Self-adaptive high-temperature damage model construction method and system

Abstract

The invention provides a self-adaptive high-temperature damage model construction method and system. Belongs to the technical field of material science and engineering calculation. The method comprises the following steps: acquiring experimental data of a material in a high-temperature environment, preprocessing the collected data through a machine learning algorithm, and extracting key damage features; constructing a basic damage prediction model based on a deep learning framework, training the model by using the preprocessed data, and enabling the basic damage prediction model to adaptively adjust internal parameters according to prediction errors through a reinforcement learning mechanism; and carrying out global search on thermodynamic parameters in the model by utilizing a genetic algorithm, searching an optimal parameter combination, and dynamically adjusting model parameters in combination with real-time temperature monitoring data. Through integration of deep learning, reinforcement learning, a genetic algorithm and a molecular dynamics simulation technology, the damage evolution process of the material in a high-temperature environment can be predicted more accurately and comprehensively.

Description

Technical Field

[0001] The invention proposes a method and system for constructing an adaptive high-temperature damage model, and belongs to the field of material science and engineering computing technology. Background Art

[0002] Most of the existing high-temperature damage models are based on static mechanics theory, which makes it difficult to fully capture the complex damage mechanisms of materials under extreme conditions such as high temperature and high pressure, such as thermal stress relaxation, creep rupture, and microstructure reconstruction. In addition, these models often lack adaptability and cannot adjust parameters according to the real-time state of the material during service, resulting in limited prediction accuracy. Therefore, it is particularly urgent to develop an innovative model that can dynamically adapt to changes in high-temperature environments and accurately predict material damage behavior. Summary of the invention

[0003] The present invention provides a method and system for constructing an adaptive high temperature damage model to solve the problems mentioned in the above background technology:

[0004] The present invention proposes a method for constructing an adaptive high temperature damage model, the method comprising:

[0005] S1. Obtain experimental data of materials under high temperature environment, pre-process the collected data through machine learning algorithm, and extract key damage features;

[0006] S2. Build a foundation damage prediction model based on a deep learning framework, use the preprocessed data to train the model, and use a reinforcement learning mechanism to enable the foundation damage prediction model to adaptively adjust internal parameters based on prediction errors;

[0007] S3, using genetic algorithms to globally search the thermodynamic parameters in the model, find the optimal parameter combination, and dynamically adjust the model parameters in combination with real-time temperature monitoring data;

[0008] S4. Based on molecular dynamics simulation technology, simulate the microstructure evolution process of materials at high temperatures, compare the simulation results with experimental data, and further optimize the model parameters through feedback mechanism;

[0009] S5. Select an independent validation data set to verify the foundation damage prediction model and evaluate its prediction performance. Based on the evaluation results, iteratively optimize the model.

[0010] The present invention proposes an adaptive high temperature damage model construction system, the system comprising:

[0011] Data acquisition module: obtains experimental data of materials in high temperature environment, pre-processes the collected data through machine learning algorithms, and extracts key damage features;

[0012] Model building module: build a foundation damage prediction model based on a deep learning framework, use preprocessed data to train the model, and use a reinforcement learning mechanism to enable the foundation damage prediction model to adaptively adjust internal parameters based on prediction errors;

[0013] Parameter adjustment module: Use genetic algorithms to perform global searches on the thermodynamic parameters in the model, find the optimal parameter combination, and dynamically adjust the model parameters in combination with real-time temperature monitoring data;

[0014] Process simulation module: Based on molecular dynamics simulation technology, the microstructure evolution process of the material at high temperature is simulated, the simulation results are compared with the experimental data, and the model parameters are further optimized through the feedback mechanism;

[0015] Iterative optimization module: select independent verification data sets to verify the foundation damage prediction model and evaluate its prediction performance. According to the evaluation results, the model is iteratively optimized.

[0016] Beneficial effects of the present invention: By integrating deep learning, reinforcement learning, genetic algorithm and molecular dynamics simulation technology, the damage evolution process of materials under high temperature environment can be predicted more accurately and comprehensively. The method combines experimental data with simulation results, and can adaptively adjust the damage model parameters under different temperatures and loading conditions, thereby improving the accuracy of damage prediction; in terms of data processing, a machine learning algorithm is used to preprocess the experimental data and extract key damage features, and further feature selection and optimization are performed through principal component analysis (PCA) and linear discriminant analysis (LDA), effectively removing noise, improving the validity of data and the learning effect of the model; in the training process, through reinforcement learning mechanism and optimization algorithm (such as genetic algorithm), the damage prediction model can dynamically adjust its internal parameters according to the prediction error, thereby realizing adaptive optimization. In addition, through regular error evaluation, learning rate decay strategy, early stopping and grid search and other optimization means, the model overfitting is effectively avoided and the generalization ability of the model is improved; the molecular dynamics simulation technology is used to simulate the microstructure of the material, which can deeply understand the damage evolution mechanism of the material at high temperature, and through comparison with experimental data, feedback optimization is performed to further improve the accuracy of the model. This process can help identify the evolution law of microscopic damage of materials in high-temperature environments and provide a theoretical basis for material design and engineering applications. Through global search of thermodynamic parameters by genetic algorithms, not only can the thermodynamic parameter combination of the model be optimized, but also the model parameters can be dynamically adjusted in combination with real-time temperature monitoring data, so that the prediction model can more accurately reflect the performance changes of materials under different working conditions. Through independent verification data sets, k-fold cross-validation, feature importance analysis and other multi-dimensional evaluation methods, the predictive performance of the model can be comprehensively analyzed, key features and parameters can be identified, and the model can be fine-tuned and continuously optimized to ensure its long-term stability and high efficiency in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a step diagram of the method of the present invention;

[0018] Figure 2 This is a system module diagram of the present invention. DETAILED DESCRIPTION

[0019] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0020] One embodiment of the present invention, as Figure 1 As shown, a method for constructing an adaptive high temperature damage model comprises:

[0021] S1. Obtain experimental data of materials under high temperature environment, pre-process the collected data through machine learning algorithm, and extract key damage features;

[0022] S2. Build a foundation damage prediction model based on a deep learning framework, use the preprocessed data to train the model, and use a reinforcement learning mechanism to enable the foundation damage prediction model to adaptively adjust internal parameters based on prediction errors;

[0023] S3, using genetic algorithms to globally search the thermodynamic parameters in the model, find the optimal parameter combination, and dynamically adjust the model parameters in combination with real-time temperature monitoring data;

[0024] S4. Based on molecular dynamics simulation technology, simulate the microstructure evolution process of materials at high temperatures, compare the simulation results with experimental data, and further optimize the model parameters through feedback mechanism;

[0025] S5. Select an independent validation data set to verify the foundation damage prediction model and evaluate its prediction performance. Based on the evaluation results, iteratively optimize the model.

[0026] The working principle of the above technical solution is as follows: obtain data from material experiments under high temperature environment, including stress-strain curve (describing the deformation of materials under stress), temperature-time curve (describing the change of temperature during the experiment), and microstructure image (showing the microstructure change of materials under high temperature); remove noise and outliers in the data to ensure the accuracy and consistency of the data; use interpolation or other algorithms to fill in the missing parts of the data to ensure the integrity of the data; extract key features that have an important impact on material damage prediction from the original data, such as stress peak, strain rate, temperature gradient, etc.; build a basic damage prediction model based on deep learning framework (such as TensorFlow, PyTorch, etc.). This model can receive preprocessed data as input and output the prediction results of material damage; use preprocessed data to train the model so that it can learn the characteristics and laws in the data; introduce reinforcement learning mechanism so that the model can automatically adjust its internal parameters according to the prediction error to improve the accuracy of the prediction; use genetic algorithm to search the thermodynamic parameters in the model globally to find the optimal parameter combination. Genetic algorithms can efficiently search the solution space by simulating natural selection and genetic mechanisms; they can dynamically adjust the model parameters in combination with real-time temperature monitoring data. This ensures that the model can maintain high prediction accuracy under different temperature conditions; based on molecular dynamics simulation technology, the microstructure evolution process of materials at high temperatures is simulated. This helps to understand the damage mechanism of materials at high temperatures; the simulation results are compared with experimental data, and the model parameters are further optimized through the feedback mechanism. This ensures that the model can more accurately reflect the actual damage of the material; an independent verification data set is selected to verify the basic damage prediction model. This can evaluate the prediction performance of the model on unknown data; the model is iteratively optimized based on the verification results. If the prediction accuracy does not meet the requirements, the model structure, parameters or data preprocessing methods are adjusted, and retraining and verification are performed until satisfactory prediction accuracy is achieved.

[0027] The effects of the above technical solution are as follows: preprocessing the collected data through machine learning algorithms, including data cleaning, missing value filling and feature selection, can effectively remove noise and outliers in the data to ensure the accuracy and consistency of the data; the extraction of key damage features helps the model to better learn the features and laws in the data, thereby improving the accuracy of the prediction; this method builds the model completely based on experimental data, avoiding the problem of relying on a large number of assumptions and simplifications in traditional methods, and improving the reliability and practicality of the model; through the reinforcement learning mechanism, the basic damage prediction model can adaptively adjust the internal parameters according to the prediction error, thereby continuously optimizing the model performance; this adaptive adjustment capability enables the model to maintain a high prediction accuracy when facing different high-temperature environments and material types; using genetic algorithms to perform a global search of the thermodynamic parameters in the model, it is possible to find the optimal parameter combination and further improve the prediction accuracy of the model; combining real-time temperature monitoring data to dynamically adjust the model parameters, so that The model can adapt to environmental changes in real time, improving the practicality and robustness of the model; based on molecular dynamics simulation technology, the microstructure evolution process of the material at high temperature is simulated, which helps to deeply understand the damage mechanism of the material at high temperature; this simulation technology can provide strong theoretical support for the construction and optimization of the model; the simulation results are compared with the experimental data, and the model parameters are further optimized through the feedback mechanism, so that the model can more accurately reflect the actual damage of the material; this feedback optimization mechanism helps to improve the prediction accuracy and generalization ability of the model; the basic damage prediction model is verified by selecting an independent verification data set, which can objectively evaluate the prediction performance of the model; this verification method helps to discover problems and deficiencies in the model and provide direction for subsequent optimization; the model is iteratively optimized according to the verification results, and the model structure, parameters or data preprocessing methods are continuously adjusted until satisfactory prediction accuracy is achieved; this iterative optimization method enables the model to be continuously improved and perfected, improving the practicality and accuracy of the model.

