Material performance prediction method and device based on multiple machine learning regression model and high-throughput experiment, medium, product and application of material performance prediction method and device

Abstract

The invention discloses a material performance prediction method and device based on a multi-machine learning regression model and a high-throughput experiment, a medium, a product and application thereof, and relates to the technical field of material design and optimization, and the method comprises the steps: obtaining preset material parameters; any performance is determined as the current performance; based on material components in the preset material parameters, determining target key features of the current performance under the preset material parameters; respectively inputting the process in the preset material parameters and the target key features of the current performance under the preset material parameters into a plurality of optimal performance prediction regression models corresponding to the current performance to obtain a plurality of sub-prediction values of the current performance under the preset material parameters; and based on the plurality of sub-predicted values of the current performance under the preset material parameters, determining a total predicted value of the current performance under the preset material parameters. The material performance prediction precision and the research and development efficiency are improved, and the research and development cost is reduced.

Description

Technical Field

[0001] The present application relates to the technical field of material design and optimization, and in particular to a material property prediction method, device, medium, product and application thereof based on multiple machine learning regression models and high-throughput experiments. Background Art

[0002] Performance prediction during traditional material preparation relies primarily on experimental trial and error based on expert experience and theoretical calculations, which presents a series of problems such as long cycles, high costs, and low efficiency. Data-driven material design and development in the era of artificial intelligence is a booming frontier of advanced science and technology, but most current performance prediction methods in this field are limited to sequential manual experiments on a single sample, a single machine learning regression prediction model, a single machine learning sampling strategy, and the optimization of a single material target property. Although this is an improvement over traditional experimental efficiency that relies on trial and error, many problems still exist, such as the small amount of experimental data, low experimental data quality, unclear convergence of machine learning algorithms, material performance optimization stuck in a local minimum, difficulty in balancing multiple comprehensive material properties, and low overall efficiency of full-chain R&D.

[0003] In summary, a material property prediction method with high prediction efficiency and high prediction accuracy is needed. Summary of the Invention

[0004] The purpose of this application is to provide a material property prediction method, device, medium, product and its application based on multiple machine learning regression models and high-throughput experiments to solve the problems of low efficiency and prediction accuracy of traditional material property prediction.

[0005] To achieve the above objectives, this application provides the following solutions:

[0006] In a first aspect, the present application provides a material property prediction method based on multiple machine learning regression models and high-throughput experiments, comprising:

[0007] Obtaining preset material parameters; the material parameters include: material composition and preparation process;

[0008] Determine any one performance as the current performance;

[0009] Based on the material composition in the preset material parameters, determine the target key characteristics of the current performance under the preset material parameters;

[0010] Input the process in the preset material parameters and the target key features of the current performance under the preset material parameters into the multiple optimal performance prediction regression models corresponding to the current performance respectively, to obtain multiple sub-prediction values of the current performance under the preset material parameters; each optimal performance prediction regression model corresponding to any performance is obtained by screening, iterative training and high-throughput experiments on multiple machine learning regression models, and the multiple machine learning regression models include: random forest regression model, reverse neural network model, extreme gradient boosting decision tree regression model and support vector machine regression model;

[0011] Based on the multiple sub-prediction values of the current performance under the preset material parameters, determine the total prediction value of the current performance under the preset material parameters.

[0012] In a second aspect, the present application provides a computer device, including: a memory, a processor, and a computer program stored on the memory and executable on the processor, and the processor executes the computer program to implement the material performance prediction method based on multiple machine learning regression models and high-throughput experiments described in any one of the above.

[0013] In a third aspect, the present application provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the material performance prediction method based on multiple machine learning regression models and high-throughput experiments described in any one of the above.

[0014] In a fourth aspect, the present application provides a computer program product, including a computer program, and when the computer program is executed by a processor, it implements the material performance prediction method based on multiple machine learning regression models and high-throughput experiments described in any one of the above.

[0015] In a fifth aspect, the present application provides an application of a material performance prediction method based on multiple machine learning regression models and high-throughput experiments, which is implemented based on the material performance prediction method based on multiple machine learning regression models and high-throughput experiments described in any one of the above, and the application of the material performance prediction method based on multiple machine learning regression models and high-throughput experiments includes:

[0016] Obtain multiple material parameters to be screened;

[0017] Determine any material parameter to be screened as the current material parameter, and determine any performance as the current performance;

[0018] Based on the material composition in the current material parameters, determine the target key features of the current performance under the current material parameters;

[0019] Input the process in the current material parameters and the target key features of the current performance under the current material parameters into multiple optimal performance prediction regression models corresponding to the current performance respectively, to obtain multiple sub-prediction values of the current performance under the current material parameters;

[0020] Based on the multiple sub-prediction values of the current performance under the current material parameters, determine the total prediction value of the current performance under the current material parameters;

[0021] Determine the material parameters to be screened with the optimal total prediction value as the optimal material parameters, to achieve the optimization of the material parameters.

[0022] According to the specific embodiments provided in the present application, the present application discloses the following technical effects:

[0023] The present application discloses a material performance prediction method, device, medium, product and its application based on a multiple machine learning regression model and high-throughput experiment. First, obtain the preset material parameters; determine any performance as the current performance; secondly, based on the material composition in the preset material parameters, determine the target key features of the current performance under the preset material parameters; then, input the process in the preset material parameters and the target key features of the current performance under the preset material parameters into multiple optimal performance prediction regression models corresponding to the current performance respectively, to obtain multiple sub-prediction values of the current performance under the preset material parameters; based on the multiple sub-prediction values of the current performance under the preset material parameters, determine the total prediction value of the current performance under the preset material parameters. The present application uses the optimal performance prediction regression models obtained by screening, iterative training and high-throughput experiment on multiple machine learning regression models to perform multiple sub-predictions of the current performance under the preset material parameters in parallel, and further determine the total prediction value according to the sub-prediction values, improving the material performance prediction efficiency and prediction accuracy, and reducing the R & D cost. Description of the Drawings

[0024] To more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following will briefly introduce the drawings required to be used in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0025] Figure 1 It is an application environment diagram of a material performance prediction method based on a multiple machine learning regression model and high-throughput experiment in an embodiment of the present application;

[0026] Figure 2 It is a flowchart of a material performance prediction method based on a multiple machine learning regression model and high-throughput experiment provided in an embodiment of the present application;

[0027] Figure 3 Schematic diagram of the determination process of multiple optimal performance prediction regression models;

[0028] Figure 4 Schematic diagram of the hierarchical feature screening process;

[0029] Figure 5 R of each machine learning regression model in Example 1 2 Curve graph;

[0030] Figure 6 RMSE curve graph of each machine learning regression model in Example 1;

[0031] Figure 7 Schematic diagram of the process for predicting the process performance of unknown performance alloy compositions by the Bootstrap method;

[0032] Figure 8 Modeling result graph of the extreme gradient boosting regression model for the hardness performance of zirconium alloy in Example 1;

[0033] Figure 9 Experimental verification result graph of the hardness performance of zirconium alloy in Example 1;

[0034] Figure 10 Modeling result graph of the extreme gradient boosting regression model for the hardness performance of zirconium alloy in Example 2;

[0035] Figure 11 Experimental verification result graph of the hardness performance of zirconium alloy in Example 2;

[0036] Figure 12 Modeling result graph of the extreme gradient boosting regression model for the hardness performance of zirconium alloy in Example 3;

[0037] Figure 13 Experimental verification result graph of the hardness performance of zirconium alloy in Example 3;

[0038] Figure 14 R of each machine learning regression model in conductivity in Example 4 2 Curve graph;

[0039] Figure 15 RMSE curve graph of each machine learning regression model in conductivity in Example 4;

[0040] Figure 16 R of each machine learning regression model in hardness in Example 4 2 Curve graph;

[0041] Figure 17 RMSE curve graph of each machine learning regression model in hardness in Example 4;

[0042] Figure 18 Prediction effect diagram of each round of conductivity prediction regression model for P1-32 under the POS strategy in Example 4;

[0043] Figure 19 Prediction effect diagram of each round of hardness prediction regression model for P1-32 under the POS strategy in Example 4;

[0044] Figure 20 Prediction effect diagram of each round of conductivity prediction regression model for P2-16 under the POS strategy in Example 4;

[0045] Figure 21 Prediction effect diagram of each round of hardness prediction regression model for P2-16 under the POS strategy in Example 4;

[0046] Figure 22 Prediction effect diagram of each round of conductivity prediction regression model for PP3-8 under the POS strategy in Example 4;

