An oriented electrical steel cold rolling texture rapid prediction method and system based on crystal plasticity finite element and sequential condition generation type deep learning agent model and a medium

Abstract

The application discloses a kind of based on crystal plastic finite element and sequence condition generation formula deep learning agent model's oriented electrical steel cold rolling texture fast prediction method, system and medium, with cold rolling process parameter set as input, with ODF slice image as output;Through the crystal plastic finite element model driven by physical mechanism, "process path-ODF" virtual labeling data pair is generated in batch in pre-set cold rolling process design space, and data set is constructed;Then using sequence condition generation formula deep learning agent model realizes the end-to-end nonlinear mapping from cold rolling process path to high-dimensional ODF image, so that the premise of not needing to call crystal plastic finite element model successively, realize the fast prediction, visualization and path regulation of cold rolling texture evolution.This application solves the problem of existing fusion crystal plastic simulation precision advantage and deep learning model high efficient generalization ability, realizes the fast and accurate prediction of cold rolling texture evolution.

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 the texture of cold-rolled oriented electrical steel based on crystal plasticity finite element method and sequential conditional generative deep learning surrogate model. Background Technology

[0002] In the preparation process of grain-oriented electrical steel (typical silicon content approximately 3.0%~3.5%), cold rolling is one of the key processes determining the texture formation path and final magnetic properties. Cold rolling introduces high-density dislocations and orientation rotation through multiple plastic deformation passes, facilitating Goss orientation ({110}) during subsequent decarburization, nitriding, and final secondary recrystallization processes. <001> The formation of the crystal provides the necessary deformation energy storage and orientation basis. The effect of cold rolling process is not only affected by parameters such as total reduction rate, reduction rate distribution of each pass, rolling speed, tension and lubrication state, but also closely coupled with the initial texture and microstructure of the normalized plate before cold rolling. If the texture evolution during the cold rolling stage is not finely controlled, it will lead to insufficient orientation selectivity in subsequent recrystallization, causing fluctuations in key properties such as magnetic induction and iron loss, thus significantly affecting product stability and yield.

[0003] Currently, the characterization, prediction, and process optimization of cold-rolled texture of grain-oriented electrical steel mainly rely on the following technical routes, but all of them have significant drawbacks:

[0004] (1) Empirical trial and error method: In industry, process parameters (such as pass reduction rate, total reduction rate, rolling speed, etc.) are often adjusted through repeated rolling tests to obtain ideal texture and performance. This method relies heavily on the accumulation of engineers' experience, has a long test cycle and high cost. Especially in the case of multiple passes and multiple parameters, it is difficult to cover the high-dimensional and nonlinear process space, and often can only obtain local optimal solutions, lacking standardized design basis that can be transferred and reused.

[0005] (2) Experimental characterization method based on static before-and-after comparison: Since the cold rolling process occurs inside the material and evolves rapidly, the dynamic evolution of microstructure and texture during the rolling process is difficult to observe in situ through conventional experiments. Existing studies generally rely on sampling before and after cold rolling and using methods such as EBSD and XRD for offline characterization, indirectly inferring the evolution mechanism through "static before-and-after comparison". This method not only suffers from time delay and incomplete information, but also cannot obtain the continuous trajectory of the evolution of secondary textures, thus making it difficult to support a fine understanding of the texture formation path and process control.

[0006] (3) Pure physical simulation methods (represented by crystal plasticity finite element method CPFEM): Crystal plasticity finite element method can describe the stress-strain distribution, crystal orientation rotation and texture formation mechanism of materials during cold rolling at the grain scale. Its simulation results usually have good consistency with experimental observations. However, this type of method has complex model construction, parameter calibration and computational overhead. When it is necessary to perform large-scale scanning of different initial microstructures, different pass reduction ratios and multiple process paths, it often takes several hours or even several days of calculation time, which is difficult to meet the needs of rapid prediction and real-time decision-making in industrial sites. At the same time, CPFEM usually has strong forward prediction ability (deriving microstructure from process), but it is difficult to reverse derive the optimal cold rolling process path from the target texture under complex constraints, and its inversion design capability is insufficient.

[0007] (4) Pure data-driven machine learning models: In recent years, machine learning methods have been used for material microstructure prediction due to their ability to fit nonlinear laws, but they essentially rely on a large amount of high-quality labeled data. In the cold rolling scenario of oriented electrical steel, the acquisition of high-quality post-cold rolling EBSD data is costly, time-consuming, and limited by on-site conditions, resulting in scarce training data and obvious batch correlation. Under small sample conditions, traditional machine learning models have poor generalization ability, and the prediction accuracy is prone to significant decline when faced with different initial microstructures or process combinations. Furthermore, it is difficult to effectively integrate physical mechanism knowledge, resulting in insufficient credibility and interpretability.

[0008] Although existing methods for predicting cold-rolled texture and designing processes have made some progress in experimental research and numerical simulation, they still cannot simultaneously meet the industry's core requirements for "fast prediction speed, accurate prediction results, and strong path design capabilities" in engineering applications, forming a significant technical bottleneck: On the one hand, cold-rolled texture evolution is highly path-dependent, and process parameters include not only scalars such as total reduction rate, but also sequential information such as the reduction rate distribution of each pass; on the other hand, texture results are usually presented in high-dimensional distribution forms such as ODF, and their changes are extremely sensitive to subsequent magnetic properties, which significantly increases the difficulty of mapping modeling "high-dimensional input - high-dimensional output".

[0009] In summary, the current prediction and control of cold-rolled texture of grain-oriented electrical steel faces the contradiction of "credible mechanism but slow calculation" and "fast calculation but scarce and unrobust data": how to integrate the accuracy advantages of crystal plasticity simulation with the efficient generalization ability of deep learning models, achieve rapid and accurate prediction of cold-rolled texture evolution under the real constraint of not being able to observe in situ, and further support the control of process paths guided by target texture has become a key problem that urgently needs to be solved in this field. Summary of the Invention

[0010] To address the shortcomings of existing technologies, the purpose of this invention is to provide a method, system, and medium for rapid prediction of cold-rolled texture of oriented electrical steel based on crystal plasticity finite element method and sequential conditional generative deep learning proxy model. This invention solves the problems of combining the accuracy advantages of existing crystal plasticity simulation with the efficient generalization ability of deep learning models, enabling rapid and accurate prediction of cold-rolled texture evolution under the real-world constraint of not being able to observe in situ, and further supporting the problem of process path control guided by target texture.

[0011] To achieve the above objectives, the present invention adopts the following technical solution:

[0012] The first aspect of this invention provides a method for rapid prediction of cold-rolled texture of oriented electrical steel based on crystal plasticity finite element method and sequential conditional generative deep learning surrogate model;

[0013] The cold rolling process parameters are expressed as a set of process vectors consisting of "pass sequence + global constraints" as input, and the output is an ODF slice image with a discrete fixed pixel size.

[0014] Using a physics-driven crystal plasticity finite element model, virtual labeled data pairs of "process path-ODF" are generated in batches within a preset cold rolling process design space. Measured texture data after cold rolling is introduced to calibrate the output of the crystal plasticity finite element model to construct a dataset.

