Design method of high-performance alloy with small sample extrapolation capability

Abstract

According to the high-performance alloy design method with the small sample extrapolation capability, component parameters, process parameters and performance indexes of an alloy are collected, and an original data set is constructed; a table variational auto-encoder is used for learning the distribution rule of material features in a potential space, and generation and characterization of complex components and process combinations are achieved; multi-scale features are extracted and optimized through feature engineering, and an extended feature set containing chemical composition, heat treatment parameters and microstructure information is formed; on the basis, a performance prediction agent model fusing a neural network model and a gradient boosting tree model is constructed, and high-precision performance prediction is achieved; through an active learning strategy based on uncertainty measurement, dynamic sampling and iterative optimization are carried out on a model prediction result, and the extrapolation capability and material discovery efficiency of the model are continuously improved. According to the method, automatic recommendation and optimization of the high-performance aluminum alloy material under the small sample condition are achieved, and the intelligence and the high-throughput level of material design are improved.

Description

A high-performance alloy design method with small sample extrapolation capability Technical Field

[0001] This invention belongs to the interdisciplinary field of materials science, artificial intelligence and computer-aided design, and relates to a high-performance alloy design method with small sample extrapolation capability. Background Technology

[0002] The optimization design of material properties is a core aspect of new material development. Especially in the field of structural materials such as high-performance aluminum alloys and nickel-based alloys, how to quickly screen out high-performance material combinations while meeting various performance constraints such as strength, plasticity, and thermal stability has always been an important research topic in the field of materials science.

[0003] Currently, materials design methods still primarily rely on researchers' domain knowledge and accumulated experience, combined with experimental trial and error to adjust composition and processes. This experience-based approach is inefficient, time-consuming, and highly dependent on expert capabilities, making it difficult to meet the demands for high-throughput, rapid, and intelligent discovery of new materials. In practical applications with limited sample sizes and complex performance requirements, traditional methods often struggle to efficiently obtain optimal material combinations that satisfy multiple objectives.

[0004] In recent years, with the development of machine learning technology, researchers have attempted to apply data-driven modeling methods to material property prediction tasks, such as using random forests, support vector machines, and neural networks to construct surrogate models between material composition, processing, and properties. However, these methods generally suffer from two prominent problems: first, most methods rely on a large number of high-quality samples, making it difficult to stably model in scenarios with scarce data; second, the models are mostly interpolation-based, lacking extrapolation capabilities, making it difficult to accurately predict the properties of new material combinations outside the training distribution, thus limiting their application in new material discovery. Summary of the Invention

[0005] To address the aforementioned technical problems, the purpose of this invention is to provide a high-performance alloy design method with small-sample extrapolation capability.

[0006] The present invention provides a high-performance alloy design method with small sample extrapolation capability, comprising:

[0007] Step 1: Obtain the composition parameters, process parameters, and performance indicators of the known alloy, construct the original dataset, and preprocess the original data;

[0008] Step 2: Use the preprocessed composition parameters and process parameters to train a tabular variational autoencoder to obtain the latent spatial distribution characteristics, and generate multiple sets of potential candidate material samples in the latent space;

[0009] Step 3: Use the SHAP method to perform feature contribution analysis on the generated candidate material samples, identify candidate material samples with obvious performance orientation, and eliminate unreasonable or meaningless combinations to obtain the filtered candidate material sample set;

[0010] Step 4: Based on prior knowledge of materials science, extract statistical features of elemental properties and construct an extended feature set for the samples to improve the expressive power of the model;

[0011] Step 5: Optimize the extended feature set using a genetic algorithm, select the three optimal extended feature combinations that contribute most significantly to the prediction performance, and add the optimal extended features to the original dataset and the candidate material sample set;

[0012] Step 6: Construct a performance prediction proxy model consisting of a neural network model and a gradient boosting tree model, and train the proxy models corresponding to the three performance predictions using the original dataset after feature enhancement.

[0013] Step 7: Using the trained surrogate model and an active learning strategy, the importance of the candidate material sample set after feature enhancement is ranked, and the optimal sample is selected from the filtered candidate material sample set according to the importance ranking.