[0047] Figure 23 Prediction effect diagram of each round of hardness prediction regression model for PP3-8 under the POS strategy in Example 4;

[0048] Figure 24 Performance distribution diagram of unknown space prediction points of P1-32 in all regions in Example 4;

[0049] Figure 25 Performance distribution diagram of unknown space prediction points of P1-32 in the target performance region in Example 4;

[0050] Figure 26 Performance distribution diagram of unknown space prediction points of P2-16 in all regions in Example 4;

[0051] Figure 27 Performance distribution diagram of unknown space prediction points of P2-16 in the target performance region in Example 4;

[0052] Figure 28 Performance distribution diagram of unknown space prediction points of PP3-8 in all regions in Example 4;

[0053] Figure 29 Performance distribution diagram of unknown space prediction points of PP3-8 in the target performance region in Example 4;

[0054] Figure 30 Experimental result diagram of conductivity in the optimization process of the POS strategy in Example 4;

[0055] Figure 31 Experimental result diagram of hardness in the optimization process of the POS strategy in Example 4;

[0056] Figure 32 It is the performance distribution diagram of the POS strategy experimental data in Example 4;

[0057] Figure 33 It is the schematic structural diagram of a computer device provided by an embodiment of the present application. Specific implementation manners

[0058] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without making creative efforts shall fall within the protection scope of the present application.

[0059] The purpose of the present application is to provide a material property prediction method, device, medium, product and its application based on a multiple machine learning regression model and high-throughput experiments, aiming to improve the prediction efficiency and accuracy of material properties.

[0060] To make the above objects, features, and advantages of the present application more obvious and understandable, the present application will be further described in detail below with reference to the accompanying drawings and specific implementation manners.

[0061] The material property prediction method based on a multiple machine learning regression model and high-throughput experiments provided by the embodiments of the present application can be applied to, for example Figure 1In the application environment shown. Among them, the terminal 102 communicates with the server 104 through the network. The data storage system can store the data that the server 104 needs to process. The data storage system can be set separately, integrated on the server 104, placed on the cloud or other servers. The terminal 102 can send the preset material parameters to the server 104. After receiving the preset material parameters, for the preset material parameters, the server 104 determines the target key features of the current performance under the preset material parameters based on the material composition in the preset material parameters; inputs the process in the preset material parameters and the target key features of the current performance under the preset material parameters into the multiple optimal performance prediction regression models corresponding to the current performance respectively, and obtains multiple sub-prediction values of the current performance under the preset material parameters; determines the total prediction value of the current performance under the preset material parameters based on the multiple sub-prediction values of the current performance under the preset material parameters. The server 104 can feedback the obtained total prediction value of the current performance under the preset material parameters to the terminal 102. In addition, in some embodiments, the material performance prediction method based on the multiple machine learning regression model and high-throughput experiment can also be implemented separately by the server 104 or the terminal 102. For example, the terminal 102 can directly perform the material performance prediction based on the multiple machine learning regression model and high-throughput experiment for the preset material parameters, or the server 104 can obtain the preset material parameters from the data storage system and perform the material performance prediction based on the multiple machine learning regression model and high-throughput experiment for the preset material parameters.

[0062] In an exemplary embodiment, as Figure 2 shown, a material performance prediction method based on a multiple machine learning regression model and high-throughput experiment is provided, including:

[0063] Step 1: Obtain the preset material parameters.

[0064] Among them, the material parameters include: material composition and preparation process.

[0065] Step 2: Determine any performance as the current performance.

[0066] Among them, the performance includes: electrical performance and mechanical performance. The electrical performance includes: conductivity. The mechanical performance includes: hardness.

[0067] Step 3: Based on the material composition in the preset material parameters, determine the target key features of the current performance under the preset material parameters.

[0068] Step 4: Input the process in the preset material parameters and the target key features of the current performance under the preset material parameters into multiple optimal performance prediction regression models corresponding to the current performance respectively, and obtain multiple sub-prediction values of the current performance under the preset material parameters.

[0069] Among them, each optimal performance prediction regression model corresponding to any performance is obtained by screening, iterative training and high-throughput experiments on multiple machine learning regression models. The multiple machine learning regression models include: Random Forest Regression (RFR), BackPropagation Neural Network (BPNN), Extreme Gradient Boosting (XGB) and Support Vector Regression (SVR).

[0070] Step 5: Based on the multiple sub-prediction values of the current performance under the preset material parameters, determine the total prediction value of the current performance under the preset material parameters.

[0071] Specifically, determine the average value of all sub-prediction values of the current performance under the preset material parameters as the total prediction value of the current performance under the preset material parameters.

[0072] As an alternative implementation, as Figure 3 shown, the determination process of multiple optimal performance prediction regression models of any current performance includes:

[0073] Step 41: Initialize the first data set to obtain an initial first data set; the first data set includes: historical data of multiple sample materials; the historical data includes historical material parameters and actual values of the current performance; the sample materials are materials that have been prepared.

[0074] Specifically, the historical data of generally no less than 30 sample materials should be included in the first data set.

[0075] Step 42: Based on the material composition in the historical material parameters of each sample material in the initial first data set, determine the initial features to be screened for the current performance; the features to be screened include extended features and composite features.

[0076] Specifically, according to the material composition in the first data set, determine the features to be screened through a publicly available feature database (such as XenonPy) and literature.

[0077] The extended features are related to the phase stability of the material, and the extended features include: atomic radius difference, electronegativity difference, valence electron concentration, mixing enthalpy, configurational entropy, intermediate parameters (parameter Ω, parameter Λ, and parameter γ), local electronegativity mismatch, cohesive energy, modulus mismatch in the strengthening model, local atomic size mismatch, energy term in the strengthening model, Peierls factor, sixth power of the work function, shear model, shear model difference, local lattice mismatch degree, and lattice distortion energy.

[0078] The calculation formulas for some extended features include:

[0079] Ω = T × ΔS / |ΔH| (1)

[0080]

[0081]

[0082] Λ = ΔS / (δr × δr) (4)

[0083]

[0084] Among them, Ω is parameter Ω; T is the average melting point of all elements; ΔS is the configurational entropy; ΔH is the mixing enthalpy; R is the gas constant; n is the total number of elements; w j is the molar ratio of the j-th element; w i is the molar ratio of the i-th element; is the mixing enthalpy of the j-th element and the i-th element; Λ is parameter Λ; δr is the atomic radius difference; r j is the atomic radius of the j-th element; r is the average atomic radius of the elements; γ is parameter γ; r min is the maximum atomic radius of all elements; r max is the minimum atomic radius of all elements.

[0085] The composite features include: basic physicochemical features and features obtained by further calculating each basic physicochemical feature using formulas (7) - (13). The basic physicochemical features include: atomic number, general atomic radius, atomic radius defined by Ram et al., atomic volume, atomic weight, boiling point, bulk modulus, C_6 dispersion coefficient, general covalent radius, covalent radius defined by Ceder et al., single-bond covalent radius defined by Pico et al., double-bond covalent radius defined by Pico et al., triple-bond covalent radius defined by Pico et al., density, dipole polarizability, electronegativity, electron affinity, Allen electronegativity scale, Ghosh electronegativity scale, Pauling electronegativity scale, enthalpy of fusion, density functional theory (DFT) band gap energy at T = 0K ground state, DFT atomic energy at T = 0K ground state (original VASP value), BCC lattice parameter estimated based on density functional theory (DFT) volume, face-centered cubic (FCC) lattice parameter estimated based on density functional theory (DFT) volume, DFT magnetic moment at T = 0K ground state, DFT atomic volume at T = 0K ground state, Herfindahl-Hirschman Index (HHI) production value, Herfindahl-Hirschman Index (HHI) reserve value, mass heat capacity, molar heat capacity, atomic volume in ICSD database, heat of vaporization, heat of formation, lattice constant, Mendeleev number, melting point, molar volume, total unfilled electrons, total valence electrons, unfilled electrons in d shell, valence electrons in d shell, unfilled electrons in f shell, valence electrons in f shell, unfilled electrons in p shell, valence electrons in p shell, unfilled electrons in s shell, valence electrons in s shell, period in the periodic table, specific heat, thermal conductivity, general van der Waals radius, van der Waals radius obtained by Alvarez et al., van der Waals radius of MM3 force field, van der Waals radius of UFF force field, speed of sound, polarizability.