[0015] The pass sequence is processed by a fixed-length unified mapping. The pass process sequence is temporally encoded by a sequence conditional generative deep learning surrogate model. Global process constraint features are fused to generate conditional vectors, which are then generated by convolutional decoding to produce the ODF slice image. This achieves an end-to-end nonlinear mapping from the cold rolling process path to the high-dimensional ODF image, thereby enabling rapid prediction, visualization, and path control of cold rolling texture evolution without having to call the crystal plasticity finite element model one by one.

[0016] Preferably, the method for rapid prediction of cold-rolled texture of grain-oriented electrical steel includes the following steps:

[0017] Step S1: Multi-source data acquisition and preprocessing;

[0018] Step S2: Generate a dataset based on the crystal plasticity finite element model;

[0019] Step S3, the construction and training of the sequence conditional generative deep learning agent model;

[0020] Step S4: Rapid prediction and visualization of the ODF slice image under the conditions of cold rolling process parameters.

[0021] Preferably, step S1 specifically includes:

[0022] Electron backscatter diffraction data of the normalized plate before cold rolling were collected, with a scanning step size of 0.5 μm to 2.0 μm;

[0023] Micro-region texture features, including initial grain topology, crystallographic orientation information, grain boundary type, and orientation difference distribution, are extracted from the electron backscatter diffraction data, and a digital characterization of the initial microstructure of cold-rolled grains is constructed accordingly.

[0024] The cold rolling process parameters are represented as a set of process vectors consisting of a "pass sequence + global constraints". .

[0025] Preferably, the cold rolling process parameters include the total reduction rate R. total Number of passes N, and sequence of reduction rates for each pass .

[0026] Preferably, the cold rolling process parameters also include rolling speed v. i Tension level σ i Lubrication / friction rating, intermediate annealing markings, and heat treatment parameters.

[0027] Preferably, the trace sequence is uniformly mapped to a fixed-length representation:

[0028] Set the maximum number of passes N max The typical value is 6, and the number of times the current path is N < N max At that time, zero-padding is performed at the end of the pass sequence to make the sequence length uniformly N. max Simultaneously, a Boolean attention mask is constructed, with the position corresponding to the valid channel being 1 and the position corresponding to the fill channel being 0, to ensure that the model only calculates the valid channel;

[0029] Number of times N > N max At that time, an equivalent sequence is generated by merging adjacent small track passes using the same pressure reduction rate;

[0030] The merging rule is as follows: adjacent passes with a reduction rate of less than 10% are merged first. The reduction rate of the merged pass is the sum of the reduction rates of all passes before merging. The process parameters are taken as the average value of all passes before merging, until the sequence length equals N. max ;

[0031] All cold rolling process parameters and output target data are standardized, normalized, and missing values ​​are processed to obtain normalized process inputs. .

[0032] Preferably, the crystal plastic finite element model in step S2 is used to simulate the stress-strain response, crystal orientation rotation, and texture evolution behavior of grain-oriented electrical steel during multi-pass cold rolling.

[0033] Preferably, the crystal plastic finite element model integrates the following physical mechanisms:

[0034] Slip systems and flow laws, based on body-centered cubic models, use viscoplastic slip rates to describe the shear strain rate of each slip system:

[0035]

[0036] in, To decompose the shear stress in the slip system, For slip resistance, and These are the reference shear rate and strain rate sensitivity indices, respectively.

[0037] Hardening evolution is updated using isotropic or latent hardening forms to characterize the accumulation of deformation energy storage and intragranular hardening behavior during cold rolling.

[0038] Orientation update and texture statistics: The crystal orientation is updated at the integration point scale by the crystal rotation update rule to obtain the orientation set after cold rolling, and the orientation distribution function ODF is calculated based on the orientation set.

[0039] Preferably, step S2 specifically includes:

[0040] The crystal plasticity finite element model is run in batches within the process design space to generate "virtual-annotated" data pairs;

[0041] For each set of process vectors The crystal plasticity finite element model outputs the crystal orientation set after cold rolling, which is used to obtain the ODF (Optical Distribution Format) using a texture calculation tool. Furthermore, the ODF slice image at a specified Euler angle section is extracted, and the ODF slice image is discretized into a two-dimensional field with a pixel size of 64*64, denoted as the true label image. It can be flattened into a 4096-dimensional vector. .

[0042] Preferably, step S2 further includes:

[0043] We used measured texture data after cold rolling to perform systematic bias calibration on the output of the crystal plasticity finite element model.

[0044] The calibration method is as follows: using the ODF calculated from the measured EBSD or XRD data as a benchmark, the intensity correction coefficient and peak position offset correction matrix of the CPFEM output ODF are obtained by fitting with the least squares method. The ODF data generated in batches by CPFEM are corrected point by point so that the relative deviation between the peak intensity and peak position of the key texture components of the virtual data and the measured data is less than 8%, thereby improving the physical fidelity and reliability of the virtual data.

[0045] Preferably, step S3 includes the following steps:

[0046] Step S31: The normalized process feature sequence of each pass. The input is fed into the embedding layer, which maps the features of each passage to a fixed-dimensional passage embedding vector. To preserve the path dependency of the cold rolling pass sequence, position encoding is superimposed on the embedding vector. Construct sequence representation ;

[0047] Step S32: The global feature vector composed of total reduction rate, number of passes, and other global process constraint parameters is generated. Input linear mapping layer to obtain and combine it with the global latent vector After concatenation, a conditional vector is obtained through nonlinear mapping. This vector is used to characterize the combined effect of "cold rolling path + global constraints" and serves as the conditional input for subsequent ODF image generation.

[0048] Step S33, condition vector The algorithm maps the data to a low-resolution two-dimensional feature map using linear layers, and then generates an ODF prediction image with a target resolution of 1×64×64 through multi-level upsampling and two-dimensional convolution. ;

[0049] Step S34: The mean squared error is used as the basic loss function to measure the predicted ODF slice image. Compared with real ODF slice images The difference between them can be expressed in the following form:

[0050]

[0051] During training, a gradient descent-based optimization algorithm is used, and the learning rate, batch size, and number of training rounds are set. The model parameters are iteratively optimized on the training set, and the loss changes are monitored on the validation set. When the loss on the validation set converges or no longer decreases significantly, the training ends and the final model weights are saved.

[0052] Preferably, in step S31, a Transformer encoder is used to process the sequence. Modeling is performed by capturing the coupling effects between passes and the global process path features through a multi-head self-attention mechanism, and the influence of fill passes is ignored by combining attention masks; the encoded result is a sequence latent representation. Then through pooling or Aggregation yields the global latent vector of the process path. ; and / or in step S33, the upsampling path is 8→16→32→64, and each stage consists of "upsampling + Conv2D + nonlinear activation";

[0053] A Sigmoid activation function is applied after the output layer to restrict pixel values ​​to the [0,1] interval, matching the numerical range of the ODF intensity after min-max normalization in step S1, thus obtaining the final prediction result. ; and / or in step S34, a structural similarity index is introduced as an auxiliary term to impose additional constraints on the local contrast and texture continuity of the predicted image, thereby improving the visual structural consistency and physical rationality of the predicted ODF slice image; or a weighted loss for key texture components is introduced to improve the prediction accuracy of key peak intensity and peak position.

