A multi-objective optimization method for material properties

Abstract

The application provides a multi-target optimization method for material performance, and belongs to the technical field of material science and artificial intelligence. The method process provided by the application takes a small sample learning model as a core, combines an active learning optimization strategy, and realizes quantitative optimization of multi-target performance indexes in material research and development in an extreme service environment. The method fully utilizes limited test data for learning, significantly reduces the number of physical tests, improves the efficiency and success rate of searching for a material design scheme meeting multi-target performance requirements through intelligent optimization search, and has important significance for accelerating development of new materials in harsh environments.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of material science and artificial intelligence, and in particular to a multi-objective optimization method for material performance. BACKGROUND

[0002] In the fields of aerospace, nuclear industry, shipbuilding, etc., many key structural materials need to work stably for a long time in extreme service environments, such as high temperature, high pressure, corrosive medium, etc. These environments require the materials to have excellent mechanical properties, thermophysical properties and corrosion resistance. For example, alloy materials working at high temperature and high pressure not only need to have sufficient tensile strength and fatigue life, but also need to have good creep resistance and elongation. At the same time, the thermophysical properties such as density and thermal expansion coefficient also need to meet the design requirements, and the materials also need to have reliable corrosion resistance in corrosive media. There is often a trade-off relationship between different performance indicators, for example, increasing the strength of the material may lead to a decrease in elongation, so how to balance multiple performance indicators for optimization is a major challenge in material research and development. Traditional material design and optimization mainly rely on expert experience and repeated testing. Researchers adjust the material composition ratio, manufacturing process parameters such as heat treatment temperature and time to improve performance, but due to the complexity of multi-objective requirements, a large number of tests are often needed to repeatedly explore, which is time-consuming and costly. In addition, material testing under extreme environments itself has high risk and high cost, and the amount of data that can be obtained is limited, making it more difficult to form a large sample data set. This limits the application of machine learning methods based on large data in this field, and ensuring prediction accuracy under small sample conditions becomes a major difficulty. In recent years, the development of small sample machine learning algorithms has provided a new way for material optimization. For example, small sample learning algorithms such as Prototypical Networks, Relation Networks, Matching Networks, Model-Agnostic Meta-Learning (MAML), Adversarially Pre-trained Transformer, TabLLM can accurately predict material performance with limited experimental data, reducing blind testing. However, performance prediction alone is not enough to directly obtain the optimal material design scheme. Active learning, as an important method in artificial intelligence, is good at learning optimal strategies through reward mechanisms in interaction with the environment. In recent years, it has achieved remarkable results in the fields of biological science, medicine, etc. If active learning is introduced into the field of material process parameter optimization, it is expected to enable the algorithm to explore different parameter combinations autonomously to meet the requirements of multi-objective performance. However, there is still a lack of methods that combine small sample high precision prediction and active learning optimization for material multi-objective optimization in the current public technology.

[0003] Existing multi-objective optimization usually adopts evolutionary algorithm, genetic algorithm or simple weight summation optimization strategy, and for the material data with few samples and complex performance targets, the existing methods have deficiencies in efficiency and effect. SUMMARY

[0004] Therefore, the present application provides a multi-objective optimization method for material performance, which solves the problems in the prior art and realizes efficient prediction and optimization of material process parameters and performance indicators under the condition of limited experimental data, greatly reducing the number of actual experiments.

[0005] The multi-objective optimization method for material performance provided by the present application adopts the following technical scheme:

[0006] A multi-objective optimization method for material performance, characterized in that it comprises:

[0007] establishing a data set;

[0008] preprocessing the data set;

[0009] constructing a small sample learning model, training the small sample learning model based on the preprocessed data set for material performance prediction;

[0010] actively learning and optimizing input parameters: setting initial design parameters, inputting the initial design parameters into the trained small sample learning model, and the small sample learning model outputs predicted performance indicators; comparing the predicted performance indicators with expected performance indicators; if the predicted performance indicators do not reach the expected performance indicators, using an active learning algorithm to analyze the difference between the predicted performance indicators and the expected performance indicators to optimize the initial design parameters to obtain optimized design parameters; inputting the optimized design parameters into the trained small sample learning model, and the small sample learning model outputs predicted performance indicators; repeatedly comparing the predicted performance indicators with the expected performance indicators, and using the active learning algorithm to analyze the difference between the predicted performance indicators and the expected performance indicators to optimize the design parameters, until the predicted performance indicators output by the small sample learning model meet the expected performance indicators, and obtaining candidate design parameters;

[0011] verifying and iteratively improving the candidate design parameters to obtain final optimized design parameters.

[0012] Optionally, the step of establishing a data set comprises collecting historical data of parameters and performance indicators of the material to be optimized.

[0013] Optionally, the preprocessing operation comprises:

[0014] determining the categories and ranges of input parameters and output performance indicators of the model;

[0015] completing missing data;

[0016] Filter out outliers and noise from the test data;

[0017] Perform feature engineering as needed;

[0018] Finally, the processed dataset is divided into training and testing sets.

[0019] Optionally, a few-shot learning model based on the AutoML framework can be built;

[0020] The steps for training a few-shot learning model include:

[0021] Build an AutoML framework that includes multiple candidate prediction models;

[0022] Based on the preprocessed training set data, the AutoML framework performs training and performance evaluation through multiple candidate prediction models, and completes the initial screening of candidate prediction models by combining algorithm adaptability indicators.

[0023] For the selected candidate prediction models, the AutoML framework performs automatic feature selection and hyperparameter optimization to train multiple candidate prediction models for performance prediction on small sample data.

[0024] Based on the performance prediction of the candidate prediction models in cross-validation, the AutoML framework integrates and combines multiple candidate prediction models in a performance-oriented manner to obtain an integrated few-shot learning model.

[0025] The performance of the ensemble few-shot learning model was validated using a test dataset.

[0026] Optionally, the candidate prediction model can be K-NearestNeighbor, Support Vector Machine, Random Forest, XGBoost, LightGBM, Multilayer Perceptron, TabNet, Prototypical Network, MetaFormer, CTFN, Model-Agnostic Meta-Learning, or Gaussian regression model.

