A material performance prediction method and related apparatus

Abstract

The application discloses a material performance prediction method and related device, obtaining a preset component and process parameters of a material; using a trained time sequence model to process the preset component and process parameters to obtain performance parameters of the material under the preset component and process parameters; the training process of the time sequence model includes: constructing a material single-step performance dataset containing the component, process parameters and performance parameters of the material; using the material single-step performance dataset to train a proxy model to obtain a trained proxy model; using the trained proxy model to predict performance parameters of the material after each step of the preparation process of the material, using the component, process parameters of the material and the predicted performance parameters of the material after each step to construct a material performance time sequence dataset; using the material performance time sequence dataset to train a time sequence model to obtain a trained time sequence model. The application can realize accurate performance prediction and further improve the material development efficiency.

Description

Technical Field

[0001] This invention belongs to the field of materials science and technology, and relates to a method and related device for predicting material properties, applicable to material systems such as titanium alloys, conductive polymers, and functional ceramics that have undergone multiple processes. Background Technology

[0002] Advanced materials, as the cornerstone of modern industrial systems, can directly drive technological iteration in strategic fields such as aerospace, energy, and biomedicine through breakthroughs in their performance. Generally, materials with superior properties require multiple processing steps for development. Their macroscopic properties are not only directly affected by the final process but also retain the long-term effects of inherited time-step effects from preceding processes. The microscopic manifestations of texture, dislocation configuration, and phase distribution generated by preceding processing profoundly influence the microstructure evolution path of subsequent processes, thereby affecting their service performance under different operating conditions. The chain response of materials—composition-process-microstructure-property—reveals the significant time-dependent and historically inherited nature of performance evolution, with a complexity far exceeding that of single-material systems. For example, high-strength titanium alloys require a series of sequential processing steps, including vacuum melting, heat treatment, and mechanical processing; functional ceramics require high-purity powder chemical synthesis, powder processing, molding, and sintering; and ceramic matrix composites achieve strengthening through fiber preform preparation, matrix deposition / impregnation, and high-temperature treatment. The sustained impact of such multi-step processes on performance is essentially a non-Markovian evolution of the material state in phase space. Multiple processing steps together constitute a high-dimensional process time series, creating a vast space of choices and unexplored areas, posing significant challenges to the design and development of high-performance materials. Traditional trial-and-error methods are limited by experimental conditions, development costs, and the domain knowledge of technical personnel. When faced with an exponentially increasing space of combinations of component variables (proportions of multiple alloying elements, composite material interfaces, etc.) and process parameters (temperature field, strain rate, holding time, etc.), experimental costs grow exponentially, posing significant difficulties for composition design, process optimization, performance prediction, and mechanism exploration.

[0003] With the rapid development of artificial intelligence and machine learning technologies, data science has provided a new paradigm for solving the aforementioned dilemmas. Based on information technology, a data-driven intelligent R&D model can achieve performance prediction through data accumulation, feature construction, and model building, thereby guiding experiments and calculations. Currently, researchers have introduced machine learning models and reverse design strategies into the field of materials science, using Gaussian process regression models to predict the strength of composite materials and designing biomedical titanium alloys with ultra-low modulus through Bayesian optimization and active learning frameworks. However, existing methods have three limitations: First, traditional machine learning models, such as support vector machines, Gaussian processes, and random forest regression models, treat composition and process features as disordered vectors, failing to characterize the sequential causal relationships of multiple continuous processes; second, while deep learning models can extract spatial features, they struggle to accurately model and train long-range time dependencies across process steps; third, when the sample size is small (n < 500 in small sample scenarios), due to redundant feature parameters, the above models are prone to overfitting, resulting in fragmented and high-noise "composition / process-performance" mapping data.