[0028] In one embodiment of the present invention, the S1 includes:

[0029] S11. According to the thermal stability range of the material, a series of temperature gradients are set to perform conditional loading, and different loading rates and stress levels are set under each loading condition;

[0030] S12, at different temperature points and loading stages, the microstructure images were obtained by SEM, and the evolution of microscopic damage was recorded;

[0031] S13. Identify outliers based on statistical methods and further verify the rationality of the data in combination with the physical properties of the material;

[0032] For time series data, such as temperature-time curves, linear interpolation or predictive interpolation based on time series models is used;

[0033] For characteristic values ​​such as stress-strain curves, neighboring point interpolation or predictive interpolation based on machine learning models is used;

[0034] S14. Extract the physical characteristics of the material such as elastic modulus, yield strength, fracture strength, thermal characteristics such as thermal expansion coefficient, thermal conductivity, heat capacity, and microstructural characteristics such as crack length, width, density, and dislocation distribution;

[0035] S15. Perform PCA analysis on the feature matrix, retain the first few principal components, and use LDA for feature selection in classification problems.

[0036] The working principle of the above technical solution is as follows: according to the thermal stability range of the material, a series of temperature gradients are set. These temperature gradients cover multiple key temperature points from room temperature to before the melting point of the material; at each temperature point, different loading condition experiments are carried out, such as static tension, compression, fatigue loading, etc. These loading conditions simulate various stress states that the material may encounter in practical applications; under each loading condition, different loading rates and stress levels are set to explore the response and damage of the material under different stress conditions; a scanning electron microscope (SEM) is used to obtain microstructure images of the material at different temperature points and after loading stages; by observing and analyzing these images, the evolution process of the microscopic damage of the material is recorded, such as the formation, expansion, generation and evolution of dislocations, etc.; based on statistical methods, the experimental data is preliminarily cleaned to identify and remove outliers. These outliers may be caused by experimental errors, instrument failures, or data recording errors; the rationality of the data is further verified in combination with the physical properties of the material (such as the change law of elastic modulus and yield strength). This helps ensure the accuracy and reliability of the data. For time series data (such as temperature-time curves), linear interpolation or predictive interpolation based on time series models is used to fill missing values ​​or smooth data. For eigenvalue data such as stress-strain curves, neighboring point interpolation or predictive interpolation based on machine learning models is used to estimate missing or abnormal data points. Physical, thermal and microstructural features related to material damage are extracted from experimental data. Physical features include elastic modulus, yield strength, fracture strength, etc., which reflect the response of materials under mechanical loading. Thermal features include thermal expansion coefficient, thermal conductivity, heat capacity, etc., which describe the thermal behavior of materials under high temperature environments. Microstructural features include crack length, width, density and dislocation distribution, which reveal the damage of materials at the microscopic level. Principal component analysis (PCA) is performed on the extracted feature matrix to retain the first few principal components. PCA is a commonly used data dimensionality reduction technique that reduces the dimensionality and redundancy of data by projecting the original data into a low-dimensional space. In classification problems, linear discriminant analysis (LDA) is used for feature selection. LDA is a supervised dimensionality reduction method that maximizes the inter-class distance and minimizes the intra-class distance based on the category label, thereby selecting the most useful features for classification; by combining PCA and LDA, the model can be further simplified, the prediction accuracy can be improved, and the computational complexity can be reduced.

[0037] The effects of the above technical solutions are as follows: according to the thermal stability range of the material, a series of temperature gradients are scientifically set, covering multiple key temperature points from room temperature to before the melting point of the material. This helps to comprehensively explore the damage behavior of the material at different temperatures; at each temperature point, a variety of loading condition experiments were carried out, such as static tension, compression, fatigue loading, etc., and different loading rates and stress levels were set. This helps to gain a deeper understanding of the response and damage mechanism of the material under different stress states; the microstructure images of the material were obtained at different temperature points and loading stages using a scanning electron microscope (SEM), which provided intuitive evidence of the microscopic damage of the material; the evolution process of the microscopic damage was recorded in detail, which helped to reveal the origin, development and evolution of material damage; outliers were effectively identified based on statistical methods to ensure the accuracy and reliability of the data; the rationality of the data was further verified in combination with the physical properties of the material (such as the change law of elastic modulus and yield strength), which enhanced the credibility of the data; for time series data (such as temperature-time curves), linear interpolation or time-based Predictive interpolation of the time series model is used, which helps to smooth the data and fill in missing values. For characteristic value data such as stress-strain curves, neighboring point interpolation or predictive interpolation based on machine learning models is used, which improves the completeness and accuracy of the data. Physical, thermal and microstructural characteristics related to material damage are extracted from the experimental data, which covers multiple aspects of material damage. Principal component analysis (PCA) is performed on the feature matrix and the first few principal components are retained, which helps to reduce the dimension and redundancy of the data and improve the computational efficiency of the model. In classification problems, linear discriminant analysis (LDA) is used for feature selection, which helps to select the most useful features for classification and improve the prediction accuracy of the model.

[0038] In one embodiment of the present invention, the S11 includes:

[0039] According to the thermal stability range of the material, the temperature gradient is initially set, and the temperature gradient is fine-tuned based on the effect of the material at different temperatures;

[0040] Design multiple loading conditions for each temperature gradient, and further refine the settings of loading rate and stress level under each loading condition;

[0041] The temperature control system is used to control the accuracy and stability of the temperature during the experiment. During the heating process, the temperature changes are monitored in real time and fine-tuned as needed.

[0042] According to the experimental plan, select appropriate loading devices and measuring equipment. For static tensile and compression experiments, use a universal material testing machine for loading and measurement; for fatigue loading experiments, use a fatigue testing machine for cyclic loading and measurement. At the same time, ensure the accuracy and stability of the loading device and measuring equipment to improve the accuracy of the experimental data;

[0043] During the experiment, the data acquisition system records key data such as temperature, stress, strain, etc. in real time, and provides feedback and adjustments to the experimental conditions based on the real-time monitoring data and experimental phenomena.

[0044] Based on the results and analysis of the preliminary experiments, the experimental conditions are further optimized; based on the adjustment and optimization of the experimental conditions, the experimental plan is iteratively updated.

[0045] The working principle of the above technical solution is: according to the thermal stability range of the material, starting from room temperature, a series of temperature gradients are set at certain temperature intervals (such as 50°C, 100°C, etc.) until it approaches but does not exceed the melting point of the material; considering the thermal expansion and thermal stress effects of the material at different temperatures, the initially set temperature gradient is fine-tuned. This helps to more accurately reflect the temperature environment of the material in real applications and optimize the experimental conditions to capture the key damage behavior of the material; for each temperature gradient, a variety of loading conditions are designed, including static stretching, compression, fatigue loading, etc., to simulate the different stress conditions of the material in actual applications; under each loading condition, the setting of loading rate and stress level is further refined. For example, different stretching rates and maximum tensile stresses are set in static stretching experiments; different number of cycles, stress amplitudes and stress ratios are set in fatigue loading experiments; a temperature control system is used to ensure the accuracy and stability of the temperature during the experiment. During the heating process, the temperature changes are monitored in real time and fine-tuned as needed to maintain the predetermined temperature gradient; a suitable loading device (such as a universal material testing machine, a fatigue testing machine, etc.) is selected to ensure the accuracy and stability of the loading process. At the same time, according to the experimental plan, the settings of the loading device are adjusted to match the predetermined loading conditions; through the data acquisition system, key data such as temperature, stress, and strain are recorded in real time. These data are the basis for subsequent analysis of material damage behavior; feedback and adjustment of experimental conditions are made based on real-time monitoring data and experimental phenomena. For example, when it is found that the temperature control is unstable or the loading device fails, timely adjustments and repairs are made to ensure the accuracy and reliability of the experimental data; based on the results and analysis of the preliminary experiments, the rationality and effectiveness of the experimental conditions are evaluated; according to the performance changes and damage mechanisms of the materials, parameters such as temperature gradient, loading rate, and stress level are adjusted. This helps to more accurately simulate the stress conditions and performance changes of the materials in actual applications; based on the adjustment and optimization of the experimental conditions, the experimental plan is iteratively updated. Through continuous trial and improvement, the optimal combination of experimental conditions is found to ensure the accuracy and reliability of the experimental data.

[0046] The effects of the above technical scheme are as follows: a preliminary temperature gradient is set according to the thermal stability range of the material, and fine-tuned based on the thermal expansion and thermal stress effects of the material at different temperatures, ensuring the scientificity and accuracy of the experimental temperature conditions; a variety of loading conditions are designed for each temperature gradient, and the settings of the loading rate and stress level are refined, which fully simulates the stress conditions of the material in actual applications and improves the accuracy and practicality of the experiment; the temperature during the experiment is precisely controlled by the temperature control system, and the temperature changes are monitored in real time and fine-tuned to ensure the accuracy and stability of the experimental temperature; appropriate loading devices and measuring equipment are selected, and their accuracy and stability are ensured, improving the accuracy and reliability of the experimental data; key data such as temperature, stress, and strain are recorded in real time by the data acquisition system, providing rich data support for subsequent analysis The experimental conditions are fed back and adjusted according to the real-time monitoring data and experimental phenomena, ensuring the smooth progress of the experiment and the accuracy of the experimental data; the experimental conditions are further optimized according to the results and analysis of the preliminary experiments, improving the accuracy and reliability of the experiment; on the basis of the adjustment and optimization of the experimental conditions, the experimental scheme is iteratively updated, and the optimal combination of experimental conditions is found through continuous trial and improvement, ensuring the accuracy and reliability of the experimental data; the experimental results obtained through this technical solution are highly practical and instructive, and can provide strong data support and theoretical basis for the research and development, design, manufacturing and application of materials; this technical solution not only improves the accuracy and reliability of the experiment, but also promotes the development of materials science research, and provides important methods and means for material performance evaluation, damage mechanism research and life prediction.