[0086]

[0087] f max,i1 = max(f j,i1 ) (11)

[0088] f min,i1 = min(f j,i1 ) (12)

[0089]

[0090] where f ave,i1 is the weighted average of the i1-th basic physicochemical feature; w j * is the normalized molar ratio of the j-th element; f j,i1 is the value of the i1-th basic physicochemical feature of the j-th element; fvar,i1 is the weighted variance of the i1-th basic physicochemical feature; f gmean,i1 is the geometric mean of the i1-th basic physicochemical feature; for f j,i1 to the power of w j ; f hmean,i1 is the harmonic mean of the i1-th basic physicochemical feature; f max,i1 is the maximum value of the i1-th basic physicochemical feature; f min,i1 is the minimum value of the i1-th basic physicochemical feature; f sum,i1 is the sum of the i1-th basic physicochemical features; w j is the molar ratio of the j-th element.

[0091] Step 43: Using statistical analysis methods, perform hierarchical feature screening on the initial features to be screened of the current performance to obtain the initial key features of the current performance.

[0092] Specifically, as Figure 4 shown, the statistical analysis methods include: variance filtering, embedding method, and correlation screening. The key features obtained according to different materials and properties are different, but are included in the extended features and composite features.

[0093] Step 44: Determine the process in the historical material parameters of each sample material in the initial first dataset, the initial key features of the current performance under the historical material parameters of each sample material, and the actual values of the current performance of each sample material as the initial second dataset.

[0094] Step 45: Based on the initial second dataset and multiple machine learning regression models, determine the initial first performance prediction regression model of the current performance.

[0095] As an optional implementation manner, step 45 includes:

[0096] Step 4501: Divide the initial second dataset to obtain an initial first training set and an initial first test set.

[0097] Specifically, in the 30 - 90% ratio interval, different division ratios are generated at 10% intervals, and the second dataset is randomly divided into a first training set and a first test set according to this ratio.

[0098] Step 4502: Use the initial first training set to train each machine learning regression model respectively to obtain multiple initially trained machine learning regression models.

[0099] Specifically, use each machine learning regression algorithm to perform cross-validation training and modeling on the first training set multiple times (e.g., 100 times). During the modeling process, first perform hyperparameter optimization, and then perform cross-validation according to a certain ratio (e.g., five-fold).

[0100] Step 4503: Use the initial first test set to evaluate each initially trained machine learning regression model, and determine the initially trained machine learning regression model that meets the evaluation criteria as the initial first performance prediction regression model; the evaluation criteria are: the minimum average root mean square error or the highest average correlation coefficient and average determination coefficient.

[0101] Specifically, after each optimization of the parameters of the machine learning regression model, perform statistical evaluation of the prediction accuracy of the machine learning model for the corresponding first test set. After comprehensive evaluation according to the criterion of the minimum average root mean square error or the highest average correlation coefficient and average determination coefficient predicted on the first test set, determine the optimal machine learning performance prediction regression model to obtain the first performance prediction regression model.

[0102] Step 46: Interpolate all historical material parameters in the initial first dataset, and based on each interpolated component, determine the interpolated key features corresponding to the current performance. Determine all interpolated processes and interpolated key features as the initial material feature parameter set.

[0103] Specifically, design the potential material composition and preparation process parameter space of the research material system according to the first dataset, which can be interpolated or extrapolated within the parameter range of the first dataset. For example, for the possible variation range of the designed material composition and preparation process, divide it at a certain interval, and then perform permutations and combinations on all possible material compositions and preparation processes. This potential material composition and preparation process parameter space will be the latent variable space for subsequent machine learning prediction and search.

[0104] Step 47: Use the Bootstrap method to perform multiple sampling with replacement on the initial second dataset to obtain multiple initial third datasets.

[0105] Specifically, perform sampling with replacement on the second dataset 1000 times by the Bootstrap method to construct 1000 third datasets.

[0106] Step 48: Based on each initial third dataset and the initial first performance prediction regression model respectively, determine multiple initial second performance prediction regression models for the current performance.

[0107] Specifically, each third dataset is first divided into a nested training set and an independent test set in a ratio of 9:1, and then the nested training set is divided into a second training set and a second test set in a ratio of 9:1. The first performance prediction regression model is used to model these second training sets, obtaining 1000 second performance prediction regression models (M1, M2, ……, M 1000 ), and the 1000 models predict the process parameter space of the component with unknown pre-set performance. Each component-process data corresponds to 1000 sub-predicted values (P1, P2, ……, P 1000 ) of the performance. Assuming that these predicted values follow a Gaussian distribution, the standard deviation of the 1000 sub-predicted values is used as the prediction accuracy error or uncertainty, and the average value of the predicted values is used as the total predicted value of the performance of this component-process. The performance of the second performance prediction regression model is evaluated by the root mean square error and the average correlation coefficient or average determination coefficient of 1000 modeling evaluations as the evaluation indicators.

[0108] Step 49: Iteratively train and optimize the initial first dataset, the initial key features, the initial second dataset, the initial first performance prediction regression model, the initial material feature parameter set, multiple initial third datasets, and multiple initial second performance prediction regression models to obtain multiple optimal performance prediction regression models corresponding to the current performance; among them, the iterative training and optimization are carried out using the first training process or the second training process.

[0109] As an alternative implementation, in step 49, the first training process includes:

[0110] Step 49101: Obtain the first dataset at the current iteration.

[0111] Step 49102: Based on the material composition in the historical material parameters of each sample material in the first dataset at the current iteration, determine the features to be screened at the current iteration for the current performance.

[0112] Step 49103: Use statistical analysis methods to perform hierarchical feature screening on the features to be screened at the current iteration for the current performance, and obtain the key features at the current iteration for the current performance.

[0113] Step 49104: Determine the process in the historical material parameters of each sample material in the first dataset at the current iteration, the initial key features of the current performance under the historical material parameters of each sample material, and the actual values of the current performance of each sample material as the second dataset at the current iteration.

[0114] Step 49105: Based on the second dataset at the current iteration and multiple machine learning regression models, determine the first performance prediction regression model at the current iteration for the current performance.

[0115] As an alternative implementation, step 49105 includes:

[0116] Step 491051: Divide the second data set at the current iteration number to obtain the first training set at the current iteration number and the first test set at the current iteration number.

[0117] Step 491052: Use the first training set at the current iteration number to train each machine learning regression model respectively to obtain multiple trained machine learning regression models at the current iteration number.

[0118] Step 491053: Use the first test set at the current iteration number to evaluate each trained machine learning regression model at the current iteration number, and determine the trained machine learning regression model at the current iteration number that meets the evaluation conditions as the first performance prediction regression model at the current iteration number.

[0119] Step 49106: Interpolate all historical material parameters in the first data set at the current iteration number, and determine the interpolated key features corresponding to the current performance based on each interpolated component. Determine all the interpolated processes and the interpolated key features as the material feature parameter set at the current iteration number.

[0120] Step 49107: Through the Bootstrap method, perform multiple sampling with replacement on the second data set at the current iteration number to obtain multiple third data sets at the current iteration number.

[0121] Step 49108: Based on each third data set at the current iteration number and the first performance prediction regression model at the current iteration number respectively, determine multiple second performance prediction regression models at the current iteration number for the current performance.

[0122] Step 49109: Input each group of the interpolated processes and the interpolated key features in the material feature parameter set at the current iteration number into each second performance prediction regression model at the current iteration number respectively to obtain multiple sub-prediction values at the current iteration number for the current performance under the corresponding interpolated material parameters.

[0123] Step 49110: Based on all the prediction values at the current iteration number for the current performance under each interpolated material parameter respectively, determine the total prediction value at the current iteration number for the current performance under each interpolated material parameter.

[0124] Step 49111: Adopt a multiple machine learning sampling strategy to sample each third data set at the current iteration number to obtain a fourth data set at the current iteration number.

[0125] Specifically, the multiple machine learning sampling strategies include, but are not limited to, reinforcement learning sampling strategy, expected learning sampling strategy, active learning sampling strategy, transfer learning sampling strategy, and Pareto optimization sampling strategy (POS) based on the target performance region. The number of material samples corresponding to each sampling strategy is determined according to the number of sampling methods, and the number of material samples corresponding to each sampling strategy is greater than 2.

[0126] (1) Reinforcement learning sampling strategy: Taking the Zr-Ni-Cu-Hf-Zr alloy as an example, by adding the alloying elements and their contents that best meet the target performance to replace the main elements, the iteration of the low-component alloy to the high-component alloy is completed, such as:

[0127]

[0128] (2) Expected learning sampling strategy: Prioritize selecting samples with prediction performance that meets the target performance for experiments.