[0004] Recurrent neural networks (RNNs) take sequential data as input and use iterative computation on the hidden states, memorizing and recursively processing the sequence's evolution. The state of each structural unit in the model at the current time step is jointly determined by the input of that time step and the output (or state) of the previous time step. This characteristic endows the model with the ability to capture temporal dependencies, providing an effective solution for solving time-series problems in materials science and optimizing development processes. To avoid gradient vanishing or exploding during model training, two variants of RNNs have emerged: Long Short-Term Memory (LSTM) networks introduce gating mechanisms, achieving selective information memorization through forget gates, input gates, and output gates; Gated recurrent units update the internal structure based on this LSM network, merging the forget gate and output gate into an update gate, achieving the same function while reducing computational complexity and realizing long-term memory. However, currently, there is limited research on constructing time-series models by approximating step-by-step processes as time series, and their structure-property relationships need further exploration. Therefore, how to construct time-series relationships based on small-sample incomplete data, learn the genetic effects in materials processing, and improve the accuracy of multi-process material performance prediction has become an urgent scientific problem to be solved. Summary of the Invention

[0005] The purpose of this invention is to solve the problems in the prior art and provide a material performance prediction method and related apparatus. This method can deeply explore the genetic effects in the processing, learn the long-term implicit mapping relationship between composition or process and performance, achieve accurate performance prediction, and further improve the efficiency of material development.

[0006] To achieve the above objectives, the present invention employs the following technical solution:

[0007] A method for predicting material properties includes the following process:

[0008] Obtain the preset composition and process parameters of the material;

[0009] The preset composition and process parameters are processed using a pre-trained time-series model to obtain the material's performance parameters under the preset composition and process parameters. The training process of the time-series model includes: constructing a single-step performance dataset of the material containing its composition, process parameters, and performance parameters; training a surrogate model using the single-step performance dataset to obtain a trained surrogate model; using the trained surrogate model to predict the material's performance parameters after each step of the material's preparation process; constructing a material performance time-series dataset using the material's composition, process parameters, and the predicted performance parameters after each step; and training the time-series model using the material performance time-series dataset to obtain a trained time-series model.

[0010] Preferably, the training process of the proxy model includes the following steps:

[0011] Using the data vector composed of the material composition and the process parameters of the preparation process in the single-step performance dataset as input and the performance parameters as output, the surrogate model is trained to obtain a trained surrogate model.

[0012] In the data vector composed of the components and process parameters of the preparation process, the process parameters before the nth step of the preparation process are set to the values ​​corresponding to the single-step performance data of the material, and the process parameters after the nth step of the preparation process are set to zero.

[0013] When training the surrogate model, each step in the preparation process is traversed n times until training is complete, resulting in a trained surrogate model.

[0014] Preferably, the performance parameter includes room temperature tensile strength.

[0015] Preferably, the timing model adopts a bidirectional gated cyclic unit model.

[0016] Preferably, the hyperparameters of the bidirectional gated recurrent unit model are configured as follows: the hidden layer is set to 2 layers, with 16 and 256 neurons in the two layers, respectively; the activation function is the tanh function, the learning rate is 0.01, the batch size is 32, and the optimizer is Adam.

[0017] Preferably, the proxy model adopts an extreme gradient boosting decision tree model.

[0018] Preferably, the parameter combination of the extreme gradient boosting decision tree model includes: 900 trees, a maximum depth of 5, a learning rate of 0.01, a sample sampling ratio of 0.8, and a feature sampling ratio of 0.6.

[0019] The present invention also provides a material property prediction system for implementing the above-described material property prediction method of the present invention, comprising:

[0020] Data acquisition unit: used to acquire the preset composition and process parameters of the material;

[0021] Prediction Unit: Used to process the preset composition and process parameters using a pre-trained time-series model to obtain the material's performance parameters under the preset composition and process parameters; wherein, the training process of the time-series model includes: constructing a single-step performance dataset of the material containing the material's composition, process parameters, and performance parameters; training a surrogate model using the single-step performance dataset to obtain a trained surrogate model; using the trained surrogate model to predict the material's performance parameters after each step in the material's preparation process; constructing a material performance time-series dataset using the material's composition, process parameters, and the predicted performance parameters after each step; and training the time-series model using the material performance time-series dataset to obtain a trained time-series model.

[0022] The present invention also provides an electronic device, comprising:

[0023] One or more processors;

[0024] A storage device on which one or more programs are stored;

[0025] When the one or more programs are executed by the one or more processors, the one or more processors implement the material property prediction method of the present invention as described above.

[0026] The present invention also provides a storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the material property prediction method of the present invention as described above.

