Material performance prediction model training method and device, equipment and storage medium

Abstract

The application discloses a material performance prediction model training method and device, equipment and a storage medium. Model training is performed based on a first candidate crystal structure and a performance parameter label corresponding to the first candidate crystal structure to obtain a first material performance prediction base model. Then, a prediction result corresponding to a second candidate crystal structure is determined based on the second candidate crystal structure. If the prediction result corresponding to the second candidate crystal structure does not satisfy a first preset condition, a derived crystal structure of the second candidate crystal structure is generated, and model training is performed based on this to obtain a second material performance prediction base model. A target material performance prediction base model is determined. The target material performance prediction base model is adjusted based on a second sample material to obtain a target material performance prediction model. In this way, the prediction efficiency of material performance is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of model training, in particular to a material performance prediction model training method and device, equipment and a storage medium. BACKGROUND

[0002] In recent years, with the deepening of the research on five-element and more-element oxide systems, the design of complex component catalysts has gradually become a new direction. Under this background, high-entropy oxides have rapidly become a research hotspot for multi-element synergistic catalytic systems.

[0003] In related technologies, a series of experiments are performed to test different combinations of variables to observe which combinations can produce the best performance. For example, in the research on high-entropy oxide catalysts, different combinations of metal elements and their proportions are tested. After each experiment, the combination of metal elements and their proportions is adjusted according to the experimental results, and the above process is repeated until satisfactory results are obtained. However, although the traditional trial-and-error method is intuitive and easy to implement, when faced with complex material systems such as high-entropy oxides, it is almost impossible to verify each combination one by one, and the efficiency is extremely low. Moreover, each experiment requires the preparation of new samples and comprehensive characterization and testing, which is time-consuming and labor-intensive, resulting in a sharp increase in costs. Moreover, experiments are usually designed based on the experience and intuition of researchers, which may limit the scope of exploration and miss the global optimal solution. Even when multiple variables such as element types, proportions, synthesis conditions, etc. are involved, the traditional trial-and-error method is difficult to effectively manage the interactions between these variables. SUMMARY

[0004] Therefore, the present application provides a material performance prediction model training method and device, equipment and a storage medium to solve the problem of low efficiency in predicting material performance in the prior art.

[0005] The purpose of the present application can be achieved by the following technical solutions:

[0006] The first aspect of the present application provides a material performance prediction model training method, comprising:

[0007] Generating a candidate crystal structure set based on the category and structure feature parameters corresponding to the first sample material, the candidate crystal structure set including a first candidate crystal structure set and a second candidate crystal structure set, the first candidate crystal structure set including a plurality of first candidate crystal structures, and the second candidate crystal structure set including a plurality of second candidate crystal structures;

[0008] Model training based on the first candidate crystal structure and the performance parameter label corresponding to the first candidate crystal structure, to obtain a first material performance prediction base model;

[0009] inputting the second candidate crystal structure into the first material performance prediction base model to obtain a predicted result corresponding to the second candidate crystal structure;

[0010] determining a first DFT result corresponding to the second candidate crystal structure;

[0011] if the predicted result corresponding to the second candidate crystal structure and the first DFT result do not satisfy a first preset condition, generating derivative crystal structures of the second candidate crystal structure under different spatial group symmetries;

[0012] training the first material performance prediction base model based on the derivative crystal structures and the performance parameter tags corresponding to the derivative crystal structures to obtain a second material performance prediction base model;

[0013] determining the second material performance prediction base model corresponding to the second candidate crystal structure with a similarity to the crystal structure of the second sample material exceeding a preset similarity threshold as a target material performance prediction base model;

[0014] performing parameter adjustment on the target material performance prediction base model based on the second sample material to obtain a target material performance prediction model.

[0015] In an optional embodiment, the method further comprises:

[0016] performing screening processing on the second sample material according to a preset element combination rule to obtain a third sample material;

[0017] performing optimization on the crystal structure of the third sample material by using an adaptive genetic algorithm to obtain a fourth sample material;

[0018] performing sampling on the fourth sample material by using parallelized Monte Carlo tree search to obtain a fifth sample material;

[0019] performing parameter adjustment on the target material performance prediction base model based on the fifth sample material to obtain a target material performance prediction model.

[0020] In an optional embodiment, the method further comprises:

[0021] inputting the material to be predicted into the target material performance prediction model to obtain a performance prediction result corresponding to the material to be predicted;

[0022] training the target material performance prediction model based on the material to be predicted and the performance prediction result corresponding to the material to be predicted satisfying a second preset condition.

[0023] In an optional embodiment, the method further comprises:

[0024] performing screening on the first sample material by using spatial clustering and structure description information.

[0025] In an optional embodiment, if the predicted result corresponding to the second candidate crystal structure does not satisfy the first preset condition with the first DFT result, a derivative crystal structure of the second candidate crystal structure under different space group symmetries is generated, including:

[0026] The uncertainty index corresponding to the second candidate crystal structure is determined.

[0027] If the predicted result corresponding to the second candidate crystal structure is less than the preset tolerance threshold with the first DFT result, the first material performance prediction base model is trained based on the second candidate crystal structure.

[0028] If the predicted result corresponding to the second candidate crystal structure is not less than the preset tolerance threshold with the first DFT result, and / or, the uncertainty index corresponding to the second candidate crystal structure is higher than the preset critical value, the DFT recalculation is performed on the second candidate crystal structure to obtain a second DFT result.

[0029] If it is determined that there is a systematic prediction bias based on the predicted result corresponding to the second candidate crystal structure and the second DFT result, a derivative crystal structure of the second candidate crystal structure under different space group symmetries is generated.

[0030] In an optional embodiment, the target material performance prediction model is obtained by adjusting parameters of the target material performance prediction base model based on the second sample material, including:

[0031] The top layer regression network of the target material performance prediction base model is adjusted by the layer selection unfreezing strategy based on the second sample material, and the target material performance prediction model is obtained.

[0032] In an optional embodiment, further comprising:

[0033] The sample material calculated based on the hybrid functional or the GW method is obtained as a sixth sample material.

[0034] The target material performance prediction model is obtained by adjusting parameters of the target material performance prediction base model based on the sixth sample material.