[0054] Preferably, step S4 includes the following steps:

[0055] Step S41: The input cold rolling process parameters are standardized / normalized in the same way as in step S1 to obtain normalized process input. And according to the maximum number of passes N max The rules complete sequence padding and attention mask construction;

[0056] Step S42, input the normalization process The input is fed into the pre-trained sequence conditional generative deep learning proxy model, and the predicted ODF slice image is obtained through the sequence encoding and convolutional decoding modules. ;

[0057] Step S43, for the predicted ODF slice image Perform denormalization to restore the original ODF intensity dimensions;

[0058] Step S44, the predicted ODF slice image As The ODF slices of the specified cross section are visualized and output, and the texture indices of Goss component intensity and key orientation area fraction are further calculated for process window evaluation and comparative analysis.

[0059] The second aspect of this invention provides a fast prediction system for the cold-rolled texture of oriented electrical steel, for implementing the fast prediction method for cold-rolled texture of oriented electrical steel based on crystal plasticity finite element method and sequential conditional generative deep learning surrogate model provided in the first aspect of this invention, comprising:

[0060] The data acquisition and virtual experiment module is used to acquire the initial electron backscatter diffraction data of the normalized plate before cold rolling, construct the crystal plasticity finite element model, and generate virtual data pairs of "process path-ODF" within the preset process design space.

[0061] The data preprocessing and encoding module is used to encode the process sequence of each pass with a uniform length, standardize / normalize the global cold rolling process parameters and ODF slice images, and discretize and store the ODF slice images to form a dataset that can be used for deep learning training.

[0062] The surrogate model training and prediction module is used to build and train a sequence conditional generative deep learning surrogate model, and after deployment, it enables fast forward prediction from cold rolling process path to ODF slice image.

[0063] The process path scanning and visualization module is used to call proxy models in batches within a given process range to generate ODF prediction results, and present them in the form of images or statistical features to assist engineers in process window evaluation, path comparison and preliminary screening.

[0064] Preferably, the rapid prediction system for the cold-rolled texture of grain-oriented electrical steel further includes:

[0065] The integration module interfaces with production line control or process design software to embed candidate process paths and predicted texture results into existing process development processes.

[0066] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed on a processor, implements the steps of the method for rapid prediction of cold-rolled texture of oriented electrical steel based on crystal plasticity finite element and sequential conditional generative deep learning surrogate model provided in the first aspect of the present invention.

[0067] The present invention provides a method, system and medium for rapid prediction of cold-rolled texture of oriented electrical steel based on crystal plasticity finite element and sequential conditional generative deep learning proxy model. It involves modeling and microstructure control of the rolling process of oriented electrical steel, for rolling process design and rapid optimization, as well as digital research and verification of the microstructure evolution mechanism, and has the following beneficial effects: (1) It significantly reduces the dependence on large-scale measured texture data and alleviates the problem of scarce cold rolling process data and the inability to observe in situ.

[0068] Unlike traditional data-driven machine learning methods that require a large amount of post-cold rolling EBSD (or even secondary EBSD) data, this invention primarily uses CPFEM to generate virtual annotation data of "process path-texture" in batches, supplemented by a small amount of measured EBSD or XRD macro-texture data for deviation calibration. Thus, without relying on high-throughput, low-cost measured annotation data that is difficult to obtain, it is still possible to construct a training set with physical credibility, significantly reducing the threshold for model building and improving the scalability of the method in industrial settings.

[0069] (2) The complex CPFEM multi-pass cold rolling forward modeling calculation is compressed into a sequence condition generation surrogate model, which greatly improves the prediction speed and process space exploration efficiency.

[0070] While CPFEM can accurately describe crystal orientation rotation and texture evolution, simulating a single multi-pass cold rolling process takes 2-8 hours, making it difficult to perform batch scanning of a large number of process paths. This invention, by training a sequence conditional generation network, "distills" the complex physical forward modeling process of multi-pass cold rolling into a surrogate model capable of rapid inference. The ODF prediction time for a single process is ≤10ms, representing a speedup of over 10 times compared to traditional CPFEM simulation. 4 It can perform batch prediction and comparative evaluation of tens of thousands of candidate cold rolling paths within 1 minute, which can fully match the timeliness requirements of industrial production cycle and process development cycle, and significantly improve the speed of process optimization and on-site response.

[0071] (3) Under the premise of maintaining the integrity of the texture information, the end-to-end mapping of "cold rolling process path sequence → full field ODF image" is realized, which naturally adapts to the path dependency problem and supports the texture target-oriented control.

[0072] The evolution of cold-rolled texture exhibits significant pass-path dependence. Traditional models simplify the process to a few scalars, resulting in texture prediction errors exceeding 20% ​​for different pass allocation schemes, failing to accurately characterize the path-dependent effect. This invention encodes the process input as a sequence of pass reduction rates plus total reduction rate (and optional global constraints), explicitly learning the coupling effects between passes through a sequence modeling mechanism. The average error in texture prediction for different pass paths is ≤5%. Simultaneously, it outputs a 64×64 ODF full-field slice, completely preserving all information about the texture distribution, unlike existing technologies that only output texture statistics. This end-to-end "sequence to two-dimensional field" mapping method not only improves the ability to express complex cold-rolling paths but also provides a unified data interface and engineering foundation for subsequent extraction of key texture indicators, process path selection, quasi-inversion control, and coupling with magnetic property models. Attached Figure Description

[0073] Figure 1 This is a flowchart illustrating the rapid prediction method for cold-rolled texture of grain-oriented electrical steel according to the present invention.

[0074] Figure 2 This is a schematic diagram of the CPFEM "virtual experiment" module structure and DoE sampling in the rapid prediction method for cold-rolled texture of oriented electrical steel of the present invention;

[0075] Figure 3 This is a schematic diagram of the deep learning prediction model architecture used in Example 2 of the rapid prediction method for cold-rolled texture of oriented electrical steel of the present invention;

[0076] Figure 4 This is a schematic diagram comparing the ODF diagram predicted by the deep learning model in Example 2 of the rapid prediction method for cold-rolled texture of oriented electrical steel of the present invention with the ODF effect obtained by actual simulation by CPFEM. Detailed Implementation

[0077] 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.

[0078] Combination Figure 1 As shown, this invention provides a fast prediction method for the cold-rolled texture of oriented electrical steel based on crystal plasticity finite element method and sequential conditional generative deep learning surrogate model.

[0079] The cold rolling process parameters are used as input, and the ODF slice image with a pixel size of 64×64 is used as output.

[0080] The physical mechanism-driven crystal plasticity finite element model (CPFEM) generates virtual labeled data pairs of "process path-ODF (64×64)" in batches within the preset cold rolling process design space (DoE). A small amount of measured EBSD or XRD macrotexture data can be introduced to calibrate the CPFEM output for systematic deviation, thereby constructing a dataset with physical credibility.

[0081] Then, a sequential conditional generative deep learning surrogate model is used to realize the end-to-end nonlinear mapping from the cold rolling process path (sequence input) to the high-dimensional ODF image (two-dimensional output field), thereby achieving rapid prediction, visualization and path control of cold rolling texture evolution without having to call the crystal plasticity finite element model one after another.