[0027] Compared with the prior art, the present invention has the following beneficial effects:

[0028] This invention provides a material performance prediction method that can accurately solve core technical challenges in the current materials field, while also bringing significant practical value. Specifically, addressing the problem of difficulty in constructing time-series relationships using small, incomplete sample data, this invention first trains a surrogate model using a single-step material performance dataset, making full use of fragmented single-step experimental data without relying on complete multi-step time-series data. Then, the trained surrogate model is used to predict the material performance after each process step, and combined with composition data to construct a complete material performance time-series dataset, filling the gap in time-series relationship construction in small-sample scenarios and providing high-quality data support for subsequent model training. Regarding the mining of genetic effects in material processing, the time-series dataset generated by this invention includes intermediate performance parameters for each step, reflecting the genetic influence of preceding processes on material performance. The time-series model, by capturing time dependencies, can automatically learn the progressive correlation between preceding processes, intermediate performance, and final performance, breaking through the limitations of traditional models that only focus on initial input and final output, and achieving long-term implicit mapping relationship learning between composition, process, and performance. In practical applications, this invention's strategy of "using a surrogate model to complete data and a time-series model to learn the correlation" makes the model more sensitive to process parameters at different stages and avoids biases caused by ignoring the cumulative effects of intermediate processes. For example, in the prediction of multi-stage processing of high-strength titanium alloys, compared with traditional machine learning regression models, the error is reduced by approximately 350 MPa, significantly improving prediction accuracy, reducing trial-and-error experiments in material development, increasing development efficiency, and reducing costs. Furthermore, this method is not limited to material type or process steps; changing the single-step dataset allows adaptation to various materials such as aluminum alloys and composite materials. It is also compatible with different recurrent neural network variants, demonstrating strong practicality and scalability. Attached Figure Description

[0029] Figure 1 This is a flowchart of the material property prediction method of the present invention;

[0030] Figure 2 This is a graph showing the performance results of the proxy model of the present invention;

[0031] Figure 3 This is a graph showing the performance results of the timing model of this invention;

[0032] Figure 4 This is a graph showing the statistical results of the time-series model's prediction of the overall process performance in the ablation experiment of this invention.

[0033] Figure 5 This is a graph showing the statistical results of the time-series model's prediction of the final process performance in the ablation experiment of this invention;

[0034] Figure 6 This is a comparison chart of the prediction results of the traditional model and the time-series model of this invention for materials with a complete process chain;

[0035] Figure 7This is a comparison chart of the prediction results of the traditional model and the time-series model for materials with a complete process chain in the incremental experiments of this invention;

[0036] Figure 8 This is a graph showing the performance results of the time series model in the incremental experiments of this invention. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0038] The material performance prediction method of this invention includes: constructing a single-step material performance dataset containing composition and process based on multiple information sources; constructing a surrogate model using the single-step material performance dataset and selecting the optimal parameter combination through grid search; using the trained surrogate model to predict the performance of the material for each process and constructing a time-series data set of material performance; constructing a time-series model using the time-series data set of material performance, selecting the optimal combination through grid search, and training to obtain a time-series model that can accurately predict the material performance for multiple processes.

[0039] See Figure 1 The material property prediction method of the present invention includes the following steps:

[0040] Step 1: Obtain a single-step performance dataset of the material, including its composition, process parameters, and one or more performance parameters.

[0041] The single-step material performance dataset includes several data vectors, each of which can be illustrated as follows:

[0042] [a1, a2, a3, a4, a5, a6, a7, a8, a9, a 10 , b1, b2, b3, c1, c2, c3, d1, d2, d3, e1, e2, e3, f1, f2, f3, g1, g2, g3, h1, h2, h3, y],

[0043] Where 'a' is a component parameter, including a1-a 10 , represents ten elements; bh represents the process parameters corresponding to different steps, a total of seven steps, each step of the process includes three process parameters; y represents the performance parameters; when a certain step is not actually operated, the three process parameters of the corresponding step are all set to zero.