[0027] Optionally, the candidate model is CTFN. The steps for using CTFN to predict performance on small sample data include:

[0028] Step S100: Generate multiple synthetic datasets based on a structured causal model;

[0029] Step S200 involves post-processing the synthetic dataset by performing nonlinear distortion, quantization, and missing value simulation.

[0030] Step S300: Construct a CTFN model with a dual-axis attention mechanism based on the Transformer network architecture, and alternately perform self-attention calculation in the feature dimension and sample dimension;

[0031] Step S400: Use the context learning mechanism to perform single-step inference training on the CTFN model;

[0032] Step S500: Apply the trained CTFN model to the regression and classification task of small sample tabular data, and output the prediction result of the target variable in a single forward inference.

[0033] Optionally, step S100 includes: constructing a structured causal model to represent the causal relationship between each feature in the tabular data and the target variable, and generating several synthetic samples by random sampling based on the structured causal model to form multiple synthetic datasets.

[0034] Optionally, in step S400, the CTFN model training specifically includes:

[0035] Each synthetic dataset is converted into a context sequence form and input into the CTFN model, where features and corresponding labels of some samples are provided as context examples, and at least one sample provides only features as the prediction target.

[0036] The CTFN model is trained to predict the target value of a sample to be predicted in a single forward inference based on contextual examples.

[0037] Optionally, in step S500, a portion of actual small sample data is used as the test set and input into the trained CTFN model in a context sequence format; step S500 includes:

[0038] Obtain a real dataset, input the features and corresponding labels of some samples in the real dataset as context examples into the trained CTFN model, and input the features to be predicted into the trained CTFN model. The trained CTFN model outputs the prediction results, and compares the prediction results with the actual results. If the prediction results meet the requirements, the trained CTFN model can be directly used for practical applications.

[0039] If the prediction results do not meet the requirements, the gradient of the trained CTFN model is updated using the real dataset.

[0040] Optionally, the AutoML framework, with a performance-oriented approach, integrates and combines multiple candidate prediction models to obtain an ensemble few-shot learning model. The steps include:

[0041] The performance of each candidate prediction model is evaluated based on the cross-validation results during the training phase. The performance evaluation metrics include mean squared error, mean absolute error, and coefficient of determination.

[0042] Based on the performance scores of the candidate prediction models, the model ensemble strategy is determined. The ensemble strategy can be simple averaging, weighted averaging based on performance scores, or stacked generalization.

[0043] Optionally, different active learning optimization algorithm strategies can be selected based on the sample size of candidate design parameter combinations in the design space input / set by the current optimization problem. When the sample size of candidate design parameter combinations in the design space is less than the threshold, a Bayesian optimization strategy based on the EHVI recommendation function of hypervolume expectation is selected; when the sample size of candidate parameter combinations in the design space is greater than or equal to the threshold, a reinforcement learning strategy based on the deep Q-network algorithm is selected. The design space sample size threshold is 50 times the sample size of the initial training set.

[0044] Optionally, when the active learning algorithm selects the EHVI recommendation function based on hypervolume expectation, in each iteration, the active learning algorithm predicts and generates the current optimization objective Pareto front in the design space based on the few-sample learning model, and recommends several candidate design parameter combinations through hypervolume expectation optimization calculation of the EHVI recommendation function.

[0045] Optionally, the specific calculation method of the EHVI recommendation function includes:

[0046] Based on a few-sample learning model, the predicted mean, predicted variance, and joint probability distribution of multiple optimization target performance values ​​corresponding to candidate points to be evaluated in the design space are calculated.

[0047] The expected hypervolume improvement after adding the candidate point to be evaluated, relative to the current Pareto front, is calculated. This expected hypervolume improvement is obtained by integrating the probability that the target performance of the candidate point is superior to the current Pareto front and the hypervolume gain caused by the current Pareto front. The specific calculation formula is as follows:

[0048] ;

[0049] in: Candidate points in the design space;

[0050] The target vector corresponds to the candidate points. The output result obtained after calculating the objective function;

[0051] Represents the m-dimensional real space;

[0052] As the current forefront of Pareto;

[0053] Represents a set The hypervolume, that is, the volume in the target space that is The volume of the dominated region;

[0054] To make the current candidate points The target vector obtained after calculating the objective function Place it in the current Pareto frontier Subsequently, the increase in hypervolume at the Pareto front;

[0055] For the surrogate model prediction At point The probability density function at that location;

[0056] Traverse the design space or use an optimizer to search and select candidate points that maximize the desired hypervolume improvement as recommended points.

[0057] Optionally, when the active learning algorithm is based on the deep Q-network algorithm, the process parameter state, parameter adjustment action and multi-objective reward function are defined on the basis of the few-shot learning model prediction. The algorithm is selected as the agent to be trained in a virtual environment. After multiple rounds of interactive learning between the agent and the environment, the optimal parameter combination is obtained through iterative search.

[0058] The specific calculation method based on the deep Q-network algorithm includes:

[0059] Determine the state, action, and reward function in the reinforcement learning process: Treat a complete set of process parameters as a state, and the algorithm changes the state by adjusting parameters / actions; Integrate multiple material performance objectives and design an active learning reward function to measure the overall performance of the parameter combination;

[0060] The selected active learning algorithm is used as the agent. The agent is trained in a virtual environment with a few-shot learning model. The few-shot learning model is used to predict its various performances. The reward is calculated based on the prediction results. The agent adjusts its strategy accordingly and continuously tries new parameter combinations to improve the cumulative reward value.

[0061] Through multiple iterative training iterations, the agent gradually learns the parameter control strategy that optimizes overall performance, resulting in one or more optimized candidate schemes for composition or process parameters, i.e., candidate design parameters. Optionally, specific steps for verifying and iteratively improving the optimization model include:

[0062] Material samples were prepared based on the candidate design parameters, and the actual performance parameters of the material samples were tested.