[0082] ODF slice images are generated from raw crystallographic orientation data acquired by EBSD or XRD. Specifically, when using EBSD data, the measured crystal orientation is first represented by Bunge Euler angles (φ1, Φ, φ2), and the discrete orientation samples are statistically analyzed. Then, the three-dimensional orientation distribution function (ODF) is calculated using kernel density estimation or equivalent orientation distribution reconstruction methods. Next, a preset fixed φ2 section is selected, the corresponding two-dimensional ODF slice is extracted, and interpolation discretization is performed on a regular grid. Finally, the discretized ODF intensity matrix is ​​normalized and converted into an image representation with a fixed pixel size.

[0083] The ODF slice image is discretized into a 64×64 two-dimensional field.

[0084] When using XRD data, ODF is first obtained by pole figure inversion, and then ODF slice images are formed according to the above-mentioned fixed section extraction, regular grid discretization and normalization methods.

[0085] The rapid prediction method for cold-rolled texture of grain-oriented electrical steel of the present invention specifically includes the following steps:

[0086] Step S1: Multi-source data acquisition and preprocessing;

[0087] Electron backscatter diffraction (EBSD) data of the pre-cooled sheet were collected with a scan step size of 0.5 μm to 2.0 μm.

[0088] Micro-region texture features such as initial grain topology, crystallographic orientation information, grain boundary type and orientation difference distribution are extracted from electron backscatter diffraction data, and a digital characterization of the initial microstructure of cold-rolled material is constructed accordingly.

[0089] The cold rolling process parameters are represented as a set of process vectors consisting of a "pass sequence + global constraints". .

[0090] Cold rolling process parameters should include at least the total reduction rate R. total Number of passes N, and sequence of reduction rates for each pass .

[0091] Cold rolling process parameters may also include rolling speed v. i Tension level σ i Lubrication / friction rating, intermediate annealing markings, and heat treatment parameters, etc.

[0092] To adapt to the input of deep learning models, this invention maps variable-length trace sequences to a fixed-length representation:

[0093] Set the maximum number of passes N max The number of times N < N maxAt that time, zero-padding is applied to the path sequence, and an attention mask is constructed simultaneously to ensure that the model only computes valid paths; when N > N max At that time, an equivalent sequence is generated by merging adjacent small passes or by resampling;

[0094] All cold rolling process parameters and output target data are standardized, normalized, and missing values ​​are processed to obtain normalized process inputs. .

[0095] Step S2, generating a dataset based on the crystal plasticity finite element model, such as... Figure 2 As shown;

[0096] A Crystal Plasticity Finite Element Method (CPFEM) model based on crystal plasticity constitutive model was constructed to simulate the stress-strain response, crystal orientation rotation, and texture evolution behavior of grain-oriented electrical steel during multi-pass cold rolling.

[0097] The finite element model of crystal plasticity integrates the following physical mechanisms:

[0098] Slip systems and flow laws (crystal plastic dynamics), based on body-centered cubic (e.g., {110}) <111> {112} <111> (etc.), the shear strain rate of each slip system is described in the form of viscoplastic slip rate:

[0099]

[0100] in, To decompose the shear stress in the slip system, This represents the slip resistance (hardening variable). and These are the reference shear rate and strain rate sensitivity indices, respectively.

[0101] Hardening evolution (dislocation density / critical shear stress evolution) is updated using isotropic or latent hardening forms to characterize the accumulation of deformation energy storage and intragranular hardening behavior during cold rolling (the hardening model commonly used by practitioners in the fields of Voce, Kocks-Mecking can be selected).

[0102] Orientation update and texture statistics: The crystal orientation is updated at the integration point scale by the crystal rotation update rule to obtain the orientation set after cold rolling, and the orientation distribution function ODF is calculated based on the orientation set.

[0103] Batch run crystal plasticity finite element models within the process design space (DoE) to generate “virtual-annotated” data pairs;

[0104] DoE space should at least cover the total downsizing rate. With the pass reduction rate allocation sequence The model can be combined in various ways; for example, it can cover a total reduction rate of 50% to 90%, a single-pass reduction rate of 5% to 40%, and a number of passes of 2 to 8 (the specific range can be set according to the production line capacity and can be expanded), and allows the introduction of additional process variables such as tension, friction and intermediate annealing to enhance the applicability of the model.

[0105] For each set of process vectors The crystal plasticity finite element model outputs the set of crystal orientations after cold rolling. The ODF is obtained using texture calculation tools (such as Euler angle statistics and kernel density estimation), and further, a specified Euler angle section (e.g., ...) is extracted. The ODF slice image (=45°) is discretized into a two-dimensional field with a pixel size of 64*64, denoted as the ground truth label image. It can be flattened into a 4096-dimensional vector. .

[0106] In step S2, a small amount of measured cold-rolled texture data (EBSD or XRD macro-texture) can be introduced to perform systematic deviation calibration on the crystal plastic finite element model or its output (such as correcting the peak intensity and peak position of key texture components) to improve the physical fidelity and reliability of the virtual data.

[0107] Step S3, construction and training of the sequence conditional generative deep learning surrogate model, such as... Figure 3 As shown;

[0108] This invention employs a sequence-conditional generative deep learning surrogate model to directly learn the mapping relationship between the cold rolling process sequence input and the high-dimensional ODF image output, thereby achieving... arrive End-to-end nonlinear prediction. The model consists of a "process sequence encoding module + global feature fusion module + two-dimensional convolutional decoding and generation module". Specifically, it includes the following steps:

[0109] Step S31, the process sequence encoding module (preferably Transformer Encoder) encodes the normalized pass process feature sequence. The input is fed into the embedding layer, which maps the features of each passage to a fixed-dimensional passage embedding vector. To preserve the path dependency of the cold rolling pass sequence, position encoding is superimposed on the embedding vector. Construct sequence representation ;

[0110] In step S31, a Transformer encoder is used to process the sequence. Modeling is performed by capturing the coupling effects between passes and the global process path features through a multi-head self-attention mechanism, and the influence of fill passes is ignored by combining attention masks; the encoded result is a sequence latent representation. Then through pooling or Aggregation yields the global latent vector of the process path. .

[0111] Step S32, the global feature fusion module combines the total reduction rate, number of passes, and other global process constraint parameters into a global feature vector. Input linear mapping layer to obtain and combine it with the global latent vector After concatenation, a conditional vector is obtained through nonlinear mapping. This vector is used to characterize the combined effect of "cold rolling path + global constraints" and serves as the conditional input for subsequent ODF image generation.

[0112] Step S33, the two-dimensional convolutional decoding generation module (generates a 64×64 ODF slice image from the conditional vector), and converts the conditional vector... The algorithm maps the data to a low-resolution two-dimensional feature map using linear layers, and then generates an ODF prediction image with a target resolution of 1×64×64 through multi-level upsampling and two-dimensional convolution. ;

[0113] In step S33, the upsampling path is 8→16→32→64, and each stage consists of "upsampling + Conv2D + nonlinear activation (ReLU / GELU)";

[0114] A Sigmoid activation function is applied after the output layer to restrict pixel values ​​to the [0,1] interval, matching the range of values ​​after min-max normalization of the ODF intensity in step S1, thus obtaining the final prediction result. .

