A method and device for constructing a material constitutive model and integrating the material constitutive model with finite elements, a computer device and a medium

Abstract

Embodiments of the present application provide a material constitutive model construction and finite element integration method, device, computer equipment and medium, relating to the technical field of computing materials, wherein the method comprises the following steps: obtaining experimental data of a target material in a wide strain rate and a wide temperature range, processing the experimental data using an automatic preprocessing algorithm, and generating a standard data set; selecting and outputting the AI algorithm with the highest fitting degree, and outputting the preprocessing parameters and training parameters corresponding to the AI algorithm; extracting strain, strain rate and temperature from the standard data set and reconstructing based on material mechanics constraints, outputting the trained material constitutive model through the reconstructed feature matrix and the constraint term reflecting the nonlinear mechanical behavior of the material; and converting the data format output by the trained material constitutive model into the input data format of the finite element software through a standardized interface. Through the constructed material constitutive model, the prediction accuracy of the constitutive model under wide working conditions is improved.

Description

Technical Field

[0001] This invention relates to the field of computational materials technology, and in particular to a method, apparatus, computer equipment, and medium for constructing a material constitutive model and integrating finite element methods. Background Technology

[0002] In strategic high-end industries such as aerospace, new energy, high-end manufacturing, medical devices, and nuclear industry, the service environment of core structural materials is becoming increasingly complex and demanding, generally facing wide strain rates (10⁻⁶). -4 s -1 ~10 4 s -1 Extreme operating conditions across a wide temperature range (-190℃ to 1800℃) are crucial. Aero-engine blades must withstand the coupling effect of high temperature and high strain rate; semiconductor packaging materials must adapt to drastic temperature fluctuations from low to normal temperatures and stress deformation at low strain rates; and composite material components must cope with complex mechanical responses under multiple directions and conditions. Material constitutive models, as the core tool for describing the stress-strain relationship of materials, are the foundation of finite element simulation, structural design, reliability assessment, and life prediction. Their prediction accuracy, applicability, and engineering practicality directly determine the rationality, safety, and economy of high-end equipment design, and are also a key support for promoting the research and development of new materials.

[0003] Currently, the construction techniques for material constitutive models are mainly divided into two categories: traditional mechanics derivation-based and artificial intelligence-driven. Traditional constitutive models rely on classical mechanics theory and simplification assumptions, making it difficult to adapt to the nonlinear mechanical behavior of complex materials under wide operating conditions. Although existing material constitutive models have overcome some limitations of traditional models, they still suffer from insufficient prediction accuracy under extreme conditions, poor adaptability, and difficulties in engineering implementation. Furthermore, they rely excessively on human experience and massive amounts of experimental data, failing to meet the practical engineering needs of multi-material, multi-condition, and high-precision applications. With the rapid iteration of high-end materials (such as novel alloys, high-performance composite materials, and functional polymer materials) and the continuous improvement of extreme service requirements, developing a material constitutive model construction method that combines strong adaptability, high prediction accuracy, high compatibility, and high intelligence has become an urgent need in the field of material mechanical property research. This has significant theoretical and engineering value for promoting the high-quality development of high-end industries, reducing R&D costs, and enhancing core competitiveness.

[0004] Traditional constitutive model construction methods are based on relevant theories of materials mechanics. They obtain data through tensile and compressive experiments, employ simplified assumptions and mathematical fitting methods to determine model parameters, and construct a constitutive model describing the stress-strain relationship of the material for predicting material mechanical properties and finite element simulations. However, these methods rely heavily on human experience for parameter calibration and model adjustment. They have a narrow scope of application, cannot adapt to wide strain rate and wide temperature range coupling scenarios, make too many assumptions, deviate significantly from actual material mechanical behavior, and involve complex and inefficient parameter calibration, relying heavily on human experience. Furthermore, the parameter fitting of traditional constitutive models depends on manual fitting, resulting in high labor costs and low efficiency.

[0005] Existing research papers on material constitutive models rely on single neural network algorithms (ANN, BP, CNN, etc.). These models collect experimental data across wide strain rates and temperature ranges, construct datasets, and train AI models to map input parameters (strain rate, temperature, strain) to output parameters (stress), thus building the constitutive model for predicting material mechanical properties. Theoretical verification is primarily achieved through ABAQUS subroutines, but standardized engineering applications have not been implemented. Existing methods suffer from insufficient prediction accuracy, poor generalization ability, excessive reliance on training data, low data utilization, poor model interpretability, lack of physical meaning support, poor adaptability, and inability to adapt to various material types. Furthermore, existing material constitutive models require the reconstruction of neural network models for different materials, consuming significant computational resources and exhibiting low efficiency. Summary of the Invention

[0006] In view of this, embodiments of the present invention provide a method for constructing a material constitutive model and integrating finite element methods to solve the technical problems of high labor costs, large consumption of computational resources, and low efficiency in the prior art. The method includes:

[0007] Experimental data of the target material over a wide strain rate and temperature range are obtained. The experimental data are then processed using an automatic preprocessing algorithm, which sequentially performs outlier removal, normalization, data augmentation, noise suppression, and data fusion to generate a standard dataset. A smart decision-making model is constructed to obtain the target material characteristics and prediction conditions input by the user. The target material characteristics and prediction conditions are compared with the material type characteristic library, performance requirement characteristic library, and working condition characteristic library. The smart decision-making model is used to select and output the AI ​​algorithm with the highest adaptability, and the preprocessing parameters and training parameters corresponding to the AI ​​algorithm are also output. Strain, strain rate, and temperature are extracted from the standard dataset and reconstructed based on material mechanics constraints. A material constitutive model is constructed based on the AI ​​algorithm using the reconstructed feature matrix and constraint terms reflecting the nonlinear mechanical behavior of the material. The material constitutive model is then optimized using the preprocessing parameters and the training parameters. The optimized material constitutive model is trained and validated using the standard dataset, and the trained material constitutive model is output. A standardized interface is constructed to convert the data format output by the trained material constitutive model into the input data format of the finite element software, and output the key parameters required for finite element simulation.

[0008] This invention also provides a device for constructing and integrating material constitutive models and finite element methods, to solve the technical problems of high labor costs, large computational resource consumption, and low efficiency in existing technologies. The device includes: A dataset module is constructed to acquire experimental data of the target material over a wide strain rate and temperature range. An automatic preprocessing algorithm is used to process the experimental data, performing outlier removal, normalization, data augmentation, noise suppression, and data fusion in sequence to generate a standard dataset. An AI algorithm module is used to build an intelligent decision-making model, obtain the target material features and prediction conditions input by the user, compare the target material features and prediction conditions with the material type feature library, performance requirement feature library, and working condition feature library, filter and output the AI ​​algorithm with the highest adaptability through the intelligent decision-making model, and output the preprocessing parameters and training parameters corresponding to the AI ​​algorithm. A material constitutive model construction module is used to extract strain, strain rate and temperature from the standard dataset and reconstruct them based on material mechanics constraints. The material constitutive model is constructed based on the AI ​​algorithm using the reconstructed feature matrix and constraint terms reflecting the nonlinear mechanical behavior of the material. The material constitutive model is then optimized using the preprocessing parameters and the training parameters. The optimized material constitutive model is trained and validated using the standard dataset, and the trained material constitutive model is output. The finite element integration module is used to build a standardized interface. Through the standardized interface, the data format output by the trained material constitutive model is converted into the input data format of the finite element software, and the key parameters required for finite element simulation are output.