[0129] (3) Active learning sampling strategy: For the Bayesian optimization sampling algorithm, the acquisition function E(I) can be defined as follows:

[0130] E(I) = E y～f(x|D) [max(0, y - f best )] = σ'[ψ(z) + zΦ(z)] (14)

[0131] where E y～f(x|D) [·] is the expectation calculation of y under the condition of the given data D; y is the test value of the performance; f best is the best performance value observed so far; ψ(z) is the probability density function; z is the difference between the best performance value observed so far and the predicted value, μ is the predicted value of the performance, σ' is the posterior standard deviation, σ is the standard deviation of the original prediction, err is the experimental error; Φ(z) is the cumulative distribution function of the standard normal distribution.

[0132] When f best = y max (y max is the maximum value of the test value of the performance), the sampling strategy is called efficient global optimization, and the acquisition function tends to select data points that meet the target performance region, that is, "exploitation".

[0133] When f best = μ max (μ max is the maximum value of the predicted value of the performance), the sampling strategy is called knowledge gradient, and the acquisition function tends to select data points that reduce the uncertainty of the model, that is, "exploration".

[0134] (4) Transfer learning sampling strategy: Use all data (source domain) to build a model, and then transfer learning (such as through hyperparameter fine-tuning) to the target performance data (target domain). The data in the target domain is selected as the top 20% sample materials in the experimental data of this generation that best match the target performance.

[0135] (5) Pareto optimization sampling strategy (POS) based on the target performance region: This strategy first delimits a target performance region on the multi-objective performance distribution of the initial dataset. Use the genetic algorithm NSGA-II to calculate the Pareto frontiers of different rankings (Rank) of the predicted unknown space performance in this region and the crowding distance (CD) of each point within the same Rank value. The sampling principle of this strategy is to prioritize Rank and then CD, that is, to sample on the Pareto frontiers with higher Rank values first, and sort the CD values of each point at the same Rank value, and select the points with larger CD values. At the same time, to prevent the experimental points selected in this region from having similar compositions and processes, the sampling process also considers feature diversity, that is, the experimental points should cover as rich a variety of material compositions and preparation processes as possible. Different from sampling on the traditional Pareto frontiers, in this work, the Pareto frontiers are used as a sampling strategy for iterative experiments, considering feature diversity. Except for the points on the Pareto frontiers, the features of the remaining points should ensure that all parameters can appear in the experimental points of the next round, and at the same time, the data points should have at least 2 parameter differences from each other.

[0136] Step 49112: Prepare based on the interpolated process, interpolated key features, and corresponding interpolated material parameters in the fourth dataset at the current iteration number to obtain multiple material samples at the current iteration number, and perform performance tests on all material samples at the current iteration number to obtain the test values of the current performance of each material sample at the current iteration number.

[0137] Specifically, when preparing to obtain multiple material samples, a process, high-throughput, multi-station, and automated experimental process is adopted. The full-process, high-throughput, multi-station, and automated experimental process includes: high-throughput automatic material powder preparation, high-throughput ball milling and powder mixing, multi-station electric press pressing into a compact block, high-throughput automatic arc melting or high-temperature powder sintering, high-throughput cold embedding or hot embedding, high-throughput automatic sample grinding and polishing, high-throughput material hardness measurement, high-throughput material composition determination, and high-throughput X-ray diffraction phase measurement, etc.

[0138] Step 49113: Determine the average root mean square error, average correlation coefficient, and average determination coefficient at the current iteration number based on the test values of the current performance of each material sample at the current iteration number and the total predicted value of the current performance at the current iteration number under the corresponding interpolated material parameters.

[0139] Step 49114: Determine whether the stop condition is met; the stop condition is: the average root mean square error at the current iteration is the smallest, or the average correlation coefficient and the average determination coefficient at the current iteration are the highest, or the preset training times are reached.

[0140] Step 49115: If so, determine the multiple second performance prediction regression models at the current iteration of the current performance as the multiple optimal performance prediction regression models of the current performance, and determine the key features at the current iteration of the current performance as the target key features.

[0141] Step 49116: If not, update the current iteration to the next iteration and return "Obtain the first data set at the current iteration".

[0142] As an optional implementation manner, in step 49, the second training process includes:

[0143] Step 49201: Determine the initial key features of the current performance as the target key features of the current performance.

[0144] Step 49202: Interpolate all historical material parameters in the initial first data set, and based on the interpolated components, determine the interpolated key features corresponding to the current performance. Determine all the interpolated processes and the interpolated key features as the material feature parameter set at the current iteration.

[0145] Step 49203: Through the Bootstrap method, perform multiple sampling with replacement on the initial second data set to obtain multiple third data sets at the current iteration.

[0146] Step 49204: Based on each third data set at the current iteration and the initial first performance prediction regression model respectively, determine multiple second performance prediction regression models at the current iteration of the current performance.

[0147] Step 49205: Input each group of the interpolated processes and the interpolated key features in the material feature parameter set at the current iteration into the second performance prediction regression models at the current iteration respectively, and obtain multiple sub-prediction values at the current iteration of the current performance under the corresponding interpolated material parameters.

[0148] Step 49206: Based on the predicted values at all current iterations of the current performance under each interpolated material parameter respectively, determine the total predicted value at the current iteration of the current performance under each interpolated material parameter.

[0149] Step 49207: Adopt a multiple machine learning sampling strategy to sample each third data set at the current iteration to obtain a fourth data set at the current iteration.

[0150] Step 49208: Prepare based on the interpolated process in the fourth dataset and the interpolated material parameters corresponding to the interpolated key features at the current iteration number, to obtain multiple material samples at the current iteration number, and perform performance tests on all the material samples at the current iteration number, to obtain the test values of the current performance of each material sample at the current iteration number.

[0151] Step 49209: Determine the average root mean square error, average correlation coefficient, and average determination coefficient at the current iteration number based on the test values of the current performance of each material sample at the current iteration number and the total predicted value of the current performance at the current iteration number under the corresponding interpolated material parameters.

[0152] Step 49210: Determine whether the stop condition is satisfied; the stop condition is: the average root mean square error at the current iteration number is the smallest, or the average correlation coefficient and average determination coefficient at the current iteration number are the highest, or the preset training number is reached.

[0153] Step 49211: If so, determine the multiple second performance prediction regression models of the current performance at the current iteration number as the multiple optimal performance prediction regression models of the current performance.

[0154] Step 49212: If not, update the current iteration number to the next iteration number, and return "Interpolate all the historical material parameters in the initial first dataset, and based on each interpolated component, determine the interpolated key features corresponding to the current performance, and determine the material feature parameter set at the current iteration number with all the interpolated processes and interpolated key features".

[0155] Furthermore, the method of the present application is also used to conduct experiments in Examples 1 - 4.

[0156] Example 1

[0157] Example 1 is used to illustrate a material performance prediction method based on a multiple machine learning regression model and high-throughput experiments of the present application, and is specifically used for the first round of iterative experiments on the hardness performance optimization of zirconium alloys. The steps are as follows:

[0158] S101: Obtain an initial 124 alloy composition process - performance dataset (i.e., the first dataset) through high-throughput experiments.

[0159] S102: According to the obtained alloy composition information, calculate the composite properties combined with the chemical composition molar ratio as the weight through different combination methods, and finally obtain multiple extended features and composite features except for the composition.

[0160] S103: Remove redundant features from the extended features and composite features by means including variance filtering, embedding method filtering, and correlation filtering, retain the key features, and obtain a dataset of composition process - key features - target performance (i.e., the second dataset). The feature screening method is as Figure 4 shown. Finally, retain the key features shown in Table 1.

[0161] Table 1 Table of Key Features and Their Physical Meanings

[0162]

[0163]

[0164] S104: Generate different partitioning ratios at intervals of 10% in the range of 30 - 90% for the second dataset obtained in step S103, and randomly partition it into a first training set and a first test set according to this ratio. Then use the random forest regression model, backpropagation neural network model, extreme gradient boosting decision tree regression model, and support vector machine regression model (including: Support Vector Regression with Polynomial kernel functions (SVR_POLY) model and Support Vector Regression with Radial Basis Function (SVR_RBF) model), and perform 100 - fold cross - validation training and modeling on the first training set respectively. In the modeling process, first perform hyperparameter optimization, and then perform five - fold cross - validation. After optimizing the model parameters each time, perform statistical evaluation of the prediction accuracy of the machine learning model for the first test set. After comprehensive evaluation according to the criterion of the minimum average root mean square error (RMSE) predicted on the first test set, or the highest average correlation coefficient (R) and average coefficient of determination, determine the optimal machine learning performance prediction regression model (i.e., the first performance prediction regression model). The model selection results are as Figure 5 and Figure 6 shown.