[0063] If the actual performance parameters meet the pre-set target performance requirements, the candidate design parameters will be used as the final design parameters, and the optimization will end.

[0064] If the actual performance parameters do not meet the pre-set target performance requirements, the candidate design parameters and actual performance are added to the dataset as supplementary data to form an updated dataset. The updated dataset is then used to train a few-shot learning model, repeatedly actively learn and optimize the input parameters to obtain new candidate design parameters, and verify and iteratively improve the new candidate design parameters until the actual performance parameters meet the pre-set target performance requirements, thus completing the final design parameter optimization.

[0065] In summary, this application includes the following beneficial technical effects:

[0066] This application combines a pre-trained few-shot learning model with an active learning algorithm for multi-objective optimization of materials. The few-shot learning model provides high-precision performance prediction, while active learning is responsible for global search optimization, achieving tight coupling between prediction and optimization and significantly reducing reliance on large amounts of experimental data.

[0067] This application's method can simultaneously consider multiple dimensions of material performance, including mechanical, thermophysical, and corrosion resistance, transforming multi-objective optimization into a single, active learning reward signal, thus achieving a comprehensive balance across various performance aspects. Compared to traditional methods of sequential optimization or simple weighted summation, this application can more effectively explore performance trade-offs and find the optimal material design that meets the comprehensive requirements of extreme environments.

[0068] Through a closed-loop iterative mechanism of "prediction-optimization-validation-feedback," the model and strategy can continuously improve themselves. Whenever the experimental validation results are unsatisfactory, new data is incorporated into the training to make the model more accurate, and the actively learned strategy is updated accordingly, thus gradually approaching the optimal solution. This adaptive iterative process increases the probability of successful optimization and reduces the risk of experimental failure.

[0069] Using the method described in this application, researchers can obtain reliable optimization solutions without conducting large-scale experiments. The small-sample learning model predicts a performance index in just seconds, and active learning can automatically execute hundreds or thousands of virtual experiments, which is more efficient than manually designing experimental plans. Ultimately, only a small number of key experiments are needed for verification, significantly shortening the R&D cycle and saving human and material costs. This is of great significance for accelerating the development of new materials under extreme service conditions. Attached Figure Description

[0070] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0071] Figure 1 This is a flowchart of a multi-objective optimization method for material properties. Detailed Implementation

[0072] The embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0073] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. This application can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0074] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this application, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0075] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this application. The illustrations only show the components related to this application and are not drawn according to the number, shape and size of the components in actual implementation. In actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0076] Furthermore, specific details are provided in the following description to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.

[0077] This application provides a multi-objective optimization method for material properties.

[0078] like Figure 1 As shown, a multi-objective optimization method for material properties includes:

[0079] Create a dataset.

[0080] Preprocess the dataset.

[0081] A few-shot learning model is constructed and trained based on the preprocessed dataset for material property prediction.

[0082] Active learning optimizes input parameters: Initial design parameters are set and input into a trained few-shot learning model. The few-shot learning model outputs a predicted performance metric. The predicted performance metric is compared with the expected performance metric. If the predicted performance metric does not meet the expected performance metric, the active learning algorithm is used to analyze the difference between the predicted and expected performance metrics and optimize the initial design parameters to obtain optimized design parameters. The optimized design parameters are then input into the trained few-shot learning model, which outputs a predicted performance metric. This process of comparing the predicted and expected performance metrics and using the active learning algorithm to analyze the difference between the predicted and expected performance metrics to optimize the design parameters continues until the predicted performance metric output by the few-shot learning model meets the expected performance metric, thus obtaining candidate design parameters.

[0083] The candidate design parameters are verified and iteratively improved to obtain the final optimized design parameters.

[0084] The method presented in this application, based on a few-shot learning model and combined with an active learning optimization strategy, achieves quantitative optimization of multi-dimensional performance indicators in the development of materials for extreme service environments. This method fully utilizes limited experimental data for learning, significantly reducing the number of physical experiments. Through intelligent optimization search, it improves the efficiency and success rate of finding material design solutions that meet multiple performance requirements, which is of great significance for accelerating the development of new materials under harsh environments.

[0085] The specific steps are as follows:

[0086] Establish a dataset: Collect historical data on the materials to be optimized and their performance indicators to construct an initial dataset. The data can come from publicly available literature, databases, simulation calculations, or a small number of preliminary experimental results. The data involved should include material composition, processing parameters, and corresponding performance index values ​​to provide basic data for subsequent modeling.

[0087] Data preprocessing: The dataset is cleaned and preprocessed. This includes, but is not limited to: determining the categories and ranges of the model's input parameters and output performance metrics; appropriately imputing missing data, such as using mean imputation, interpolation, or prediction methods based on similar samples; filtering out outliers and noise from the experimental data, such as using statistical discrimination methods to remove measurements that deviate from the normal range; performing feature engineering as needed, such as synthesizing, reducing dimensionality, or normalizing various related parameters; and finally, dividing the processed dataset into training and testing sets to prepare for model training and validation.

[0088] A few-shot learning model based on the AutoML framework is constructed. Using preprocessed training set data, the training set is input into the few-shot learning model for learning, resulting in a model that can map the relationship between material structure, composition, process parameters, and performance indicators. The prediction accuracy of the few-shot learning model is verified using a test set to ensure its ability to accurately predict the performance of unseen parameter combinations. AutoML stands for Automated Machine Learning.

[0089] Specifically, the steps for training a few-shot learning model include:

[0090] An AutoML framework containing multiple candidate prediction models is constructed. Based on preprocessed training data, the AutoML framework performs lightweight training and performance evaluation on multiple candidate prediction models, and completes the initial screening of candidate prediction models by combining algorithm suitability indicators. For the selected candidate prediction models, the AutoML framework performs automatic feature selection and hyperparameter optimization to train multiple candidate prediction models for performance prediction on small sample data. Based on the performance prediction of the candidate prediction models in cross-validation, the AutoML framework integrates and combines multiple candidate prediction models with performance as the guide to obtain an ensemble few-sample learning model to improve its prediction accuracy and robustness. The performance of the ensemble few-sample learning model is validated using a test dataset to evaluate its prediction ability on unknown data. Candidate prediction models can be K-NearestNeighbor, Support Vector Machine, Random Forest, XGBoost, LightGBM, MultilayerPerceptron, TabNet, Prototypical Network, MetaFormer, CTFN, Model-Agnostic Meta-Learning, or Gaussian regression models.