[0035] The second aspect of the present application provides a material performance prediction model training device, including:

[0036] The first generation module is configured to generate a candidate crystal structure set based on the category and structure feature parameters corresponding to the first sample material, the candidate crystal structure set including a first candidate crystal structure set and a second candidate crystal structure set, the first candidate crystal structure set including a plurality of first candidate crystal structures, and the second candidate crystal structure set including a plurality of second candidate crystal structures.

[0037] The first training module is configured to perform model training based on the first candidate crystal structure and the performance parameter label corresponding to the first candidate crystal structure, to obtain a first material performance prediction base model;

[0038] The input module is configured to input the second candidate crystal structure into the first material performance prediction base model, to obtain a predicted result corresponding to the second candidate crystal structure;

[0039] The first determination module is configured to determine a first DFT result corresponding to the second candidate crystal structure;

[0040] The second generation module is configured to, if the predicted result corresponding to the second candidate crystal structure and the first DFT result do not satisfy a first preset condition, generate a derivative crystal structure of the second candidate crystal structure under different spatial group symmetries;

[0041] The second training module is configured to perform training on the first material performance prediction base model based on the derivative crystal structure and the performance parameter label corresponding to the derivative crystal structure, to obtain a second material performance prediction base model;

[0042] The second determination module is configured to determine, as a target material performance prediction base model, the second material performance prediction base model corresponding to the second candidate crystal structure whose similarity to the crystal structure of the second sample material exceeds a preset similarity threshold value;

[0043] The adjustment module is configured to perform parameter adjustment on the target material performance prediction base model based on the second sample material, to obtain a target material performance prediction model.

[0044] The third aspect of the present application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor executes the computer program to implement the method of the first aspect.

[0045] The fourth aspect of the present application provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the method of the first aspect.

[0046] Compared with the prior art, the material performance prediction model training method provided in the application generates a candidate crystal structure set based on the category and structure feature parameters corresponding to the first sample material; model training is performed based on the first candidate crystal structure and the performance parameter label corresponding to the first candidate crystal structure, to obtain a first material performance prediction base model, and then the prediction result corresponding to the second candidate crystal structure is determined based on the second candidate crystal structure; if the prediction result corresponding to the second candidate crystal structure does not satisfy the first preset condition, a derived crystal structure of the second candidate crystal structure is generated, and model training is performed based on this to obtain a second material performance prediction base model; a target material performance prediction base model is determined; and the target material performance prediction model is obtained by performing parameter adjustment on the target material performance prediction base model based on the second sample material. In this way, the prediction efficiency of material performance is improved. BRIEF DESCRIPTION OF DRAWINGS

[0047] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0048] Figure 1 A flowchart of a material performance prediction model training method provided by an embodiment of the present application;

[0049] Figure 2 Another flowchart of a material performance prediction model training method provided by an embodiment of the present application;

[0050] Figure 3 A structural block diagram of a material performance prediction model training device provided by an embodiment of the present application;

[0051] Figure 4 A structural block diagram of an electronic device for implementing a material performance prediction model training method provided by an embodiment of the present application. DETAILED DESCRIPTION

[0052] In order to enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should be within the scope of protection of the present application.

[0053] It should be noted that the terms "first", "second" and the like in the description and in the claims of the present application and the above-mentioned accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances so that the embodiments of the present application described herein can be implemented in other than the order illustrated or described herein. In addition, the terms "comprise" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device that includes a list of steps or units as an element does not have to be limited to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to such processes, methods, products or devices.

[0054] It should be understood that in the embodiments of the present application, "at least one" means one or more, and "multiple" means two or more. "And / or" is only a description of the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B can mean that A exists alone, A and B exist together, and B exists alone. The character " / " generally represents that the associated objects before and after it are in an "or" relationship. "Including A, B and / or C" means including any one or any two or three of A, B and C.

[0055] It should be understood that in the embodiments of the present application, "B corresponding to A", "B corresponding to A", "A corresponding to B" or "B corresponding to A" means that B is associated with A, and B can be determined according to A. Determining B according to A does not mean that B is determined only according to A, but also can be determined according to A and / or other information.

[0056] To solve the technical problems existing in the related art, the embodiments of the present application provide a material performance prediction model training method, device, equipment and storage medium.

[0057] The material performance prediction model training method provided by the embodiments of the present application can be executed by an electronic device, which can be a terminal or a server and the like. The terminal can be a terminal device such as a smart phone, a tablet computer, a notebook computer and the like. The server can be a physical server, a server cluster or a distributed system composed of multiple physical servers, a cloud server providing cloud services, cloud databases, cloud computing, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and basic cloud computing services such as big data and artificial intelligence platforms. It can be understood that the present application does not limit the specific subject executing the material performance prediction model training method.

[0058] Before describing the solution in this application in detail, the terms used in this application will be explained:

[0059] High-entropy oxides: These are oxides with a configurational entropy greater than or equal to 1.5R. The specific calculation formula is as follows:

[0060]

[0061] Where S is the mixture configuration entropy, and N is the number of element types. Let be the mole fraction of the i-th component, and R be the universal gas constant.

[0062] GNN (Graph Neural Network): refers to a general term for algorithms that use graph neural networks to learn graph-structured data, extract and discover features and patterns in graph-structured data, and meet the needs of graph learning tasks such as clustering, classification, prediction, segmentation, and generation.

[0063] First principles: In physics, first principles, or ab initio calculations, refer to derivations and calculations that start from fundamental physical laws without adding assumptions or empirical fitting. For example, using the Schrödinger equation to solve for electronic structure using some approximation methods, without fitting the model to experimental data, is a kind of ab initio calculation method.

[0064] High-throughput screening: High-throughput screening is a technical strategy that uses automated and parallel experimental or computational methods to rapidly test and analyze massive amounts of candidate samples in a short time. Its core objective is to efficiently identify materials, compounds, or biomolecules that meet specific performance requirements from a large candidate library.

[0065] The technical solution of this application will be described in detail below through specific embodiments. It should be noted that the following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments described below are used to explain the technical solution of this application and are not intended to limit actual use.

[0066] To address the technical problems existing in related technologies, embodiments of this application provide a method for training a material property prediction model, such as... Figure 1 As shown, Figure 1 This is an example flowchart of a material property prediction model training method provided in an embodiment of this application. It should be noted that the steps shown may be executed in a different logical order than that shown in the flowchart. The method may include the following steps S101 to S108.

[0067] Step S101: Generate a set of candidate crystal structures based on the category and structural feature parameters corresponding to the first sample material.