[0047] In one embodiment of the present invention, the S2 includes:

[0048] S21. Select a deep learning framework. If it is used to process time series data, select the LSTM model, such as temperature-time curve and stress-strain curve. If it is used to process long sequence data, select the Transformer model.

[0049] S22. Combined with the physical properties of the material, a loss function containing multiple sub-items is designed, such as elastic modulus prediction error, yield strength prediction error, and L1 or L2 regularization terms are added;

[0050] S23. According to the amount of data and model complexity, a batch size is selected, such as 64 or 128, and the data is randomly shuffled before each training cycle; the amount of data D is obtained by the following formula:

[0051]

[0052] Among them, n represents the number of features; f irepresents the weight factor of the i-th feature, which is set according to the importance of the feature; d i represents the dimension of the i-th feature; N represents the total number of data points; N0 represents a benchmark number of data points, which is used to adjust the impact of the data volume on the results;

[0053] The model complexity C is obtained by the following formula:

[0054]

[0055] Among them, L represents the number of model layers; pl represents the number of parameters in the lth layer; α represents the complexity factor of the activation function, and different activation functions have different values; al represents the additional complexity weight of the activation function in the lth layer; γ represents the layer number influence factor; L n Indicates the number of baseline layers, used to adjust the contribution of the number of layers to complexity;

[0056] S24, for the fully connected layer and LSTM layer, use the Xavier initialization method, for the layer with ReLU activation function, use the He initialization method;

[0057] S25. When the model prediction error is greater than a preset threshold, the model parameters are adjusted to make the prediction result closer to the true value, and during the model training process, the model parameters are dynamically adjusted according to the reward and punishment signals of the prediction results.

[0058] The working principle of the above technical solution is as follows: the LSTM model is suitable for processing time series data, such as temperature-time curves, stress-strain curves, etc. LSTM (Long Short-Term Memory) is a special recursive neural network (RNN) that solves the gradient vanishing problem of ordinary RNN in long sequence data processing by introducing memory units (CellState). LSTM can capture dependencies in long sequences and is very suitable for processing time series data. The Transformer model is suitable for processing long sequence data. The Transformer model was proposed by Google and is a neural network model based on the self-attention mechanism for processing sequence data. Compared with the traditional recurrent neural network model, the Transformer model has better parallel performance and shorter training time, and can capture long-distance dependencies; combined with the physical properties of the material, a loss function containing multiple sub-items is designed, such as elastic modulus prediction error, yield strength prediction error, etc. These sub-items reflect the accuracy of the model in predicting different physical properties; adding L1 or L2 regularization terms is used to prevent the model from overfitting and improve the generalization ability of the model by constraining the complexity of the model parameters. L1 regularization (Lasso regression) and L2 regularization (Ridge regression) are two commonly used regularization methods; According to the amount of data and model complexity, choose an appropriate batch size (BatchSize), such as 64, 128, etc. The choice of batch size will affect the training speed and convergence effect of the model; Before each training cycle, randomly shuffle the data: This helps the model learn the global characteristics of the data and avoid falling into the local optimal solution; For the fully connected layer and LSTM layer, use the Xavier initialization method. The Xavier initialization method initializes the weights according to the input and output dimensions of the layer, which helps to maintain the consistency of data distribution between layers; For the layer of the ReLU activation function, use the He initialization method. The He initialization method takes into account the characteristics of the ReLU activation function and can more effectively maintain the data distribution; When the model prediction error is greater than the preset threshold, adjust the model parameters (such as learning rate, weight, etc.) to make the prediction result closer to the true value. This can be achieved through optimization algorithms such as gradient descent; During the model training process, dynamically adjust the model parameters according to the reward and punishment signals of the prediction results (such as accuracy, loss value, etc.). This can be achieved through methods such as reinforcement learning, which enables the model to better adapt to changes in the data.

[0059] The effects of the above technical solutions are as follows: the solution can flexibly select deep learning frameworks, such as LSTM models are suitable for time series data (such as temperature-time curves, stress-strain curves), while Transformer models are suitable for processing long sequence data. This flexibility ensures that the model can optimally process different types of data; by selecting the most suitable deep learning framework, the prediction performance and accuracy of the model can be significantly improved, thereby better meeting the needs of actual applications; the design of the loss function combines the physical properties of the material, such as elastic modulus prediction error, yield strength prediction error, etc., which helps the model to more accurately capture the mechanical behavior of the material; by adding L1 or L2 regularization terms, the model can be effectively prevented from overfitting and the generalization ability of the model can be improved; selecting an appropriate batch size based on the amount of data and model complexity can balance the training speed and convergence effect of the model to ensure that the model achieves optimal performance within a reasonable time; randomly shuffling the data before each training cycle helps the model learn the global characteristics of the data and avoid falling into local optimal solutions; using the Xavier initialization method and the He initialization method respectively for Initializing the weights of the fully connected layers, LSTM layers, and ReLU activation function layers helps maintain the consistency of data distribution between layers, thereby improving the training efficiency and performance of the model; when the model prediction error is greater than the preset threshold, the model parameters are adjusted to make the prediction result closer to the true value, which helps to ensure the prediction accuracy of the model; during the model training process, the model parameters are dynamically adjusted according to the reward and punishment signals of the prediction results. This strategy can further improve the adaptability and accuracy of the model; through the comprehensive application of the above technical solutions, the prediction accuracy of the model can be significantly improved, providing more reliable data support for materials science research; the flexibility and dynamism of the solution help to quickly adjust and optimize the model, thereby shortening the material research and development cycle and improving research and development efficiency; this technical solution provides new ideas and methods for materials science research, and helps to promote technological innovation and industrial upgrading. Through the data volume calculation formula, the required data volume processing D can be dynamically calculated based on the number of features, the importance of features, the dimension of features, the total number of data points and a benchmark number of data points; this method takes into account the impact of feature importance and dimension on data volume, and uses a logarithmic function to smoothly adjust the data volume to adapt to data sets of different sizes; the model complexity formula is used to evaluate the complexity processing C of the model, taking into account the number of model layers processing L, the number of parameters per layer, the activation function complexity factor processing α, the additional complexity weight of the activation function per layer, and the layer number influence factor processing γ and the benchmark layer number processing L0; this evaluation method not only takes into account the structure of the model (the number of layers and the number of parameters per layer), but also takes into account the complexity of the activation function and the impact of the number of layers on the overall complexity, so as to more accurately reflect the complexity of the model; according to the data volume and model complexity, the batch size (such as 64, 128, etc.) can be reasonably selected.Larger batch sizes can speed up the training process, but may lead to increased memory consumption and less refined gradient updates; smaller batch sizes can improve the generalization ability of the model, but may increase training time; randomly shuffling the data before each training cycle helps the model learn the global distribution of the data, avoid overfitting, and improve the generalization performance of the model. Processing these formulas and methods provides flexibility and adaptability, and can be adjusted according to different data sets and model structures. By adjusting parameters such as feature weight factors, the number of benchmark data points, and the complexity factor of the activation function, the calculation of data volume and model complexity can be optimized to obtain better training results and performance; through reasonable data volume calculation and model complexity evaluation, as well as the selection and optimization of batch size, the training efficiency can be improved, the training time can be reduced, and the performance of the model can be maintained or improved. This is especially important for large-scale data sets and complex models. In summary, the above formulas and methods for selecting data batch sizes have significant beneficial effects in the fields of machine learning and deep learning, and can dynamically adjust the data volume, accurately evaluate the model complexity, and optimize the selection of batch size, thereby improving training efficiency and model performance.

[0060] In one embodiment of the present invention, the S25 includes:

[0061] After each training cycle, the prediction error of the model on the validation set is calculated, and the prediction performance of the model is evaluated by comparing the difference between the predicted value and the true value; the prediction error is calculated using the following formula:

[0062]

[0063] Among them, E total represents the total prediction error on the validation set; N represents the number of samples on the validation set; M represents the number of other key physical properties except elastic modulus and yield strength; w e and process w y They represent the weights of the prediction errors of elastic modulus (Elastic processes Modulus) and yield strength (Yield processes Strength); represents the weight of the prediction error of other physical characteristics in the jth process; It represents the prediction error of the elastic modulus of the i-th sample and can be defined as in represents the true value, represents the predicted value; represents the yield strength prediction error of the i-th sample, which is defined in the same way as similar; It represents the prediction error of other physical properties of the jth sample in the ith sample, and is defined in the same way as in the processing similar;

[0064] According to the size and distribution of prediction errors, formulate a strategy for adjusting model parameters; during model training, update model parameters through the gradient descent algorithm; dynamically adjust model parameters according to the reward and punishment signals of the prediction results; process

[0065] Through the learning rate decay strategy, a larger learning rate is set at the beginning of training to achieve rapid convergence; and as the training progresses, the learning rate is gradually reduced to adjust the model parameters;

[0066] Add L1 or L2 regularization terms to the loss function and dynamically adjust the weight of the regularization term based on the model's predictive performance;

[0067] During the model training process, the changes in the loss function, the prediction error on the validation set, and the update of the model parameters are monitored in real time. If the prediction error on the validation set does not decrease significantly after multiple consecutive training cycles, or shows an upward trend, an early stopping strategy is adopted to avoid overfitting of the model.

[0068] Tune the model's hyperparameters (such as learning rate, batch size, number of network layers, etc.) through grid search; find the optimal hyperparameter combination by comparing the model performance under different hyperparameter combinations;

[0069] Evaluate the performance of the model on the test set, select the optimal model architecture by comparing the performance of different models (such as LSTM, Transformer, etc.), and iteratively optimize the model based on the evaluation results on the test set.