[0091] Among them, K-NearestNeighbor is a proximity algorithm; Support Vector Machine is a support vector machine; XGBoost stands for eXtreme Gradient Boosting; LightGBM stands for Light Gradient Boosting Machine; Multilayer Perceptron is a feedforward artificial neural network model; TabNet stands for Tabular Data Neural Network, a neural network architecture designed specifically for tabular data; Prototypical Network is a prototype network; MetaFormer is a meta-framework transformer; CTFN stands for Causal Tabular Foundation Network; and Model-Agnostic Meta-Learning is a model-agnostic meta-learning method.

[0092] In this embodiment of the application, the candidate model contains CTFN, and the steps for using CTFN to predict the performance of small sample data include:

[0093] Step S100: Generate multiple synthetic datasets based on the structured causal model.

[0094] Step S200: Post-processing of the synthetic dataset, including nonlinear distortion, quantization, and missing value simulation.

[0095] Step S300: Construct a CTFN model with a dual-axis attention mechanism based on the Transformer network, alternately performing self-attention computation on the feature dimension and the sample dimension. The Transformer is a neural network architecture based on a self-attention mechanism, widely used in natural language processing tasks.

[0096] Step S400: The CTFN model is trained using a context learning mechanism in a single-step inference manner, enabling the CTFN model to learn patterns in small-sample tabular data without relying on multiple iterative optimizations. The context learning mechanism is called In-Context Learning, or ICL for short.

[0097] Step S500: Apply the trained CTFN model to the regression and classification task of small sample tabular data, and output the prediction result of the target variable in a single forward inference. Here, a portion of the actual small sample data is used as the test set and input into the trained CTFN model in the context sequence format.

[0098] Step S100 includes: constructing a structured causal model to represent the causal relationship between each feature and the target variable in the tabular data; generating a large number of synthetic samples based on the structured causal model through random sampling to form multiple synthetic datasets. The structured causal model is a directed acyclic graph, where each node corresponds to a feature or target variable, and each node is associated with a structural equation or conditional probability distribution used to generate the data for that node. The synthetic samples are generated by sampling the directed acyclic graph in topological order.

[0099] Step S200: Post-processing of the synthetic dataset, including nonlinear distortion, quantization, and missing value simulation.

[0100] In one specific embodiment, this application constructs a priori generation process for synthetic data based on the principle of causal reasoning: a causal graph structure for the tabular data is randomly set, including the causal relationship between feature variables and target variables; according to the set causal structure, a functional form and probability distribution, such as linear or nonlinear functional relationship, are assigned to each variable, and noise terms are added, thereby generating simulated tabular datasets that satisfy different causal relationship assumptions. In this way, millions of small-sample tabular data containing various relationship patterns can be synthesized for pre-training the model. This synthetic data covers different task types such as classification and regression, greatly enriching the diversity and complexity of training data and providing a source of prior knowledge for the model.

[0101] In step S300, the network structure of the CTFN model is designed to include layers that alternately perform self-attention calculations along the feature dimension and along the sample dimension, thus forming a dual-axis attention mechanism. Through this dual-axis attention mechanism, the model can simultaneously model the correlation between features and the patterns between samples in tabular data. This application designs the CTFN model based on the Transformer network architecture and introduces a dual-axis attention mechanism to adapt to the two-dimensional structure of tabular data. The dual-axis attention mechanism means that the model's self-attention is performed separately along the feature dimension and the sample dimension, that is, modeling the relationships between features and the relationships between samples separately, similar to axial attention. In this way, the model can capture the dependencies between different feature columns in the table and extract potential causal associations; on the other hand, it can use the patterns between training samples for inference and learn generalization rules from a small number of samples. Unlike the traditional Transformer which directly applies attention to flattened sequences, dual-axis attention greatly reduces the noise interference of the attention mechanism in the case of small samples and learns the internal structure of the data more effectively. The CTFN model of this application is based on the Transformer encoder and combines techniques such as position embedding to process tabular data, enabling the model to perceive the feature field to which each value belongs. With the above structure, the model in this application has the ability to perform causal reasoning and pattern recognition on tabular data with small sample sizes.

[0102] In step S400, the CTFN model training specifically includes:

[0103] Each synthetic dataset is converted into a context sequence and input into the CTFN model, where features and corresponding labels of some samples are provided as context examples, and at least one sample provides only features as the prediction target; the CTFN model is trained to predict the target value of the sample to be predicted in a single forward inference based on the context examples.

[0104] For small-sample tasks in real-world applications, a small amount of real-labeled data is input into the pre-trained CTFN model, which then leverages its existing knowledge to quickly predict the target. In most cases, the CTFN model can directly output high-precision predictions without extensive additional training. For particularly complex tasks, accuracy can be further improved by making minor adjustments to the model on small samples.

[0105] Lightweight fine-tuning specifically includes: obtaining a real validation dataset, using the features and corresponding labels of a small subset of samples in the validation dataset as context examples to input into the trained CTFN model, inputting the features to be predicted into the trained CTFN model to output the prediction results, and comparing the prediction results with the actual results.

[0106] If the prediction results meet the requirements, the trained CTFN model can be used directly for practical applications without the need for model training and fine-tuning. The regression or classification prediction results for each sample in the dataset can be output after a single-step forward calculation.

[0107] If the prediction results do not meet the requirements, perform a small number of gradient updates on the trained CTFN model using a real validation dataset.