[0044] Step 2: Based on the aforementioned single-step data set of material properties, construct a surrogate model and perform hyperparameter mesh search to achieve material property prediction without time-series information;

[0045] Based on the aforementioned material single-step performance dataset, the material preparation process is divided into several time steps with each process step as a node, resulting in several material sub-data that are gradually accumulated in the process steps; the composition and process parameters of the several material sub-data are used as inputs, and a trained surrogate model is used to predict the single-step performance.

[0046] Merge several sub-data sets of the same material to obtain a time-series dataset of material properties with a complete process chain; for non-existent process steps, set the feature value to zero and set the performance parameter to the result of the preceding step, indicating that the step has no actual operation and is only used as a placeholder.

[0047] In other words, the surrogate model is trained using a data vector composed of the material composition and process parameters of the preparation process from the material single-step performance dataset as input and performance parameters as output, to obtain a trained surrogate model. In the data vector composed of the composition and process parameters of the preparation process (i.e., the material sub-data), the process parameters before the nth step of the preparation process are set to the corresponding values ​​in the material single-step performance dataset, and the process parameters after the nth step of the preparation process are set to zero. When training the surrogate model, each step of the preparation process is traversed n times until training is completed, to obtain a trained surrogate model.

[0048] Step 3

[0049] Based on the aforementioned material property time series dataset, the composition is used as the starting state of material processing, each process pass is used as the subsequent process, each pass is used as a time step, the composition or process parameters are used as input, and the material properties are used as output. The same data partitioning method as the surrogate model is used to train time series models such as gated recurrent units.

[0050] The optimal hyperparameter combination of time series models such as grid search gated cyclic units is obtained to obtain a time series model with embedded optimal hyperparameter combination;

[0051] A trained time series model is obtained to predict the performance of materials with known processes, and the prediction error is used as an evaluation index of the model's prediction accuracy.

[0052] The above-mentioned solution of the present invention can deeply explore the genetic effects in the processing, learn the long-term implicit mapping relationship between composition or process and performance, prove the temporal effectiveness of time series model and the necessity of previous historical information, realize accurate prediction of material performance in multi-pass process, and further improve the development efficiency of materials.

[0053] Example

[0054] See Figure 1This embodiment of the material performance prediction method is based on the room temperature tensile strength of high-strength titanium alloys, constructing a single-step material performance dataset. A surrogate model is trained using composition and process parameters as features and room temperature tensile strength as the target attribute. Based on the trained surrogate model, time-series data is generated to construct a material performance time-series dataset. The time-series model is then trained to explore the implicit mapping relationship between composition and process parameters and room temperature tensile strength, achieving accurate prediction of material performance across multiple processes.

[0055] Based on information from multiple sources, including literature, public databases, and in-group experimental data, a single-step performance dataset for the room temperature tensile strength of titanium alloys was constructed. The dataset contains 31 dimensions of features, including chemical composition and process parameters; and one dimension of the target attribute, namely room temperature tensile strength. The database exhibits strong overall heterogeneity, with significant differences between samples in the high-dimensional space. A small amount of data shows local similarity, and some samples are closely related in the high-dimensional space. The broad data coverage is beneficial for the model to explore a wide range of compositional and process dimensions.

[0056] See Figure 2 Eight traditional machine learning models (SVR-RBF, XGBoost, Lasso, Decision Tree, GPR, ENR, Bayesian Ridge, and KNN) were selected as surrogate models. Ten-fold cross-validation was performed based on a dataset of single-step performance of room-temperature tensile strength of titanium alloys. The optimal hyperparameter combination was selected through grid search, and the test set coefficient of determination (R²) was used as the surrogate model. 2 The root mean square error (RMSE) and root mean square error (RMSE) are used as model evaluation metrics, with the shaded area representing the 95% confidence interval. Among them, the extreme gradient boosting decision tree model (XGBoost) performs well on the test set R. 2 With a value greater than 0.82 and an RMSE less than 110 MPa, it exhibits the best prediction accuracy and generalization ability. Its parameter combination is: 900 trees, maximum depth of 5, learning rate of 0.01, sample ratio of 0.8, and feature ratio of 0.6.