[0070] The working principle of the above technical solution is as follows: after each training cycle, the model is predicted using the validation set, and the prediction error of key physical properties (such as elastic modulus, yield strength, etc.) is calculated. By comparing the difference between the predicted value and the true value, the prediction performance of the model can be evaluated. This step is an important part of model performance monitoring and helps to find problems with the model in a timely manner. According to the size and distribution of the prediction error, a strategy for adjusting the model parameters is formulated. If the prediction error is generally large, it may mean that the model complexity is insufficient. At this time, you can consider increasing the number of network layers or the number of neurons to improve the fitting ability of the model. If the prediction error is large in certain specific intervals, you may need to make more detailed parameter adjustments for these intervals, such as adjusting the weight initialization method, learning rate, etc. During the model training process, the model parameters are updated by the gradient descent algorithm. The gradient descent algorithm dynamically adjusts the model parameters according to the reward and punishment signal of the prediction result (that is, the gradient of the prediction error) to make the prediction result closer to the true value. This step is the core of model optimization. By continuously iteratively updating the parameters, the prediction error can be gradually reduced. The learning rate is one of the key factors affecting the speed and effect of model training. In the early stage of training, setting a larger learning rate can accelerate the convergence of the model. However, as the training progresses, an excessively large learning rate may cause the model to oscillate near the optimal solution and fail to converge stably. Therefore, adopting a learning rate decay strategy to gradually reduce the learning rate can help the model adjust parameters more finely in the later stage of training and achieve better convergence. Adding L1 or L2 regularization terms to the loss function can effectively prevent the model from overfitting. The weight of the regularization term can be dynamically adjusted according to the prediction performance of the model. When the prediction error of the model on the validation set is large, the weight of the regularization term can be appropriately increased to limit the complexity of the model parameters; when the prediction error of the model on the validation set is small, the weight of the regularization term can be appropriately reduced to allow the model parameters to be adjusted more flexibly. During the model training process, the changes in the loss function, the prediction error on the validation set, and the update of the model parameters are monitored in real time. If the prediction error on the validation set does not decrease significantly after multiple consecutive training cycles, or shows an upward trend, an early stopping strategy is adopted to avoid model overfitting. The early stopping strategy is an effective way to prevent model overfitting. It can stop training in time before the model performance begins to decline, thereby retaining the best performance of the model. The model's hyperparameters (such as learning rate, batch size, number of network layers, etc.) are tuned through grid search. Grid search is an exhaustive search method that tries all possible hyperparameter combinations and compares the model performance under different combinations to find the optimal hyperparameter combination. Although this step is computationally intensive, it can ensure that the global optimal solution or a parameter combination close to the global optimal solution is found. The performance of the model is evaluated on the test set, including indicators such as prediction accuracy and stability.By comparing the performance of different models (such as LSTM, Transformer, etc.), the optimal model architecture is selected. According to the evaluation results on the test set, the model is iteratively optimized, including adjusting model parameters, optimizing network structure, etc., to further improve the performance of the model.

[0071] The effects of the above technical solution are as follows: after each training cycle, the prediction error of the model on the validation set is calculated and evaluated, which helps to promptly discover problems with the model and take corresponding optimization measures; according to the size and distribution of the prediction error, a sophisticated parameter adjustment strategy is formulated, such as increasing the complexity of the model or adjusting parameters for a specific interval, which can significantly improve the prediction performance of the model; by gradually reducing the learning rate, it can ensure that the model converges quickly in the early stages of training, and adjust the parameters more finely in the later stages, thereby improving the stability and accuracy of the model; reasonable learning rate management helps to avoid the model from oscillating near the optimal solution, while reducing the risk of overfitting; adding L1 or L2 regularization terms to the loss function can effectively prevent the model from overfitting and improve the generalization ability of the model; dynamically adjusting the weight of the regularization term according to the prediction performance of the model can further optimize The performance of the model; during the model training process, the changes in the loss function, the prediction error on the validation set, and the update of the model parameters are monitored in real time, which helps to discover and solve problems in a timely manner; an early stopping strategy is adopted, and when the prediction error on the validation set does not decrease significantly or shows an upward trend after multiple consecutive training cycles, training is stopped in time to avoid model overfitting; the hyperparameters of the model are tuned through grid search to find the optimal hyperparameter combination, thereby improving the performance of the model; the performance of different models is evaluated on the test set, and the optimal model architecture is selected, which helps to ensure the accuracy and stability of the model in practical applications; according to the evaluation results on the test set, the model is iteratively optimized to continuously improve the performance of the model; by constantly trying new model architectures, parameter adjustment strategies, etc., the model can be continuously improved to adapt to different application scenarios and data characteristics. The above formula not only considers the two important physical properties of elastic modulus and yield strength, but also considers other possible key physical properties, so that the prediction performance of the model on different physical properties can be more comprehensively evaluated; through the weight factor, the prediction errors of different physical properties can be weighted to different degrees according to the needs of different application scenarios, so as to achieve customized evaluation of model performance; the prediction error in the formula is in the form of relative error (that is, the absolute value of the difference between the predicted value and the true value divided by the true value), which helps to compare between physical properties of different dimensions and quantify the accuracy of model prediction; by calculating the total prediction error, the prediction performance of the model on different physical properties can be clearly understood, thereby guiding subsequent model optimization work. For example, if it is found that the prediction error of the model on a certain physical property is large, the model structure or training strategy can be adjusted in a targeted manner to improve the prediction accuracy of the property; comprehensively considering the prediction errors of multiple physical properties and weighted summing them can help improve the reliability and stability of the model.Because even if the model's prediction on a certain characteristic is biased, it can be compensated to a certain extent by the prediction errors of other characteristics, thereby reducing the overall prediction error; the formula provides a unified framework to compare and evaluate the performance of different models on the same validation set. By calculating the total prediction error, the prediction accuracy of different models on different physical characteristics can be intuitively compared, so that the model with better performance can be selected for subsequent application. In summary, the above formula has significant beneficial effects in comprehensively evaluating model performance, providing flexibility and customizability, quantifying prediction errors, guiding model optimization, improving model reliability, and facilitating comparison and evaluation.

[0072] In one embodiment of the present invention, S3 includes:

[0073] S31, converting the thermodynamic parameter into a binary string, each bit representing a parameter value, and encoding it using the real value of the thermodynamic parameter;

[0074] S32. Combine multiple thermodynamic indicators of materials to design a multi-objective fitness function and assign weights according to the importance of each indicator;

[0075] S33, by selecting temperature sensors arranged at key positions of the material, such as thermocouples, thermal resistors, etc., real-time temperature data is collected, and the collected temperature data is preprocessed;

[0076] S34. Design parameter update rules based on the temperature changes monitored in real time, such as updating the model parameters every time the temperature rises within a certain range, and setting the upper and lower limits of parameter adjustment based on the optimal parameter combination obtained by the genetic algorithm.

[0077] The working principle of the above technical solution is as follows: first, convert the thermodynamic parameters (such as thermal expansion coefficient, thermal conductivity, etc.) into binary strings. Each binary bit represents a parameter value, which is encoded by the real value of the thermodynamic parameter. The encoding method can ensure that each parameter can be uniquely represented as a binary string; combine multiple thermodynamic indicators of the material (such as thermal expansion coefficient prediction error, thermal conductivity prediction error, etc.) to design a multi-objective fitness function. This function will be used to evaluate the performance of the model under different parameter combinations. According to the importance of each indicator, weights are assigned to ensure that the performance of each thermodynamic indicator can be balanced during the optimization process; temperature sensors (such as thermocouples, thermal resistors, etc.) are arranged at key positions of the material to collect temperature data in real time. The collected temperature data is preprocessed, including filtering, denoising, calibration and other steps to ensure the accuracy and reliability of the data; according to the temperature changes monitored in real time, parameter update rules are designed. For example, it can be set that the model parameters are updated once every time the temperature rises within a certain range. At the same time, according to the optimal parameter combination obtained by the genetic algorithm, the upper and lower limits of the parameter adjustment are set to avoid excessive parameter adjustment leading to model instability. During the parameter updating process, the parameter values ​​can be gradually adjusted until the optimal parameter combination is found.

[0078] The effects of the above technical solution are: converting thermodynamic parameters into binary strings, each bit represents a parameter value, and this encoding method can greatly improve the processing efficiency of parameters; binary encoding has the characteristics of discretization, which can accurately represent the value range of parameters, thereby avoiding the precision loss that may be caused by real number encoding; using real values ​​of thermodynamic parameters for encoding can ensure the accuracy and authenticity of parameters; real value encoding can also provide more possibilities for subsequent genetic algorithm operations (such as crossover and mutation), which helps to find a better parameter combination; combining multiple thermodynamic indicators of materials (such as thermal expansion coefficient, thermal conductivity, etc.) to design a multi-objective fitness function can comprehensively evaluate the performance of the model; this design method helps to find parameter combinations that can perform well in multiple thermodynamic indicators, thereby improving the overall performance of the model; weight allocation according to the importance of each indicator can ensure that the performance of each thermodynamic indicator can be balanced during the optimization process; the flexibility of weight allocation can also be adjusted according to actual needs to adapt to different application scenarios The data characteristics of the scene and data; the temperature data are collected in real time by temperature sensors (such as thermocouples, thermal resistors, etc.) arranged at key positions of the material, which can ensure the timeliness and accuracy of the data; the real-time collected data can also provide strong support for the subsequent model parameter update; preprocessing (such as filtering, denoising, etc.) the collected temperature data can improve the reliability and stability of the data; the preprocessed data can also reduce the impact of noise on the model performance, thereby improving the prediction accuracy of the model; designing parameter update rules based on the temperature changes monitored in real time can ensure that the model can adapt to different temperature change conditions; for example, when the temperature rises within a certain range, the model parameters are updated once, and this design method can ensure that the model can maintain stable performance when the temperature changes; setting the upper and lower limits of parameter adjustment according to the optimal parameter combination obtained by the genetic algorithm can avoid the problem of model instability caused by excessive parameter adjustment; this measure can also ensure that the model can maintain stable performance during the parameter update process, thereby improving the reliability and stability of the model.