[0115] Step S34: The mean squared error (MSE) is used as the basic loss function to measure the predicted ODF tile image. Compared with real ODF slice images The difference between them can be expressed in the following form:

[0116]

[0117] In step S34, a structural similarity (SSIM) index can be introduced as an auxiliary term to impose additional constraints on the local contrast and texture continuity of the predicted image, thereby improving the visual structural consistency and physical rationality of the predicted ODF slice image; or a weighted loss for key texture components (such as the Goss peak region) can be introduced to improve the prediction accuracy of key peak intensity and peak position.

[0118] During training, an optimization algorithm based on gradient descent (such as AdamW / Adam) is used, and appropriate learning rate, batch size and number of training rounds are set. The model parameters are iteratively optimized on the training set, and the loss changes are monitored on the validation set. When the loss on the validation set converges or no longer decreases significantly, the training ends and the final model weights are saved.

[0119] Step S4: Rapid prediction and visualization of ODF slice images under cold rolling process parameters;

[0120] After the surrogate model is trained, for any given cold rolling process path (total reduction rate and pass reduction rate sequence, etc.), the following steps can be used to quickly predict the ODF slice image after cold rolling:

[0121] Step S41: Standardize / normalize the input cold rolling process parameters using the same method as in step S1 to obtain the normalized process input. And according to the maximum number of passes N max The rules complete sequence padding and attention mask construction;

[0122] Step S42, input the normalization process. The input is fed into a pre-trained sequence conditional generative deep learning proxy model, and the predicted ODF slice image is obtained through sequence encoding and convolutional decoding modules. ;

[0123] Step S43, for the predicted ODF slice image Perform denormalization to restore the original ODF intensity dimensions;

[0124] Step S44: The predicted ODF slice image As The ODF slices of the specified cross section are visualized and output, and the texture indices of Goss component intensity and key orientation area fraction are further calculated for process window evaluation and comparative analysis.

[0125] Through the above steps, this invention can quickly obtain ODF slice prediction results under different cold rolling paths without repeating large-scale CPFEM calculations or a large number of trial rolling experiments, providing process engineers with an efficient tool for cold rolling path design, texture control and performance stability improvement.

[0126] The specific architecture of the deep learning proxy model in this invention adopts a sequence conditional generative deep learning proxy model composed of a process sequence encoder and a two-dimensional convolutional decoder, wherein:

[0127] The process sequence encoder is preferably a Transformer Encoder (or can be replaced by Bi-LSTM / 1D-CNN) to normalize the pass-by-pass process sequence. As input, the pass order information is preserved through pass embedding layer and position encoding, and the coupling effect between passes is learned through self-attention mechanism to obtain the potential vector representing the cold rolling path. ;

[0128] The global feature fusion module integrates global process constraint vectors such as total reduction rate and number of passes. Mapped to and with Fusion generates conditional vectors This is used to guide the generation of subsequent ODF slice images;

[0129] The two-dimensional convolutional decoder will convert the conditional vector First, it is mapped to a low-resolution two-dimensional feature map (e.g., C0×8×8), and then through multiple levels of upsampling and two-dimensional convolution, an ODF prediction map with a resolution of 64×64 is generated step by step. This allows for explicit modeling of the spatial correlation and local continuity of the ODF field;

[0130] Preferably, a Sigmoid activation function is applied after the output layer to limit the predicted pixel values ​​to a numerical range consistent with the normalization method of the training data (e.g., [0,1]), and the ODF slice image is output in a reshape / reconstruction manner, which facilitates engineering visualization and index extraction.

[0131] The data representation and loss function design in this invention specifically include:

[0132] The cold rolling process parameters are uniformly encoded as a hybrid input structure of "pass sequence + global features":

[0133] A zero-padding and attention mask mechanism is used to achieve a unified length input for variable-length pass sequences, thereby supporting process path modeling for different pass numbers;

[0134] Scalar features such as total downtime and number of passes are standardized or interval normalized and then fused with sequence coding features;

[0135] The ODF cross section after cold rolling is discretized into a 64×64 pixel two-dimensional image with a fixed spatial layout. The ODF intensity can be logarithmically transformed or normalized according to the maximum value / min-max before training to improve numerical stability and learning difficulty of peak region.

[0136] Mean squared error (MSE) is used as the basic loss function to measure the difference between the predicted ODF image and the real ODF image. Preferably, structural similarity (SSIM), L1 loss or key texture component region weighted loss can be superimposed to improve the ability to preserve the local peak intensity, peak position and texture structure of ODF, thereby enhancing the physical rationality and engineering usability of the prediction results.

[0137] The process path scanning and quasi-inversion control based on the surrogate model in this invention specifically includes:

[0138] By utilizing a pre-trained sequence conditional generative deep learning surrogate model, high-density sampling and batch inference are performed on the total reduction rate and pass reduction rate allocation sequences in a continuous process space to quickly obtain ODF slice prediction maps corresponding to a large number of candidate cold rolling paths.

[0139] By screening and sorting specific texture features of the predicted ODF (such as Goss component intensity, area fraction of specific orientation intervals, texture uniformity index, etc.), a set of process paths that meet the preset texture target is identified, providing a candidate solution set for subsequent fine optimization, coupling with magnetic performance models, or production line verification.

[0140] This enables the search and control of process paths that support near-reverse design with low computational cost, allowing for rapid process path recommendations oriented towards cold-rolled texture without large-scale CPFEM scanning calculations.

[0141] The present invention also provides a rapid prediction system for the cold-rolled texture of grain-oriented electrical steel for implementing the rapid prediction method of the present invention, comprising:

[0142] The data acquisition and virtual experiment module is used to acquire the initial electron backscatter diffraction data of the normalized plate before cold rolling, construct the crystal plasticity finite element model and generate virtual data pairs of "process path-ODF" in the preset process design space, and can also perform deviation calibration with a small amount of measured texture data.

[0143] The data preprocessing and encoding module is used to perform uniform length encoding (padding + mask) on the process sequence of each pass, standardize / normalize the global cold rolling process parameters and ODF slice images, and discretize and store the ODF slice images to form a dataset that can be used for deep learning training.

[0144] The surrogate model training and prediction module is used to build and train a sequence conditional generative deep learning surrogate model (Transformer encoding + convolutional decoding), and after deployment, it enables fast forward prediction from the cold rolling process path to the ODF slice image.

[0145] The process path scanning and visualization module is used to call proxy models in batches within a given process range to generate ODF prediction results, and present them in the form of images or statistical features to assist engineers in process window evaluation, path comparison and preliminary screening.

[0146] The rapid prediction system for cold-rolled texture of grain-oriented electrical steel of the present invention also includes:

[0147] The integrated module interfaces with production line control or process design software to embed candidate process paths and predicted texture results into the existing process development process, thereby realizing the engineering application of the fast prediction method for cold-rolled texture of oriented electrical steel in industrial settings.