[0014] Step 8: Conduct experimental verification on the selected optimal sample, feed the obtained performance data back to the surrogate model for retraining and updating, and repeat steps 2 to 7 until a material design scheme that meets the target performance requirements is obtained.

[0015] The present invention provides a high-performance alloy design method with small-sample extrapolation capability, which has the following beneficial effects:

[0016] (1) Under limited sample conditions, the generative model automatically constructs alloy design schemes that meet physical and technological constraints in the potential space, reducing the reliance on human experience and improving the systematicness and consistency of the design process;

[0017] (2) A set of high-dimensional extended general feature set generation method for alloy performance prediction was constructed, which covers the basic physical and chemical characteristics of alloys and has the ability to be adapted to different alloy systems.

[0018] (3) A feature selection strategy that balances computational efficiency and modeling effectiveness is proposed to quickly determine a subset of features suitable for modeling from an expanded feature set, thereby reducing the impact of redundant features on model stability.

[0019] (4) An alloy performance prediction model integrating latent variables was established. While ensuring the prediction accuracy within the domain, the generalization stability of the model in the region outside the data distribution was improved, so that the alloy design process can explore the parameter space outside the domain more reliably.

[0020] (5) Combining Bayesian optimization ideas, candidate alloy schemes are sorted and screened based on the prediction mean and uncertainty, thereby improving the exploration efficiency under limited experimental or high-precision computing resources. Attached Figure Description

[0021] Figure 1 is a flowchart of a high-performance alloy design method with small sample extrapolation capability according to the present invention. Detailed Implementation

[0022] As shown in Figure 1, a high-performance alloy design method with small-sample extrapolation capability according to the present invention includes:

[0023] Step 1: Obtain the composition parameters, process parameters and performance indicators of the known alloy, construct the original dataset, and preprocess the original data.

[0024] The composition parameters include: elemental composition and their mass ratio; the process parameters include: solution temperature, aging temperature, aging time, forming process, heat treatment medium, strain hardening method, aging type, post-treatment process, and heat treatment type; the performance indicators include: tensile strength, yield strength, and elongation.

[0025] The preprocessing includes data cleaning, outlier removal, and missing value imputation. For samples with missing values, a joint imputation method based on K-nearest neighbor interpolation (KNN-Imputation) combined with domain prior constraints is used to reduce the impact of data anomalies on model training and improve data quality and model stability. Six categories of features—forming process, heat treatment medium, strain hardening method, aging type, post-processing process, and heat treatment type—are digitally encoded, and three continuous features—solution temperature, aging temperature, and aging time—are normalized.

[0026] In this embodiment, 828 sets of original samples were obtained by collecting chemical composition parameters, heat treatment process parameters, and mechanical property indicators of aluminum alloy samples. Input features included 19 elemental compositions and their mass ratios, namely Si, Fe, Cu, Mn, Mg, Cr, Zn, V, Ti, Zr, Li, Ni, Ga, Bi, Pb, B, Be, Sc, and Al. Nine process parameter features were collected from existing laboratory data and literature. Output variables were tensile strength (UTS) and elongation (EL).

[0027] Step 2: Train a tabular variational autoencoder (TVAE) using the preprocessed composition and process parameters to obtain the latent spatial distribution features and generate multiple sets of potential candidate material samples in this latent space. Compared with traditional autoencoder models, the TVAE described in this invention can capture multimodal distributions while taking into account the continuity and interpretability of the material composition space, thereby improving the physical rationality of the generated samples.

[0028] The table variational autoencoder is trained using a combination of maximum likelihood reconstruction loss and KL divergence to simultaneously capture the potential distributional relationship between components and process features. The loss function is as follows:

[0029]

[0030]

[0031]

[0032]