[0079] In one embodiment of the present invention, the S31 includes:

[0080] Select key parameters that have a significant impact on material properties from the material's thermodynamic parameter library, such as thermal expansion coefficient, thermal conductivity, heat capacity, etc., and determine the effective value range of each parameter based on the physical meaning of these parameters and experimental data; process

[0081] For each selected thermodynamic parameter, a binary encoding scheme is designed; and the number of bits of the encoding is determined according to the value range and accuracy requirements of the parameter; for example, if the value range of a parameter is between 0 and 1 and needs to be accurate to four decimal places, a sufficiently long binary string can be used to represent this parameter;

[0082] Establish the mapping relationship between binary coding and real values ​​of thermodynamic parameters; analyze the impact of different coding bit numbers on parameter representation accuracy; evaluate whether the coding accuracy meets the preset requirements by comparing experimental data and encoded parameter values;

[0083] The designed binary coding scheme is applied to the encoding and decoding process of actual thermodynamic parameters, and the accuracy and reliability of the coding scheme are evaluated by comparing the original data with the encoded data.

[0084] Based on the test results, the coding scheme is optimized for performance, and based on actual application requirements and feedback, the coding scheme is continuously iterated and optimized.

[0085] The working principle of the above technical solution is: from the thermodynamic parameter library of the material, select the key parameters that have a significant impact on the material performance. These parameters usually include thermal expansion coefficient, thermal conductivity, heat capacity, etc., which have an important influence on the thermal stability, thermal conductivity and thermal response of the material; according to the physical meaning and experimental data of these parameters, determine the effective value range of each parameter. For example, the thermal expansion coefficient may have a stable value within a specific temperature range, while the thermal conductivity may be affected by factors such as material composition and structure. For example, suppose that the thermal expansion coefficient of a certain material has a value range of 1×10-5 (1 / ℃) at room temperature, the thermal conductivity has a value range of 1 to 5 (W / m·K) at room temperature, and the heat capacity has a value range of 500 to 1000 (J / kg·℃) at room temperature. These value ranges will serve as the basis for the subsequent coding scheme design. For each selected thermodynamic parameter, design a binary coding scheme. The number of bits of the code is determined according to the value range and accuracy requirements of the parameter. For example, if the value range of a parameter is between 0 and 1, and it needs to be accurate to four decimal places, a sufficiently long binary string can be used to represent this parameter. The specific number of bits can be determined by calculation. For example, a 16-bit binary number can represent integers between 0 and 65535. If 0 to 65535 is linearly mapped to between 0 and 1 (and four decimal places are retained), the required accuracy can be achieved. The coding scheme can use fixed-length or variable-length binary strings to represent different thermodynamic parameters. Fixed-length coding schemes are suitable for parameters with relatively stable value ranges, while variable-length coding schemes can be dynamically adjusted according to the value range of the parameters; during the encoding process, the real values ​​of the thermodynamic parameters need to be quantized and converted into binary strings. The quantization process can be uniform quantization or non-uniform quantization, depending on the value range and distribution characteristics of the parameters; during the decoding process, the binary string needs to be restored to the real value of the thermodynamic parameter. The restoration process can be achieved by inverse quantization, that is, according to the quantization rules and mapping relationships during encoding, the binary string is converted into the corresponding real value; the influence of different coding bits on the accuracy of parameter representation can be determined by theoretical analysis and experimental verification. Generally speaking, the more coding bits there are, the higher the accuracy of parameter representation, but the computational complexity and storage requirements will also increase accordingly; the accuracy of the coding scheme can be evaluated by comparing the experimental data and the encoded parameter values. If the difference between the encoded parameter value and the experimental data is within an acceptable range, the coding scheme is considered to have sufficient accuracy; according to the preset accuracy requirements (such as four decimal places), evaluate whether the coding scheme meets the requirements. If the coding scheme cannot meet the preset accuracy requirements, it is necessary to redesign the coding scheme or adjust the number of coding bits; apply the designed binary coding scheme to the encoding and decoding process of actual thermodynamic parameters.By comparing the original data and the encoded data, the accuracy and reliability of the coding scheme can be evaluated; based on the test results, the performance of the coding scheme can be optimized. This includes adjusting the number of coding bits, improving the coding algorithm, or introducing new coding technologies. By continuously optimizing the coding scheme, its accuracy and reliability can be improved and meet the needs of actual applications; based on actual application needs and feedback, the coding scheme is continuously iterated and optimized. This includes collecting more experimental data, analyzing the performance bottlenecks of the coding scheme, and exploring new coding technologies and methods. Through continuous iteration and optimization, it can be ensured that the coding scheme is always in the best state.

[0086] The effects of the above technical scheme are as follows: by screening out key parameters that have a significant impact on material properties, such as thermal expansion coefficient, thermal conductivity, heat capacity, etc., from the thermodynamic parameter library of the material, the amount of data for subsequent processing can be greatly reduced, and the efficiency of parameter processing can be improved; at the same time, these key parameters can more accurately reflect the thermodynamic properties of the material, and provide more valuable information for subsequent analysis and optimization; using a binary coding scheme to encode thermodynamic parameters can make full use of the advantages of computer processing binary data and improve the speed and efficiency of data processing; by determining the number of bits of the encoding according to the value range and accuracy requirements of the parameters, it can be ensured that the encoded data can accurately represent the value of the original parameter, avoiding information loss and precision reduction; the number of bits of the encoding scheme can be dynamically adjusted according to the value range and accuracy requirements of the parameters. This flexibility enables the encoding scheme to adapt to the needs of different parameters and different precision requirements, and improves the adaptability of the encoding scheme; by establishing a mapping relationship between binary coding and the real value of the thermodynamic parameter, the accuracy and reliability of the encoding and decoding process can be ensured. At the same time, this mapping relationship is also helpful to analyze the influence of different coding bits on the accuracy of parameter representation, and provide strong support for the optimization of coding schemes; by comparing experimental data and encoded parameter values, it is possible to evaluate whether the accuracy of the coding scheme meets the preset requirements. If it does not meet the requirements, the coding scheme can be adjusted and optimized to improve the coding accuracy; this accuracy evaluation and optimization process can ensure the accuracy and reliability of the coding scheme in practical applications, and provide strong guarantees for subsequent model parameter updates and genetic algorithm optimization; according to the test results, the performance of the coding scheme is optimized, including adjusting the number of coding bits, improving the coding algorithm, or introducing new coding technologies. These optimization measures can further improve the efficiency and accuracy of the coding scheme; at the same time, by continuously iterating and optimizing the coding scheme, it can be made more in line with actual application requirements and feedback, and improve the practicality and reliability of the coding scheme; this technical solution provides strong support for the analysis and optimization of the thermodynamic properties of materials. By accurately representing and processing thermodynamic parameters, the research and development process of materials can be accelerated, and the performance and quality of materials can be improved; this technical solution can also promote technological innovation and development in related fields. For example, in industries such as energy, aerospace, and automobiles, there are high requirements for the thermodynamic properties of materials. By adopting this technical solution, technological progress and innovative development in these industries can be promoted.

[0087] In one embodiment of the present invention, the S4 includes:

[0088] S41. Determine the type and number of atoms in the simulation based on the chemical composition of the material, and set the relevant parameters of the simulation, such as initial temperature, pressure, boundary conditions, and select the potential function based on the characteristics of the material, such as Lennard-Jones potential, Morse potential, etc.; process

[0089] S42, the temperature during the simulation is controlled by a Nose-Hoover heat bath, and the pressure during the simulation is controlled by a Parrinello-Rahman method;

[0090] S43. According to the simulation accuracy and computational efficiency, the time step is selected, and during the simulation process, the microstructure image is output regularly to observe the evolution process of the microscopic damage; the simulation accuracy is obtained by the following formula:

[0091]

[0092] Where T represents the quantity of physical properties to be compared (e.g. elastic modulus, yield strength, etc.); p i represents the simulated predicted value of the physical property of the i-th treatment; e i represents the experimental measurement value of the physical property of the i-th treatment; R i represents the acceptable error range or reference value w of the physical characteristic of the i-th treatment i represents the weight factor of the ith physical characteristic, reflecting the importance of this characteristic in the overall accuracy assessment;

[0093] The calculation efficiency is obtained by the following formula:

[0094]

[0095] Where E represents the computational efficiency; T r Indicates the time actually required to complete the simulation; T p represents the simulation and ideal or expected completion time; C represents the consumption coefficient of computing resources;

[0096] S44, comparing the simulated microstructure image with the experimental microstructure image to analyze the similarities and differences of microscopic damage;

[0097] S45. Compare the differences between the simulated physical properties (such as elastic modulus, yield strength, etc.) and the experimental data, and adjust the model parameters by gradient descent according to the differences between the simulation results and the experimental data.