[0165] S105: Construct a parameter space of potential material compositions for the Zr - Sn - Nb - Fe - Cr - Mo system in the design study.

[0166] S106: Bootstrap method is used to perform sampling with replacement 1000 times on the second data set obtained in step S103 to construct 1000 third data sets. Each third data set is first divided into a nested training set and an independent test set according to a ratio of 9:1, and then the nested training set is further divided into a second training set and a second test set according to a ratio of 9:1. The optimal model is used to model these third data sets to obtain 1000 machine learning models (i.e., the second performance prediction regression models) (M1, M2, ……, M 1000 ), and these 1000 models are used to predict the component process parameter space with unknown pre-set performance. Each component-process data corresponds to 1000 predicted values (P1, P2, ……, P 1000 ) (i.e., sub-predicted values). Assuming that these sub-predicted values follow a Gaussian distribution, the standard deviation of the 1000 sub-predicted values is used as the prediction accuracy error or uncertainty, and the average value of the sub-predicted values is used as the total predicted value of the performance. The model performance is evaluated by the average root mean square error and the average correlation coefficient of 1000 modellings. The Bootstrap method is as Figure 7 shown, and the model performance is as Figure 8 shown. The RMSE of the second training set for 1000 modellings is 6.2 HV, and R is 0.980. The RMSE of the second test set for 1000 modellings is 23.8 HV, and R is 0.540. The RMSE of the independent test set for 1000 modellings is 24.2 HV, and R is 0.504.

[0167] S107: Four sampling strategies, including reinforcement learning sampling, expected learning sampling, active learning sampling (where active learning sampling includes two algorithms: efficient global optimization and knowledge gradient), and transfer learning sampling, are used to sample the set of alloy composition process - predicted values. There are 8 samples for each sampling strategy, and a total of 32 samples are selected, which can be run through the 32 workstations of the high-throughput arc melting system.

[0168] Reinforcement learning: Based on the low-hardness Zr-Sn-Nb alloy, each alloying element composition in the binary alloy with lower hardness is added, and 8 experimental points are taken.

[0169] Expected learning: The hardness properties corresponding to the pre-set space components are sorted from smallest to largest. Components with a higher ranking are preferentially selected, and alloy systems that have never appeared are preferably selected. Considering the composition differences for alloys in the same system, a total of 8 experimental points are selected.

[0170] Active learning: It is calculated through the acquisition function E(I), as shown in formula (14) specifically. Among them, for efficient global optimization (EGO), take f best = y min , where y min is the minimum value of the test values of the performance, and for knowledge gradient (KG), take is the minimum predicted value of performance. According to the calculation results, preferentially select the components with higher rankings, and try to select alloy systems that have never appeared. For alloys in the same system, consider the compositional differences. Select 4 experimental points for both EGO and KG, and a total of 8 experimental points are selected.

[0171] 4. Transfer learning: Use all the data (source domain) to build a model, and then transfer learn (hyperparameter fine-tuning) to the target performance data (target domain). Select the top 20% of the samples in the current generation of experimental data that best match the target performance as the target domain data. According to the prediction results of the transfer learning model, sort the hardness performance corresponding to the preset space components from smallest to largest. Preferentially select the components with higher rankings, and try to select alloy systems that have never appeared. For alloys in the same system, consider the compositional differences, and a total of 8 experimental points are selected.

[0172] S108: Carry out the preparation of high-throughput experimental samples for the alloy samples selected in step S107. The experimental preparation process includes: high-throughput automatic material powder preparation, high-throughput ball milling and powder mixing, multi-station electric press to press into a compact block, high-throughput automatic arc melting, high-throughput cold embedding or hot embedding, and high-throughput automatic sample grinding and polishing. And carry out high-throughput material hardness measurement and high-throughput material composition determination on the experimental samples.

[0173] S109: Compare the experimental performance of the samples obtained in step S108 with the predicted performance of the samples in S106, and use the root mean square error and the correlation coefficient as the evaluation indicators for the prediction effect.

[0174] The experimental verification results are as Figure 9 shown. The RMSE between the experimental results and the predicted results of the samples sampled by reinforcement learning is 64.0 HV, and R is -0.483; the RMSE between the experimental results and the predicted results of the samples sampled by expected learning is 60.1 HV, and R is 0.187; the RMSE between the experimental results and the predicted results of the samples sampled by active learning is 80.8 HV, and the correlation coefficient (R) is 0.662; the RMSE between the experimental results and the predicted results of the samples sampled by transfer learning is 95.9 HV, and R is -0.570;

[0175] In Example 1, among the four sampling strategies, the RMSE between the experimental results and the predicted results of the samples sampled by expected learning is the smallest, and the R of active learning is the smallest.

[0176] S110: Evaluate whether the model performance and the material performance (hardness performance) meet the requirements.

[0177] In Example 1, both the model performance and the material performance (hardness performance) do not meet the requirements, so continue with the next round of iterative experiments.

[0178] Example 2

[0179] Example 2 is used to illustrate the material property prediction method based on multiple machine learning regression models and high-throughput experiments of the present application, and is used for the second-round iterative experiment of zirconium alloy hardness property optimization. The steps are as follows:

[0180] Example 2 is basically the same as Example 1, with the special features being:

[0181] The first dataset in Example 2 contains 124 + 32 = 156 pieces of data.

[0182] Example 2 retains the results of S102 - S105 in Example 1 and starts from step S206. The model performance is as Figure 10 shown. The RMSE of the training set for 1000 - time modeling is 7.7 HV, R is 0.978. The RMSE of the test set for 1000 - time modeling is 26.8 HV, R is 0.695. The RMSE of the independent test set for 1000 - time modeling is 27.9 HV, R is 0.646.

[0183] The sampling strategy selection in S107 in Example 2 is the same as that in Example 1, and the full - process high - throughput preparation and characterization method in S108 is the same. The final experimental verification results are as Figure 11 shown. The RMSE between the experimental results and the predicted results of the samples sampled by reinforcement learning is 55.3 HV, R is 0.220; the RMSE between the experimental results and the predicted results of the samples sampled by expected learning is 25.6 HV, R is 0.500; the RMSE between the experimental results and the predicted results of the samples sampled by active learning is 19.2 HV, R is 0.616; the RMSE between the experimental results and the predicted results of the samples sampled by transfer learning is 14.1 HV, R is 0.804.

[0184] In Example 2, among the four sampling strategies, the RMSE between the experimental results of the transfer - learning samples and the predicted results is the smallest, and R is also the smallest.

[0185] Compared with Example 1, in the final experimental verification results of Example 2, the RMSE between the experimental results of the samples of the four sampling strategies and the predicted results decreases, and R increases. It shows that the iterative experiment can improve the model performance.

[0186] In Example 2, both the model performance and the material properties (hardness properties) do not meet the requirements. Therefore, the next - round iterative experiment is continued.

[0187] Example 3

[0188] Example 3 is used to illustrate the material property prediction method based on multiple machine learning regression models and high - throughput experiments of the present application, and is used for the third - round iterative experiment of zirconium alloy hardness property optimization. The steps are as follows:

[0189] Example 3 is basically the same as Example 1, with the special features being:

[0190] In Example 3, the initial data set contains 124 + 32 + 32 = 188 pieces of data.

[0191] Example 3 retains the results of S102 - S105 in Example 1 and starts from step S306. The model performance is as Figure 12 shown. The RMSE of the training set for 1000 - time modeling is 7.3 HV, R is 0.980. The RMSE of the test set for 1000 - time modeling is 27.6 HV, R is 0.761. The RMSE of the independent test set for 1000 - time modeling is 27.4 HV, R is 0.662.

[0192] In Example 3, the sampling strategy selection in S106 is the same as that in Example 1, and the full - process preparation and characterization method in S108 is the same. The final experimental verification results are as Figure 13 shown. The RMSE between the experimental results and the predicted results of the reinforcement learning sampling samples is 9.8 HV, R is 0.930; the RMSE between the experimental results and the predicted results of the expected learning sampling samples is 7.3 HV, R is 0.695; the RMSE between the experimental results and the predicted results of the active learning sampling samples is 23.6 HV, R is 0.307; the RMSE between the experimental results and the predicted results of the transfer learning sampling samples is 19.6 HV, R is 0.897.

[0193] In Example 3, among the four sampling strategies, the RMSE between the experimental results and the predicted results of the expected learning samples is the smallest, and the R of the reinforcement learning is the largest.