[0033] The total loss function L is composed of the continuous feature reconstruction error. Discrete feature reconstruction error and latent space regularization term It consists of three parts; among them, the reconstruction loss of continuous numerical features adopts the mean square error. , The total number of samples, and Indicates the first The true and predicted values ​​of each sample; the reconstruction loss for discrete process schemes is calculated using weighted binary cross-entropy. , Indicates the number of discrete features. Indicates the first The number of categories for a discrete variable. Indicates the first Among the discrete features, the first one is... The weights of each category are used to mitigate the class imbalance problem. In other words, the lower the frequency of a certain category, the higher its corresponding weight and the greater the penalty. Indicates the first Among the discrete variables, the first one is the... The true labels of the samples in each category , The output represents the discrete features of the reconstructed model; regularization term. The KL divergence of each dimension of independent Gaussian distributions relative to the standard normal distribution is used to constrain the latent spatial distribution, where For the first The squared mean of each latent dimension, output by the encoder, represents the cost of deviating from the prior center. Indicates the first The variance of dimension, when At that time, situations arise where losses increase, penalty variance is too large, or noise is excessive. It is used to penalize cases where the variance approaches 0 and degenerates into deterministic encoding, while ensuring the continuity and sampleability of the latent space;

[0034] After training, Gaussian distribution sampling is performed in the latent space mapped by the encoder of the tabular variational autoencoder. The sampled latent variables are input into the decoder of the tabular variational autoencoder, and the decoder outputs the generated new alloy element mass fraction and corresponding process parameters.

[0035] In this embodiment, the TVAE encoder employs a two-layer fully connected structure with hidden layer dimensions of 256 and 64, respectively, and the activation function is ReLU. The encoder output latent vector obtains the latent distribution parameters—mean and log-variance—through two sets of linear mappings. The TVAE decoder is a three-layer fully connected network that progressively reconstructs the latent variable z back to the original feature space. Training results on the aluminum alloy dataset show that after 5000 training iterations, the reconstruction accuracy for discrete variables reaches 0.83, and the mean absolute error (MAE) for continuous variables is 0.12, indicating that TVAE can effectively capture the latent distribution structure of the composition-process space.

[0036] After training the TVAE on the original dataset, Gaussian distribution sampling is performed in its latent space to generate 10,000 new potential candidate points. Subsequently, these latent vectors are input into the trained TVAE decoder to map from the latent space back to the original feature space, generating new alloy element mass fractions and corresponding process parameters.

[0037] Step 3: Use the SHAP method to perform feature contribution analysis on the generated candidate material samples, identify candidate material samples with obvious performance orientation, and eliminate unreasonable or meaningless combinations to obtain the filtered candidate material sample set, specifically:

[0038] Step 3.1: Calculate candidate material samples No. SHAP value of each feature , The sign and absolute value of the value reflect the direction and intensity of the influence of this feature on performance, respectively; for candidate material samples No. One characteristic, when At that time, it was considered that this feature had a positive promoting effect on the performance of the current candidate material samples.

[0039] Step 3.2: Statistics on the condition that... Conditions, No. The value distribution interval of a feature across all candidate material samples is defined as the feature. The effective positive contribution interval.

[0040] Step 3.3: During the candidate material sample screening process, when the value of a certain feature in the candidate material sample to be screened falls into its corresponding positive contribution interval, the feature is determined to meet the screening condition; when all features of the candidate material sample simultaneously meet its positive contribution interval constraint, the candidate material sample is determined to be a sample that meets the target screening criteria, and the screened candidate material sample set is obtained.

[0041] In this embodiment, the SHAP algorithm is used to analyze the feature contribution of the candidate material samples generated by TVAE, calculating the marginal impact of each feature on UTS and EL. Based on the feature range where both UTS and EL contribute positively, candidate ranges are defined: for example, Cu element mass fraction of 0.5%–5.5%, Si of 0–2%, and Fe of 0–0.7%. Features exhibiting monotonically positive gains within the existing data range are not constrained. Finally, the 10,000 candidate material samples generated by TVAE are filtered using SHAP to obtain a set of 376 filtered candidate material samples. This step effectively eliminates unreasonable or physically meaningless combinations, improving the rationality and interpretability of the generated samples.

[0042] Step 4: Based on prior knowledge of materials science, extract statistical features of elemental properties and construct an extended feature set of samples to improve the expressive power of the model.