[0009] This invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the construction of any of the above-mentioned material constitutive models and finite element integration methods, thereby solving the technical problems of high labor costs, large consumption of computing resources, and low efficiency in the prior art.

[0010] This invention also provides a computer-readable storage medium storing a computer program that executes any of the above-described methods for constructing material constitutive models and integrating finite element data, in order to solve the technical problems of high labor costs, large consumption of computing resources, and low efficiency in the prior art.

[0011] Compared with the prior art, the beneficial effects that at least one technical solution adopted in the embodiments of this specification can achieve include at least: By constructing a material constitutive model, the prediction accuracy of the constitutive model under a wide range of working conditions is improved, and the problems of excessive assumptions and large deviations from the actual mechanical behavior of materials in traditional constitutive models are solved. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart of a method for constructing a material constitutive model and integrating finite elements, provided in an embodiment of the present invention. Figure 2 This is a flowchart of a method for constructing and integrating the above-described material constitutive model according to an embodiment of the present invention; Figure 3 This is a flowchart of the multi-AI algorithm integration and adaptive matching provided in the embodiments of the present invention; Figure 4 This is a flowchart of AI model construction and training provided in an embodiment of the present invention; Figure 5 This is a flowchart of the finite element adaptation and interface call provided in the embodiments of the present invention; Figure 6 This is a structural block diagram of a computer device provided in an embodiment of the present invention; Figure 7 This is a structural block diagram of a material constitutive model construction and finite element integration device provided in an embodiment of the present invention. Detailed Implementation

[0014] The embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0015] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. This application can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0016] In this embodiment of the invention, a method for constructing a material constitutive model and integrating finite element methods is provided, such as... Figure 1 As shown, the method includes: Step S101: Obtain experimental data of the target material in a wide strain rate and wide temperature range, and process the experimental data using an automatic preprocessing algorithm, performing outlier removal, normalization, data augmentation, noise suppression and data fusion in sequence to generate a standard dataset; Step S102: Construct an intelligent decision-making model, obtain the target material features and prediction conditions input by the user, compare the target material features and prediction conditions with the material type feature library, performance requirement feature library, and working condition feature library, filter and output the AI ​​algorithm with the highest adaptability through the intelligent decision-making model, and output the preprocessing parameters and training parameters corresponding to the AI ​​algorithm. Step S103: Extract strain, strain rate and temperature from the standard dataset and reconstruct them based on material mechanics constraints. Construct a material constitutive model based on the AI ​​algorithm using the reconstructed feature matrix and constraint terms reflecting the nonlinear mechanical behavior of the material. Optimize the hyperparameters of the material constitutive model using the preprocessing parameters and the training parameters. Train and validate the optimized material constitutive model using the standard dataset and output the trained material constitutive model. Step S104: Construct a standardized interface to convert the data format output by the trained material constitutive model into the input data format of the finite element software, and output the key parameters required for finite element simulation.

[0017] In specific implementation, the following steps are used to compare the target material characteristics and prediction conditions with the material type feature library, performance requirement feature library, and working condition feature library. The intelligent decision-making model then selects and outputs the AI ​​algorithm with the highest suitability, and outputs the preprocessing parameters and training parameters corresponding to the AI ​​algorithm: A standardized algorithm library is constructed by integrating multiple AI algorithms, including Artificial Neural Networks (ANN), Backpropagation Neural Networks (BP), Convolutional Neural Networks (CNN), Long Short-Term Memory Networks (LSTM), and CNN-LSTM hybrid networks with attention mechanisms. The system acquires the target material type, material microstructure parameters, predicted mechanical properties, and operating conditions input by the user as the input feature vector of the intelligent decision-making model. This input feature vector is compared with a material type feature library, a performance requirement feature library, and an operating condition feature library. The intelligent decision-making model then performs a weighted score on each AI algorithm in the standardized algorithm library based on its suitability, accuracy, and efficiency, selecting the AI ​​algorithm with the highest weighted score as the chosen algorithm. Finally, the system outputs the identifier of the AI ​​algorithm and simultaneously outputs the corresponding preprocessing parameters and training parameters.

[0018] In specific implementation, the following steps are used to extract strain, strain rate, and temperature from the standard dataset and reconstruct them based on material mechanics constraints. Using the reconstructed feature matrix and constraint terms reflecting the nonlinear mechanical behavior of the material, a material constitutive model is constructed based on the AI ​​algorithm: Based on the AI ​​algorithm's identifier, a material constitutive model framework is constructed, comprising a feature extraction component, a prediction component, and a constraint component. The feature extraction component receives input features, the prediction component calculates stress, and the constraint component stores a database of fundamental mechanical parameters for various materials. Strain, strain rate, and temperature are extracted from the standard dataset as initial input features. These initial input features are reconstructed using the material mechanics constraints to generate a feature matrix embedding mechanical laws. This feature matrix is ​​used as input to the feature extraction component, which outputs the extracted feature vector. The extracted feature vector is then input to the prediction component to calculate the predicted stress value. The yield criterion, plastic hardening law, and temperature softening effect are transformed into constraint terms, which are added to the loss function of the material constitutive model. The fundamental mechanical parameters corresponding to the target material type stored in the constraint component are used as hidden layer constraints for the prediction component. Finally, a material constitutive model comprising the feature extraction component, prediction component, and constraint component is output.

[0019] In specific implementation, the following steps are used to extract strain, strain rate, and temperature as original input features from the standard dataset; reconstruct the original input features using the material mechanics constraints to generate a feature matrix embedding mechanical laws; use the feature matrix as input to the feature extraction component and output the extracted feature vector; input the extracted feature vector to the prediction component to calculate the predicted stress value; and transform the yield criterion, plastic hardening law, and temperature softening effect into constraint terms, which are then added to the loss function of the material constitutive model. According to the type of target material, the original input features are reconstructed using the corresponding material mechanics formula. If it is a metallic material, the input features are reconstructed using Johnson-Cook constitutive model, Zener-Hollomon parameters or Arrhenius equation. The reconstructed first feature matrix is ​​used as the feature matrix, wherein the first feature matrix is ​​used to characterize embedded strain hardening, strain rate strengthening and temperature softening effects. For composite materials, the input features are reconstructed using the Hashin failure criterion or classical laminate theory. The reconstructed second feature matrix is ​​used as the feature matrix, where the second feature matrix characterizes the anisotropy of the stiffness in the embedded fiber direction and the interlaminar coupling effect. The feature matrix is ​​used as the input to the feature extraction component, and the extracted feature vector is output. In the prediction component, an elastic sub-network corresponding to the elastic segment of the material and a plastic sub-network corresponding to the plastic segment of the material are established, and the stress-strain relationship of the elastic segment and the plastic segment are independently fitted respectively. The extracted feature vector is input to the elastic sub-network and the plastic sub-network, and the predicted stress values ​​of the elastic segment and the plastic segment are output respectively. For metallic materials, the monotonicity constraint that the predicted stress value decreases with increasing temperature and the monotonicity constraint that the predicted stress value increases with increasing strain rate are used as constraint terms. For composite materials, the proportional constraint of stress components in different directions and the upper limit constraint that the strain does not exceed the failure strain threshold are set as constraint terms. The constraint terms are added to the loss function of the material constitutive model.