[0148] Example 1

[0149] In this embodiment 1, normalized plates of oriented electrical steel from the same heat number were selected, and EBSD scanning was performed in the thickness direction of the plate with a step size of 1 μm to obtain representative initial orientation distribution and grain morphology information. Based on this, a CPFEM grain-scale initial orientation field was constructed (for subsequent texture statistical benchmarks and model parameter settings).

[0150] The design space for the cold rolling process is set as follows:

[0151] Total reduction rate range: =60%~88%;

[0152] Number of passes: N=3~7;

[0153] Single-pass reduction rate range: =8%~35%, and meets the following conditions ;

[0154] The cold rolling process parameters are encoded as process inputs consisting of "pass sequence + global features". :

[0155] The sequence of tracks is Global features are .

[0156] In this embodiment 1, the maximum number of passes N is set. max =8, zero-padding is performed on samples where the number of times N < 8, and an attention mask is constructed simultaneously to ensure that the model only calculates the true number of times. , Continuous variables are normalized to the [0,1] interval using min-max normalization to obtain normalized process inputs. .

[0157] Within the aforementioned design space, 600 sets of cold rolling process path samples were generated using a hybrid sampling method combining Latin hypercube sampling and regular grid sampling. For each process path, multi-pass cold rolling simulations were performed sequentially in the CPFEM model to obtain the orientation set after cold rolling, and the ODF was calculated. =45° cross section, and uniformly discretize the ODF slices into 64×64 pixel images as output labels. To alleviate the problem of a large dynamic range in ODF intensity, this embodiment 1 first performs a logarithmic transformation on the ODF intensity (taking log(ODF+0.01)) and then performs min-max normalization to improve numerical stability and the difficulty of learning the peak region. The CPFEM output result is directly used as the "true value" to train the surrogate model without introducing actual measurement calibration.

[0158] This embodiment 1 adopts a preferred sequence conditional generative deep learning proxy model structure:

[0159] Process sequence encoder: Transformer Encoder, embedding dimension d model =128, the number of encoding layers is 3, the number of attention heads is 4, the feedforward layer dimension is 256, and padding mask is enabled;

[0160] Global feature fusion: After mapping to 128 dimensions and concatenating with sequence features, the conditional vector is obtained through a linear layer and GELU activation. ;

[0161] Two-dimensional convolutional decoder: The initial feature map is mapped to C0×8×8, where C0=64. Then, it is upsampled sequentially to 8→16→32→64. Each level contains "upsampling + 3×3 convolution + ReLU". The output layer uses 1×1 convolution to obtain a 1-channel 64×64 image, and then uses Sigmoid to limit the output to [0,1].

[0162] The training configuration involved dividing 600 sets of dummy data into training, validation, and test sets in a 7:1.5:1.5 ratio; the loss function used was MSE, the optimizer was AdamW, and the learning rate was set to 3×10⁻⁶. -4 The batch size is 16, the maximum number of training rounds is 250, and early stopping is used on the validation set.

[0163] After training, the average pixel-level MSE of the test set remained in the range of 0.002 to 0.006, enabling rapid prediction and visualization of ODF slices under different cold rolling paths, and supporting batch scanning and preliminary screening of candidate process paths.

[0164] Example 2

[0165] In this embodiment 2, normalized plates of oriented electrical steel from the same two heat numbers were selected, and EBSD scanning was performed in the thickness direction of the plates with a step size of 0.8 μm to cover the initial texture differences. Additionally, a small number of XRD macrotexture data of cold-rolled samples were collected for calibration (8 sets per heat number, 16 sets in total) to enhance the physical consistency and cross-batch generalization ability of the virtual data.

[0166] The design space for the cold rolling process is set as follows:

[0167] Total reduction rate range: =55%~92%;

[0168] Number of passes range: N=2~8, maximum number of passes set to N max =10;

[0169] Single-pass reduction rate range: =5%~40%;

[0170] The following process variables are added in this Example 2:

[0171] Rolling speed grade (Low / Medium / High levels, coded as 0 / 0.5 / 1);

[0172] Tension level (None / Medium / High, coded as 0 / 0.5 / 1);

[0173] Friction / Lubrication Grade (Excellent / Medium / Poor, as global discrete features);

[0174] Intermediate annealing mark (Whether intermediate annealing exists is used as a global feature; if a=1, add the annealing temperature and time as input and normalize them).

[0175] The process input is encoded into a sequence of pass feature vectors. With global features Continuous variables are normalized using min-max normalization, while discrete levels are normalized using interval mapping to obtain the normalized input. For lane change samples, a padding+mask mechanism is used to unify the input length.

[0176] Within this design space, 900 process paths were generated using Latin hypercube sampling, and corresponding ODF slice labels were obtained by running CPFEM multi-pass simulations. To improve the fidelity of the virtual data, this embodiment 2 uses 16 sets of measured XRD macrotexture data to calibrate the CPFEM output for deviation: by adjusting key texture components (such as Goss, {111}) <112> {112} <111> The peak intensity and peak position of the corresponding regions are calibrated to make the macroscopic texture characteristics of the CPFEM output ODF consistent with the measured trend, thereby forming a "calibrated" virtual training set.

[0177] In terms of output representation, this embodiment 2 does not use logarithmic transformation, but directly normalizes the ODF intensity to the maximum value and then performs min-max normalization, so that the peak region occupies a higher weight in terms of numerical value, thereby strengthening the learning of key textures.

[0178] Combination Figure 3 As shown, this embodiment 2 employs a stronger preferred sequence conditional generative deep learning proxy model structure:

[0179] Transformer Encoder: Embedding dimension d model =192, the number of encoding layers is 4, the number of attention heads is 6, the feedforward layer dimension is 384, and a dropout of 0.1 is added after the encoder output to suppress overfitting;

[0180] Convolutional decoder: The initial feature map is set to C0×8×8, where C0=96, the upsampling path is 8→16→32→64, and a 3×3 convolutional residual block is added at the 32×32 layer to enhance texture representation; the output layer is connected to Sigmoid.

[0181] The loss function uses a composite loss:

[0182]

[0183] SSIM is used to constrain local structural similarity, improving the peak shape and texture continuity of ODF images. The training configuration involves dividing 900 datasets into training / validation / test sets in an 8:1:1 ratio; the optimizer used is AdamW with a learning rate of 2×10⁻⁶. -4 The batch size is 12, the maximum number of training rounds is 300, and early stopping is used on the validation set.

[0184] After training, the MSE of the test set in Example 2 decreased by approximately 15% to 30% compared to Example 1, while the SSIM index in the key texture peak region was significantly improved. ODF microstructure prediction was performed on a 6-pass, 90% reduction cold rolling process; ODF comparisons are attached. Figure 4As shown, the overall pixel-level MSE is comparable to the average level of the test set. Furthermore, it maintains stable prediction accuracy in cross-furnace number tests, demonstrating that the optimal approach of "a small amount of experimental calibration + expanded process variables" can enhance the model's robustness and engineering applicability under real-world production variations.