[0068] It should be noted that the candidate crystal structure set includes a first candidate crystal structure set and a second candidate crystal structure set, the first candidate crystal structure set includes a plurality of first candidate crystal structures, and the second candidate crystal structure set includes a plurality of second candidate crystal structures.

[0069] In an optional embodiment, the first sample material can be a metal oxide, a perovskite material, a two-dimensional material, an alloy material, an organic-inorganic hybrid material, a semiconductor material, a nanocomposite material, and a superconducting material, etc., wherein the metal oxide can be a high-entropy oxide. The present application does not limit the specific first sample material.

[0070] In an optional embodiment, the category of the first sample material can include at least one of the following categories: a space group category, a crystal system category, a point group category, a coordination number category, a Wyckoff position category, a thermodynamic stability category, and a functional property category, etc. The present application does not limit this.

[0071] In a specific embodiment, when the first sample material is a high-entropy oxide, the structural characteristic parameters of the first sample material include at least one of the following data: composition, crystal structure, electrical conductivity, catalytic activity, thermal stability, formation energy, band structure, and oxygen vacancy formation energy of the high-entropy oxide, etc. The present application does not limit this. It should be noted that the composition, structure, electrical conductivity, catalytic activity, and thermal stability are obtained by experimental measurement or theoretical calculation prediction, and the formation energy, band structure, and oxygen vacancy formation energy are usually obtained by DFT (first principle, Density Functional Theory) calculation.

[0072] In a more specific embodiment, the first sample material can be obtained from an open source database such as Materials Project.

[0073] In another optional embodiment, before generating the candidate crystal structure set based on the category and the structural characteristic parameters corresponding to the first sample material, the symmetry feature of the first sample material can be extracted; the symmetry feature of the first sample material is input into a preset classification prediction model to obtain the coordination environment category of the doping atom of the first sample material, wherein the symmetry feature can include at least one of the following features: space group number, Wyckoff position symmetry level, coordination polyhedron type, coordination number of the doping atom, and bond length standard deviation. In this way, the candidate crystal structure set satisfying the preset structure constraint condition can be generated according to the predicted category subsequently.

[0074] It should be noted that the training process of the classification prediction model: based on the class label and the structure feature parameter corresponding to the sample material in the training set corresponding to the classification prediction model, the model training is carried out, and the classification prediction model is obtained. The specific training process is similar to the training process of the classification prediction model in the related art, and the present application does not make a detailed description here.

[0075] In an alternative embodiment, the classification prediction model can be a graph neural network.

[0076] In an alternative embodiment, the preset structure constraint condition includes at least one of the following conditions: point group symmetry operation, crystal system feature and Wyckoff position.

[0077] In this step, by generating a candidate crystal structure set satisfying the preset structure constraint condition, the invalid sampling problem existing in the traditional random sampling method is effectively avoided, the structure search space dimension is greatly reduced, and the efficiency and accuracy of subsequent machine learning training are improved.

[0078] In another alternative embodiment, the material performance prediction model training method provided by the embodiments of the present application further includes the following steps:

[0079] The first sample material is screened by using the spatial cluster and the structure description information.

[0080] In a specific embodiment, the structure description information includes at least one of the following information: Wyckoff position information, atomic coordinates, atomic coordination environment, point group and structure motif.

[0081] In a specific embodiment, the candidate crystal structures are classified according to the space group number, and all candidate crystal structures belonging to the same space group are classified into one category. Wyckoff position matching: within the same space group, the Wyckoff position configuration of each candidate crystal structure is further checked. If the Wyckoff positions of two candidate crystal structures are the same or only simple translation, rotation and other symmetry operations exist, they are considered redundant. For the redundant candidate crystal structures found, one of them is retained as a representative, and the rest is removed. By removing the structural redundancy samples through symmetry analysis, the structural representativeness and distribution rationality of the training samples are significantly improved.

[0082] Step S102: based on the first candidate crystal structure and the performance parameter label corresponding to the first candidate crystal structure, model training is performed to obtain a first material performance prediction base model.

[0083] It should be noted that the material performance prediction base model is a model for predicting the performance of a crystal structure based on the crystal structure.

[0084] In an optional embodiment, the performance parameter comprises at least one of the following parameters: mixing enthalpy and formation energy, etc.

[0085] In an optional embodiment, the material performance prediction base model can be a graph neural network.

[0086] In an optional embodiment, during the model training process, statistical analysis can also be performed on the prediction results, so as to determine the probability distribution range of the target parameter. It should be noted that the target parameter is at least one of the key thermodynamic parameters such as mixing enthalpy and formation energy. By determining the probability distribution range of the target parameter, the uncertainty of the model prediction result can be understood, and then it is determined whether to further study the material; new materials with ideal thermodynamic properties can also be screened out; by analyzing the probability distribution range of the target parameter, the area with large model prediction error can be identified, and the data samples in this area can be increased accordingly, so as to improve the overall performance of the model.

[0087] Step S103: inputting the second candidate crystal structure into the first material performance prediction base model to obtain a prediction result corresponding to the second candidate crystal structure.

[0088] Step S104: determining a first DFT result corresponding to the second candidate crystal structure.

[0089] It should be noted that the DFT result refers to various physical properties obtained after performing electronic structure calculation by using DFT.

[0090] In an optional embodiment, the DFT result comprises at least one of the following data: total energy, band gap, state density, charge density distribution and stress tensor.

[0091] Step S105: if the prediction result corresponding to the second candidate crystal structure and the first DFT result do not satisfy the first preset condition, generating derivative crystal structures of the second candidate crystal structure under different spatial group symmetries.

[0092] In an optional embodiment, if the predicted result corresponding to the second candidate crystal structure does not satisfy the first preset condition with the first DFT result, a derivative crystal structure of the second candidate crystal structure under different space group symmetries is generated, including the following steps: determining an uncertainty index corresponding to the second candidate crystal structure; if the predicted result corresponding to the second candidate crystal structure is less than a preset tolerance threshold from the first DFT result, training the first material performance prediction based model based on the second candidate crystal structure; if the predicted result corresponding to the second candidate crystal structure is not less than the preset tolerance threshold from the first DFT result, and / or, the uncertainty index corresponding to the second candidate crystal structure is higher than a preset critical value, performing DFT recalculation on the second candidate crystal structure to obtain a second DFT result; if it is determined that there is a systematic prediction bias based on the predicted result corresponding to the second candidate crystal structure and the second DFT result, generating a derivative crystal structure of the second candidate crystal structure under different space group symmetries.