[0020] In specific implementation, a standardized interface is constructed through the following steps: the data format output by the trained material constitutive model is converted into the input data format of the finite element software through the standardized interface, and the key parameters required for finite element simulation are output. The deep learning framework in the trained material constitutive model is stripped and exported as a lightweight model, which contains only forward computation logic and weight parameters. Based on the type of the target finite element software, a corresponding user material subroutine is selected from a preset template library. Through the user material subroutine, the strain tensor, strain increment, temperature, and time step transmitted by the target finite element software during simulation are received. The strain tensor is converted into Mises equivalent strain, the strain increment is divided by the time step to obtain the equivalent strain rate, the temperature is normalized to generate a normalized temperature, and the Mises equivalent strain and the equivalent strain are... The strain rate and the normalized temperature are combined and used as the input scalar of the lightweight model; the input scalar is input into the lightweight model, forward calculation is performed, equivalent stress is output, the equivalent stress is restored to a six-component stress tensor, and the six-component stress tensor is returned to the finite element software; stress limits, strain rate limits, temperature limits, and stress increment limits are set in the user material subroutine; elastic parameters, constitutive characteristic parameters, weights and biases of the lightweight model are input into the user material subroutine in a fixed order through the parameter fields of the finite element software, and the path of the compiled user material subroutine dynamic library is configured to the finite element software.

[0021] In specific implementation, the following steps are performed: The user material subroutine receives the strain tensor, strain increment, temperature, and time step transmitted by the target finite element software during simulation; the strain tensor is converted into Mises equivalent strain; the strain increment is divided by the time step to obtain the equivalent strain rate; the temperature is normalized to generate a normalized temperature; and the Mises equivalent strain, the equivalent strain rate, and the normalized temperature are combined and used as the input scalar of the lightweight model. The six components of the strain tensor are received from the finite element software, and the Mises equivalent strain is calculated according to the Mises equivalent strain formula. ,in, Let i be the row number of the strain tensor and j be the column number of the component. The six components of the strain increment tensor are received from the finite element software, and the equivalent strain increment is calculated using the Mises equivalent strain increment method. ,in, The components of the strain increment tensor; the equivalent strain rate is obtained by dividing the equivalent strain increment by the time step. ; Receive the temperature value T at the current integration point, and determine the material's reference temperature. and melting point Normalization is performed to obtain the normalized temperature. ,in, The actual temperature at the current integration point; the equivalent strain Equivalent variation rate and normalized temperature Combined into a feature vector [ , , ], and serve as the input scalar of the lightweight model.

[0022] In specific implementation, the following steps are used to automatically adjust the parameters of the constraint components of the material constitutive model when a new material is added, and to adjust the number of network layers, the number of neurons, and the training parameters in the prediction component of the material constitutive model, so as to output a material constitutive model adapted to the new material: Experimental data on the new material over a wide strain rate and temperature range are acquired, and a standard dataset for the new material is generated based on the experimental data. The standard dataset of the new material is input into the feature extraction component of the trained material constitutive model, and a feature vector of the new material is output. The feature vector of the new material is compared with a database of basic mechanical parameters of various materials stored in the constraint component, and the material type and mechanical properties of the new material are identified by the intelligent decision model. According to the identified material type and mechanical properties, the corresponding basic mechanical parameters are retrieved from the database of basic mechanical parameters and the parameters in the constraint component are updated. The basic mechanical parameters include elastic modulus, Poisson's ratio, yield strength, hardening exponent, and coefficient of thermal expansion. According to the feature vector of the new material and the updated constraint component parameters, the number of network layers and neurons in the prediction component of the material constitutive model are fine-tuned, and the learning rate and number of iterations in the training parameters are adaptively adjusted to obtain an optimized model structure and training parameters. The adjusted constraint component parameters, the optimized model structure, and the training parameters are embedded into the trained material constitutive model, and a material constitutive model adapted to the new material is output.

[0023] Example 1: This embodiment provides a method for constructing a material rate-temperature coupled constitutive model based on artificial intelligence algorithms. This method can automate the entire process from experimental data to finite element simulation and supports multi-material adaptive adaptation. Figure 2 As shown, the method mainly includes the following steps.

[0024] Step 1: Experimental data collection.

[0025] High-precision tensile and compression testing equipment is used, such as an electronic universal testing machine (for quasi-static loading, strain rate range 10). -4 S -1 ~10 0 S -1) and split Hopkinson bars (SHPB, for dynamic loading, strain rate range 10) 2 s 1 ~10 4 s 1) Systematically acquire experimental data of the target material over a wide strain rate and temperature range.

[0026] In this embodiment, the strain rate is set with 5 or more gradients, and the temperature is set with 6 or more gradients to ensure data coverage across all operating conditions. The core data collected includes: strain rate. The module records temperature (T), strain (ε), and stress (σ), while simultaneously recording environmental parameters (humidity, air pressure, loading rate) during the experiment, providing support for subsequent noise suppression and data optimization. This module supports experimental data acquisition for various materials (metals, composite materials, polymers, etc.), adapting to diverse material testing needs.

[0027] Step 2: Intelligent Data Preprocessing The process employs an AI-powered automatic preprocessing algorithm to achieve full automation, requiring no manual intervention. Specifically, it includes the following sub-steps: Step 2.1, Outlier Removal: Based on the Isolation Forest AI anomaly detection algorithm, outlier data caused by testing errors and environmental interference is automatically identified and removed. The number of trees in the Isolation Forest is set to 100, the subsampling size to 256, and the anomaly score threshold to 0.6. Compared with the traditional 3σ criterion, this method has higher recognition accuracy for high-dimensional, non-normally distributed data.

[0028] Step 2.2, Normalization: The min-max normalization algorithm is used to map the strain rate, temperature, strain, and stress data to the [0,1] interval to eliminate the influence of dimensions. The normalization formula is: .in, The normalized dimensionless value, ranging from [0,1]. These are the original data values ​​(e.g., strain rate, temperature, strain, stress). This is the minimum value of the variable across all training samples. This represents the maximum value of the variable across all training samples.

[0029] Step 2.3, Data Augmentation: An AI data augmentation algorithm combining Generative Adversarial Networks (GANs) and linear interpolation is employed. First, a conditional GAN ​​(cGAN) is trained using the original experimental data to generate new stress-strain curves based on strain rate and temperature. Then, linear interpolation is performed on data points from adjacent operating conditions to expand the data for intermediate operating conditions. This method can expand a limited sample size (e.g., 3-5 curves per operating condition) to 5-10 times the original data volume, improving data coverage.

[0030] Step 2.4, Noise Suppression: A combination of wavelet denoising algorithm and AI noise filtering algorithm is used. First, the original stress signal is subjected to wavelet transform (Daubechies 4 wavelet (i.e., db4 wavelet), with 5 decomposition layers) and soft thresholding for denoising. Then, the denoised signal is input into a trained autoencoder for secondary noise filtering, and clean data is output.