[0108] The AutoML framework, with a performance-oriented approach, integrates multiple candidate prediction models to obtain an ensemble few-shot learning model. The steps include:

[0109] The performance of each candidate prediction model is evaluated based on the cross-validation results during the training phase. The performance evaluation metrics include mean squared error, mean absolute error, and coefficient of determination. Based on the performance scores of the candidate prediction models, the model ensemble strategy is determined. The ensemble strategy can be simple averaging, weighted averaging based on performance scores, or stacked generalization.

[0110] The weighted average refers to assigning ensemble weights to each candidate prediction model based on its performance score in cross-validation, with higher weights for models with superior performance. The stacked generalization strategy refers to constructing a two-level fusion model, which takes the prediction results of each candidate prediction model as input and learns the optimal combination of candidate prediction models through a neural network or tree model. The selection of the ensemble strategy can be based on the overall performance on the validation set or the reserved dataset, and the best option can be selected.

[0111] Different active learning optimization algorithm strategies are selected based on the sample size of candidate design parameter combinations in the design space input / defined by the current optimization problem. When the sample size of candidate parameter combinations in the design space is less than a certain threshold, a Bayesian optimization strategy based on the Hypervolume Expectation (EHVI) recommendation function is selected; when the sample size of candidate parameter combinations in the design space is greater than or equal to a certain threshold, a reinforcement learning strategy based on a deep Q-network algorithm is selected. The design space sample size threshold is 50 times the sample size of the initial training set. Here, EHVI stands for Expected Hypervolume Improvement.

[0112] When the active learning algorithm selects the EHVI recommendation function based on hypervolume expectation, in each iteration, the active learning algorithm predicts and generates the current optimization target Pareto front in the design space based on the few-shot learning model, and recommends several candidate design parameter combinations through hypervolume expectation optimization calculation using the EHVI recommendation function. For the active learning input parameter optimization step of this application, based on the fast prediction and evaluation provided by the few-shot learning model, the EHVI recommendation function is introduced to automatically optimize and search for material composition or process parameters. First, based on the distribution of the parameter combinations in the current training set in the target performance space, a Pareto front representing the current optimal solution set is constructed. Subsequently, the performance prediction distribution of all parameter combinations in the design space is obtained using the few-shot learning model, including the predicted values ​​and their uncertainties. The improvement / optimization amount and the realization probability / optimization probability of all parameter combinations to the current Pareto front hypervolume metric are evaluated, and then the expected hypervolume improvement EHVI index is calculated, and the EHVI values ​​of the parameter combinations are ranked. Through multiple iterations, one or more candidate solutions of components or process parameters with high EHVI values ​​are gradually learned, i.e., candidate design parameters.

[0113] If a reinforcement learning strategy based on a deep Q-network algorithm is chosen, based on the prediction of the few-shot learning model, the process parameter state, parameter adjustment action, and multi-objective reward function are defined. The algorithm is selected as the agent and trained in a virtual environment. After multiple rounds of interactive learning between the agent and the environment, the optimal parameter combination is obtained through iterative search. For the active learning input parameter optimization step of this application, based on the fast prediction and evaluation provided by the few-shot learning model, an active learning algorithm is introduced to automatically optimize the material composition or process parameters. First, the state space and action space of active learning are determined: a complete set of process parameters can be regarded as a state, and the algorithm changes the state by adjusting parameters / actions; or the process of gradually adjusting parameters can be modeled as a sequential decision-making process, where the state space: each state A complete set of material composition or process parameters is represented by a vector in the state space. Description, in which, Representing each individual material composition or process parameter; motion space: using motion This represents the discretized adjustment of parameters through a finite set of actions. Achieve state transition. Design an active learning reward function to measure the overall performance of a combination of parameters. This reward function needs to integrate multiple objectives: for example, assign positive rewards to performances that need to be improved as much as possible, such as strength and lifespan, and set penalties for deviations from the target for indicators that need to be controlled within a certain range, such as density and coefficient of thermal expansion. Alternatively, a threshold can be set for each performance indicator to determine whether the solution is qualified, thereby transforming multi-dimensional performance indicators into a single scalar reward value. Reward function: transforms a state... Below, material performance indicators are integrated into scalar rewards. :

[0114] ;

[0115] in: Represents the reward function, used to evaluate states in reinforcement learning. The degree of superiority or inferiority. Represented by the current state. The reward value, This indicates the performance metric that needs to be maximized / optimized. This indicates the performance index that must meet the constraints. For the first The weight coefficients corresponding to each objective function value For the first The objective function has several values, and it is a monotonically increasing function. In order to be with the first Penalty terms related to each constraint condition; if the constraint target exceeds the range, the penalty term is triggered. , As weight, To tolerate bias, This represents the actual performance index that satisfies the constraints. This represents the preset target performance index that satisfies the constraints.

[0116] Then, an appropriate active learning algorithm is selected as the agent, such as value function-based methods or policy gradient methods. The agent is trained in a "virtual environment" using a few-shot learning model: for a given candidate parameter combination, the few-shot learning model predicts its various performance characteristics, calculates the reward based on the prediction results, and adjusts its policy accordingly, continuously trying new parameter combinations to increase the cumulative reward value. Through multiple iterations of training, the agent gradually learns the parameter control strategy that optimizes the overall performance, obtaining one or more sets of optimized candidate schemes for component or process parameters, i.e., candidate design parameters. The specific calculation formula is as follows, and the Bellman equation to be solved is updated as follows:

[0117] ;

[0118] in, In the state Execute action The expected value of long-term cumulative rewards that can be obtained afterward; Representing state After the action The next state after that; In the state Execute action The single-step reward is obtained immediately afterward; Discount factor; Indicates the target network parameters; To perform the action After, state Transition to the next state And in the next state Next, execute candidate actions This allows for the maximization of future benefits.

[0119] Optimization process loss function Defined as:

[0120] ;

[0121] in, This represents the expectation operator, used to solve for the sample in state. Execute action State transition to Get instant rewards back, and The expected value of the mean square error between them; The target Q value is defined as the state. Execute action The expected value of long-term cumulative rewards obtained later; Indicates the predicted Q value, The parameters of the policy network represent the current state. Execute action Then, while following a certain strategy, the predicted value of the accumulated reward is calculated over the long term.