[0194] Compared with Example 1 and Example 2, in the final experimental verification results of Example 3, the RMSE between the experimental results and the predicted results of the reinforcement learning sampling strategy and the expected learning sampling strategy gradually decreases. However, the RMSE of the active learning sampling strategy and the transfer learning sampling strategy shows a trend of first decreasing and then increasing; the R of the reinforcement learning sampling strategy, the expected learning sampling strategy, and the transfer learning sampling strategy gradually increases, while the R of the active learning sampling strategy gradually decreases.

[0195] After multiple rounds of iterative experiments, Zr alloy compositions with low, medium, and high hardnesses were obtained according to the target. The experimental hardnesses of the different - hardness zirconium alloys obtained using Example 3 are shown in Table 2.

[0196] Table 2 Experimental hardness table of zirconium alloys with different hardnesses

[0197]

[0198]

[0199] As shown in Table 2, the model performance and material properties (hardness properties) in Example 3 both meet the requirements, so the iterative experiment is terminated.

[0200] Example 4

[0201] Example 4 is used to illustrate the material property prediction method based on a multiple machine learning regression model and high-throughput experiment of the present application, and is used for the multi-objective three-round iterative optimization process of the hardness and electrical conductivity of Cu-Zr-Cr alloys. The steps are as follows:

[0202] S401: Obtain an initial dataset of 128 alloy compositions - processes (solution temperature T1, time t1, aging temperature T2, time t2) - properties (electrical conductivity and hardness) through literature data collection.

[0203] S402: According to the obtained alloy composition information, calculate the composite properties combined with the chemical composition molar ratio as the weight through different combination methods, and finally obtain multiple extended features and composite features except for the composition.

[0204] S403: Use means including variance filtering, embedding method filtering, and correlation filtering to remove redundant features from the extended features and composite features, retain the key features, and obtain a dataset of composition process - key features - target properties. The key features of the hardness prediction regression model and the key features of the electrical conductivity prediction regression model are shown in Table 3 and Table 4 respectively.

[0205] Table 3 Key features of the hardness prediction regression model and their physical meanings

[0206] Key Feature Physical Meaning differenceofelectronegativity Electronegativity Difference parameterΩ Parameter Ω localmodelmismatch Local Model Mismatch

[0207] Table 4 Key features of the electrical conductivity prediction regression model and their physical meanings

[0208]

[0209]

[0210] S404: For the dataset obtained in step S403, different partitioning ratios are generated at intervals of 10% within the range of 30 - 90%, and it is randomly partitioned into an initial training set and a test set according to this ratio. Then, multiple machine learning regression algorithms (including but not limited to Random Forest Regression (RFR), Back Propagation Neural Network (BPNN), Extreme Gradient Boosting Decision Tree Regression (XGB), Support Vector Machine Regression (SVR_RBF, SVR_POLY)) are used to perform cross - validation training and modeling on the initial training set data multiple times (such as 100 times). During the modeling process, hyperparameter optimization is first carried out, and then cross - validation is performed according to a certain ratio (such as five - fold). After optimizing the model parameters each time, statistical evaluation of the prediction accuracy of the machine learning model is carried out for the corresponding test set. After comprehensive evaluation based on the criterion of the minimum average root mean square error or the highest average correlation coefficient and average determination coefficient predicted on the test set, the optimal machine learning performance prediction regression model is determined.

[0211] The selection results of each machine learning regression model in Example 4 are as Figures 14 - 17 shown. For the conductivity and hardness models, the XGB model has a higher R 2 and a lower RMSE when the training set: test set = 9:1. Select this algorithm and the training set - test set partitioning ratio for subsequent modeling and prediction.

[0212] S405: Construct the process parameter space of the Cu - Zr - Cr alloy composition.

[0213] S406: Using the XGBoost algorithm, when the training set: test set = 9:1, draw the scatter plots of model prediction - literature data for conductivity and hardness under their respective models, as Figures 18 - 23 shown. From the perspective of the model modeling effect, for the conductivity model and the hardness model, the prediction accuracy on the training sets of each round of modeling is very good, with R 2 close to 1 and a low RMSE. However, on the test set, R 2 decreases and the RMSE increases, indicating a certain degree of overfitting.

[0214] It can be seen that after introducing the experimental data of P1 - 32 (32 samples in the first round) as the initial dataset in the initial literature dataset for P2 - 16 (16 samples in the second round), the model effect decreases in both the conductivity and hardness models. And after introducing the experimental data of P2 - 16, the model effect of PP3 - 8 (8 samples in the third round) improves to varying degrees in terms of conductivity and hardness performance.

[0215] S407: For the set of alloy composition processes - predicted values, a Pareto optimization sampling strategy (POS) based on the target performance region is adopted. This strategy first delimits a desired target performance region on the conductivity and hardness performance distributions of the initial data set. In the Cu-Cr-Zr alloy, the goal is the performance region of high hardness and high conductivity. The genetic algorithm NSGA-II is used to calculate the Pareto frontiers of different rankings (Rank) of the predicted unknown space performance on this region and the crowding distance of each point within the same Rank value. The sampling principle of this strategy is to prioritize Rank and then CD, that is, to sample first on the Pareto front with a higher Rank value, sort the CD values of each point at the same Rank value, and select the point with a larger CD value. At the same time, in order to prevent the experimental points selected in this region from having similar composition - processes, the sampling process also considers feature diversity, that is, the experimental points should cover as rich a variety of material compositions and preparation processes as possible. Experimental points are selected according to the above method to conduct experiments.

[0216] S407.1: As Figures 24 - 25 shown, in the target performance region (EC ≥ 80% IACS and HD ≥ 145 HV), select the experimental points of P1-32 (32 samples in the first round) using the following principles:

[0217] Points on the Pareto performance frontier line, that is, Rank = 1. All points on Rank = 1 are selected. When the predicted performances of different composition - process points are the same, select the material composition and preparation process points with a larger difference in composition - process from other selected points.

[0218] Diversity of experimental point composition - process. Starting from Cr, Zr, T1, t1, T2, t2, select the parameters that do not appear on Rank = 1 in turn on Rank ≥ 2, based on the principle of prioritizing Rank and then the crowding distance CD, that is, preferentially select the point with a smaller Rank value. When the Rank values are the same, preferentially select the point with a larger CD value; select the point with fewer occurrences of the corresponding feature parameters and more features. For example, if all the values corresponding to the features of prediction point A do not appear in Rank = 1, point A is preferentially selected. [[ID=1,3]]

[0219] According to the principle, 8 points with Rank = 1 are selected, and a total of 24 experimental points outside Rank = 1, including 7 (Cr) + 3 (Zr) + 2 (T1) + 5 (t1) + 3 (t2) + 4 (diversity), form the 32 experimental points of P1-32.

[0220] S407.2: As Figures 26 - 27As shown, select the P2-16 experimental points in the target area. Select points according to the previous principle. First, select the points on the Pareto front, and then give priority to ensuring the diversity of points in the target area. Starting from Cr, Zr, T1, t1, T2, t2, select the parameters that do not appear on Rank = 1 on Rank ≥ 2 in turn. According to the principle of first Rank and then crowding distance CD, that is, give priority to selecting the ones with a small Rank value. When the Rank values are the same, give priority to selecting the ones with a large CD value. Finally, eliminate a point with a relatively small number of occurrences of a component-process parameter, and select 16 points as the experimental points of P2-16.

[0221] S407.3: As Figures 28 - 29 shown, all the sampling points of PP3-8 are on the Pareto front of the full-performance area, that is, on the front line of Rank = 1. It can be seen from the figure that there are no experimental points that meet the requirements in the target performance area for this round of predicted points. Although the diversity of characteristic parameters is taken into account during the sampling process, the predicted points still deviate from the target performance area due to the performance range of the target performance area.

[0222] During the sampling process of PP3-8, appropriately reduce the boundary value of the target performance area, and take around 85% of the two maximum performance values in the literature data (EC ≥ 78% IACS and HD ≥ 137 HV). The Pareto sampling strategy optimization selects 8 experimental points on the Pareto front for experiments in the last round. Select 4, 2, and 2 points (the internal yellow solid points in the figure) in the appropriately reduced target performance area, high-conductivity area, and high-hardness area respectively. The selection interval is the performance point whose absolute value of the conductivity of the predicted points in each area is closest to the conductivity equal division point in each area.