[0093] It should be noted that the uncertainty index is used to quantify the uncertainty in the predicted result, so as to ensure that the understanding of the model output is more comprehensive and accurate.

[0094] In an optional embodiment, the uncertainty includes at least one of the following uncertainties: statistical uncertainty, model uncertainty, data uncertainty, and algorithmic uncertainty.

[0095] Among them, the statistical uncertainty includes standard error, confidence interval, variance and standard deviation. The model uncertainty includes Bayesian uncertainty, model bias and variance. The data uncertainty includes noise, missing data and outliers.

[0096] The preset tolerance threshold and the preset critical value can be set according to actual conditions, and the present application does not limit this. For example, the preset tolerance threshold is RMSE (Root Mean Square Error) ≤ 20 meV (milli-electron volts) / atom.

[0097] The systematic prediction bias refers to the prediction error continuously exhibited by the model under certain conditions or for certain types of input data. This error is usually related to the limitations of the model, the bias of the training data, or the shortcomings in feature engineering.

[0098] In another optional embodiment, a full space group DFT calculation is performed on the derivative crystal structure to obtain a performance corresponding to the derivative crystal structure.

[0099] In another alternative embodiment, sample weights can also be applied to the sample data corresponding to the derived crystal structure, so that the model's root mean square error in predicting the crystal structure converges stably below 10 meV / atom. This achieves targeted allocation of computing resources and targeted strengthening of weak points in the model, improving training efficiency by more than 40% compared to traditional methods.

[0100] Step S106: Training the first material performance prediction base model based on the derived crystal structure and the performance parameter label corresponding to the derived crystal structure, to obtain a second material performance prediction base model.

[0101] Step S107: Determining the second material performance prediction base model corresponding to the second candidate crystal structure with a similarity to the crystal structure of the second sample material exceeding a preset similarity threshold as a target material performance prediction base model.

[0102] In an alternative embodiment, the similarity of the crystal structure can be calculated by a cosine similarity formula, or can be calculated based on another similarity formula, which is not limited herein.

[0103] Step S108: Parameter adjustment of the target material performance prediction base model based on the second sample material to obtain a target material performance prediction model.

[0104] In an alternative embodiment, parameter adjustment of the target material performance prediction base model based on the second sample material to obtain a target material performance prediction model includes the following steps: parameter adjustment of the top layer regression network of the target material performance prediction base model by a layer-selective unfreezing strategy based on the second sample material to obtain a target material performance prediction model.

[0105] In this step, by using a model that has been pre-trained on a large-scale related task, the amount of labeled data required on the target task can be significantly reduced. Not only time is saved, but also consumption of computing resources is reduced.

[0106] In another alternative embodiment, the material performance prediction model training method provided by the embodiments of the present application further includes: obtaining a sample material calculated based on a hybrid functional or GW (Green's function and screened Coulomb interaction) method as a sixth sample material; and parameter adjustment of the target material performance prediction base model based on the sixth sample material to obtain a target material performance prediction model.

[0107] In this embodiment, in order to correct the systematic deviation of the DFT functional, such as the underestimation of the band gap and the overestimation of the formation energy caused by the PBE (Perdew-Burke-Ernzerhof) functional, a certain number of high-precision data sets based on hybrid functionals or GW calculations are selected to construct a residual learning module to learn the system deviation mapping function. Gradient clipping constraint (threshold can be 0.1, of course, it can also be other numerical values) and small batch training are used to realize fixed-point accurate optimization, so that the root mean square error (RMSE) of the band gap prediction of the target material is stably reduced to a preset numerical range, for example, below 28 meV. Compared with the traditional end-to-end training, this method saves 90% of high-precision calculation resources, and realizes the prediction error of the band gap-overpotential correlation of less than 5% in the design of perovskite electrocatalysts.

[0108] In another optional embodiment, the material performance prediction model training method provided in the embodiments of the present application further includes: performing screening processing on the second sample material according to a preset element combination rule to obtain a third sample material; optimizing the crystal structure of the third sample material by using an adaptive genetic algorithm to obtain a fourth sample material; sampling the fourth sample material by using a parallelized Monte Carlo tree search to obtain a fifth sample material; and adjusting parameters of the base model based on the fifth sample material to obtain a target material performance prediction model.

[0109] In a specific embodiment, the preset element combination rule includes but is not limited to: an atomic radius difference ≤ 15%, an electronegativity difference ≤ 0.4, and / or a valence electron number matching degree ≥ 80%. It should be noted that the above numerical values can be set according to actual needs, which are not limited by the present application.

[0110] In a specific embodiment, the parameter setting of the adaptive genetic algorithm can include: setting a crossover probability ≥ 0.65 and a mutation probability ≤ 0.05. It should be noted that the parameter setting of the adaptive genetic algorithm is self-defined according to the actual situation, which is not limited by the present application.

[0111] In a specific embodiment, the Monte Carlo tree search performs equilibrium state structure sampling > 1000 times per component. It should be noted that the above numerical values can be set according to actual needs, which are not limited by the present application.

[0112] In the above embodiment, the search space is compressed to less than 5% of the original space by the dual-algorithm cooperation.

[0113] In another optional embodiment, during the model training process, the training set can be divided into a 70% training set, a 15% validation set, and a 15% test set. The training set, the validation set, and the test set can also be divided in other proportions, which are not limited by the present application.

[0114] In another alternative embodiment, the model training can be performed using mean square error, cross-entropy loss or Huber loss.

[0115] In another alternative embodiment, the optimizer can be Adam (Adaptive Moment Estimation).

[0116] In another alternative embodiment, regularization methods such as Dropout, L2 regularization and early stopping can be used to effectively prevent model overfitting and improve its prediction accuracy on new data.

[0117] In another alternative embodiment, the k-fold cross-validation method is used to evaluate the model performance.

[0118] In another alternative embodiment, the material performance prediction model training method provided by the embodiments of the present application further comprises: inputting the to-be-predicted material into the target material performance prediction model to obtain the performance prediction result corresponding to the to-be-predicted material; and training the target material performance prediction model based on the to-be-predicted material and the performance prediction result corresponding to the to-be-predicted material that satisfy the preset condition.