[0043] To address the challenge that traditional features are insufficient to fully characterize the multi-scale relationships within materials, this invention introduces a feature enhancement strategy combining domain knowledge and data-driven approaches in the basic feature construction stage. Specifically, by integrating materials science mechanism cognition with information theory analysis methods, composite features with clear physical meaning and interpretability are constructed from multiple dimensions, including alloy composition ratios, strengthening phase distribution, impurity control, grain structure characteristics, and process parameters, to more comprehensively depict the intrinsic relationship mechanism between composition, process, and performance. The newly added extended features specifically include:

[0044] New features at the atomic scale include: atomic radius, atomic volume, covalent radius, ionic radius, and their weighted average, range, and variance in the composition;

[0045] Electronic structure characteristics include: the number of valence electrons, the number of electron shells, the number of d orbitals, atomic electronegativity and their combination differences and ratios;

[0046] Thermodynamic and binding energy characteristics include: melting point, vapor pressure, heat of vaporization, atomic binding energy, first ionization energy, and their statistics in multi-component systems;

[0047] Prior characteristics of the domain include: Mg / Si ratio, equivalent Mg2Si content, excess on the Mg-rich side, excess on the Si-rich side, total strengthening elements of the 7xxx series, ratio characteristics of the 6xxx series and 2xxx series, comparison indicators of mixed strengthening mechanisms, impurity control and risk of brittle phase formation, total impurity index, interaction term between grain refinement and strengthening mechanism, and recrystallization inhibition characteristics.

[0048] Information theory features include: compositional distribution entropy, which is used to quantify the diversity of alloying elements. The higher the value, the more complex the compositional system and the more significant the entropy mixing strengthening trend.

[0049] Atomic-scale derived features include: atomic radius mismatch parameters, which reflect the degree of lattice distortion caused by differences in the atomic size of different alloying elements;

[0050] Process-derived characteristics include: the temperature difference between solution treatment and aging, the product of aging temperature and time, the logarithm and reciprocal of aging time, the logarithm of aging temperature, and the logarithm of the temperature-time product.

[0051] Step 5: Optimize the extended feature set using a genetic algorithm, selecting the three optimal extended feature combinations that contribute most significantly to the prediction performance, and then adding these optimal extended features to the original dataset and the candidate material sample set. Specifically:

[0052] Step 5.1: Evaluate the contribution of each extended feature using the ablation method to achieve preliminary screening of extended features.

[0053] Step 5.1.1: Use the original feature set constructed from the component parameters and process parameters in Step 1. Training yields baseline models for each performance level. And obtain the predicted performance index. .

[0054] Step 5.1.2: For each new extended feature ,Will Add the model to the original feature set and retrain the corresponding baseline model to obtain the model. And obtain the predicted performance index. .

[0055] Step 5.1.3: According to the formula Calculate extended features Independent marginal contribution Preliminary screening The extended features are then used to obtain three sets of candidate extended features corresponding to the three performance levels, which are then used for the next step of feature optimization.

[0056] Step 5.2: Perform global search optimization on the three candidate extended feature sets using a genetic algorithm: Encode the three selected candidate extended feature sets into binary form, where each extended feature corresponds to one bit in the encoded sequence, with a value of 1 indicating the selected feature. Then, use this encoded sequence as the initial population for the genetic algorithm, with a population size of 30. Use the model's coefficient of determination R² as the main indicator for fitness evaluation, and adopt a tournament strategy. During the evolution process, the crossover probability is set to 0.8, the mutation probability is set to 0.1, and the number of iterations is 50. Output the three optimal extended feature combinations corresponding to the three performance levels.

[0057] In this embodiment, a combination of 7 features for UTS was finally optimized, among which the constructed neighborhood prior Cu / MgSi contributed the most to the model improvement; for EL, an extended feature combination of length 39 was optimized, among which the constructed Cu / Mg and sum(Zn+Mg) ranked first and second in terms of contribution, which fully demonstrates the rationality of our feature construction.

[0058] The optimization results of the genetic algorithm show that, for the elongation (EL) target performance, the R² index of the LightGBM model improved by 0.023 after adding 39 optimal features; for the tensile strength (UTS) target performance, the R² index improved by 0.004 after adding 7 features, verifying the effectiveness of the feature enhancement and combined optimization strategy.