[0223] S408: Carry out the preparation of full-process high-throughput experimental samples for the alloy samples selected in step S407. The experimental preparation process includes: high-throughput automatic material powder preparation, high-throughput ball milling and powder mixing, multi-station electric press to press into a compact block, high-throughput automatic arc melting, high-throughput cold embedding or hot embedding, and high-throughput automatic sample grinding and polishing. Conduct eddy current conductivity measurement, high-throughput material hardness measurement, and high-throughput material composition determination on the experimental samples.

[0224] S409: Compare the experimental performance of the samples obtained in step S408 with the predicted performance of the samples in S406.

[0225] Conduct experiments on the above sampling experimental points, test the conductivity and hardness of the experimental points, and sort the experimental results of each round in descending order. The results are as Figures 30 - 31As shown. In terms of conductivity performance, most of the predicted values of P1-32 are lower than the experimental values in the high-conductivity region and higher than the experimental values in the low-conductivity region. The experimental values of P2-16 are higher than the predicted values, and the predicted values of PP3-8 are consistent with the experimental values in trend. In terms of hardness performance, the predicted values of P1-32 are all higher than the experimental values, and the differences at most points are relatively large. The predicted values of P2-16 are also higher than the experimental values. By the time of PP3-8, the predicted values are already consistent with the experimental values in trend. The optimized sampling strategy based on the target performance region has been improved in terms of the prediction accuracy of the model after two rounds of experimental iterations. In terms of conductivity, the errors of the three rounds of experiments gradually decrease. The MAE and MRE of PP3-8 are reduced by 47.61% and 47.33% respectively compared with P1-32. In terms of hardness, the MAE and MRE of PP3-8 are reduced by 83.05% and 87.11% respectively compared with P1-32.

[0226] Summarize the experimental points of each round and draw the scatter distribution of the HD-EC performance of the experimental points of each round, as Figure 32 shown. It can be seen from the figure that the points of P1-32 are mainly concentrated in the high-conductivity and low-hardness region. The experimental points of P2-16 rise towards the high-hardness direction, but overall they are still in the high-conductivity and medium-hardness region. Four experimental points of PP3-8 meet the conditions of the initially set target performance region. The experimental points of PP3-8 are near the front of the target performance region.

[0227] The method of the present application has the following advantages: (1) Most of the current technical methods in this field are limited to sequential manual experiments of single samples, single machine learning regression prediction models, single machine learning sampling strategies and optimization of single material target properties. Although the efficiency of material experiments has been improved compared to traditional empirical trial and error, there are still many problems such as small amount of experimental data, low quality of experimental data, unclear convergence of machine learning algorithms, material performance optimization falling into local minimum, difficulty in balancing multiple comprehensive material properties and low overall efficiency of full-chain research and development. The method of the present application provides a material performance prediction method (HT-MID) based on multiple machine learning regression models and high-throughput experiments, which can efficiently and accurately design and implement as few experiments as possible in a huge material candidate parameter space to achieve the target performance of new materials. In particular, it solves the dependence of machine learning models on incremental experimental data and dynamic optimization problems, and accelerates the full-chain research and development of materials efficiently and at low cost. (2) Breaking through the bottleneck of slow and expensive experimental R&D of traditional materials, the high-throughput intelligent material design (HT-MID) method combines high-throughput experimental acceleration (more than 10-100 times) and intelligent design acceleration (more than 100-1000 times) to improve the overall R&D efficiency of materials by a thousand to a hundred thousand times, and the R&D cost of a single material is greatly reduced (1 / 2-1 / 10). The application of the high-throughput intelligent material design (HT-MID) method can greatly accelerate the design and development of new materials, thereby providing important basic material guarantees for the high-quality development of advanced manufacturing. (3) Although this patent uses alloy design as an implementation case, the proposed design method is in principle applicable to the R&D process of any material and has wide applicability and scalability.

[0228] In an exemplary embodiment, a computer device is provided, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a material property prediction method based on multiple machine learning regression models and high-throughput experiments.

[0229] In an exemplary embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, a material property prediction method based on multiple machine learning regression models and high-throughput experiments is implemented.

[0230] In an exemplary embodiment, a computer program product is provided, comprising a computer program that, when executed by a processor, implements a material property prediction method based on multiple machine learning regression models and high-throughput experiments.

[0231] In an exemplary embodiment, an application of a material property prediction method based on a multiple machine learning regression model and high-throughput experiments is provided. It is implemented based on any of the above-mentioned material property prediction methods based on a multiple machine learning regression model and high-throughput experiments. The application of the material property prediction method based on a multiple machine learning regression model and high-throughput experiments includes:

[0232] Obtain a plurality of material parameters to be screened.

[0233] Determine any one of the material parameters to be screened as the current material parameter, and determine any one of the properties as the current property.

[0234] Based on the material composition in the current material parameter, determine the target key features of the current property under the current material parameter.

[0235] Input the process in the current material parameter and the target key features of the current property under the current material parameter into the multiple optimal performance prediction regression models corresponding to the current property respectively, and obtain a plurality of sub-prediction values of the current property under the current material parameter.

[0236] Based on the plurality of sub-prediction values of the current property under the current material parameter, determine the total prediction value of the current property under the current material parameter.

[0237] Determine the material parameter to be screened with the optimal total prediction value as the optimal material parameter to achieve the optimization of the material parameter.

[0238] In an exemplary embodiment, a computer device is provided. The computer device can be a server or a terminal, and its internal structure diagram can be as Figure 33 shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O), and a communication interface. Among them, the processor, the memory, and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The input / output interface of the computer device is used for the processor to exchange information with external devices. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, it implements a material property prediction method based on a multiple machine learning regression model and high-throughput experiments.

[0239] Those skilled in the art can understand, Figure 33The structure shown is only a block diagram of some structures related to the solution of this application, and does not constitute a limitation on the computer device to which the solution of this application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine some components, or have different component arrangements.

[0240] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, database, or other medium used in the embodiments provided in this application can include at least one of non-volatile and volatile memories. Non-volatile memories can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memories can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0241] The databases involved in the embodiments provided in this application can include at least one of relational databases and non-relational databases. Non-relational databases can include distributed databases based on blockchain, etc., and are not limited thereto. The processors involved in the embodiments provided in this application can be general-purpose processors, central processors, graphics processors, digital signal processors, programmable logics, data processing logics based on quantum computing, etc., and are not limited thereto.

[0242] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use, and processing of relevant data need to comply with relevant regulations.

[0243] The technical features of the above embodiments can be combined arbitrarily. For the sake of concise description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope described in this specification.

[0244] Specific examples are used in this article to elaborate on the principles and implementation manners of the present application. The description of the above embodiments is only used to help understand the method and its core idea of the present application; at the same time, for those of ordinary skill in the art, according to the idea of the present application, there will be changes in the specific implementation manners and application scopes. In summary, the content of this specification should not be construed as a limitation to the present application.

Claims

1. A material property prediction method based on a multiple machine learning regression model and high-throughput experiments, characterized in that The material property prediction method based on a multiple machine learning regression model and high-throughput experiments includes: Obtain preset material parameters; the material parameters include: material composition and preparation process; Determine any property as the current property; Based on the material composition in the preset material parameters, determine the target key features of the current property under the preset material parameters; Input the process in the preset material parameters and the target key features of the current property under the preset material parameters into multiple optimal property prediction regression models corresponding to the current property respectively, and obtain multiple sub-prediction values of the current property under the preset material parameters; each optimal property prediction regression model corresponding to any property is obtained by screening, iterative training and high-throughput experiments on multiple machine learning regression models, and the multiple machine learning regression models include: random forest regression model, backpropagation neural network model, extreme gradient boosting decision tree regression model and support vector machine regression model; Based on the multiple sub-prediction values of the current property under the preset material parameters, determine the total prediction value of the current property under the preset material parameters.

2. The material property prediction method based on a multiple machine learning regression model and high-throughput experiments according to claim 1, characterized in that, The determination process of multiple optimal property prediction regression models for any current property includes: Initialize the first data set to obtain the initial first data set; the first data set includes: historical data of multiple sample materials; the historical data includes historical material parameters and actual values of the current property; the sample materials are materials that have been prepared; Based on the material composition in the historical material parameters of each sample material in the initial first data set, determine the initial features to be screened of the current property; the features to be screened include extended features and composite features; Use statistical analysis methods to perform hierarchical feature screening on the initial features to be screened of the current property to obtain the initial key features of the current property; Determine the process in the historical material parameters of each sample material in the initial first data set, the initial key features of the current property under the historical material parameters of each sample material, and the actual values of the current property of each sample material as the initial second data set; Based on the initial second data set and multiple machine learning regression models, determine the initial first property prediction regression model of the current property; Interpolate all the historical material parameters in the initial first data set, and based on the interpolated components, determine the interpolated key features corresponding to the current property, and determine all the interpolated processes and interpolated key features as the initial material feature parameter set; Through the Bootstrap method, perform multiple sampling with replacement on the initial second data set to obtain multiple initial third data sets; Based on each initial third data set and the initial first property prediction regression model respectively, determine multiple initial second property prediction regression models of the current property; Perform iterative training and optimization on the initial first data set, initial key features, initial second data set, initial first property prediction regression model, initial material feature parameter set, multiple initial third data sets and multiple initial second property prediction regression models to obtain multiple optimal property prediction regression models corresponding to the current property; among them, use the first training process or the second training process for iterative training and optimization.