[0122] After training converges, the action with the largest Q-value is selected to generate the parameter adjustment strategy.

[0123] ;

[0124] in, The optimal action is indicated by the state. Below, based on the current situation The value function represents the optimal action chosen. Indicates all actions Perform a traversal to find the one that makes The biggest movement; Indicates the predicted Q value, The parameters of the policy network represent the current state. Execute action Then, while following a certain strategy, the predicted value of the accumulated reward is calculated over the long term.

[0125] The optimal parameter combination can then be output. And its predictive performance.

[0126] The specific steps for validating and iteratively improving the optimization model include:

[0127] Material samples are prepared based on candidate design parameters, and the actual performance parameters corresponding to the material samples are tested, such as mechanical, thermophysical, and corrosion resistance values. If the actual performance parameters meet the pre-set target performance requirements, the candidate design parameters are used as the final design parameters, and the optimization ends. If the actual performance parameters do not meet the pre-set target performance requirements, the candidate design parameters and actual performance parameters are included in the dataset to form an updated dataset. The updated dataset is used to train a few-shot learning model, repeatedly actively learn and optimize the input parameters to obtain new candidate design parameters, and verify and iteratively improve the new candidate design parameters until the actual performance parameters meet the pre-set target performance requirements, and the final design parameters are obtained, thus ending the optimization.

[0128] By continuously feeding new experimental data into the model and iteratively training it, the model's prediction accuracy and its relevance to the optimization objectives will continuously improve. Active learning can also explore new parameter spaces in an updated model environment. This process of repeated training and optimization gradually approaches the ideal multi-objective optimal solution. When the experimental verification results show that all target performance indicators meet the design requirements, the final optimal composition or process parameter scheme is confirmed.

[0129] If one or more performance parameters fail to meet the standards, the actual performance parameters are deemed not to meet the pre-set target performance requirements.

[0130] This application uses a high-temperature alloy material as an example to demonstrate how to apply the method of this application to optimize its multi-objective performance in a high-temperature corrosive environment:

[0131] Step 1: Establish a dataset: Collect performance data of the alloy material under different formulations and processing conditions. Data sources include experimental reports from publicly available literature and a small number of pre-conducted experiments. The data covers the chemical composition ratio of the material, heat treatment process parameters, and corresponding measured mechanical, thermophysical, and corrosion resistance indicators. The collected raw data are then compiled into an initial dataset. This dataset includes heat treatment process parameters such as quenching temperature, holding time, and aging process; mechanical properties such as tensile strength, yield strength, elongation, and fatigue life; thermophysical properties such as density and coefficient of thermal expansion; and corrosion resistance properties such as high-temperature oxidation weight gain and corrosion rate.

[0132] Step 2, data preprocessing: The initial data is cleaned and organized. If some data records are missing values ​​such as elongation or corrosion rate, appropriate methods are used to fill in the missing values, such as interpolation estimation based on other similar process samples. For obviously abnormal test data points, such as a test intensity value that is much lower than the level of similar samples, statistical discrimination is performed to eliminate possible data anomalies. Based on the knowledge of the materials field, key features that have a significant impact on performance are selected from many features, and some highly correlated features are merged and simplified to reduce the model dimensionality. Finally, the cleaned data is randomly divided into training and testing sets, with most of the data used for model training and a small portion retained to verify the model's predictive ability.

[0133] Step 3, perform few-shot learning model training: Construct a candidate prediction model pool containing multiple candidate prediction models such as decision trees, neural networks, Gaussian regression, and pre-trained models; based on the feature dimensions, sample size, and task type of the few-shot dataset, initially select model structures with strong adaptability. Subsequently, the AutoML framework performs lightweight training on each candidate prediction model and combines cross-validation to evaluate its prediction performance on the training set to determine its adaptability to the current task.

[0134] During the model selection process, the system simultaneously performs automatic feature selection and hyperparameter tuning: by analyzing the correlation between each input feature and the performance output, redundant or low-contribution features are eliminated; in terms of hyperparameter optimization, the system adopts a grid search strategy to find the hyperparameter combination that optimizes the cross-validation performance.

[0135] After multiple rounds of evaluation and adjustments, the AutoML framework outputs several trained candidate prediction models and their performance under different feature subsets and hyperparameter combinations. Finally, guided by the model's prediction accuracy, the system integrates the multiple candidate prediction models using methods such as weighted averaging or stacked generalization to form a few-shot learning model with higher generalization ability and robustness.

[0136] After training, the test set was input into the model for prediction, and the results were compared with the actual values. The results showed that the prediction error of the model for each performance index was within an acceptable range, proving that the model can be used for the next step of optimization decision support.

[0137] Subsequently, target performance prediction and active learning optimization are performed: the few-shot learning model and the active learning optimization module constitute a closed-loop system. The few-shot learning model is used to quickly predict multi-objective performance under given process parameters, while the active learning module continuously adjusts the input parameters using feedback from the model to find the optimal solution. In the specific implementation, the optimization objective is first determined—assuming the alloy is expected to simultaneously meet the following conditions at 800℃: tensile strength not less than 1000 MPa, creep life not less than 500 hours, elongation not less than 5%, and weight gain after 100 hours of high-temperature oxidation not exceeding a certain limit. Due to the limited design space in this case, this embodiment uses an EHVI recommendation function based on hypervolume expectation improvement for the active learning algorithm; a Pareto front for the target performance is generated based on the current training set; the expected improvement EHVI of the new samples for the hypervolume front is calculated using the prediction distribution of the few-shot learning model; and candidate parameter combinations are recommended based on EHVI ranking.