[0098] The working principle of the above technical solution is as follows: first determine the type and number of atoms involved in the simulation according to the chemical composition of the material. This is the basis of the simulation, because different atoms have different physical and chemical properties, which will directly affect the simulation results; set the initial temperature, pressure and boundary conditions of the simulation. These parameters will determine the physical state of the simulation environment, thereby affecting the microstructure and performance of the material; the selection of the initial temperature should be based on the actual working environment or research purpose of the material; the setting of the pressure should consider the mechanical properties of the material and the influence of the external environment; the selection of boundary conditions should reflect the actual constraints of the material, such as free boundaries, fixed boundaries or periodic boundaries; according to the characteristics of the material, select a suitable potential function to describe the interaction between atoms. The potential function is a mathematical model that describes the interaction force between atoms in the simulation, which is crucial to the accuracy of the simulation results; the Lennard-Jones potential is suitable for describing the interaction between inert gas atoms; the Morse potential is often used to describe the interaction between metal atoms; when selecting the potential function, its applicability, accuracy and computational efficiency should be considered. For example, suppose we are studying the microscopic damage process of a certain metal material. According to the chemical composition of the material, we determined that the type of atoms involved in the simulation is metal atoms, and set the initial temperature to room temperature (such as 300K), the pressure to atmospheric pressure, and the boundary condition to periodic boundary. When selecting the potential function, we chose the Morse potential to describe the interaction between metal atoms. The temperature during the simulation is controlled by the Nose-Hoover heat bath. The Nose-Hoover heat bath is a commonly used temperature control method. It limits the movement of atoms by introducing a non-Newtonian term in the equation of motion, so that the temperature fluctuates around the expected value; the pressure during the simulation is controlled by the Parrinello-Rahman method. The Parrinello-Rahman method is a pressure control method that simultaneously considers the changes in atomic positions and simulation unit size; it adjusts the size and shape of the simulation unit to adapt to the changes in internal pressure, thereby maintaining pressure stability during the simulation; according to the accuracy and computational efficiency of the simulation, the appropriate time step is selected. The time step is the step size of time advancement in the simulation, which determines the precision and amount of calculation of the simulation; a smaller time step can improve the accuracy of the simulation, but increase the amount of calculation; a larger time step can improve the calculation efficiency, but may reduce the accuracy of the simulation; therefore, when selecting the time step, it is necessary to weigh the relationship between accuracy and efficiency; during the simulation process, microstructure images are output regularly. These images can reflect the microstructure and damage evolution of the material during the simulation process; by observing these images, we can understand the microscopic damage mechanism of the material, the damage extension process, and the impact of damage on the material performance; compare the simulated microstructure images with the experimental microstructure images.This comparison can verify the accuracy and reliability of the simulation results; by comparing the similarities and differences of the microstructures, the effectiveness of the simulation method can be evaluated and a basis for subsequent optimization and improvement can be provided; the physical properties obtained by simulation (such as elastic modulus, yield strength, etc.) are compared with experimental data. This comparison can further verify the accuracy of the simulation results and evaluate the ability of the simulation method to predict material properties; according to the difference between the simulation results and the experimental data, the model parameters are adjusted through optimization algorithms such as gradient descent. These parameters include the parameters of the potential function, initial conditions, boundary conditions, etc.; by adjusting these parameters, the difference between the simulation results and the experimental data can be reduced, and the precision and accuracy of the simulation can be improved; according to the accuracy of the simulation results, the initial conditions and potential functions of the simulation can be adjusted. These adjustments can further optimize the simulation process and improve the reliability and accuracy of the simulation results; for example, if it is found that there is a large difference between the simulation results and the experimental data, it can be considered to replace the potential function or adjust the initial conditions.

[0099] The effects of the above technical solution are as follows: by determining the type and number of atoms in the simulation according to the chemical composition of the material, setting precise simulation parameters (such as initial temperature, pressure, boundary conditions) and selecting appropriate potential functions, the technical solution can accurately simulate the microstructure of the material; this helps to deeply understand the microscopic properties of the material and provide strong support for the design and optimization of the material; by controlling the temperature and pressure in the simulation process by the Nose-Hoover heat bath and Parrinello-Rahman methods respectively, the technical solution can ensure that the simulation environment fluctuates stably within the set range; this helps to reduce the uncertainty in the simulation process and improve the reliability and accuracy of the simulation results; by selecting an appropriate time step and regularly outputting microstructure images during the simulation process, the technical solution can efficiently observe the evolution of microscopic damage; this helps to timely discover potential damage in the material and provide an important basis for the life prediction and maintenance of the material; by comparing the simulated microstructure images with the experimental microstructure images, and comparing the differences between the simulated physical properties and the experimental data, the technical solution can verify the simulation results The accuracy of the simulation method can be ensured, which can help ensure the effectiveness of the simulation method and provide reliable guarantee for subsequent material research and application. According to the difference between the simulation results and the experimental data, the model parameters are adjusted by optimization algorithms such as gradient descent, and the initial conditions and potential functions of the simulation are adjusted according to the accuracy of the simulation results. This technical solution can continuously improve the accuracy of the simulation. This can help further optimize the simulation method, improve the reliability and accuracy of the simulation results, and provide more powerful support for the research and application of materials science. The technical solution can be used to study the relationship between the microstructure and performance of materials, and reveal the microscopic damage mechanism and evolution law of materials. This can help provide a scientific basis for the design, preparation and optimization of materials and promote the development of materials science. In engineering applications, the technical solution can be used to evaluate the durability and reliability of materials and predict the life and performance changes of materials. This can help provide an important reference for engineering design and maintenance and ensure the safety and reliability of engineering. The technical solution can also be used for the development and research of new materials, and predict their performance in practical applications by simulating the microstructure and performance of new materials. This can help accelerate the research and development process of new materials and promote the development of the new materials industry. The above formulas are used to evaluate the accuracy and computational efficiency of the simulation, respectively, and the time step selection and regular output of microstructure images are performed during the simulation based on these evaluation results.The accuracy calculation formula of the simulation can comprehensively evaluate the overall accuracy of the simulation by considering the difference between the simulation prediction values ​​and the experimental measurement values ​​of multiple physical properties, and combining the acceptable error range and weight factor of each property; the weight factor reflects the importance of each physical property in the overall accuracy evaluation, so that the evaluation results are more in line with the needs of practical applications; through accuracy evaluation, problems in the simulation can be discovered in time, such as the large difference between the prediction value and the experimental measurement value, so that corresponding measures can be taken to improve the reliability of the simulation; the accuracy evaluation results can be used as an important basis for selecting the time step. When the simulation accuracy is low, the time step can be appropriately reduced to improve the accuracy; conversely, when the simulation accuracy is high, the time step can be appropriately increased to improve the computational efficiency. The computational efficiency evaluation formula can quantitatively evaluate the computational efficiency of the simulation by comparing the actual time required to complete the simulation with the expected completion time, and combining the consumption coefficient of computing resources; the computational efficiency evaluation results can be used as a basis for optimizing the allocation of computing resources. When the computational efficiency is low, you can consider increasing computing resources or optimizing algorithms to improve efficiency; conversely, when the computational efficiency is high, you can reasonably allocate resources to perform more simulation or analysis work. By regularly evaluating computational efficiency, bottleneck problems in the simulation process, such as insufficient computing resources or inefficient algorithms, can be discovered in a timely manner, so that corresponding measures can be taken to improve them. The comprehensive application of accuracy and computational efficiency evaluation formulas can optimize computational efficiency while improving simulation accuracy, thereby obtaining higher quality simulation results. During the simulation process, microstructure images are regularly output, and combined with the evaluation results of accuracy and computational efficiency, the evolution of microscopic damage can be more accurately observed and analyzed, providing strong support for research in the fields of materials science and engineering.

[0100] In one embodiment of the present invention, the S41 includes:

[0101] Determine the elemental composition of the material through chemical composition analysis (such as mass spectrometry, energy spectrum, etc.); Based on the elemental composition, identify the types of atoms that need to be included in the simulation and determine the quantity ratio of each atom;

[0102] Set the initial temperature, pressure and boundary conditions of the simulation according to the actual working environment of the material;

[0103] Select the appropriate potential function according to the crystal structure of the material and the interaction characteristics between atoms; for common potential functions such as Lennard-Jones potential and Morse potential, its parameters (such as potential well depth, equilibrium distance, etc.) need to be adjusted according to the characteristics of the material;

[0104] By changing the initial conditions (such as temperature, pressure) and potential function parameters, analyze the impact of these changes on the simulation results and determine the parameters that have a significant impact on the simulation results;

[0105] Use known experimental data or theoretical results to verify the simulation; compare the simulation results with the experimental data or theoretical predictions to evaluate the accuracy and reliability of the simulation; adjust the simulation parameters based on the verification results;

[0106] Generate the initial positions of atoms based on the crystal structure and density of the material; assign initial velocities to atoms based on the initial temperature and the principles of statistical mechanics;

[0107] Refine the boundary conditions to ensure the accuracy and stability of the simulation results;

[0108] Select appropriate simulation software and algorithms based on the complexity of the simulation and computing resources; reasonably allocate computing resources (such as CPU, GPU, memory, etc.) based on the scale of the simulation and computing requirements.

[0109] The working principle of the above technical solution is: use mass spectrometry, energy spectrum and other analytical methods to determine the elemental composition of the material. These methods infer the elemental composition by measuring the energy of specific wavelengths of light or particles emitted or absorbed by the material; determine the type and quantity of atoms: identify the type of atoms that need to be included in the simulation based on the elemental composition, and calculate the number ratio of each atom. This is usually based on the chemical formula or molecular formula of the material. For example, if the material is an aluminum alloy, the main atomic type is aluminum (Al) and other alloying elements that may exist (such as copper Cu, magnesium Mg, etc.). Through chemical composition analysis, the content of each element can be determined, thereby determining the number ratio of each atom in the simulation. According to the actual working environment of the material, the initial temperature, pressure and boundary conditions of the simulation are reasonably set. The selection of the initial temperature needs to take into account the phase transition point and thermal stability of the material to ensure that no unexpected phase transition occurs during the simulation. The initial pressure should reflect the stress state of the material in the real application. The boundary conditions are reasonably set according to the specific problems of the simulation (such as tension, compression, shear, etc.) to simulate the real physical process; according to the crystal structure of the material and the interaction characteristics between atoms, select the appropriate potential function. For commonly used potential functions such as Lennard-Jones potential and Morse potential, their parameters (such as potential well depth, equilibrium distance, etc.) need to be adjusted according to the characteristics of the material to ensure the accuracy of the simulation results; for example, for metal materials, Morse potential may be a suitable choice because Morse potential can better describe the interaction between metal atoms. When adjusting the parameters, the parameters of the potential function can be optimized by comparing the simulation results with known experimental data or theoretical predictions; the simulation is verified using known experimental data or theoretical results. The accuracy and reliability of the simulation are evaluated by comparing the differences between the simulation results and the experimental data or theoretical predictions; according to the verification results, the simulation parameters (such as initial conditions, potential function parameters, etc.) are adjusted to improve the accuracy and reliability of the simulation; the initial positions of atoms are generated according to the crystal structure and density of the material. This usually involves the calculation of lattice constants and the arrangement of atoms in the lattice; the initial velocities are assigned to atoms according to the initial temperature and the principles of statistical mechanics. This usually involves the relationship between temperature and kinetic energy and the application of Maxwell-Boltzmann distribution; the boundary conditions are refined to ensure the accuracy and stability of the simulation results. For example, under periodic boundary conditions, it is necessary to ensure that the interactions between atoms at the boundaries are consistent with those of the internal atoms to avoid the influence of boundary effects on the simulation results; according to the complexity of the simulation and the computing resources, choose the appropriate simulation software and algorithm. This involves the evaluation of the simulation scale, computing requirements, and software and algorithm performance; for example, for large-scale molecular dynamics simulations, it may be necessary to use high-performance computing clusters and parallel algorithms to improve computing efficiency. When choosing simulation software, factors such as the functionality, ease of use, and compatibility with other tools need to be considered.