[0057] Based on a single-step performance dataset of room temperature tensile strength of titanium alloys, a time-series dataset of room temperature tensile strength of titanium alloys is constructed. The specific implementation logic is as follows: If a material involves multiple processes, each process step is used as a node to divide the time into several sub-data points. Each sub-data point is progressively accumulated with one process step based on the preceding process. A trained Extreme Gradient Boosting Decision Tree (XGBoost) model is used to predict the single-step performance result of each sub-data point after splitting. Its output value is regarded as the approximate measured data under this process path. All predicted sub-data points of the same material are merged with the original data according to the process sequence to form the time-series performance data of the complete process chain. For non-existent process steps or unprocessed time steps, the feature values ​​are set to zero, and the performance parameters of this step inherit the results of the preceding steps. Based on literature data, experimental measurement data, and surrogate model prediction results, this invention constructs a time-series dataset of room temperature tensile strength of titanium alloys.

[0058] See Figure 3 The time-series dataset of room temperature tensile strength of titanium alloys was partitioned using the same method as the single-step performance dataset of room temperature tensile strength of titanium alloys. A bidirectional gated cyclic unit (GRU) model with 8 time steps was selected, and the test set determination coefficient (R²) was used. 2 The root mean square error (RMSE) and root mean square error were used as model evaluation metrics, with the shaded area representing the 95% confidence interval. The model hyperparameters were configured as follows: two hidden layers with 16 and 256 neurons respectively; the activation function was tanh, the learning rate was 0.01, the batch size was 32, the optimizer was Adam, and the training epochs were 500. The model performed well on the test set. 2 Greater than 0.83 and RMSE less than 110 MPa.

[0059] To verify the temporal validity of the time series model, this invention designs the following ablation experiment: 50% of the temporal information is removed sequentially, and the stability of the fitting results is ensured through cyclic traversal using multiple random seeds. (Refer to...) Figure 4 The model's prediction accuracy for the overall process performance decreases significantly with the reduction of retained time-series information. (Refer to...) Figure 5 The model's prediction accuracy for the final process performance maintains the same trend, meaning that the less time-series information is retained, the more the model's performance degrades.

[0060] Based on a trained time-series model (GRU) and a traditional machine learning model (SVR), predictions are made on a set of experimentally validated material property data covering the entire process. (Refer to...) Figure 6 The prediction results of the GRU model follow the same trend as the experimental data, while the SVR model is not sensitive to process parameter characteristics, and its prediction error is significantly higher than that of the GRU model. Moreover, the prediction error shows a dynamic fluctuation and a continuous growth trend as the process chain lengthens.

[0061] To verify the necessity of prior historical information, this invention designs the following incremental experiment: Real experimental data is divided into eight equal-length sub-data sets according to process steps. In each iteration, one sub-data set is selected to supplement the training set, with a learning rate of 0.001. The updated training set is used to fine-tune the model, and performance metrics for the last time step are predicted. (Refer to...) Figure 7 The prediction accuracy of the GRU model dynamically improves as prior historical information accumulates, while the prediction results of the SVR model do not show significant changes at each time step. This phenomenon simulates the actual operational logic of technicians—predicting based on the current state and guiding subsequent process adjustments. This dynamic learning characteristic exhibited by the GRU model indicates that it effectively captures key characterizing information from historical process data. (Refer to...) Figure 8 The GRU model in the incremental experimental test set R 2 The dynamic increase and the decrease in RMSE volatility further confirm the rationality of the "gradual monitoring" operation.

[0062] As can be seen from the above scheme, this invention, through surrogate model construction, time-series data generation, and time-series model construction methods, deeply explores the genetic effects in the processing, learns the long-term implicit mapping relationship between composition or process and performance, and significantly improves the prediction accuracy of material properties with long process flows. In specific experimental verification, compared with traditional machine learning regression models, the time-series model is sensitive to the characteristics of process parameters at different stages, and reduces the prediction error of high-strength titanium alloys that have undergone multiple heat treatments and mechanical treatments by about 350 MPa, achieving accurate performance prediction and further improving the efficiency of material development.