3. The material property prediction method based on a multiple machine learning regression model and high-throughput experiments according to claim 2, characterized in that, Based on the initial second dataset and multiple machine learning regression models, determining the initial first performance prediction regression model of the current performance, including: Dividing the initial second dataset to obtain an initial first training set and an initial first test set; Using the initial first training set to train each machine learning regression model respectively to obtain multiple initially trained machine learning regression models; Using the initial first test set to evaluate each initially trained machine learning regression model respectively, and determining the initially trained machine learning regression model that meets the evaluation conditions as the initial first performance prediction regression model; the evaluation conditions are: the minimum average root mean square error or the highest average correlation coefficient and average determination coefficient.

4. The material property prediction method based on multiple machine learning regression models and high-throughput experiments according to claim 2, wherein The first training process includes: Obtaining the first dataset at the current iteration; Based on the material composition in the historical material parameters of each sample material in the first dataset at the current iteration, determining the features to be screened at the current iteration of the current performance; Using statistical analysis methods to perform hierarchical feature screening on the features to be screened at the current iteration of the current performance to obtain the key features at the current iteration of the current performance; Determining the process in the historical material parameters of each sample material in the first dataset at the current iteration, the initial key features of the current performance under the historical material parameters of each sample material, and the actual values of the current performance of each sample material as the second dataset at the current iteration; Based on the second dataset at the current iteration and multiple machine learning regression models, determining the first performance prediction regression model of the current performance at the current iteration; Interpolating all the historical material parameters in the first dataset at the current iteration, and based on the interpolated components, determining the interpolated key features corresponding to the current performance, and determining all the interpolated processes and interpolated key features as the material feature parameter set at the current iteration; Through the Bootstrap method, performing multiple sampling with replacement on the second dataset at the current iteration to obtain multiple third datasets at the current iteration; Based on each third dataset at the current iteration and the first performance prediction regression model at the current iteration respectively, determining multiple second performance prediction regression models of the current performance at the current iteration; Inputting each group of the interpolated processes and interpolated key features in the material feature parameter set at the current iteration into each second performance prediction regression model at the current iteration respectively to obtain multiple sub-prediction values of the current performance under the corresponding interpolated material parameters at the current iteration; Based on all the prediction values of the current performance under each interpolated material parameter at the current iteration respectively, determining the total prediction value of the current performance under each interpolated material parameter at the current iteration; Adopting a multiple machine learning sampling strategy to sample each third dataset at the current iteration to obtain a fourth dataset at the current iteration; Prepare based on the interpolated process in the fourth dataset and the interpolated material parameters corresponding to the interpolated key features at the current iteration number to obtain multiple material samples at the current iteration number, and perform performance tests on all material samples at the current iteration number to obtain the test values of the current performance of each material sample at the current iteration number; Based on the test values of the current performance of each material sample at the current iteration number and the total predicted values of the current iteration number of the current performance under the corresponding interpolated material parameters, determine the average root mean square error, average correlation coefficient, and average determination coefficient at the current iteration number; Judge whether the stop condition is satisfied; the stop condition is: the average root mean square error at the current iteration number is the smallest, or the average correlation coefficient and average determination coefficient at the current iteration number are the highest, or the preset training number is reached; If so, determine the multiple second performance prediction regression models of the current performance at the current iteration number as the multiple optimal performance prediction regression models of the current performance, and determine the key features at the current iteration number of the current performance as the target key features; If not, update the current iteration number to the next iteration number and return "Obtain the first dataset at the current iteration number".

5. The material property prediction method based on a multiple machine learning regression model and high-throughput experiments according to claim 4, wherein Based on the second dataset at the current iteration number and multiple machine learning regression models, determine the first performance prediction regression model of the current performance at the current iteration number, including: Divide the second dataset at the current iteration number to obtain the first training set at the current iteration number and the first test set at the current iteration number; Use the first training set at the current iteration number to train each machine learning regression model respectively to obtain multiple trained machine learning regression models at the current iteration number; Use the first test set at the current iteration number to evaluate each trained machine learning regression model at the current iteration number, and determine the trained machine learning regression model at the current iteration number that meets the evaluation conditions as the first performance prediction regression model at the current iteration number.

6. The material property prediction method based on multiple machine learning regression models and high-throughput experiments according to claim 2, wherein The second training process includes: Determine the initial key features of the current performance as the target key features of the current performance; Interpolate all historical material parameters in the initial first dataset, and based on each interpolated component, determine the interpolated key features corresponding to the current performance, and determine all interpolated processes and interpolated key features as the material feature parameter set at the current iteration number; Through the Bootstrap method, perform multiple sampling with replacement on the initial second dataset to obtain multiple third datasets at the current iteration number; Based on each third dataset at the current iteration number and the initial first performance prediction regression model respectively, determine multiple second performance prediction regression models of the current performance at the current iteration number; Input each group of interpolated processes and interpolated key features in the material feature parameter set at the current iteration number into each second performance prediction regression model at the current iteration number to obtain multiple sub-predicted values of the current performance under the corresponding interpolated material parameters at the current iteration number; Based on the predicted values at all current iteration counts under each interpolated material parameter respectively based on the current performance, determine the total predicted value of the current performance at the current iteration count under each interpolated material parameter; Adopt a multiple machine learning sampling strategy to sample the third dataset at each current iteration count to obtain the fourth dataset at the current iteration count; Prepare based on the interpolated process and the interpolated material parameters corresponding to the interpolated key features in the fourth dataset at the current iteration count to obtain multiple material samples at the current iteration count, and perform performance tests on all material samples at the current iteration count to obtain the test values of the current performance of each material sample at the current iteration count; Based on the test values of the current performance of each material sample at the current iteration count and the total predicted value of the current performance at the current iteration count under the corresponding interpolated material parameters, determine the average root mean square error, average correlation coefficient, and average determination coefficient at the current iteration count; Judge whether the stop condition is satisfied; the stop condition is: the average root mean square error at the current iteration count is the smallest, or the average correlation coefficient and average determination coefficient at the current iteration count are the highest, or the preset training count is reached; If so, determine the multiple second performance prediction regression models of the current performance as the multiple optimal performance prediction regression models of the current performance; If not, update the current iteration count to the next iteration count, and return "Interpolate all historical material parameters in the initial first dataset, and based on each interpolated component, determine the interpolated key features corresponding to the current performance, and determine the material feature parameter set at the current iteration count with all the interpolated processes and interpolated key features".

7. A computer device, comprising: A memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor executes the computer program to implement the material performance prediction method based on a multiple machine learning regression model and high-throughput experiments according to any one of claims 1-6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the material performance prediction method based on a multiple machine learning regression model and high-throughput experiments according to any one of claims 1-6.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the material performance prediction method based on a multiple machine learning regression model and high-throughput experiments according to any one of claims 1-6.

10. Application of a method for predicting material properties based on a multiple machine learning regression model and high-throughput experiments, implemented based on the method for predicting material properties based on a multiple machine learning regression model and high-throughput experiments according to any one of claims 1-6, characterized in that The application of the material performance prediction method based on a multiple machine learning regression model and high-throughput experiments includes: Obtain multiple material parameters to be screened; Determine any one of the material parameters to be screened as the current material parameter, and determine any one of the performances as the current performance; Based on the material composition in the current material parameter, determine the target key features of the current performance under the current material parameter; Input the process in the current material parameter and the target key features of the current performance under the current material parameter into the multiple optimal performance prediction regression models corresponding to the current performance respectively to obtain multiple sub-predicted values of the current performance under the current material parameter; Based on the multiple sub-predicted values of the current performance under the current material parameter, determine the total predicted value of the current performance under the current material parameter; Determine the material parameter to be screened with the optimal total predicted value as the optimal material parameter to achieve the optimization of the material parameter.

Images