[0119] In a specific embodiment, the optimized candidate crystal structure is input into the pre-trained performance prediction neural network, and the formation energy, band gap and target functional property prediction values are output synchronously to establish a comprehensive scoring system in multiple dimensions. After sorting the candidate materials in descending order of scores, the top 5% of the high-priority candidate set (Top-N, N≥50 and ≤10% of the total sample size) with the highest scores is selected and output to the experimental verification link. It should be noted that the above numerical values can be set according to actual needs, and the present application does not limit this.

[0120] In a more specific embodiment, the to-be-predicted materials that satisfy the preset condition are synthesized by solid phase reaction or chemical vapor deposition, and the performance of the to-be-predicted materials that satisfy the preset condition can also be tested to train the target material performance prediction model in the future.

[0121] In a more specific embodiment, before training the target material performance prediction model based on the to-be-predicted materials that satisfy the preset condition and the performance prediction results corresponding to the to-be-predicted materials, the prediction materials can also be redundantly merged. The specific process is described in the foregoing process, and the present application will not be repeated here.

[0122] In a more specific embodiment, a weight of a preset multiple is applied to the training sample material with a deviation between the measured performance and the prediction value less than a certain threshold, and an adversarial verification is used to identify the distribution deviation area of the training set / test set.

[0123] In a more specific embodiment, based on the Bayesian neural network output prediction variance, when the prediction variance is higher than the corresponding preset threshold value, it is determined that it is a high uncertainty region, and the structure data corresponding to the target crystal structure of the region is obtained by DFT, a plurality of derived structures are generated in the preset formation energy interval by using Latin hypercube sampling, the derived structures are added to the training set, and the target material performance prediction model is trained.

[0124] In the embodiments of the present application, the traditional method takes a long time to synthesize and characterize sample materials, usually several days to several weeks. The trained graph neural network can predict the performance (such as formation energy, ionic conductivity) of the candidate material within seconds without physical synthesis. And the candidate component combination of the sample material increases exponentially with the number of elements, which cannot be exhausted by the traditional method. Through enumerating the element permutation and combination space, the graph neural network can screen potential candidates in the full composition range. Experimental design is limited by researchers' prior knowledge, which may miss non-intuitive potential advantage materials. The graph neural network automatically learns the interatomic interaction (such as bond length, coordination environment, etc.), without the need for artificial design of complex features.

[0125] Moreover, the traditional experimental trial-and-error method produces a large number of experimental synthesis substandard samples (such as amorphous phase or mixed phase), resulting in waste of precious metals and energy, while the present scheme only needs to verify the candidates (such as the top 1%) predicted by the graph neural network, saving more than 90% of resources. The traditional experimental trial-and-error method relies on expensive characterization equipment (such as synchrotron radiation and spherical aberration electron microscopy), while the present scheme uses Active Learning (Active Learning) to preferentially synthesize samples that are most valuable for model improvement, avoiding redundant experiments.

[0126] Assuming that 1000 high-entropy oxide components are screened: the cost of the traditional experimental trial-and-error method is about 1000x2000 yuan (synthesis + characterization) = 20 million yuan; the screening cost of the present scheme is model training (1.5 million yuan) + experimental verification (10x2000 yuan) = 17 million yuan, which is lower in cost.

[0127] Graph structure modeling: The multi-element synergistic effect of high-entropy oxides is difficult to describe with a simple thermodynamic model, making it difficult for the traditional experimental trial-and-error method to predict and characterize fine structures. The present scheme represents the crystal of high-entropy oxide as an atomic node + chemical bond, directly encoding multi-body interactions, such as the influence of 5 kinds of metal atoms on oxygen coordination.

[0128] Dynamic feature learning: The theoretical guidance of the traditional experimental trial-and-error method often ignores the dynamic effects of defect formation or changes in local chemical environment. Through the Message Passing (Message Passing) mechanism, the present scheme can capture long-range lattice distortion and local electronic structure features, thereby achieving accurate description of special structure sites on the surface.

[0129] Multi-task learning capability: Traditional experimental trial-and-error methods usually focus on a single performance and are difficult to balance multiple index conflicts. The model of this solution can predict multiple performance indicators at the same time and find the optimal solution through frontier analysis, thereby supporting complex constraint conditions.

[0130] Pre-trained model: The data obtained by traditional experimental trial-and-error methods is isolated and scattered, and the synthesis conditions of different research groups differ, making it difficult to reuse the data. This solution uses large-scale data sets of ordinary oxides (such as binary / ternary oxides) to pre-train the graph neural network, and then uses model fine-tuning to improve small sample data of high-entropy oxides.

[0131] Atomic contribution analysis: Due to the complexity and uncontrollable factors of the experimental process, it is difficult to locate the specific reasons for failure in traditional experimental trial-and-error methods. This solution can visualize the impact of key atoms on performance through gradient backpropagation and track the key role of active sites in catalytic OER (Oxygen Evolution Reaction) reactions.

[0132] Structure sensitivity analysis: Traditional experimental trial-and-error methods are limited by characterization methods and are difficult to identify complex features such as superlattice structures in materials, leading to misjudgments of factors affecting catalytic performance. This solution can extract unique characteristics at higher scales and larger ranges in generated structures, thereby avoiding the "black box" dilemma of traditional experimental trial-and-error and the phenomenon of experimental irreproducibility caused by incorrect attribution.

[0133] Active learning closed loop: The linear process of experiment → characterization → analysis in traditional experimental trial-and-error methods cannot provide real-time feedback for optimization. This solution uses an initial model to predict candidate materials, uses experimental verification to feed back new data, and iteratively updates the model to reduce prediction errors, ultimately achieving dynamic adjustment of the search space, efficient search for element combinations with high experimental feasibility, and bridging the gap between theory and industrial batch manufacturing.

[0134] Experimental verification and closed-loop optimization are the core bridge connecting theoretical prediction and practical application. Even if the model prediction accuracy is high, if there is no experimental verification, the screening results will lose credibility; the feedback mechanism can continuously correct the model bias, forming a "prediction-verification-iteration" positive cycle. This step directly determines whether the entire technical route can be transformed from theoretical exploration to actual material discovery capability.

[0135] Automatically convert the prediction results of the graph neural network into high-throughput synthesis schemes, and realize the parallel preparation and characterization of hundreds of samples per day through automated equipment, greatly improving experimental efficiency. Use the real performance data measured by experiments to feed back the model, calibrate the prediction bias, and dynamically optimize the screening threshold to ensure the convergence of model prediction and experimental results. Combine domain knowledge, introduce application scenario constraints, and gradually narrow down the range of candidate materials to achieve progressive optimization from coarse screening to fine screening.