[0031] Step 2.5, Data Fusion: A weighted fusion algorithm is used to intelligently fuse experimental data from different batches and under different experimental conditions. For data from multiple experiments under the same operating condition, weights are assigned according to their signal-to-noise ratio (SNR), with data having a higher SNR receiving greater weight.

[0032] Step 3: Integration of multiple AI algorithms and adaptive matching.

[0033] Integrating multiple advanced AI algorithms, a standardized algorithm library is constructed, including: ANN, BP neural network, CNN, LSTM, and a CNN-LSTM hybrid network with an attention mechanism. Simultaneously, an intelligent decision-making model is built, combining material type feature libraries, performance requirement feature libraries, and operating condition feature libraries to achieve automatic matching of the optimal algorithm. The specific process is as follows: Step 3.1: Obtain user input.

[0034] Users only need to input the target material type (such as metal, composite material, polymer material), material microstructure parameters (such as grain size, fiber content), the mechanical properties to be predicted (such as yield strength, fracture toughness, dynamic strain response), and operating conditions (strain rate range, temperature range).

[0035] Step 3.2: Feature comparison and weighted scoring.

[0036] like Figure 3As shown, the intelligent decision-making model compares the user-input feature vector with three major feature libraries and assigns a weighted score to each AI algorithm in the algorithm library. In this embodiment, the weights are set as follows: fitness 40%, accuracy 30%, and efficiency 30%. Fitness is determined by matching the material type with the algorithm's historical applicability; for example, CNN or CNN-LSTM+attention mechanism is preferred for composite materials. Accuracy is based on the average prediction error (R² or RMSE) of existing models under similar conditions. Efficiency considers model training time and inference time.

[0037] Step 3.3, Automatic Invocation.

[0038] After selecting the AI ​​algorithm with the highest score, the system automatically calls the algorithm to build the model structure and simultaneously matches the corresponding preprocessing parameters (such as data augmentation factor and noise suppression strength) and training parameters (such as learning rate, number of iterations, and batch size).

[0039] For example, for the dynamic response of metallic materials under wide strain rate and wide temperature range, the system usually selects a CNN-LSTM hybrid network + attention mechanism as the core algorithm, and automatically configures the CNN feature extraction layer as 2 convolutional layers and 1 pooling layer, the LSTM temporal capture layer as 2 LSTM layers, and the attention mechanism weight allocation as 60% for coupled features and 40% for single features.

[0040] Step 4: AI model building and intelligent training optimization, such as Figure 4 As shown.

[0041] Step 4.1: Model framework construction.

[0042] The AI ​​algorithm, based on matching, automatically builds the model structure. For a single algorithm (ANN, BP, CNN, LSTM), it automatically matches the optimal number of network layers (ANN / BP: 3-5 layers, CNN: 5-7 layers, LSTM: 4-6 layers) and the number of neurons (32-128). For hybrid algorithms (CNN-LSTM + attention mechanism), it automatically optimizes the CNN feature extraction layer (2-3 convolutional layers, 1-2 pooling layers) and the LSTM temporal capture layer (2-3 LSTM layers).

[0043] Step 4.2: Integration of materials mechanics theory.

[0044] The fundamental laws of classical materials mechanics and plasticity are embedded into the computational logic of artificial intelligence models, specifically including the following three levels: (1) Feature layer reconstruction (input layer).

[0045] Instead of using the original strain, strain rate, and temperature as direct inputs, the input characteristics are reconstructed using material mechanics formulas.

[0046] For metallic materials, the dimensionless terms in the Johnson-Cook constitutive equations are used for reconstruction, or the Zener-Hollomon parameters are used.

[0047] For composite materials, the fiber direction stiffness matrix and interlaminar stress coupling terms are reconstructed using the Hashin failure criterion or the classical laminate theory (CLT) as input features.

[0048] (2) Hidden layer segmented fitting.

[0049] In the prediction component, an elastic sub-network corresponding to the elastic segment and a plastic sub-network corresponding to the plastic segment are established for the metallic material. The stress-strain relationships of the elastic and plastic segments are independently fitted, respectively. The two sub-networks can be in parallel, sharing the output of the feature extraction component, and then outputting the stresses of the elastic and plastic segments separately. The final stress is obtained through a weighted combination. This segmented network structure avoids the discontinuity problem at the junction of the elastic and plastic segments in traditional fully connected layers.

[0050] In the loss function, in addition to the usual mean squared error (MSE) data fitting term, a constraint term reflecting the true mechanical behavior of the material is added: For metallic materials, a monotonicity constraint is added that the predicted stress decreases with increasing temperature, and a monotonicity constraint increases with increasing strain rate. This can be specifically represented as a penalty term. ,in, This is a physical constraint penalty term, added to the total loss function to quantify the degree to which the model predicts violations of the laws governing material mechanics. The weighting coefficient (positive number) for the temperature monotonicity constraint controls the degree of influence of this term in the total loss, and is usually determined through hyperparameter optimization. Predict stress for the model Temperature The partial derivative of stress represents the rate of change of stress with temperature. Physically, metallic materials typically undergo thermal softening at high temperatures, meaning that stress should decrease as temperature increases. This is the weighting coefficient (positive number) for the strain rate monotonicity constraint, which controls the degree of influence of this term in the total loss. Predict stress for the model strain rate The partial derivative of represents the rate of change of stress with strain rate. Physically, metallic materials typically exhibit positive strain rate sensitivity, meaning that stress increases with increasing strain rate. `Max()` is a maximum value function that ensures a penalty is applied only in directions that violate physical constraints (positive values), without constraining directions that conform to physical laws. This penalty term allows the model to automatically learn during training to avoid outputting stress predictions that violate the fundamental laws of metal thermal softening and strain rate hardening.

[0051] For composite materials, add proportional constraints on stress components in different directions (e.g., satisfying orthogonal anisotropic constitutive relations) and upper limit constraints on strain not exceeding the failure strain threshold.

[0052] Step 4.3: Automatic parameter initialization and optimization.

[0053] The Xavier initialization method is used to automatically set the initial weights and biases of the network based on the input and output dimensions. For layers with ReLU activation functions, He initialization is used.

[0054] Hyperparameter optimization: Grid search + Bayesian optimization algorithm is used. Search range: learning rate 0.001~0.01, number of iterations 500~2000, batch size 32~128.

[0055] During training, the model parameters are automatically fine-tuned based on the changes in the validation set loss (for example, by using a learning rate decay strategy, the learning rate is halved when the validation loss does not decrease for 5 consecutive rounds).

[0056] Step 4.4, Intelligent Training and Validation.

[0057] The preprocessed standard dataset is automatically divided into training, validation, and test sets in a ratio of 7:2:1.

[0058] The Adam optimizer was used for training with an initial learning rate of 0.001. An early stopping strategy was employed during training: training was automatically stopped when the validation set loss function did not decrease for 10 consecutive iterations.

[0059] After training, the model's prediction accuracy is verified using a test set. If the prediction error (e.g., RMSE) exceeds a preset threshold (e.g., 5%), parameter fine-tuning or data augmentation optimization processes are automatically initiated, and training is repeated until the error reaches the target. The final output is the trained material constitutive model.