[0110] The effects of the above technical solution are: determining the elemental composition of the material through chemical composition analysis, and identifying the type and quantity ratio of atoms that need to be included in the simulation based on this, which can accurately reflect the microstructure of the material; setting the initial temperature, pressure and boundary conditions according to the actual working environment of the material, and selecting a suitable potential function and adjusting its parameters, can more accurately simulate the physical and chemical behavior of the material; by changing the initial conditions and potential function parameters, and analyzing the impact of these changes on the simulation results, the parameters that significantly affect the simulation results can be determined, thereby optimizing the simulation settings; using known experimental data or theoretical results to verify the simulation, the accuracy and reliability of the simulation can be evaluated, and the simulation parameters can be adjusted according to the verification results to further improve the reliability of the simulation results; according to the complexity of the simulation and computing resources, selecting suitable simulation software and algorithms, and reasonably allocating computing resources (such as CPU, GPU, memory, etc.) can ensure the efficient operation of the simulation process while avoiding waste of resources; refining the boundary conditions For example, ensuring that the interaction between atoms at the boundary under periodic boundary conditions is consistent with that of internal atoms can reduce the impact of boundary effects on simulation results and enhance the stability and accuracy of simulation results. This technical solution provides a powerful tool for materials science research, which can deeply explore the relationship between the microstructure and performance of materials and provide a scientific basis for the design and development of new materials. At the same time, this technical solution can also provide reliable simulation results for engineering applications, guide the selection of materials and performance optimization, and improve product quality and reliability. In the field of material development, this technical solution can accelerate the research and development process of new materials, predict the performance and behavior of materials through simulation, and reduce experimental costs and time. In engineering applications, this technical solution can evaluate the durability and reliability of materials, predict the performance changes of materials under specific working environments, and provide important references for engineering design and maintenance. In the field of scientific research, this technical solution can reveal the microscopic damage mechanism and evolution law of materials, and provide a theoretical basis for the optimization and improvement of material performance.

[0111] In one embodiment of the present invention, S5 includes:

[0112] S51, divide the original data set into a training set, a validation set and a test set according to the time sequence; use the early data as the training set, the mid-term data as the validation set, and the late data as the test set;

[0113] S52, check the consistency of data distribution among the training set, validation set and test set, and select evaluation indicators such as accuracy, recall, F1 score, and mean square error (MSE) according to the prediction target of the model;

[0114] S53, and through k-fold cross validation, comprehensively evaluate the prediction performance of the model, analyze the influence of model parameters on the prediction results, identify key parameters; analyze the importance of features to the prediction results, and identify key features;

[0115] S54. According to the analysis results, adjust the model's hierarchy, activation function, loss function, etc., and optimize the feature selection and extraction strategy;

[0116] S55. Set multiple convergence criteria, such as the prediction performance improvement after multiple consecutive iterations is less than the set threshold, the model parameter changes tend to be stable, etc. When the set convergence criteria are met, stop the iteration and output the final model.

[0117] The working principle of the above technical solution is as follows: the original data set is divided into a training set, a validation set and a test set to ensure the independence of model training, validation and testing, so as to obtain a more accurate model performance evaluation; the data set is divided according to the time order to ensure the uniformity of data distribution. The early data is used as the training set for model learning; the mid-term data is used as the validation set to adjust the parameters of the model and select the best model; the late data is used as the test set to evaluate the final performance of the model; the data distribution between the training set, validation set and test set is ensured to be consistent to avoid the deviation of model performance evaluation caused by the difference in data distribution; according to the prediction target of the model, appropriate evaluation indicators are selected, such as accuracy, recall rate, F1 score (for classification problems) or mean square error (MSE, for regression problems), etc. These indicators can objectively reflect the prediction performance of the model; the training set is further divided into k complementary subsets, and k-1 subsets are selected as training data each time, and the remaining subset is used as validation data to evaluate the model. Repeat this process k times, each time selecting a different subset as validation data, and finally calculate the average value of the k groups of test indicators as an estimate of model accuracy; through k-fold cross-validation, the prediction performance of the model is comprehensively evaluated. At the same time, analyze the degree of influence of model parameters on the prediction results and identify key parameters; analyze the importance of features to the prediction results and identify key features; according to the analysis results of k-fold cross-validation, adjust the model's hierarchy, activation function, loss function, etc. to improve the model's prediction performance; optimize feature selection and extraction strategies, select features that are more important to the prediction results for model training to improve the efficiency and accuracy of the model; set multiple convergence criteria, such as the prediction performance improvement after multiple consecutive iterations is less than the set threshold, and the model parameter changes tend to be stable. These criteria are used to determine whether the model has converged to the optimal solution; when the set convergence criteria are met, stop the iteration process and output the final model. This can avoid the problem of model overfitting caused by overtraining.

[0118] The effects of the above technical solution are as follows: by dividing the training set, validation set and test set according to chronological order and ensuring the uniformity of data distribution, data leakage and overfitting problems can be effectively avoided, so that the model can be better generalized to unseen data; the practice of using early data as training sets, mid-term data as validation sets, and late data as test sets helps to simulate the flow and changes of data in the real world, making the model more adaptable to actual application scenarios; checking the consistency of data distribution between training sets, validation sets and test sets can ensure that the performance of the model on different data sets is consistent and reliable; selecting appropriate evaluation indicators according to the prediction target of the model, such as accuracy, recall rate, F1 score and mean square error (MSE), can more accurately measure the performance of the model and provide strong support for model optimization; through k-fold cross-validation, the predictive power of the model can be It can conduct a comprehensive evaluation and reduce the performance evaluation bias caused by improper data division; analyze the importance of model parameters and features, which helps to identify key parameters and key features, and provide an important basis for model optimization; according to the analysis results of k-fold cross-validation, the model's hierarchy, activation function, loss function, etc. are adjusted to significantly improve the model's performance; optimizing feature selection and extraction strategies can reduce the impact of redundant features on model performance and improve the efficiency and accuracy of the model; setting multiple convergence criteria, such as the prediction performance improvement after multiple consecutive iterations is less than the set threshold, and the model parameter changes tend to be stable, can ensure that the model stops iterating at the right time to avoid overfitting problems caused by over-training; when the iteration is stopped and the final model is output, it can ensure that the model has achieved the optimal performance on the given data set, providing reliable prediction results for practical applications.

[0119] One embodiment of the present invention, as Figure 2 As shown, a self-adaptive high temperature damage model construction system, the system comprising:

[0120] Data acquisition module: obtains experimental data of materials in high temperature environment, including stress-strain curve, temperature-time curve, microstructure image, and pre-processes the collected data through machine learning algorithm, including data cleaning, missing value filling and feature selection, and extracts key damage features;

[0121] Model building module: build a foundation damage prediction model based on a deep learning framework, use preprocessed data to train the model, and use a reinforcement learning mechanism to enable the foundation damage prediction model to adaptively adjust internal parameters based on prediction errors;

[0122] Parameter adjustment module: Use genetic algorithms to perform global searches on the thermodynamic parameters in the model, find the optimal parameter combination, and dynamically adjust the model parameters in combination with real-time temperature monitoring data;

[0123] Process simulation module: Based on molecular dynamics simulation technology, the microstructure evolution process of the material at high temperature is simulated, the simulation results are compared with the experimental data, and the model parameters are further optimized through the feedback mechanism;

[0124] Iterative optimization module: select independent verification data sets to verify the foundation damage prediction model and evaluate its prediction performance. According to the evaluation results, the model is iteratively optimized.

[0125] The working principle of the above technical solution is as follows: obtain data from material experiments under high temperature environment, including stress-strain curve (describing the deformation of materials under stress), temperature-time curve (describing the change of temperature during the experiment), and microstructure image (showing the microstructure change of materials under high temperature); remove noise and outliers in the data to ensure the accuracy and consistency of the data; use interpolation or other algorithms to fill in the missing parts of the data to ensure the integrity of the data; extract key features that have an important impact on material damage prediction from the original data, such as stress peak, strain rate, temperature gradient, etc.; build a basic damage prediction model based on deep learning framework (such as TensorFlow, PyTorch, etc.). This model can receive preprocessed data as input and output the prediction results of material damage; use preprocessed data to train the model so that it can learn the characteristics and laws in the data; introduce reinforcement learning mechanism so that the model can automatically adjust its internal parameters according to the prediction error to improve the accuracy of the prediction; use genetic algorithm to search the thermodynamic parameters in the model globally to find the optimal parameter combination. Genetic algorithms can efficiently search the solution space by simulating natural selection and genetic mechanisms; they can dynamically adjust the model parameters in combination with real-time temperature monitoring data. This ensures that the model can maintain high prediction accuracy under different temperature conditions; based on molecular dynamics simulation technology, the microstructure evolution process of materials at high temperatures is simulated. This helps to understand the damage mechanism of materials at high temperatures; the simulation results are compared with experimental data, and the model parameters are further optimized through the feedback mechanism. This ensures that the model can more accurately reflect the actual damage of the material; an independent verification data set is selected to verify the basic damage prediction model. This can evaluate the prediction performance of the model on unknown data; the model is iteratively optimized based on the verification results. If the prediction accuracy does not meet the requirements, the model structure, parameters or data preprocessing methods are adjusted, and retraining and verification are performed until satisfactory prediction accuracy is achieved.