[0059] Step 6: Construct a performance prediction proxy model consisting of a neural network model (TabPFN) and a gradient boosting tree model (LightGBM). Train the proxy models corresponding to the three performance predictions using the original dataset after feature enhancement.

[0060] In practice, the LightGBM model fine-tunes key hyperparameters using a grid search approach and selects the optimal parameter combination on the validation set. The TabPFN model, a Transformer-based tabular data prediction model, directly performs inference using pre-trained weights, eliminating the need for large-scale parameter training on the current task; it only requires formatting the input data to complete the prediction. Finally, the prediction results of the two models are evaluated separately and used for subsequent comparative analysis.

[0061] Step 7: Using the trained surrogate model and an active learning strategy, rank the candidate material sample set after feature enhancement by importance, and select the optimal sample from the filtered candidate material sample set according to the importance ranking; specifically:

[0062] Step 7.1: Input the screened candidate material sample set into the encoder of the trained tabular variational autoencoder to extract the latent variable representation z of the candidate samples, and input the original data into the encoder of the trained tabular variational autoencoder to obtain the latent space distribution center. Calculate the latent variable representations z and Euclidean distance .

[0063] Step 7.2: Based on the relative position of the Euclidean distance within the candidate sample distribution interval, dynamically adjust the weighting ratio of the neural network model and the gradient boosting tree model, and calculate the adaptive distance weights according to the following formula:

[0064]

[0065] in, Representative sample The Euclidean distance from the data center in a two-dimensional potential space. and This represents the minimum and maximum values ​​of the Euclidean distance.

[0066] Step 7.3: Input the three enhanced candidate material sample sets into the corresponding surrogate models to obtain the expected predicted values ​​of the corresponding performance:

[0067]

[0068] in, The prediction results of the neural network model. The prediction results of the gradient boosting tree model, To predict the expected performance of the enhanced candidate material samples using the surrogate model, we obtain the expected performance values ​​for three properties based on three sets of enhanced candidate material samples.

[0069] In this embodiment, each enhanced candidate material sample set contains 376 samples, which are input into an adaptively weighted TabPFN–LightGBM ensemble surrogate model for performance prediction. This model dynamically assigns weights based on the Euclidean distance between samples in the TVAE latent variable space, thereby achieving an optimal balance between interpolation accuracy and extrapolation capability. Prediction results show that in five-fold cross-validation, the surrogate model significantly outperforms the traditional machine learning model, with the coefficient of determination R for tensile strength (UTS) being significantly higher. 2=0.965, coefficient of determination R for elongation (EL) 2 =0.840. The results show that the surrogate model has high consistency and stability in prediction performance on both near-field and far-field samples.

[0070] Step 7.4: Introduce an uncertainty measurement mechanism to assess the reliability of the prediction results. This uncertainty is represented by a combination of the model's uncertainty estimate and the latent space distance, where the uncertainty estimate is defined as follows:

[0071]

[0072]

[0073]

[0074] in, This represents an uncertainty estimate of the predicted performance of the enhanced candidate material sample. This represents the uncertainty of the gradient boosting tree model. This represents the uncertainty of the neural network model; The uncertainty is estimated by its quantile regression mechanism. For sample index, Indicates the first The samples are at the quantile of Prediction results under the given conditions The difference between the quantiles is the prediction interval; the wider the interval, the higher the uncertainty of the model prediction. The uncertainty is measured by the statistical dispersion induced by the predicted distribution of the model output, where k represents the discrete binning index, used to identify the k value intervals obtained after the target variable is divided. This represents the total number of discrete bins. Indicates the first The predicted result for the sample falls into the first... The probability of each bin is obtained, and this probability is output by the model. Indicates the first The center value of each bin is used as the representative of the prediction for that bin; finally, the uncertainty measure of the sample is normalized.

[0075] Step 7.5: The hypervolume improvement algorithm is used to calculate the sample priority based on the prediction uncertainty estimate and prediction performance value, specifically as follows:

[0076] Step 7.5.1: Let the Pareto optimal performance solution set formed by the original data be... .

