image is 64×64 or 128×128; The forward prediction network F and / or the reverse design network G include 2 to 6 fully connected layers, with 64 to 1024 hidden units per layer.

11. A system for rapidly predicting the evolution of high-temperature annealed microstructure and reverse engineering the process of oriented electrical steel based on cellular automata and bidirectional consistency deep learning surrogate model as described in any one of claims 1-10, characterized in that, include: The data acquisition and preprocessing module is used to acquire and organize the initial state characteristics of the oriented electrical steel before entering high-temperature annealing and the corresponding annealing process information, forming a consistent input for subsequent modeling. The CA virtual experiment data generation module is used to run cellular automata mechanism models in batches within a preset process design space and construct a "virtual labeled" training dataset. The bidirectional surrogate model training and inference module is used to build, train and deploy bidirectional consistency deep learning surrogate models to achieve rapid forward prediction and verifiable reverse process inference of high-temperature annealing texture. The target texture process inversion and constraint optimization module is used to realize the reverse design from the target texture to the executable annealing process path, and output the optimal process solution under multiple process constraints and quality risk constraints; The closed-loop verification and online incremental learning module is used to apply the inverted optimal process. The experiment is performed in an experimental furnace or production line, and the measured results are used to perform closed-loop verification and continuous updates on the CA virtual experimental data generation module and the bidirectional agent model training and inference module, thereby improving cross-batch robustness and long-term availability.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed on the processor, it implements the steps of the method for rapid prediction of microstructure evolution and reverse engineering of high-temperature annealing of oriented electrical steel based on cellular automata and bidirectional consistency deep learning agent model as described in any one of claims 1-10.

Images