[0060] Step 5: Finite element adaptation and standardized interface, such as Figure 5 As shown.

[0061] Step 5.1: Lightweight model export.

[0062] By stripping deep learning frameworks (such as TensorFlow and PyTorch) from the trained AI constitutive model, a lightweight model containing only forward computation logic and weight parameters is derived. This model stores weights and biases in the form of C++ arrays or structures and does not depend on any deep learning libraries.

[0063] Step 5.2: Writing the user materials subroutine.

[0064] Select the corresponding user material subroutine template based on the type of target finite element software: Abaqus: VUMAT (explicit) or UMAT (implicit); ANSYS: USERMAT or USERANISOPL; LS-DYNA: MAT_USER_DEFINED; MSC.Marc: USERMATERIAL.

[0065] This embodiment uses LS-DYNA's VUMAT as an example. In the VUMAT subroutine, firstly through... The MAT_USER_DEFINED card passes in the elastic parameters (elastic modulus E, Poisson's ratio ν), constitutive characteristic parameters (such as reference strain rate, melting point), and the weights and biases of the lightweight model in a fixed order.

[0066] Step 5.3, Physical quantity reception and conversion.

[0067] The subroutine receives the strain tensor ε (6 components), strain increment tensor Δε, temperature T, and time step Δt from the finite element solver at each integration point and time step. It then converts these into scalar inputs for a lightweight model. Mises equivalent effect: ,in, The components of the strain tensor are i, which is the row number of the component, and j, which is the column number of the component. The values ​​of i and j are 1, 2, and 3, respectively, representing the x, y, and z directions in three-dimensional space. Equivalent change increment: , The components of the strain increment tensor. This leads to the equivalent strain rate. ; Normalized temperature: ; Step 5.4, AI forward calculation and stress reduction.

[0068] Input scalar [ , , The lightweight model is input, and forward computation (forward propagation of a multi-layer neural network) is performed, outputting the equivalent stress. Then, based on the current strain state, the equivalent stress is reduced to a 6-component stress tensor σ. For anisotropic materials, the transformation is performed based on the stiffness matrix in the constraint components.

[0069] Step 5.5: Numerical stabilization.

[0070] Incorporating stress limits, strain rate limits, temperature limits, and stress increment limits helps prevent divergence in finite element calculations.

[0071] Step 5.6: Dynamic library configuration.

[0072] By configuring the path of the compiled VUMAT dynamic library (.so or .dll) in the LS-DYNA solver environment, a complete interface between the AI ​​constitutive model and the finite element solver can be achieved. During the simulation, the solver calls this subroutine at each incremental step to obtain the stress response of the material.

[0073] This invention also supports multi-material adaptive adaptation without retraining the entire model. The specific implementation is as follows: When users need to adapt to new materials, they only need to provide a small amount of experimental data on the new material's wide strain rate and wide temperature range (no fewer than 5 different operating conditions). The system will automatically execute the following process: 1. Using the same preprocessing algorithm as in step 2, generate a standard dataset for the new material.

[0074] 2. Input the standard dataset of the new material into the feature extraction component of the pre-trained AI constitutive model, and output the feature vector of the new material.

[0075] 3. Compare the feature vector with the basic mechanical parameter library of various materials stored in the constraint component (basic parameters of metals, composite materials, and polymer materials, such as elastic modulus, Poisson's ratio, yield strength, hardening index, coefficient of thermal expansion, etc.), and identify the type and mechanical properties of the new material through the AI ​​classifier.

[0076] 4. Based on the identification results, retrieve the corresponding basic mechanical parameters from the basic mechanical parameter library and update the parameters in the constraint component.

[0077] 5. Based on the new material feature vector and the updated constraint component parameters, fine-tune the number of network layers and neurons in the prediction component (e.g., add or remove a few nodes in the last fully connected layer), and adaptively adjust the learning rate and number of iterations in the training parameters (e.g., use a smaller learning rate for a small number of rounds of transfer learning).

[0078] 6. Embed the adjusted constraint component parameters, optimized model structure, and training parameters into the original AI constitutive model to output an AI constitutive model adapted to the new material. The entire process eliminates the need to re-execute steps from experimental data acquisition to multi-AI algorithm integration and adaptive matching, significantly improving the model's versatility and adaptation efficiency.

[0079] This invention achieves fully automated control of the entire process from data acquisition, data preprocessing, algorithm matching, model building, training optimization to finite element method adaptation, without manual intervention. The system also includes a built-in intelligent monitoring and early warning module to monitor the operational status of each step in real time. Data anomaly warning: When the original data missing rate exceeds 5% or the proportion of outliers exceeds 10%, the system will automatically prompt the user to supplement the experiment or adjust the collection parameters.

[0080] Model training anomaly warning: When the validation set loss does not decrease for 20 consecutive rounds and early stopping is not triggered, the system automatically judges it as slow convergence and suggests increasing the learning rate or adjusting the network structure.

[0081] Interface adaptation exception warning: When a subroutine fails to compile or a dynamic library fails to load, the system outputs an error log and recommends a repair solution.

[0082] In addition, it supports batch model building and prediction for multiple materials and multiple working conditions, and can simultaneously handle the material constitutive model building needs of 3-5 materials, greatly improving research and design efficiency.

[0083] In this embodiment, a computer device is provided, such as... Figure 6 As shown, it includes a memory 601, a processor 602, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the construction of any of the above-mentioned material constitutive models and the finite element integration method.

[0084] Specifically, the computer device can be a computer terminal, a server, or a similar computing device.

[0085] In this embodiment, a computer-readable storage medium is provided, which stores a computer program that performs the above-described method for constructing any of the material constitutive models and integrating finite element methods.

[0086] Specifically, computer-readable storage media include both permanent and non-permanent, removable and non-removable media, which can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer-readable storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable storage media do not include transient media, such as modulated data signals and carrier waves.

[0087] Based on the same inventive concept, this invention also provides a device for constructing and integrating a material constitutive model with finite element analysis, as described in the following embodiments. Since the principle of the device for constructing and integrating a material constitutive model with finite element analysis is similar to that of the method for constructing and integrating a material constitutive model with finite element analysis, the implementation of the device can refer to the implementation of the method for constructing and integrating a material constitutive model with finite element analysis, and will not be repeated. As used below, the terms "unit" or "module" can refer to a combination of software and / or hardware that performs a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0088] Figure 7 This is a structural block diagram of a material constitutive model construction and finite element integration device according to an embodiment of the present invention, such as... Figure 7 As shown, it includes: a dataset construction module 701, an AI algorithm adaptation module 702, a material constitutive model construction module 703, and a finite element integration module 704. The structure is described below.