[0185] Example 3

[0186] In this embodiment 3, normalized plates of oriented electrical steel from the same heat number were selected, and the EBSD scan step size was 2μm. The cold rolling process was fixed at 5 passes (N=5), and only the reduction ratio distribution of each pass and the total reduction ratio were adjusted. The design space was set as follows:

[0187] Total reduction rate range: =65%~85%;

[0188] Single-pass reduction rate range: =10%~30%, and meets the total reduction rate constraint;

[0189] Since the number of passes is fixed, in this embodiment 3, the input is directly encoded into a fixed-length vector:

[0190]

[0191] And obtained by min-max normalization No padding or mask is needed.

[0192] Within this design space, 500 process paths were generated using a combination of regular grids and random sampling. Each path was then subjected to a multi-pass CPFEM simulation to obtain ODF slice labels. The output representation was maintained as... For a 64×64 pixel image of the ODF at a 45° cross-section, this embodiment does not perform logarithmic transformation, but only min-max normalization.

[0193] This embodiment 3 employs an encoder replacement scheme: the preferred Transformer encoder is replaced with a Bi-LSTM sequence encoder to verify the substitutability of the sequence modeling module.

[0194] Bi-LSTM encoder: hidden dimension 128, number of layers 2, trace sequence Encode into sequence features and pool to obtain ;

[0195] Global feature fusion: and splicing ;

[0196] Convolutional decoder: Employs a lighter structure, The mapping is C0×8×8, where C0=48, and the upsampling is 8→16→32→64. Each stage uses only one 3×3 convolutional layer and ReLU; the output layer is connected to Sigmoid.

[0197] Regarding the loss function, to enhance the prediction accuracy of key texture components, this embodiment 3 introduces a key region-weighted MSE: Regions of interest (ROIs) corresponding to the Goss orientation are predefined in the ODF image. Pixel errors within the ROI are assigned higher weights (e.g., weight 3), while non-ROI regions have a weight of 1, thus constructing a weighted loss.

[0198]

[0199] in, Set the value to 3 in the ROI region and 1 in the rest.

[0200] The training configuration was to split 500 datasets into training / validation / test sets in a 7:1.5:1.5 ratio; the optimizer used was Adam, with a learning rate of 5×10⁻⁶. -4 The batch size is 20, the maximum number of training rounds is 200, and early stopping is used on the validation set.

[0201] After training, the overall MSE in the test set of Example 3 is slightly higher than that of Example 1, but the error in the Goss ROI region is significantly reduced, which can meet the engineering screening requirement of "target texture peak priority". Example 3 proves that under the conditions of fixed number of passes and simplified process variables, even with the non-preferred solution of using Bi-LSTM to replace Transformer, it is still possible to quickly predict the ODF image from the cold rolling process, and the target texture can be enhanced and guided through the weighted loss of key regions.

[0202] 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 fast prediction method for cold-rolled texture of oriented electrical steel based on crystal plasticity finite element method and sequential conditional generative deep learning surrogate model, characterized in that: The cold rolling process parameters are expressed as a set of process vectors consisting of "pass sequence + global constraints" as input, and the output is an ODF slice image with a discrete fixed pixel size. Using a physics-driven crystal plasticity finite element model, virtual labeled data pairs of "process path-ODF" are generated in batches within a preset cold rolling process design space. Measured texture data after cold rolling is introduced to calibrate the output of the crystal plasticity finite element model to construct a dataset. The pass sequence is processed by a fixed-length unified mapping. The pass process sequence is temporally encoded by a sequence conditional generative deep learning surrogate model. Global process constraint features are fused to generate conditional vectors, which are then generated by convolutional decoding to produce the ODF slice image. This achieves an end-to-end nonlinear mapping from the cold rolling process path to the high-dimensional ODF image, thereby enabling rapid prediction, visualization, and path control of cold rolling texture evolution without having to call the crystal plasticity finite element model one by one.

2. The method for rapid prediction of cold-rolled texture of oriented electrical steel based on crystal plasticity finite element method and sequential conditional generative deep learning surrogate model according to claim 1, characterized in that, The rapid prediction method for cold-rolled texture of grain-oriented electrical steel includes the following steps: Step S1: Multi-source data acquisition and preprocessing; Step S2: Generate a dataset based on the crystal plasticity finite element model; Step S3, the construction and training of the sequence conditional generative deep learning agent model; Step S4: Rapid prediction and visualization of the ODF slice image under the conditions of cold rolling process parameters.

3. The method for rapid prediction of cold-rolled texture of oriented electrical steel based on crystal plasticity finite element method and sequential conditional generative deep learning surrogate model according to claim 2, characterized in that, Step S1 specifically includes: Electron backscatter diffraction data of the normalized plate before cold rolling were collected, with a scanning step size of 0.5 μm to 2.0 μm; Micro-region texture features, including initial grain topology, crystallographic orientation information, grain boundary type, and orientation difference distribution, are extracted from the electron backscatter diffraction data, and a digital characterization of the initial microstructure of cold-rolled grains is constructed accordingly. The cold rolling process parameters are represented as a set of process vectors consisting of a "pass sequence + global constraints". .

4. The method for rapid prediction of cold-rolled texture of oriented electrical steel based on crystal plasticity finite element and sequential conditional generative deep learning surrogate model according to claim 3, characterized in that: The cold rolling process parameters include the total reduction rate R. total Number of passes N, and sequence of reduction rates for each pass .

5. The method for rapid prediction of cold-rolled texture of oriented electrical steel based on crystal plasticity finite element method and sequential conditional generative deep learning surrogate model according to claim 4, characterized in that: The cold rolling process parameters also include the rolling speed v. i Tension level σ i Lubrication / friction rating, intermediate annealing markings, and heat treatment parameters.

6. The method for rapid prediction of cold-rolled texture of oriented electrical steel based on crystal plasticity finite element method and sequential conditional generative deep learning surrogate model according to claim 3, characterized in that, The trace sequence is uniformly mapped to a fixed-length representation: Set the maximum number of passes N max The number of times N < N max At that time, zero-padding is performed at the end of the pass sequence to make the sequence length uniformly N. max Simultaneously, a Boolean attention mask is constructed, with the position corresponding to the valid channel being 1 and the position corresponding to the fill channel being 0, to ensure that the model only calculates the valid channel; Number of times N > N max At that time, an equivalent sequence is generated by merging adjacent small track passes using the same pressure reduction rate; The merging rule is as follows: adjacent passes with a reduction rate of less than 10% are merged first. The reduction rate of the merged pass is the sum of the reduction rates of all passes before merging. The process parameters are taken as the average value of all passes before merging, until the sequence length equals N. max ; All cold rolling process parameters and output target data are standardized, normalized, and missing values ​​are processed to obtain normalized process inputs. .

7. The method for rapid prediction of cold-rolled texture of oriented electrical steel based on crystal plasticity finite element method and sequential conditional generative deep learning surrogate model according to claim 2, characterized in that: The crystal plastic finite element model described in step S2 is used to simulate the stress-strain response, crystal orientation rotation, and texture evolution behavior of grain-oriented electrical steel during multi-pass cold rolling.