[0136] In the embodiments of the present application, the general oxide pre-training model is adapted to the high-entropy oxide system through feature space mapping technology to solve the problem of data scarcity of high-entropy oxides. A unified physical and chemical property database is established by combining experimental data (synthesis conditions, catalytic activity) and DFT calculation data (formation energy, oxygen vacancy energy). Composite input feature design: simultaneously integrating element concentration ratio, local chemical environment parameters (lattice distortion rate, oxygen coordination polyhedron twist angle) and thermodynamic parameters, a multi-dimensional feature vector is constructed. Based on Shapley Value (Shapley Value) to evaluate the contribution of features, dynamically optimize the input feature set, for example, remove features with a contribution of <5%. Joint Huber loss (regression task) and cross-entropy loss (classification task) to balance outlier sensitivity and class imbalance problems. Simultaneous application of Dropout (for example, p=0.3), weight decay (for example, λ=0.001) and early stopping mechanism to build a triple overfitting defense system. Embedding atomic radius difference (Δr<15%) and electronegativity difference (Δχ<0.5) hard constraints in the element combination generation stage, compressing the invalid search space by more than 70%. Based on the frontier analysis of prediction performance and uncertainty (such as prediction variance), intelligent screening of Top-N candidate materials is realized. When the experimental performance deviates from the predicted value by more than 20%, local model retraining is automatically triggered, and data areas with a prediction uncertainty of >0.1 (Bayesian variance) are preferentially updated. The prediction results are directly converted into automated synthesis parameters (such as precursor molar ratio, calcination temperature gradient) to realize the parallel verification of hundreds of samples per day.

[0137] The embodiments of the present application also provide a flowchart of a material performance prediction model training method, as shown in Figure 2 , specifically comprising the following steps:

[0138] Step one: high-throughput calculation parameter adjustment;

[0139] Step two: batch modeling;

[0140] Step three: structure optimization;

[0141] Step four: single-point energy calculation;

[0142] Step five: data processing;

[0143] It should be noted that steps one to five belong to the link of training set construction.

[0144] Step six: model training;

[0145] Step seven: validity verification of the model.

[0146] It should be noted that steps six and seven are mainly the adjustment of the training set and the hyperparameter, and belong to the model construction and application.

[0147] Step eight: crystal structure screening;

[0148] Step nine: material synthesis and characterization;

[0149] Step ten: performance test.

[0150] It should be noted that steps eight to ten belong to the experimental verification link.

[0151] Corresponding to the material performance prediction model training method provided in the embodiments of the present application, the embodiments of the present application also provide a material performance prediction model training device, as shown in Figure 3 The material performance prediction model training device comprises:

[0152] The first generation module 301 is configured to generate a candidate crystal structure set based on the category and structure feature parameters corresponding to the first sample material, wherein the candidate crystal structure set comprises a first candidate crystal structure set and a second candidate crystal structure set, the first candidate crystal structure set comprises a plurality of first candidate crystal structures, and the second candidate crystal structure set comprises a plurality of second candidate crystal structures.

[0153] The first training module 302 is configured to perform model training based on the first candidate crystal structure and the performance parameter label corresponding to the first candidate crystal structure, to obtain a first material performance prediction base model.

[0154] The input module 303 is configured to input the second candidate crystal structure into the first material performance prediction base model, to obtain a prediction result corresponding to the second candidate crystal structure.

[0155] The first determination module 304 is configured to determine a first DFT result corresponding to the second candidate crystal structure.

[0156] The second generation module 305 is configured to generate a derivative crystal structure of the second candidate crystal structure under different spatial group symmetries, if the prediction result corresponding to the second candidate crystal structure and the first DFT result do not satisfy a first preset condition.

[0157] The second training module 306 is configured to perform training on the first material performance prediction base model based on the derivative crystal structure and the performance parameter label corresponding to the derivative crystal structure, to obtain a second material performance prediction base model.

[0158] The second determination module 307 is configured to determine the second material performance prediction base model corresponding to the second candidate crystal structure with a similarity to the crystal structure of the second sample material exceeding a preset similarity threshold value as a target material performance prediction base model.

[0159] The adjusting module 308 is configured to perform parameter adjustment on the target material performance prediction base model based on the second sample material, to obtain a target material performance prediction model.

[0160] In an optional embodiment, the method further comprises the following module:

[0161] The second sample material is filtered according to a preset element combination rule, to obtain a third sample material.

[0162] The crystal structure of the third sample material is optimized by using an adaptive genetic algorithm, to obtain a fourth sample material.

[0163] The fourth sample material is sampled by using a parallelized Monte Carlo tree search, to obtain a fifth sample material.

[0164] The target material performance prediction base model is adjusted based on the fifth sample material, to obtain a target material performance prediction model.

[0165] In an optional embodiment, the method further comprises the following module:

[0166] The material to be predicted is input into the target material performance prediction model, to obtain a performance prediction result corresponding to the material to be predicted.

[0167] The target material performance prediction model is trained based on the material to be predicted and the performance prediction result corresponding to the material to be predicted that satisfy a second preset condition.

[0168] In an optional embodiment, the method further comprises the following module:

[0169] The first sample material is filtered by using spatial clustering and structure description information.

[0170] In an optional embodiment, the second generating module is configured to:

[0171] An uncertainty index corresponding to the second candidate crystal structure is determined.

[0172] If the prediction result corresponding to the second candidate crystal structure is less than a preset tolerance threshold from the first DFT result, the first material performance prediction base model is trained based on the second candidate crystal structure.

[0173] If the prediction result corresponding to the second candidate crystal structure is not less than the preset tolerance threshold from the first DFT result, and / or the uncertainty index corresponding to the second candidate crystal structure is higher than a preset critical value, the second candidate crystal structure is subjected to DFT recalculation, to obtain a second DFT result.

[0174] If it is determined that there is systematic prediction deviation based on the prediction result corresponding to the second candidate crystal structure and the second DFT result, a derivative crystal structure of the second candidate crystal structure under different spatial group symmetries is generated.

[0175] In an optional embodiment, the adjusting module is configured to:

[0176] Based on the second sample material, the top layer regression network of the target material performance prediction base model is adjusted in parameter through the layer-selective thawing strategy, and a target material performance prediction model is obtained.