[0089] A dataset construction module 701 is used to acquire experimental data of the target material in a wide strain rate and wide temperature range. The experimental data is processed by an automatic preprocessing algorithm, which sequentially performs outlier removal, normalization, data augmentation, noise suppression and data fusion to generate a standard dataset. The AI ​​algorithm module 702 is used to build an intelligent decision-making model, obtain the target material features and prediction conditions input by the user, compare the target material features and prediction conditions with the material type feature library, performance requirement feature library, and working condition feature library, filter and output the AI ​​algorithm with the highest adaptability through the intelligent decision-making model, and output the preprocessing parameters and training parameters corresponding to the AI ​​algorithm. The material constitutive model construction module 703 is used to extract strain, strain rate and temperature from the standard dataset and reconstruct them based on material mechanics constraints. Through the reconstructed feature matrix and constraint terms reflecting the nonlinear mechanical behavior of the material, the material constitutive model is constructed based on the AI ​​algorithm. The material constitutive model is optimized by hyperparameters using the preprocessing parameters and the training parameters. The optimized material constitutive model is trained and validated using the standard dataset, and the trained material constitutive model is output. The finite element integration module 704 is used to construct a standardized interface, through which the data format output by the trained material constitutive model is converted into the input data format of the finite element software, and the key parameters required for finite element simulation are output.

[0090] In one embodiment, the AI ​​algorithm module is adapted, including: A standardized algorithm library unit is constructed to integrate multiple AI algorithms and build a standardized algorithm library. The AI ​​algorithms include artificial neural networks (ANN), backpropagation neural networks (BP neural networks), convolutional neural networks (CNN), long short-term memory networks (LSTM), and CNN-LSTM hybrid networks with attention mechanisms. An input feature vector unit is constructed to obtain the target material type, material microstructure parameters, mechanical properties to be predicted, and working conditions input by the user, which serve as the input feature vector of the intelligent decision-making model. The AI ​​algorithm unit is used to compare the input feature vector with the material type feature library, performance requirement feature library, and working condition feature library. The intelligent decision-making model performs a weighted score on each AI algorithm in the standardized algorithm library based on its adaptability, accuracy, and efficiency, and selects the AI ​​algorithm with the highest weighted score as the selected AI algorithm. The output unit is used to output the identifier of the AI ​​algorithm and simultaneously output the preprocessing parameters and training parameters corresponding to the AI ​​algorithm.

[0091] In one embodiment, constructing a material constitutive model module includes: A framework unit is constructed to build a material constitutive model framework, including a feature extraction component, a prediction component, and a constraint component, based on the identifier of the AI ​​algorithm. The feature extraction component is used to receive input features, the prediction component is used to calculate stress, and the constraint component is used to store a library of basic mechanical parameters for various materials. The model building unit is used to extract strain, strain rate, and temperature as original input features from the standard dataset, reconstruct the original input features through the material mechanics constraints, generate a feature matrix embedded with mechanical laws, use the feature matrix as input to the feature extraction component, and output the extracted feature vector. The extracted feature vector is then input to the prediction component to calculate the stress prediction value. The yield criterion, plastic hardening law, and temperature softening effect are transformed into constraint terms, and the constraint terms are added to the loss function of the material constitutive model. Construct constraint units to use the basic mechanical parameters corresponding to the target material type stored in the constraint components as hidden layer constraints of the prediction components; The first output model unit is used to output a material constitutive model that includes the feature extraction component, prediction component, and constraint component.

[0092] In one embodiment, the model building unit is further configured to reconstruct the original input features according to the type of the target material using corresponding material mechanics formulas. If the material is metallic, the input features are reconstructed using Johnson-Cook constitutive models, Zener-Hollomon parameters, or Arrhenius equations, and the reconstructed first feature matrix is ​​used as the feature matrix. The first feature matrix characterizes embedded strain hardening, strain rate strengthening, and temperature softening effects. If the material is a composite material, the input features are reconstructed using the Hashin failure criterion or classical laminate theory, and the reconstructed second feature matrix is ​​used as the feature matrix. The second feature matrix characterizes the anisotropy of embedded fiber stiffness and interlaminar coupling effects. The feature matrix is ​​then used as the... The feature extraction component takes input as input and outputs extracted feature vectors. In the prediction component, an elastic sub-network corresponding to the elastic segment of the material and a plastic sub-network corresponding to the plastic segment are established, and the stress-strain relationship of the elastic and plastic segments is independently fitted respectively. The extracted feature vectors are input to the elastic and plastic sub-networks, and the predicted stress values ​​of the elastic and plastic segments are output respectively. For metallic materials, the monotonicity constraint that the predicted stress value decreases with increasing temperature and the monotonicity constraint that the predicted stress value increases with increasing strain rate are used as constraint terms. For composite materials, the proportional constraint of stress components in different directions and the upper limit constraint that the strain does not exceed the failure strain threshold are set as constraint terms. These constraint terms are added to the loss function of the material constitutive model.

[0093] In one embodiment, the finite element integration module includes: The lightweight model unit is used to extract the deep learning framework from the trained material constitutive model and export it as a lightweight model, wherein the lightweight model only contains forward computation logic and weight parameters. The subroutine selection unit is used to select the corresponding user material subroutine from the preset template library according to the type of the target finite element software. An input scalar unit is constructed to receive the strain tensor, strain increment, temperature, and time step transmitted by the target finite element software during the simulation process through the user material subroutine. The strain tensor is converted into Mises equivalent strain, the strain increment is divided by the time step to obtain the equivalent strain rate, the temperature is normalized to generate a normalized temperature, and the Mises equivalent strain, the equivalent strain rate, and the normalized temperature are combined and used as the input scalar of the lightweight model. A six-component stress tensor element is obtained, which is used to input the input scalar into the lightweight model, perform forward calculation, output equivalent stress, restore the equivalent stress to a six-component stress tensor, and return the six-component stress tensor to the finite element software; The setting limit unit is used to set stress limit, strain rate limit, temperature limit and stress increment limit in the user material subroutine; Configure the finite element software unit to input the elastic parameters, constitutive characteristic parameters, weights and biases of the lightweight model into the user material subroutine in a fixed order through the parameter fields of the finite element software, and configure the dynamic library path of the compiled user material subroutine to the finite element software.

[0094] In one embodiment, the construction of the input scalar element is also used to receive the six components of the strain tensor from the finite element software and calculate the Mises equivalent strain according to the Mises equivalent strain formula. ,in, Let i be the row number of the strain tensor and j be the column number of the component. The six components of the strain increment tensor are received from the finite element software, and the equivalent strain increment is calculated using the Mises equivalent strain increment method. ,in, The components of the strain increment tensor; the equivalent strain rate is obtained by dividing the equivalent strain increment by the time step. ; Receive the temperature value T at the current integration point, and determine the material's reference temperature. and melting point Normalization is performed to obtain the normalized temperature. ,in, The actual temperature at the current integration point. At room temperature The melting point temperature of the material; the equivalent strain Equivalent variation rate and normalized temperature Combined into a feature vector [ , , ], and serve as the input scalar of the lightweight model.

[0095] In one embodiment, the device further includes a new material addition module.