[0138] Specifically, the calculation method of the EHVI recommendation function includes:

[0139] Based on a few-shot learning model, the predicted mean, predicted variance, and joint probability distribution of multiple optimization target performance values ​​corresponding to candidate points to be evaluated in the design space are calculated. The expected hypervolume improvement relative to the current Pareto front after the addition of a candidate point is calculated. The expected hypervolume improvement is obtained by integrating the probability that the target performance value of the candidate point is better than the current Pareto front and the resulting hypervolume gain. The specific calculation formula is as follows:

[0140] ;

[0141] in: Candidate points in the design space; The target vector corresponds to the candidate points. The output result obtained after calculating the objective function; Represents the m-dimensional real space; As the current forefront of Pareto; Represents a set The hypervolume, that is, the volume in the target space that is The volume of the dominated region; To make the current candidate points The target vector obtained after calculating the objective function Place it in the current Pareto frontier Subsequently, the increase in hypervolume at the Pareto front; For the surrogate model prediction At point The probability density function at that location.

[0142] Traverse the design space or use an optimizer to search and select candidate points that maximize the desired hypervolume improvement as recommended points.

[0143] Step 4: Enter the experimental verification and iterative improvement stage: Verify the optimized scheme output by the active learning in the laboratory. Prepare alloy samples according to the scheme and conduct standard tensile tests, creep tests, elongation tests, and high-temperature oxidation tests to obtain actual performance data. Compare the actual test results with the optimization target: If all performances meet or exceed the target values, it proves that the scheme successfully meets the design requirements, and this optimization design is completed; if a certain performance is found to be substandard, such as creep life slightly below 500 hours, it indicates that the model prediction has a bias or there is still room for improvement. At this time, the experimental data, such as the actual creep life value of the optimized scheme, is added to the dataset to update the small sample learning model, making its prediction more accurate in that area. Then, the active learning module is restarted to further search for improvement schemes under the guidance of the new model. For example, the agent may increase the content of relevant elements or adjust the heat treatment regime to address the problem of low creep life, and provide a new combination of parameters. The new scheme is then verified by sample preparation and testing, and so on. With each new round of data feedback, the model's predictive ability becomes more perfect, and the search direction of active learning becomes more accurate. After several rounds of iteration, an optimized solution was finally obtained, which enabled all the tested performance indicators of the alloy material to meet the target requirements at 800℃, thus completing the optimized design.

[0144] As can be seen from the above description of the implementation methods, the method of this application can achieve synergistic optimization of multiple properties of materials even with extremely limited real experimental data. Leveraging the rapid and accurate predictive capabilities of the small-sample learning model and its autonomous exploration and decision-making capabilities through proactive learning, researchers only need a few rounds of experimental verification to converge and obtain the ideal process parameter configuration. Compared with traditional methods involving numerous trial-and-error experiments, this application significantly improves R&D efficiency, reduces costs, and has practical engineering application value for the design of materials for extreme service environments.

[0145] The method of this application is applicable to a variety of material types, especially structural materials that serve in extreme environments such as high temperature, high pressure, and corrosion. By introducing intelligent algorithms, it simultaneously optimizes multi-dimensional performance indicators such as mechanical properties, thermophysical properties, and corrosion resistance to obtain the optimal combination of process parameters that meet service requirements.

[0146] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A multi-objective optimization method for material properties, characterized in that, include: The steps to establish a dataset include: collecting historical data on the parameters and performance indicators of the materials to be optimized. The data content covers the chemical composition ratio of the materials, heat treatment process parameters, and the corresponding measured mechanical properties, thermophysical properties, and corrosion resistance indicators. Preprocess the dataset; Construct a few-shot learning model and train it on a preprocessed dataset for material property prediction. Active learning optimizes input parameters: Initial design parameters are set and input into a trained few-shot learning model. The few-shot learning model outputs a predicted performance index. The predicted performance index is compared with the expected performance index. If the predicted performance index does not meet the expected performance index, the active learning algorithm is used to analyze the difference between the predicted and expected performance indices and optimize the initial design parameters to obtain optimized design parameters. The optimized design parameters are then input into the trained few-shot learning model, which outputs a predicted performance index. This process of comparing the predicted and expected performance indices and using the active learning algorithm to analyze the difference between the predicted and expected performance indices to optimize the design parameters continues until the predicted performance index output by the few-shot learning model meets the expected performance index, thus obtaining candidate design parameters. The candidate design parameters are verified and iteratively improved to obtain the final optimized design parameters. Different active learning optimization algorithm strategies are selected based on the sample size of candidate design parameter combinations in the design space input / defined by the current optimization problem. When the sample size of candidate design parameter combinations in the design space is less than the threshold, a Bayesian optimization strategy based on the EHVI recommendation function of hypervolume expectation is selected; when the sample size of candidate parameter combinations in the design space is greater than or equal to the threshold, a reinforcement learning strategy based on the deep Q-network algorithm is selected. The design space sample size threshold is 50 times the sample size of the initial training set. When the active learning algorithm selects the EHVI recommendation function based on hypervolume expectation, in each iteration, the active learning algorithm predicts and generates the current optimization target Pareto front in the design space based on the few-sample learning model, and recommends several candidate design parameter combinations through hypervolume expectation optimization calculation of the EHVI recommendation function. The specific calculation method for the EHVI recommendation function includes: Based on a few-sample learning model, the predicted mean, predicted variance, and joint probability distribution of multiple optimization target performance values ​​corresponding to candidate points to be evaluated in the design space are calculated. The expected hypervolume improvement after adding the candidate point to be evaluated, relative to the current Pareto front, is calculated. This expected hypervolume improvement is obtained by integrating the probability that the target performance of the candidate point is superior to the current Pareto front and the hypervolume gain caused by the current Pareto front. The specific calculation formula is as follows: ； in, Candidate points in the design space; The target vector corresponds to the candidate points. The output result obtained after calculating the objective function; Represents the m-dimensional real space; As the current forefront of Pareto; Represents a set The hypervolume, that is, the volume in the target space that is The volume of the dominated region; To make the current candidate points The target vector obtained after calculating the objective function Place it in the current Pareto frontier Subsequently, the increase in hypervolume at the Pareto front; For the surrogate model prediction At point The probability density function at that location; Traverse the design space or use an optimizer to search and select candidate points that maximize the desired hypervolume improvement as recommended points.