[0177] In an optional embodiment, the adjusting module is configured to:

[0178] The sample material calculated based on the hybrid functional or the GW method is obtained as the sixth sample material.

[0179] The target material performance prediction base model is adjusted in parameter based on the sixth sample material, and a target material performance prediction model is obtained.

[0180] Corresponding to the material performance prediction model training method provided in the embodiments of the present application, the embodiments of the present application further provide an electronic device for executing the material performance prediction model training method, as shown in Figure 4 The electronic device includes a processor 401 and a memory 402 for storing the program of the material performance prediction model training method. After the device is powered on and the processor runs the program of the material performance prediction model training method, the following steps are performed:

[0181] Based on the category and structure feature parameters corresponding to the first sample material, a candidate crystal structure set is generated, and the candidate crystal structure set includes a first candidate crystal structure set and a second candidate crystal structure set. The first candidate crystal structure set includes a plurality of first candidate crystal structures, and the second candidate crystal structure set includes a plurality of second candidate crystal structures.

[0182] Based on the first candidate crystal structure and the performance parameter label corresponding to the first candidate crystal structure, a model is trained to obtain a first material performance prediction base model.

[0183] The second candidate crystal structure is input into the first material performance prediction base model to obtain a prediction result corresponding to the second candidate crystal structure.

[0184] A first DFT result corresponding to the second candidate crystal structure is determined.

[0185] If the prediction result corresponding to the second candidate crystal structure and the first DFT result do not satisfy a first preset condition, a derivative crystal structure of the second candidate crystal structure under different spatial group symmetries is generated.

[0186] train the first material performance prediction base model based on the derived crystal structure and the performance parameter label corresponding to the derived crystal structure, to obtain a second material performance prediction base model;

[0187] determine the second material performance prediction base model corresponding to the second candidate crystal structure with a similarity to the crystal structure of the second sample material exceeding a preset similarity threshold as a target material performance prediction base model;

[0188] perform parameter adjustment on the target material performance prediction base model based on the second sample material, to obtain a target material performance prediction model.

[0189] Corresponding to the material performance prediction model training method provided in the embodiments of the present application, the embodiments of the present application also provide a computer readable storage medium storing a program of a material performance prediction model training method, the program being run by a processor to perform the following steps:

[0190] generate a candidate crystal structure set based on the category and structure feature parameters corresponding to the first sample material, the candidate crystal structure set including a first candidate crystal structure set and a second candidate crystal structure set, the first candidate crystal structure set including a plurality of first candidate crystal structures, and the second candidate crystal structure set including a plurality of second candidate crystal structures;

[0191] perform model training based on the first candidate crystal structure and the performance parameter label corresponding to the first candidate crystal structure, to obtain a first material performance prediction base model;

[0192] input the second candidate crystal structure into the first material performance prediction base model, to obtain a prediction result corresponding to the second candidate crystal structure;

[0193] determine a first DFT result corresponding to the second candidate crystal structure;

[0194] if the prediction result corresponding to the second candidate crystal structure and the first DFT result do not satisfy a first preset condition, generate derived crystal structures of the second candidate crystal structure under different spatial group symmetries;

[0195] train the first material performance prediction base model based on the derived crystal structure and the performance parameter label corresponding to the derived crystal structure, to obtain a second material performance prediction base model;

[0196] determine the second material performance prediction base model corresponding to the second candidate crystal structure with a similarity to the crystal structure of the second sample material exceeding a preset similarity threshold as a target material performance prediction base model;

[0197] perform parameter adjustment on the target material performance prediction base model based on the second sample material, to obtain a target material performance prediction model.

[0198] Corresponding to the material performance prediction model training method provided in the embodiments of the present application, the embodiments of the present application also provide a computer program containing instructions, when the program is executed by a computer, the instructions cause the computer to execute the following steps:

[0199] generate a candidate crystal structure set based on the category and structure feature parameters corresponding to the first sample material, the candidate crystal structure set including a first candidate crystal structure set and a second candidate crystal structure set, the first candidate crystal structure set including a plurality of first candidate crystal structures, and the second candidate crystal structure set including a plurality of second candidate crystal structures;

[0200] perform model training based on the first candidate crystal structure and the performance parameter label corresponding to the first candidate crystal structure, to obtain a first material performance prediction base model;

[0201] input the second candidate crystal structure into the first material performance prediction base model, to obtain a prediction result corresponding to the second candidate crystal structure;

[0202] determine a first DFT result corresponding to the second candidate crystal structure;

[0203] if the prediction result corresponding to the second candidate crystal structure and the first DFT result do not satisfy a first preset condition, generate derivative crystal structures of the second candidate crystal structure under different spatial group symmetries;

[0204] perform training on the first material performance prediction base model based on the derivative crystal structures and the performance parameter labels corresponding to the derivative crystal structures, to obtain a second material performance prediction base model;

[0205] determine the second material performance prediction base model corresponding to the second candidate crystal structure with a similarity to the crystal structure of the second sample material exceeding a preset similarity threshold as a target material performance prediction base model;

[0206] perform parameter adjustment on the target material performance prediction base model based on the second sample material, to obtain a target material performance prediction model.

[0207] It should be noted that the detailed description of the material performance prediction model training device, the electronic device, the computer readable storage medium and the computer program provided in the embodiments of the present application can refer to the related description of the material performance prediction model training method provided in the embodiments of the present application, which will not be repeated here.

[0208] Although the present application is disclosed with reference to the preferred embodiments above, it is not intended to limit the present application, and any person skilled in the art can make possible variations and modifications without departing from the spirit and scope of the present application. Therefore, the scope of the present application should be defined by the appended claims.

[0209] In one typical configuration, the electronic device includes one or more processors (Central Processing Unit), input / output interfaces, network interfaces, and memory.

[0210] The memory can include non-persistent memory, Random Access Memory (RAM), and / or non-volatile memory, such as Read Only Memory (ROM) or flash memory, in a computer readable medium. The memory is an example of computer readable media.