[0077] Step 7.5.2: Based on the predicted values ​​of the three properties, the three uncertainty estimates, and the optimal performance solution set obtained from the enhanced candidate material samples, the following steps are taken: ; Calculate the hypervolume improvement value (EHVI) of the corresponding enhanced candidate material sample.

[0078] Step 7.5.3: Sort the candidate material sample set after feature enhancement by the hypervolume improvement value (EHVI) from largest to smallest.

[0079] Step 7.5.4: Select the top ten best samples from the screened candidate material sample set according to the importance ranking.

[0080] In this embodiment, the candidate material sample with the highest score predicts a UTS of 401.59 MPa and an EL of 22.89%. The uncertainty of the EL prediction at this point is extremely high (variance value of 0.38), indicating that the system is greatly affected by uncertainty in the early stage of iteration. The model prioritizes the selection of samples that can improve its known domain for experimentation, and is therefore selected by the active learning algorithm for experimentation. Secondly, in the first iteration, the model also includes the point that breaks through the Pareto front in the experimental scope, indicating that the method can effectively identify high-potential material combinations and realize model-driven intelligent evolutionary design.

[0081] Step 8: Conduct experimental verification on the selected optimal sample, feed the obtained performance data back to the surrogate model for retraining and updating, and repeat steps 2 to 7 until a material design scheme that meets the target performance requirements is obtained.

[0082] The above description is only a preferred embodiment of the present invention and is not intended to limit the ideas of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A high-performance alloy design method with small sample extrapolation capability, characterized in that, This includes: Step 1: Obtain the composition parameters, process parameters, and performance indicators of the known alloy, construct the original dataset, and preprocess the original data; Step 2: Train a tabular variational autoencoder using preprocessed composition and process parameters to obtain latent spatial distribution characteristics and generate multiple sets of potential candidate material samples within this latent space; Step 3: Use the SHAP method to analyze the feature contribution of the generated candidate material samples, identify candidate material samples with obvious performance orientation, and eliminate unreasonable or meaningless combinations to obtain a filtered set of candidate material samples; Step 4: Based on prior knowledge of materials science, extract statistical features of elemental properties to construct an extended feature set for the samples to improve the model's expressive power; Step 5: Optimize the extended feature set using a genetic algorithm, selecting the three optimal extended feature combinations that contribute most significantly to the prediction of three performance characteristics, and... Step 6: Add the optimal extended features to the original dataset and candidate material sample set; Step 7: Construct a performance prediction proxy model composed of a neural network model and a gradient boosting tree model, and train the proxy model corresponding to the three performance predictions using the original dataset after feature enhancement; Step 8: Using the trained proxy model, adopt an active learning strategy to rank the importance of the candidate material sample set after feature enhancement, and select the optimal sample from the selected candidate material sample set according to the importance ranking; Step 9: Perform experimental verification on the selected optimal sample, and feed the obtained performance data back to the proxy model for retraining and updating. Repeat steps 2 to 7 until a material design scheme that meets the target performance requirements is obtained.

2. The high-performance alloy design method with small sample extrapolation capability according to claim 1, characterized in that, The composition parameters include: elemental composition and their mass ratio; the process parameters include: solution temperature, aging temperature, aging time, forming process, heat treatment medium, strain hardening method, aging type, post-treatment process, and heat treatment type; the performance indicators include: tensile strength, yield strength, and elongation.

3. The high-performance alloy design method with small sample extrapolation capability according to claim 2, characterized in that, The preprocessing includes: data cleaning, outlier removal, and missing value filling; digital encoding of six categories of features: forming process, heat treatment medium, strain hardening method, aging type, post-treatment process, and heat treatment type; and normalization of three continuous features: solution temperature, aging temperature, and aging time.