2. The multi-objective optimization method for material properties according to claim 1, characterized in that, Preprocessing operations include: Determine the types and ranges of the model's input parameters and output performance metrics; Complete the missing data; Filter out outliers and noise from the test data; Perform feature engineering as needed; Finally, the processed dataset is divided into training and testing sets.

3. The multi-objective optimization method for material properties according to claim 2, characterized in that, Construct a few-shot learning model based on the AutoML framework; The steps for training a few-shot learning model include: Build an AutoML framework that includes multiple candidate prediction models; Based on the preprocessed training set data, the AutoML framework performs training and performance evaluation through multiple candidate prediction models, and completes the initial screening of candidate prediction models by combining algorithm adaptability indicators. For the selected candidate prediction models, the AutoML framework performs automatic feature selection and hyperparameter optimization to train multiple candidate prediction models for performance prediction on small sample data. Based on the performance prediction of the candidate prediction models in cross-validation, the AutoML framework integrates and combines multiple candidate prediction models in a performance-oriented manner to obtain an integrated few-shot learning model. The performance of the ensemble few-shot learning model was validated using a test dataset.

4. The multi-objective optimization method for material properties according to claim 3, characterized in that, Candidate prediction models can be K-NearestNeighbor, Support Vector Machine, Random Forest, XGBoost, LightGBM, Multilayer Perceptron, TabNet, Prototypical Network, MetaFormer, CTFN, Model-Agnostic Meta-Learning, or Gaussian regression models.

5. The multi-objective optimization method for material properties according to claim 4, characterized in that, The candidate model is CTFN. The steps for using CTFN to predict performance on small sample data include: Step S100: Generate multiple synthetic datasets based on a structured causal model; Step S200 involves post-processing the synthetic dataset by performing nonlinear distortion, quantization, and missing value simulation. Step S300: Construct a CTFN model with a dual-axis attention mechanism based on the Transformer network architecture, and alternately perform self-attention calculation in the feature dimension and sample dimension; Step S400: Use the context learning mechanism to perform single-step inference training on the CTFN model; Step S500: Apply the trained CTFN model to the regression and classification task of small sample tabular data, and output the prediction result of the target variable in a single forward inference.

6. The multi-objective optimization method for material properties according to claim 5, characterized in that, Step S100 includes: constructing a structured causal model to represent the causal relationship between each feature in the tabular data and the target variable, and generating several synthetic samples by random sampling based on the structured causal model to form multiple synthetic datasets.

7. The multi-objective optimization method for material properties according to claim 5, characterized in that, In step S400, the CTFN model training specifically includes: Each synthetic dataset is converted into a context sequence form and input into the CTFN model, where features and corresponding labels of some samples are provided as context examples, and at least one sample provides only features as the prediction target. The CTFN model is trained to predict the target value of a sample to be predicted in a single forward inference based on contextual examples.

8. The multi-objective optimization method for material properties according to claim 7, characterized in that, In step S500, a portion of actual small sample data is used as the test set and input into the trained CTFN model in a context sequence format; step S500 includes: Obtain a real dataset, input the features and corresponding labels of some samples in the real dataset as context examples into the trained CTFN model, and input the features to be predicted into the trained CTFN model. The trained CTFN model outputs the prediction results, and compares the prediction results with the actual results. If the prediction results meet the requirements, the trained CTFN model can be directly used for practical applications. If the prediction results do not meet the requirements, the gradient of the trained CTFN model is updated using the real dataset.

9. The multi-objective optimization method for material properties according to claim 3, characterized in that, The AutoML framework, with a performance-oriented approach, integrates multiple candidate prediction models to obtain an ensemble few-shot learning model. The steps include: The performance of each candidate prediction model is evaluated based on the cross-validation results during the training phase. The performance evaluation metrics include mean squared error, mean absolute error, and coefficient of determination. Based on the performance scores of the candidate prediction models, the model ensemble strategy is determined. The ensemble strategy can be simple averaging, weighted averaging based on performance scores, or stacked generalization.

10. The multi-objective optimization method for material properties according to claim 1, characterized in that, When using a deep Q-network-based active learning algorithm, based on the prediction of a few-shot learning model, process parameter states, parameter adjustment actions, and multi-objective reward functions are defined. The algorithm is selected as the agent and trained in a virtual environment. After multiple rounds of interactive learning between the agent and the environment, the optimal parameter combination is obtained through iterative search.

11. The multi-objective optimization method for material properties according to claim 10, characterized in that, The specific calculation method based on the deep Q-network algorithm includes: Determine the state, action, and reward function in the reinforcement learning process: Treat a complete set of process parameters as a state, and the algorithm changes the state by adjusting parameters / actions; Integrate multiple material performance objectives and design an active learning reward function to measure the overall performance of the parameter combination; The selected active learning algorithm is used as the agent. The agent is trained in a virtual environment with a few-shot learning model. The few-shot learning model is used to predict its various performances. The reward is calculated based on the prediction results. The agent adjusts its strategy accordingly and continuously tries new parameter combinations to improve the cumulative reward value. Through multiple iterations of training, the agent gradually learns the parameter control strategy that optimizes the overall performance, and obtains one or more sets of optimized candidate schemes for composition or process parameters, i.e., candidate design parameters.

12. The multi-objective optimization method for material properties according to claim 1, characterized in that, The specific steps for validating and iteratively improving the optimization model include: Material samples were prepared based on the candidate design parameters, and the actual performance parameters of the material samples were tested. If the actual performance parameters meet the pre-set target performance requirements, the candidate design parameters will be used as the final design parameters, and the optimization will end. If the actual performance parameters do not meet the pre-set target performance requirements, the candidate design parameters and actual performance are added to the dataset as supplementary data to form an updated dataset. The updated dataset is then used to train a few-shot learning model, repeatedly actively learn and optimize the input parameters to obtain new candidate design parameters, and verify and iteratively improve the new candidate design parameters until the actual performance parameters meet the pre-set target performance requirements, thus completing the final design parameter optimization.

Images