[0096] In one embodiment, the new material addition module includes: A standard dataset unit for new materials is constructed to obtain experimental data of new materials over a wide strain rate and temperature range, and a standard dataset of new materials is generated based on the experimental data. The new material feature vector unit is used to input the standard dataset of the new material into the feature extraction component of the trained material constitutive model and output the new material feature vector. The material identification unit is used to compare the feature vector of the new material with the basic mechanical parameter library of various materials stored in the constraint component, and to identify the material type and mechanical properties of the new material through the intelligent decision model; The parameter update unit is used to retrieve the corresponding basic mechanical parameters from the basic mechanical parameter library according to the identified material type and mechanical properties, and update the parameters in the constraint component. The basic mechanical parameters include elastic modulus, Poisson's ratio, yield strength, hardening index and coefficient of thermal expansion. An adaptive adjustment unit is used to fine-tune the number of network layers and neurons in the prediction component of the material constitutive model according to the feature vector of the new material and the updated constraint component parameters, and to adaptively adjust the learning rate and number of iterations in the training parameters to obtain the optimized model structure and training parameters. The second output model unit is used to embed the adjusted constraint component parameters, the optimized model structure, and the training parameters into the trained material constitutive model, and output a material constitutive model adapted to the new material.

[0097] The embodiments of the present invention achieve the following technical effects: This invention constructs a method that can stably cover 10 -4 s - ¹~10 4 s -¹A material constitutive model with a wide strain rate range and a wide temperature range of -190℃ to 1800℃ is adapted to extreme and complex working conditions, effectively reducing prediction errors caused by changes in working conditions and significantly improving the prediction accuracy of the model under a wide range of working conditions, meeting the mechanical performance prediction requirements of high-end equipment in extreme environments. By significantly reducing the model's simplification assumptions, advanced AI technology is used to fully capture the coupling effect of strain rate and temperature under a wide strain rate and temperature range, accurately reflecting the nonlinear mechanical behavior of materials such as yielding, hardening, and fracture, minimizing the deviation between the model and the actual mechanical behavior of materials, especially suitable for complex materials such as composite materials and new alloys, solving the problem of excessive assumptions and large deviations from the actual mechanical behavior of materials in traditional constitutive models. The model parameter calibration process is simplified, and the automatic optimization and calibration of parameters are achieved through AI automatic learning capabilities, reducing the dependence on a large amount of experimental data and significantly improving the efficiency and consistency of parameter calibration; at the same time, the model's adaptability to different materials is enhanced, without the need to rebuild the model, only requiring the input of a small amount of experimental data for the corresponding material to complete the parameter adaptation, reducing the dependence on human experience and improving research and design efficiency. The AI ​​algorithm structure is optimized to enhance the model's ability to capture coupling effects, significantly reducing prediction errors and improving its generalization ability. This ensures the model maintains high-precision predictions even under untrained conditions and with minor material composition variations, meeting the accuracy requirements of engineering applications. Efficient data augmentation and fusion algorithms are designed to fully utilize limited experimental data and improve data utilization. Simultaneously, noise suppression techniques are introduced to effectively suppress noise interference in experimental data, preventing overfitting and improving the model's stability and reliability in practical applications, significantly reducing experimental costs and time. Addressing the interface requirements of mainstream finite element software, the model's output format and computational efficiency are systematically optimized to achieve seamless integration with finite element software. The model can be directly embedded into the finite element simulation process without additional data conversion or programming, meeting the needs of engineering applications.

[0098] Obviously, those skilled in the art should understand that the modules or steps of the above-described embodiments of the present invention can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those presented here, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the embodiments of the present invention are not limited to any particular hardware and software combination.

[0099] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations can be made to the embodiments of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for constructing a material constitutive model and integrating finite element methods, characterized in that, include: Experimental data of the target material over a wide strain rate and temperature range are obtained. The experimental data are then processed using an automatic preprocessing algorithm, which sequentially performs outlier removal, normalization, data augmentation, noise suppression, and data fusion to generate a standard dataset. A smart decision-making model is constructed to obtain the target material characteristics and prediction conditions input by the user. The target material characteristics and prediction conditions are compared with the material type characteristic library, performance requirement characteristic library, and working condition characteristic library. The smart decision-making model is used to select and output the AI ​​algorithm with the highest adaptability, and the preprocessing parameters and training parameters corresponding to the AI ​​algorithm are also output. Strain, strain rate, and temperature are extracted from the standard dataset and reconstructed based on material mechanics constraints. A material constitutive model is constructed based on the AI ​​algorithm using the reconstructed feature matrix and constraint terms reflecting the nonlinear mechanical behavior of the material. The material constitutive model is then optimized using the preprocessing parameters and the training parameters. The optimized material constitutive model is trained and validated using the standard dataset, and the trained material constitutive model is output. A standardized interface is constructed to convert the data format output by the trained material constitutive model into the input data format of the finite element software, and output the key parameters required for finite element simulation.

2. The method for constructing a material constitutive model and integrating finite element methods as described in claim 1, characterized in that, The target material characteristics and prediction conditions are compared with material type feature databases, performance requirement feature databases, and operating condition feature databases. The intelligent decision-making model then selects and outputs the AI ​​algorithm with the highest suitability, and outputs the preprocessing parameters and training parameters corresponding to the AI ​​algorithm, including: A standardized algorithm library is constructed by integrating multiple AI algorithms, including Artificial Neural Network (ANN), Backpropagation Neural Network (BP), Convolutional Neural Network (CNN), Long Short-Term Memory Network (LSTM), and CNN-LSTM hybrid network with attention mechanism. The target material type, material microstructure parameters, mechanical properties to be predicted, and working conditions input by the user are obtained and used as the input feature vector of the intelligent decision-making model. The input feature vector is compared with the material type feature library, performance requirement feature library, and working condition feature library. The intelligent decision-making model performs a weighted score on each AI algorithm in the standardized algorithm library based on its adaptability, accuracy, and efficiency, and selects the AI ​​algorithm with the highest weighted score as the selected AI algorithm. Output the identifier of the AI ​​algorithm, and simultaneously output the preprocessing parameters and training parameters corresponding to the AI ​​algorithm.

3. The method for constructing a material constitutive model and integrating finite element methods as described in claim 1, characterized in that, Strain, strain rate, and temperature are extracted from the standard dataset and reconstructed based on material mechanics constraints. Using the reconstructed feature matrix and constraint terms reflecting the nonlinear mechanical behavior of the material, a material constitutive model is constructed based on the AI ​​algorithm, including: Based on the identifier of the AI ​​algorithm, a material constitutive model framework is constructed, which includes a feature extraction component, a prediction component, and a constraint component. The feature extraction component is used to receive input features, the prediction component is used to calculate stress, and the constraint component is used to store a library of basic mechanical parameters for various materials. Strain, strain rate, and temperature are extracted from the standard dataset as original input features. The original input features are reconstructed by the material mechanics constraints to generate a feature matrix embedded with mechanical laws. The feature matrix is ​​used as the input of the feature extraction component and the extracted feature vector is output. The extracted feature vector is input to the prediction component to calculate the stress prediction value. The yield criterion, plastic hardening law, and temperature softening effect are transformed into constraint terms and added to the loss function of the material constitutive model. The basic mechanical parameters corresponding to the target material type stored in the constraint component are used as the hidden layer constraints of the prediction component. The output is a material constitutive model that includes the feature extraction component, prediction component, and constraint component.