4. The high-performance alloy design method with small sample extrapolation capability according to claim 1, characterized in that, In step 2, the tabular variational autoencoder is trained using a joint optimization of maximum likelihood reconstruction loss and KL divergence. The loss function is as follows: The total loss function L is composed of the continuous feature reconstruction error. Discrete feature reconstruction error and latent space regularization term It consists of three parts; among them, the reconstruction loss of continuous numerical features adopts the mean square error. ， The total number of samples, and Indicates the first The true and predicted values ​​of each sample; the reconstruction loss for discrete process schemes is calculated using weighted binary cross-entropy. ， Indicates the number of discrete features. Indicates the first The number of categories for a discrete variable. Indicates the first Among the discrete features, the first one is... The weights of each category are used to mitigate the class imbalance problem. In other words, the lower the frequency of a certain category, the higher its corresponding weight and the greater the penalty. Indicates the first Among the discrete variables, the first one is the... The true labels of the samples in each category ， The output represents the discrete features of the reconstructed model; regularization term. The KL divergence of each dimension of independent Gaussian distributions relative to the standard normal distribution is used to constrain the latent spatial distribution, where For the first The squared mean of each latent dimension, output by the encoder, represents the cost of deviating from the prior center. Indicates the first The variance of dimension, when At that time, situations arise where losses increase, penalty variance is too large, or noise is excessive. It is used to penalize the case where the variance approaches 0 and degenerates into deterministic encoding, and to ensure the continuity and sampleability of the latent space. After training, Gaussian distribution sampling is performed in the latent space mapped by the encoder of the tabular variational autoencoder. The sampled latent variables are input into the decoder of the tabular variational autoencoder, and the decoder outputs the generated new alloy element mass fraction and corresponding process parameters.

5. The high-performance alloy design method with small sample extrapolation capability according to claim 1, characterized in that, Step 3 specifically includes: Step 3.1: Calculating candidate material samples. The SHAP value of each feature ， The sign and absolute value of the value reflect the direction and intensity of the influence of this feature on performance, respectively; for candidate material samples The One characteristic, when When this feature is considered to have a positive promoting effect on performance in the current candidate material sample; Step 3.2: Statistical analysis is performed when the following conditions are met. Under the conditions, The value distribution interval of a feature across all candidate material samples is defined as the feature. The effective positive contribution interval; Step 3.3: In the process of screening candidate material samples, when the value of a certain feature in the candidate material sample to be screened falls into its corresponding positive contribution interval, it is determined that the feature meets the screening conditions; when all features of the candidate material sample simultaneously meet its positive contribution interval constraints, the candidate material sample is determined to be a sample that meets the target screening criteria, and the screened candidate material sample set is obtained.

6. The high-performance alloy design method with small sample extrapolation capability according to claim 1, characterized in that, The extended features in step 4 include: new atomic-scale features such as atomic radius, atomic volume, covalent radius, ionic radius, and their weighted average, range, and variance in the components; electronic structure features such as the number of valence electrons, number of electron shells, number of d orbitals, atomic electronegativity, and their combination differences and ratios; thermodynamic and binding energy features such as melting point, vapor pressure, heat of vaporization, atomic binding energy, first ionization energy, and their statistics in multi-component systems; and prior features such as the Mg / Si ratio, equivalent Mg2Si content, Mg-rich side excess, Si-rich side excess, total strengthening elements in the 7xxx series, and the 6xxx and 2xx series. The characteristics of the x-series proportions, comparative indicators of mixed strengthening mechanisms, impurity control and risk of brittle phase formation, total impurity index, interaction term between grain refinement and strengthening mechanism, and recrystallization inhibition characteristics; information theory characteristics include: compositional distribution entropy, used to quantify the diversity of alloying elements, the higher the value, the more complex the compositional system and the more significant the entropy mixed strengthening trend; atomic scale derived characteristics include: atomic radius mismatch parameter, used to reflect the degree of lattice distortion caused by the difference in atomic size of different alloying elements; process derived characteristics include: solid solution and aging temperature difference, aging temperature-time product, logarithm and reciprocal of aging time, logarithm of aging temperature, and logarithm of temperature-time product.