8. The method for rapid prediction of cold-rolled texture of oriented electrical steel based on crystal plasticity finite element method and sequential conditional generative deep learning surrogate model according to claim 7, characterized in that: The crystal plastic finite element model integrates the following physical mechanisms: Slip systems and flow laws, based on body-centered cubic models, use viscoplastic slip rates to describe the shear strain rate of each slip system: ； in, To decompose the shear stress in the slip system, For slip resistance, and These are the reference shear rate and strain rate sensitivity indices, respectively. Hardening evolution is updated using isotropic or latent hardening forms to characterize the accumulation of deformation energy storage and intragranular hardening behavior during cold rolling. Orientation update and texture statistics: The crystal orientation is updated at the integration point scale by the crystal rotation update rule to obtain the orientation set after cold rolling, and the orientation distribution function ODF is calculated based on the orientation set.

9. The method for rapid prediction of cold-rolled texture of oriented electrical steel based on crystal plasticity finite element method and sequential conditional generative deep learning surrogate model according to claim 8, characterized in that, Step S2 specifically includes: The crystal plasticity finite element model is run in batches within the process design space to generate "virtual-annotated" data pairs; For each set of process vectors The crystal plasticity finite element model outputs the crystal orientation set after cold rolling, which is used to obtain the ODF (Optical Distribution Format) using a texture calculation tool. Furthermore, the ODF slice image at a specified Euler angle section is extracted, and the ODF slice image is discretized into a two-dimensional field with a pixel size of 64*64, denoted as the true label image. It can be flattened into a 4096-dimensional vector. .

10. The method for rapid prediction of cold-rolled texture of oriented electrical steel based on crystal plasticity finite element method and sequential conditional generative deep learning surrogate model according to claim 9, characterized in that, Step S2 further includes: We used measured texture data after cold rolling to perform systematic bias calibration on the output of the crystal plasticity finite element model. The calibration method is as follows: using the ODF calculated from the measured EBSD or XRD data as a benchmark, the intensity correction coefficient and peak position offset correction matrix of the CPFEM output ODF are obtained by fitting with the least squares method. The ODF data generated in batches by CPFEM are corrected point by point so that the relative deviation between the peak intensity and peak position of the key texture components of the virtual data and the measured data is less than 8%, thereby improving the physical fidelity and reliability of the virtual data.

11. The method for rapid prediction of cold-rolled texture of oriented electrical steel based on crystal plasticity finite element and sequential conditional generative deep learning surrogate model according to claim 9, characterized in that, Step S3 includes the following steps: Step S31: The normalized process feature sequence of each pass. The input is fed into the embedding layer, which maps the features of each passage to a fixed-dimensional passage embedding vector. To preserve the path dependency of the cold rolling pass sequence, position encoding is superimposed on the embedding vector. Construct sequence representation ; Step S32: The global feature vector composed of total reduction rate, number of passes, and other global process constraint parameters is generated. Input linear mapping layer to obtain and combine it with the global latent vector After concatenation, a conditional vector is obtained through nonlinear mapping. This vector is used to characterize the combined effect of "cold rolling path + global constraints" and serves as the conditional input for subsequent ODF image generation. Step S33, condition vector The algorithm maps the data to a low-resolution two-dimensional feature map using linear layers, and then generates an ODF prediction image with a target resolution of 1×64×64 through multi-level upsampling and two-dimensional convolution. ; Step S34: The mean squared error is used as the basic loss function to measure the predicted ODF slice image. Compared with real ODF slice images The differences between them; During training, a gradient descent-based optimization algorithm is used, and the learning rate, batch size, and number of training rounds are set. The model parameters are iteratively optimized on the training set, and the loss changes are monitored on the validation set. When the loss on the validation set converges or no longer decreases significantly, the training ends and the final model weights are saved.

12. The method for rapid prediction of cold-rolled texture of oriented electrical steel based on crystal plasticity finite element method and sequential conditional generative deep learning surrogate model according to claim 11, characterized in that: A loss function is used to constrain the difference between the predicted ODF slice image and the real ODF slice image, and a gradient descent-based optimization algorithm is used to train the sequence conditional generative deep learning surrogate model. Training ends and the model weights are saved when the validation set loss converges or no longer decreases significantly.

13. The method for rapid prediction of cold-rolled texture of oriented electrical steel based on crystal plasticity finite element method and sequential conditional generative deep learning surrogate model according to claim 11, characterized in that: The loss function includes a structural similarity index (SSIM) constraint term; The loss function includes weighted constraint terms for the peak intensity and / or peak position of key texture components.

14. The method for rapid prediction of cold-rolled texture of oriented electrical steel based on crystal plasticity finite element method and sequential conditional generative deep learning surrogate model according to claim 11, characterized in that: Step S4 includes the following steps: Step S41: The input cold rolling process parameters are standardized / normalized in the same way as in step S1 to obtain normalized process input. And according to the maximum number of passes N max The rules complete sequence padding and attention mask construction; Step S42, input the normalization process The input is fed into the pre-trained sequence conditional generative deep learning proxy model, and the predicted ODF slice image is obtained through the sequence encoding and convolutional decoding modules. ; Step S43, for the predicted ODF slice image Perform denormalization to restore the original ODF intensity dimensions; Step S44, the predicted ODF slice image As The ODF slices of the specified cross section are visualized and output, and the texture indices of Goss component intensity and key orientation area fraction are further calculated for process window evaluation and comparative analysis.

15. A rapid prediction system for the texture of cold-rolled oriented electrical steel, used to implement the rapid prediction method for the texture of cold-rolled oriented electrical steel based on crystal plasticity finite element and sequential conditional generative deep learning surrogate model as described in any one of claims 1-14, characterized in that, include: The data acquisition and virtual experiment module is used to acquire the initial electron backscatter diffraction data of the normalized plate before cold rolling, construct the crystal plasticity finite element model, and generate virtual data pairs of "process path-ODF" within the preset process design space. The data preprocessing and encoding module is used to encode the process sequence of each pass with a uniform length, standardize / normalize the global cold rolling process parameters and ODF slice images, and discretize and store the ODF slice images to form a dataset that can be used for deep learning training. The surrogate model training and prediction module is used to build and train a sequence conditional generative deep learning surrogate model, and after deployment, it enables fast forward prediction from cold rolling process path to ODF slice image. The process path scanning and visualization module is used to call proxy models in batches within a given process range to generate ODF prediction results, and present them in the form of images or statistical features to assist engineers in process window evaluation, path comparison and preliminary screening.

16. The rapid prediction system for cold-rolled texture of grain-oriented electrical steel according to claim 15, characterized in that, The rapid prediction system for cold-rolled texture of oriented electrical steel also includes: An integrated module that interfaces with production line control or process design software to embed candidate process paths and predicted texture results into existing process development processes; The fast prediction system for cold-rolled texture of oriented electrical steel further includes a processor and a memory. The memory stores a computer program, which, when executed by the processor, implements the steps of the fast prediction method for cold-rolled texture of oriented electrical steel based on crystal plasticity finite element and sequential conditional generative deep learning surrogate model as described in any one of claims 1-13.

17. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed on a processor, it implements the steps of the fast prediction method for cold-rolled texture of oriented electrical steel based on crystal plastic finite element and sequential conditional generative deep learning surrogate model as described in any one of claims 1-14.

Images