4. The method for constructing and integrating the material constitutive model with the finite element method as described in claim 3, characterized in that, Strain, strain rate, and temperature are extracted as raw input features from the standard dataset. These raw input features are then reconstructed using the material mechanics constraints to generate a feature matrix embedding mechanical laws. This feature matrix is ​​used as input to the feature extraction component, which outputs the extracted feature vector. The extracted feature vector is then input to the prediction component to calculate the predicted stress value. The yield criterion, plastic hardening law, and temperature softening effect are transformed into constraint terms, which are added to the loss function of the material constitutive model. This includes: According to the type of target material, the original input features are reconstructed using the corresponding material mechanics formula. If it is a metallic material, the input features are reconstructed using Johnson-Cook constitutive model, Zener-Hollomon parameters or Arrhenius equation. The reconstructed first feature matrix is ​​used as the feature matrix, wherein the first feature matrix is ​​used to characterize embedded strain hardening, strain rate strengthening and temperature softening effects. If it is a composite material, the input features are reconstructed by the Hashin failure criterion or the classical theory of laminates, and the reconstructed second feature matrix is ​​used as the feature matrix. The second feature matrix is ​​used to characterize the anisotropy of the stiffness of the embedded fiber and the interlayer coupling effect. The feature matrix is ​​used as the input to the feature extraction component, and the extracted feature vector is output. In the prediction component, an elastic subnetwork corresponding to the elastic segment of the material and a plastic subnetwork corresponding to the plastic segment of the material are established, and the stress-strain relationship of the elastic segment and the plastic segment are independently fitted respectively. The extracted feature vectors are input into the elastic subnetwork and the plastic subnetwork, and the stress prediction values ​​of the elastic segment and the plastic segment are output respectively. If it is a metallic material, the monotonicity constraint that the predicted stress value decreases with increasing temperature and the monotonicity constraint that the predicted stress value increases with increasing strain rate are used as constraint terms. If it is a composite material, set the proportional constraints of stress components in different directions and the upper limit constraint that the strain does not exceed the failure strain threshold as constraint terms; The constraint term is added to the loss function of the material constitutive model.

5. The method for constructing a material constitutive model and integrating finite element methods as described in claim 1, characterized in that, A standardized interface is constructed to convert the data format output by the trained material constitutive model into the input data format of the finite element software, and to output the key parameters required for finite element simulation, including: The deep learning framework in the trained material constitutive model is stripped and exported as a lightweight model, wherein the lightweight model contains only forward computation logic and weight parameters. Based on the type of the target finite element software, select the corresponding user material subroutine from the preset template library; The user material subroutine receives the strain tensor, strain increment, temperature, and time step transmitted by the target finite element software during the simulation process. It converts the strain tensor into Mises equivalent strain, divides the strain increment by the time step to obtain the equivalent strain rate, normalizes the temperature to generate a normalized temperature, and combines the Mises equivalent strain, the equivalent strain rate, and the normalized temperature as the input scalar of the lightweight model. The input scalar is input into the lightweight model, forward calculation is performed, equivalent stress is output, the equivalent stress is restored to a six-component stress tensor, and the six-component stress tensor is returned to the finite element software; In the user material subroutine, stress limits, strain rate limits, temperature limits, and stress increment limits are set; The elastic parameters, constitutive characteristic parameters, weights and biases of the lightweight model are input into the user material subroutine in a fixed order through the parameter fields of the finite element software, and the dynamic library path of the compiled user material subroutine is configured to the finite element software.

6. The method for constructing and integrating the material constitutive model with the finite element method as described in claim 5, characterized in that, The user material subroutine receives the strain tensor, strain increment, temperature, and time step transmitted by the target finite element software during simulation. It converts the strain tensor into Mises equivalent strain, divides the strain increment by the time step to obtain the equivalent strain rate, normalizes the temperature to generate a normalized temperature, and combines the Mises equivalent strain, the equivalent strain rate, and the normalized temperature as input scalars for the lightweight model, including: The six components of the strain tensor are received from the finite element software, and the Mises equivalent strain is calculated according to the Mises equivalent strain formula. ,in, Let i be the row number of the strain tensor and j be the column number of the component. The six components of the strain increment tensor are received from the finite element software, and the equivalent strain increment is calculated according to the Mises equivalent strain increment method. ,in, These are the components of the strain increment tensor; The equivalent strain rate is obtained by dividing the equivalent strain increment by the time step. ; Receive the temperature value T at the current integration point, based on the material's reference temperature. and melting point Normalization is performed to obtain the normalized temperature. ,in, The actual temperature at the current integration point; The equivalent strain Equivalent variation rate and normalized temperature Combined into a feature vector [ , , ], and serve as the input scalar of the lightweight model.

7. The method for constructing and integrating the material constitutive model with the finite element method as described in any one of claims 1 to 6, characterized in that, Also includes: When a new material is added, the parameters of the constraint components of the material constitutive model are automatically adjusted, and the number of network layers, the number of neurons, and the training parameters in the prediction component of the material constitutive model are also adjusted to output a material constitutive model adapted to the new material, including: Obtain experimental data for the new material over a wide strain rate and temperature range, and generate a standard dataset for the new material based on the experimental data; The standard dataset of the new material is input into the feature extraction component of the trained material constitutive model, and the new material feature vector is output. The feature vector of the new material is compared with the basic mechanical parameter library of various materials stored in the constraint component, and the material type and mechanical properties of the new material are identified by the intelligent decision model. Based on the identified material type and mechanical properties, the corresponding basic mechanical parameters are retrieved from the basic mechanical parameter library and the parameters in the constraint component are updated. The basic mechanical parameters include elastic modulus, Poisson's ratio, yield strength, hardening index and coefficient of thermal expansion. Based on the new material feature vector and the updated constraint component parameters, the number of network layers and neurons in the prediction component of the material constitutive model are fine-tuned, and the learning rate and number of iterations in the training parameters are adaptively adjusted to obtain the optimized model structure and training parameters. The adjusted constraint component parameters, optimized model structure, and training parameters are embedded into the trained material constitutive model, and a material constitutive model adapted to the new material is output.

8. A device for constructing a material constitutive model and integrating finite element methods, characterized in that, include: A dataset module is constructed to acquire experimental data of the target material over a wide strain rate and temperature range. An automatic preprocessing algorithm is used to process the experimental data, performing outlier removal, normalization, data augmentation, noise suppression, and data fusion in sequence to generate a standard dataset. An AI algorithm module is used to build an intelligent decision-making model, obtain the target material features and prediction conditions input by the user, compare the target material features and prediction conditions with the material type feature library, performance requirement feature library, and working condition feature library, filter and output the AI ​​algorithm with the highest adaptability through the intelligent decision-making model, and output the preprocessing parameters and training parameters corresponding to the AI ​​algorithm. A material constitutive model construction module is used to extract strain, strain rate and temperature from the standard dataset and reconstruct them based on material mechanics constraints. The material constitutive model is constructed based on the AI ​​algorithm using the reconstructed feature matrix and constraint terms reflecting the nonlinear mechanical behavior of the material. The material constitutive model is then optimized using the preprocessing parameters and the training parameters. The optimized material constitutive model is trained and validated using the standard dataset, and the trained material constitutive model is output. The finite element integration module is used to build a standardized interface. Through the standardized interface, the data format output by the trained material constitutive model is converted into the input data format of the finite element software, and the key parameters required for finite element simulation are output.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method for constructing the material constitutive model and integrating the finite element method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that performs the method for constructing a material constitutive model and integrating finite elements according to any one of claims 1 to 7.

Images