7. The high-performance alloy design method with small sample extrapolation capability according to claim 6, characterized in that, Step 5 specifically comprises: Step 5.1: Evaluating the contribution of each extended feature using the ablation method to achieve preliminary screening of extended features; Step 5.1.1: Using the original feature set composed of the component parameters and process parameters from Step 1. Training yields baseline models for each performance level. And obtain the predicted performance index. Step 5.1.2: For each new extended feature ,Will Add the model to the original feature set and retrain the corresponding baseline model to obtain the model. And obtain the predicted performance index. Step 5.1.3: According to the formula Calculate extended features Independent marginal contribution Preliminary screening The extended features are obtained, and three sets of candidate extended features corresponding to three performance levels are obtained respectively for the next step of feature optimization; Step 5.2: Global search optimization is performed on the three sets of candidate extended features using a genetic algorithm: the three sets of candidate extended features after screening are binary encoded, where each extended feature corresponds to one bit of the encoding sequence, and a value of 1 indicates that the feature is selected; then the encoding sequence is used as the initial population of the genetic algorithm, the population size is set to 30, the coefficient of determination R² of the model is used as the main indicator, and the fitness evaluation is based on the tournament strategy; during the evolution process, the crossover probability is set to 0.8, the mutation probability is set to 0.1, the number of iterations is 50, and the three sets of optimal extended feature combinations corresponding to three performance levels are output respectively.

8. The high-performance alloy design method with small sample extrapolation capability according to claim 1, characterized in that, Step 7 specifically involves: Step 7.1: Inputting the selected candidate material sample set into the encoder of the trained tabular variational autoencoder to extract the latent variable representation z of the candidate samples, and inputting the original data into the encoder of the trained tabular variational autoencoder to obtain the latent space distribution center. Calculate the latent variable representations z and Euclidean distance Step 7.2: Based on the relative position of the Euclidean distance within the candidate sample distribution interval, dynamically adjust the weighting ratio of the neural network model and the gradient boosting tree model, and calculate the distance adaptive weights according to the following formula: in, Representative sample The Euclidean distance from the data center in a two-dimensional potential space. and The minimum and maximum values ​​of the Euclidean distance are represented; Step 7.3: Input the three enhanced candidate material sample sets into the corresponding surrogate models to obtain the expected predicted values ​​of the corresponding performance: in, The prediction results of the neural network model. The prediction results of the gradient boosting tree model, To predict the expected performance of the enhanced candidate material samples using the surrogate model, three expected performance values ​​are obtained based on three sets of enhanced candidate material samples. Step 7.4: An uncertainty measurement mechanism is introduced to evaluate the reliability of the prediction results. This uncertainty is represented by a combination of the model's uncertainty estimate and the latent space distance, where the uncertainty estimate is defined as follows: in, This represents an estimate of the uncertainty in the predicted performance of the enhanced candidate material sample. This represents the uncertainty of the gradient boosting tree model. This represents the uncertainty of the neural network model; The uncertainty is estimated by its quantile regression mechanism. For sample index, Indicates the first The samples are at the quantile of Prediction results under the given conditions The difference between the quantiles is the prediction interval; the wider the interval, the higher the uncertainty of the model prediction. The uncertainty is measured by the statistical dispersion induced by the predicted distribution of the model output, where k represents the discrete binning index, used to identify the k value intervals obtained after the target variable is divided. This represents the total number of discrete bins. Indicates the first The predicted result for the sample falls into the first... The probability of each bin is obtained, and this probability is output by the model. Indicates the first The center value of each bin is used as the representative of the prediction for that bin; finally, the uncertainty measure of the sample is normalized. Step 7.5: The hypervolume improvement algorithm is used to calculate the sample priority based on the prediction uncertainty estimate and the prediction performance value.

9. The high-performance alloy design method with small sample extrapolation capability according to claim 8, characterized in that, Step 7.5 specifically includes: Step 7.5.1: Let the Pareto optimal performance solution set formed by the original data be... ； Step 7.5.2: Based on the predicted values ​​of the three properties, the three uncertainty estimates, and the optimal performance solution set obtained from the enhanced candidate material samples, the following steps are taken: ; Calculate the hypervolume improvement value (EHVI) of the corresponding enhanced candidate material samples; Step 7.5.3: Sort the candidate material sample set after feature enhancement according to the hypervolume improvement value (EHVI) from largest to smallest; Step 7.5.4: Select the top ten best samples from the screened candidate material sample set according to the importance ranking.

Images