[0126] The effects of the above technical solution are as follows: preprocessing the collected data through machine learning algorithms, including data cleaning, missing value filling and feature selection, can effectively remove noise and outliers in the data to ensure the accuracy and consistency of the data; the extraction of key damage features helps the model to better learn the features and laws in the data, thereby improving the accuracy of the prediction; this method builds the model completely based on experimental data, avoiding the problem of relying on a large number of assumptions and simplifications in traditional methods, and improving the reliability and practicality of the model; through the reinforcement learning mechanism, the basic damage prediction model can adaptively adjust the internal parameters according to the prediction error, thereby continuously optimizing the model performance; this adaptive adjustment capability enables the model to maintain a high prediction accuracy when facing different high-temperature environments and material types; using genetic algorithms to perform a global search of the thermodynamic parameters in the model, it is possible to find the optimal parameter combination and further improve the prediction accuracy of the model; combining real-time temperature monitoring data to dynamically adjust the model parameters, so that The model can adapt to environmental changes in real time, improving the practicality and robustness of the model; based on molecular dynamics simulation technology, the microstructure evolution process of the material at high temperature is simulated, which helps to deeply understand the damage mechanism of the material at high temperature; this simulation technology can provide strong theoretical support for the construction and optimization of the model; the simulation results are compared with the experimental data, and the model parameters are further optimized through the feedback mechanism, so that the model can more accurately reflect the actual damage of the material; this feedback optimization mechanism helps to improve the prediction accuracy and generalization ability of the model; the basic damage prediction model is verified by selecting an independent verification data set, which can objectively evaluate the prediction performance of the model; this verification method helps to discover problems and deficiencies in the model and provide direction for subsequent optimization; the model is iteratively optimized according to the verification results, and the model structure, parameters or data preprocessing methods are continuously adjusted until satisfactory prediction accuracy is achieved; this iterative optimization method enables the model to be continuously improved and perfected, improving the practicality and accuracy of the model.

[0127] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A method for constructing an adaptive high temperature damage model, characterized in that: The method comprises: S1. Obtain experimental data of materials under high temperature environment, pre-process the collected data through machine learning algorithm, and extract key damage features; S2. Build a foundation damage prediction model based on a deep learning framework, use preprocessed data to train the model, and use a reinforcement learning mechanism to enable the foundation damage prediction model to adaptively adjust internal parameters based on prediction errors; S3, using genetic algorithms to globally search the thermodynamic parameters in the model, find the optimal parameter combination, and dynamically adjust the model parameters in combination with real-time temperature monitoring data; S4. Based on molecular dynamics simulation technology, simulate the microstructure evolution process of materials at high temperatures, compare the simulation results with experimental data, and further optimize the model parameters through feedback mechanism; S5. Select an independent validation data set to verify the foundation damage prediction model and evaluate its prediction performance. Based on the evaluation results, iteratively optimize the model.

2. The method for constructing an adaptive high temperature damage model according to claim 1, characterized in that: Said S1 comprises: S11. According to the thermal stability range of the material, set the temperature gradient, perform conditional loading, and set different loading rates and stress levels under each loading condition; S12, at different temperature points and loading stages, the microstructure images were obtained by SEM, and the evolution of microscopic damage was recorded; S13. Identify outliers based on statistical methods and further verify the rationality of the data in combination with the physical properties of the material; S14, extracting the physical characteristics, thermal characteristics and microstructure characteristics of the material; S15. Perform PCA analysis on the feature matrix, retain the principal components, and use LDA for feature selection in classification problems.

3. The method for constructing an adaptive high temperature damage model according to claim 2, characterized in that: The S11 includes: According to the thermal stability range of the material, the temperature gradient is initially set, and the temperature gradient is fine-tuned based on the effect of the material at different temperatures; Design multiple loading conditions for each temperature gradient, and further refine the settings of loading rate and stress level under each loading condition; The temperature control system is used to control the accuracy and stability of the temperature during the experiment. During the heating process, the temperature changes are monitored in real time and fine-tuned as needed. During the experiment, the key data is recorded in real time through the data acquisition system, and the experimental conditions are fed back and adjusted according to the real-time monitored data and experimental phenomena; Based on the results and analysis of the preliminary experiments, the experimental conditions are further optimized; based on the adjustment and optimization of the experimental conditions, the experimental plan is iteratively updated.

4. The method for constructing an adaptive high temperature damage model according to claim 1, characterized in that: The S2 comprises: S21. Select a deep learning framework. If it is used to process time series data, select the LSTM model. If it is used to process long sequence data, select the Transformer model. S22. Combined with the physical properties of the material, a loss function containing multiple sub-items is designed, and L1 or L2 regularization terms are added; S23. Select the batch size based on the amount of data and model complexity, and randomly shuffle the data before each training cycle; S24, for the fully connected layer and LSTM layer, use the Xavier initialization method, for the layer with ReLU activation function, use the He initialization method; S25. When the model prediction error is greater than a preset threshold, the model parameters are adjusted to make the prediction result closer to the true value, and during the model training process, the model parameters are dynamically adjusted according to the reward and punishment signals of the prediction results.

5. The method for constructing an adaptive high temperature damage model according to claim 4, characterized in that: The S25 comprises: After each training cycle, the prediction error of the model on the validation set is calculated, and the prediction performance of the model is evaluated by comparing the difference between the predicted value and the true value; Formulate strategies for adjusting model parameters based on the size and distribution of prediction errors; update model parameters through gradient descent algorithm during model training; dynamically adjust model parameters based on reward and punishment signals of prediction results; Through the learning rate decay strategy, in the early stage of training, the learning rate is set to converge quickly; and as the training progresses, the learning rate is gradually reduced to adjust the model parameters; Add L1 or L2 regularization terms to the loss function and dynamically adjust the weight of the regularization term based on the model's predictive performance; During the model training process, the changes in the loss function, the prediction error on the validation set, and the update of the model parameters are monitored in real time. If the prediction error on the validation set does not decrease significantly after multiple consecutive training cycles, or shows an upward trend, an early stopping strategy is adopted to avoid overfitting of the model. Tune the model's hyperparameters through grid search; find the optimal hyperparameter combination by comparing the model performance under different hyperparameter combinations; Evaluate the performance of the model on the test set, select the optimal model architecture by comparing the performance of different models, and iteratively optimize the model based on the evaluation results on the test set.

6. The method for constructing an adaptive high temperature damage model according to claim 1, characterized in that: The S3 includes: S31, converting the thermodynamic parameter into a binary string, each bit representing a parameter value, and encoding it using the real value of the thermodynamic parameter; S32. Combine multiple thermodynamic indicators of materials to design a multi-objective fitness function and assign weights according to the importance of each indicator; S33, collecting temperature data in real time by selecting temperature sensors arranged at key positions of the material, and preprocessing the collected temperature data; S34. According to the temperature changes monitored in real time, parameter updating rules are designed, and the upper and lower limits of parameter adjustment are set according to the optimal parameter combination obtained by the genetic algorithm.

7. The method for constructing an adaptive high temperature damage model according to claim 6, characterized in that: The S31 includes: Select key parameters that have a significant impact on material performance from the material's thermodynamic parameter library, and determine the effective value range of each parameter based on the physical meaning of these parameters and experimental data; For each selected thermodynamic parameter, a binary coding scheme is designed; and the number of bits of the coding is determined according to the value range and accuracy requirements of the parameter; Establish the mapping relationship between binary coding and real values ​​of thermodynamic parameters; analyze the impact of different coding bit numbers on parameter representation accuracy; evaluate whether the coding accuracy meets the preset requirements by comparing experimental data and encoded parameter values; The designed binary coding scheme is applied to the encoding and decoding process of actual thermodynamic parameters, and the accuracy and reliability of the coding scheme are evaluated by comparing the original data with the encoded data. Based on the test results, the coding scheme is optimized for performance, and based on actual application requirements and feedback, the coding scheme is continuously iterated and optimized.

8. The method for constructing an adaptive high temperature damage model according to claim 1, characterized in that: The S4 comprises: S41. Determine the type and quantity of atoms in the simulation according to the chemical composition of the material, set relevant parameters of the simulation, and select the potential function according to the characteristics of the material; S42, the temperature during the simulation is controlled by a Nose-Hoover heat bath, and the pressure during the simulation is controlled by a Parrinello-Rahman method; S43, selecting a time step according to the simulation accuracy and computational efficiency, and regularly outputting microstructure images during the simulation process to observe the evolution of microscopic damage; S44, comparing the simulated microstructure image with the experimental microstructure image to analyze the similarities and differences of microscopic damage; S45. Compare the differences between the simulated physical properties and the experimental data, and adjust the model parameters by gradient descent according to the differences between the simulation results and the experimental data.

9. The method for constructing an adaptive high temperature damage model according to claim 1, characterized in that: The S5 comprises: S51, dividing the original data set into a training set, a validation set and a test set according to the time sequence, taking the early data as the training set, the mid-term data as the validation set, and the late data as the test set; S52, checking the consistency of data distribution among the training set, validation set and test set, and selecting evaluation indicators according to the prediction target of the model; S53, and through k-fold cross validation, comprehensively evaluate the prediction performance of the model, analyze the influence of model parameters on the prediction results, identify key parameters; analyze the importance of features to the prediction results, and identify key features; S54. According to the analysis results, the model is adjusted and the feature selection and extraction strategy is optimized; S55. Set multiple convergence criteria. When the set convergence criteria are met, stop iteration and output the final model.

10. An adaptive high temperature damage model construction system, characterized in that: The system comprises: Data acquisition module: obtains experimental data of materials in high temperature environment, pre-processes the collected data through machine learning algorithms, and extracts key damage features; Model building module: build a foundation damage prediction model based on a deep learning framework, use preprocessed data to train the model, and use a reinforcement learning mechanism to enable the foundation damage prediction model to adaptively adjust internal parameters based on prediction errors; Parameter adjustment module: Use genetic algorithms to perform global searches on the thermodynamic parameters in the model, find the optimal parameter combination, and dynamically adjust the model parameters in combination with real-time temperature monitoring data; Process simulation module: Based on molecular dynamics simulation technology, the microstructure evolution process of the material at high temperature is simulated, the simulation results are compared with the experimental data, and the model parameters are further optimized through the feedback mechanism; Iterative optimization module: select independent verification data sets to verify the foundation damage prediction model and evaluate its prediction performance. According to the evaluation results, the model is iteratively optimized.

Images