[0063] Furthermore, embodiments of the present invention also provide a material property prediction system for implementing the material property prediction method described above. This system includes:

[0064] Data acquisition unit: used to acquire the preset composition and process parameters of the material;

[0065] Prediction Unit: Used to process the preset composition and process parameters using a pre-trained time-series model to obtain the material's performance parameters under the preset composition and process parameters; wherein, the training process of the time-series model includes: constructing a single-step performance dataset of the material containing the material's composition, process parameters, and performance parameters; training a surrogate model using the single-step performance dataset to obtain a trained surrogate model; using the trained surrogate model to predict the material's performance parameters after each step in the material's preparation process; constructing a material performance time-series dataset using the material's composition, process parameters, and the predicted performance parameters after each step; and training the time-series model using the material performance time-series dataset to obtain a trained time-series model.

[0066] The embodiments of the present invention also provide corresponding electronic devices and computer-readable storage media for implementing the solutions provided in the embodiments of the present invention.

[0067] The electronic device includes a storage device and one or more processors. The storage device stores instructions or code, and the processors execute the instructions or code to enable the device to perform the material property prediction method described in any embodiment of this application.

[0068] The storage medium stores a computer program, which, when executed by a processor, implements the material property prediction method described in any embodiment of this application.

[0069] Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort should fall within the scope of protection of the present invention.

[0070] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for predicting material properties, characterized in that, The process includes the following: Obtain the preset composition and process parameters of the material; The preset composition and process parameters are processed using a pre-trained time-series model to obtain the material's performance parameters under the preset composition and process parameters. The training process of the time-series model includes: constructing a single-step performance dataset of the material containing its composition, process parameters, and performance parameters; training a surrogate model using the single-step performance dataset to obtain a trained surrogate model; using the trained surrogate model to predict the material's performance parameters after each step of the material's preparation process; constructing a material performance time-series dataset using the material's composition, process parameters, and the predicted performance parameters after each step; and training the time-series model using the material performance time-series dataset to obtain a trained time-series model. The proxy model employs an extreme gradient boosting decision tree model. The training process of the proxy model includes the following steps: Using the data vector composed of the material composition and the process parameters of the preparation process in the single-step performance dataset as input and the performance parameters as output, the surrogate model is trained to obtain a trained surrogate model. In the data vector composed of the components and process parameters of the preparation process, the process parameters before the nth step of the preparation process are set to the values ​​corresponding to the single-step performance data of the material, and the process parameters after the nth step of the preparation process are set to zero. When training the surrogate model, each step in the preparation process is traversed n times until training is complete, resulting in a trained surrogate model.

2. The material property prediction method according to claim 1, characterized in that, The performance parameters include room temperature tensile strength.

3. The material property prediction method according to claim 1, characterized in that, The timing model adopts a bidirectional gated cyclic unit model.

4. The material property prediction method according to claim 3, characterized in that, The hyperparameters of the bidirectional gated recurrent unit model are configured as follows: two hidden layers with 16 and 256 neurons respectively; the activation function is tanh, the learning rate is 0.01, the batch size is 32, and the optimizer is Adam.

5. The material property prediction method according to claim 1, characterized in that, The parameter combination of the extreme gradient boosting decision tree model includes: 900 trees, a maximum depth of 5, a learning rate of 0.01, a sample sampling ratio of 0.8, and a feature sampling ratio of 0.

6.

6. A material property prediction system, characterized in that, The method for predicting material properties according to any one of claims 1-5 includes: Data acquisition unit: used to acquire the preset composition and process parameters of the material; Prediction Unit: Used to process the preset composition and process parameters using a pre-trained time-series model to obtain the material's performance parameters under the preset composition and process parameters; wherein, the training process of the time-series model includes: constructing a single-step performance dataset of the material containing the material's composition, process parameters, and performance parameters; training a surrogate model using the single-step performance dataset to obtain a trained surrogate model; using the trained surrogate model to predict the material's performance parameters after each step in the material's preparation process; constructing a material performance time-series dataset using the material's composition, process parameters, and the predicted performance parameters after each step; and training the time-series model using the material performance time-series dataset to obtain a trained time-series model.

7. An electronic device, characterized in that, include: One or more processors; A storage device on which one or more programs are stored; When the one or more programs are executed by the one or more processors, the one or more processors implement the material property prediction method as described in any one of claims 1 to 5.

8. A storage medium, characterized in that, It stores a computer program, wherein the computer program, when executed by a processor, implements the material property prediction method as described in any one of claims 1-5.

Images