[0211] Computer readable media includes permanent and non-permanent, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (Programmable Analog Module), static random access memory (Static Random Access Memory), dynamic random access memory (Dynamic Random Access Memory), other types of random access memory (Random Access Memory), read only memory (Read Only Memory), electrically erasable programmable read only memory (Electrically Erasable Programmable Read Only Memory), flash memory or other memory technologies, compact disc read only memory (Compact Disc Read-Only Memory), digital video disc (Digital Video Disc) or other optical storage, magnetic cassette tape, magnetic tape disk storage or other magnetic storage devices or any other non-transmission medium that can be used to store information accessible to a computing device. According to the definition herein, computer readable media does not include non-transitory computer readable media (transitory media), such as modulated data signals and carriers.

[0212] Those skilled in the art will appreciate that embodiments of the present application can be readily used as a method, a system or a computer program product. Accordingly, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product on one or more computer-usable storage media (including, but not limited to, disk storage, Compact Disc Read-Only Memory, optical storage memory and the like) embodying computer-readable program code.

[0213] Although the present application has been disclosed in its preferred embodiments with reference to the accompanying drawings, it is to be understood that the application is not limited to those precise embodiments, and that changes and modifications can be effected therein by one skilled in the art without departing from the scope or spirit of the application. That scope should be limited only by the claims as set forth and equivalents thereof.

Claims

1. A method for training a material property prediction model, characterized in that, The method comprises the following steps: generating a candidate crystal structure set based on the category and structural feature parameters corresponding to the first sample material, wherein the candidate crystal structure set comprises a first candidate crystal structure set and a second candidate crystal structure set, the first candidate crystal structure set comprises a plurality of first candidate crystal structures, and the second candidate crystal structure set comprises a plurality of second candidate crystal structures; training a model based on the first candidate crystal structure and the performance parameter label corresponding to the first candidate crystal structure to obtain a first material performance prediction base model; inputting the second candidate crystal structure into the first material performance prediction base model to obtain a prediction result corresponding to the second candidate crystal structure; determining a first DFT result corresponding to the second candidate crystal structure; if the prediction result corresponding to the second candidate crystal structure and the first DFT result do not satisfy a first preset condition, generating a derivative crystal structure of the second candidate crystal structure under different space group symmetries; training the first material performance prediction base model based on the derivative crystal structure and the performance parameter label corresponding to the derivative crystal structure to obtain a second material performance prediction base model; determining the second material performance prediction base model corresponding to the second candidate crystal structure with a similarity to the crystal structure of the second sample material exceeding a preset similarity threshold as a target material performance prediction base model; adjusting the target material performance prediction base model based on the second sample material to obtain a target material performance prediction model.

2. The material property prediction model training method of claim 1, wherein Further comprising: screening the second sample material according to a preset element combination rule to obtain a third sample material; optimizing the crystal structure of the third sample material by using an adaptive genetic algorithm to obtain a fourth sample material; sampling the fourth sample material by using a parallelized Monte Carlo tree search to obtain a fifth sample material; adjusting the target material performance prediction base model based on the fifth sample material to obtain a target material performance prediction model.

3. The material property prediction model training method of claim 1, wherein Further comprising: inputting a material to be predicted into the target material performance prediction model to obtain a performance prediction result corresponding to the material to be predicted; training the target material performance prediction model based on the material to be predicted satisfying a second preset condition and the performance prediction result corresponding to the material to be predicted.

4. The material property prediction model training method of claim 1, wherein, Further comprising: screening the first sample material by using spatial clustering and structure description information.

5. The material performance prediction model training method of claim 1, wherein, If the prediction result corresponding to the second candidate crystal structure and the first DFT result do not satisfy the first preset condition, the method further comprises the following steps: determining an uncertainty index corresponding to the second candidate crystal structure; if the prediction result corresponding to the second candidate crystal structure and the first DFT result are less than a preset tolerance threshold, training the first material performance prediction base model based on the second candidate crystal structure. If the prediction result corresponding to the second candidate crystal structure is not less than the first DFT result by the preset tolerance threshold, and / or the uncertainty index corresponding to the second candidate crystal structure is higher than the preset threshold, the DFT re-calculation is performed on the second candidate crystal structure to obtain a second DFT result; If it is determined that there is a systematic prediction bias based on the prediction result corresponding to the second candidate crystal structure and the second DFT result, the derived crystal structure of the second candidate crystal structure under different space group symmetries is generated.

6. The material performance prediction model training method of claim 1, wherein, The parameter adjustment of the target material performance prediction model based on the second sample material includes: The parameter adjustment of the top layer regression network of the target material performance prediction model based on the second sample material through the layer selection thawing strategy is performed to obtain the target material performance prediction model.

7. The material performance prediction model training method of claim 1, wherein, Further comprising: Obtaining a sample material calculated based on a hybrid functional or a GW method as a sixth sample material; The parameter adjustment of the target material performance prediction model based on the sixth sample material is performed to obtain the target material performance prediction model.

8. A material property prediction model training apparatus characterized by comprising: Including: The first generation module is configured to generate a candidate crystal structure set based on the category and structure feature parameters corresponding to the first sample material, wherein the candidate crystal structure set includes a first candidate crystal structure set and a second candidate crystal structure set, the first candidate crystal structure set includes a plurality of first candidate crystal structures, and the second candidate crystal structure set includes a plurality of second candidate crystal structures; The first training module is configured to perform model training based on the first candidate crystal structure and a performance parameter label corresponding to the first candidate crystal structure to obtain a first material performance prediction base model; The input module is configured to input the second candidate crystal structure into the first material performance prediction base model to obtain a prediction result corresponding to the second candidate crystal structure; The first determination module is configured to determine a first DFT result corresponding to the second candidate crystal structure; The second generation module is configured to generate a derived crystal structure of the second candidate crystal structure under different space group symmetries if the prediction result corresponding to the second candidate crystal structure does not satisfy a first preset condition with the first DFT result; The second training module is configured to perform training on the first material performance prediction base model based on the derived crystal structure and a performance parameter label corresponding to the derived crystal structure to obtain a second material performance prediction base model; The second determination module is configured to determine a second material performance prediction base model corresponding to a second candidate crystal structure with a similarity to a crystal structure of a second sample material exceeding a preset similarity threshold as a target material performance prediction base model; The adjustment module is configured to perform parameter adjustment on the target material performance prediction base model based on the second sample material to obtain a target material performance prediction model.

9. An electronic device, comprising: The computer readable storage medium stores a computer program, and the computer program is executed by the processor to implement the material performance prediction model training method in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the computer program is executed by the processor to implement the material performance prediction model training method in any one of claims 1-7.

Images