CN-122024954-A - Electrode material design method and system based on high entropy doping screening and attention

CN122024954ACN 122024954 ACN122024954 ACN 122024954ACN-122024954-A

Abstract

The application provides an electrode material design method and system based on high-entropy doping screening and attention, wherein the method comprises the steps of determining a substrate structure of an electrode material, screening based on preset physicochemical rules to form a high-entropy doping element library, carrying out hierarchical clustering based on the high-entropy doping element library to obtain a high-entropy doping scheme, obtaining a high-entropy electrode material data set based on the high-entropy doping scheme, obtaining graph structure input data, constructing a graph neural network model, pre-training the graph neural network model to obtain a pre-trained graph neural network model, carrying out optimization training on the pre-trained graph neural network model based on the high-entropy electrode material data set and hierarchical self-distillation to obtain a high-entropy electrode material design model, and screening out the high-entropy doping scheme meeting preset performance by using the high-entropy electrode material design model. The application realizes end-to-end reasoning of performances such as lithium ion diffusion energy barrier and the like, improves the prediction speed, and enables efficient virtual design of electrode materials to be possible.

Inventors

YANG SHENGMIN
WEI ZHIYUN

Assignees

上海灵纭科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260122

Claims (10)

1. The electrode material design method based on high entropy doping screening and attention is characterized by comprising the following steps of: obtaining a matrix structure of an electrode material, and screening based on a preset physicochemical rule to form a high-entropy doping element library; performing high-entropy doping, structure optimization and energy barrier calculation on the matrix structure based on the high-entropy doping scheme to obtain a high-entropy electrode material data set; acquiring graph structure input data based on an open source material database, constructing a graph neural network model based on an attention mechanism, pre-training the graph neural network model based on the graph structure input data, and obtaining a pre-trained graph neural network model; Optimizing and training the pre-trained graphic neural network model based on a high-entropy electrode material data set and hierarchical self-distillation to obtain a high-entropy electrode material design model; and predicting the performance of the high-entropy doping scheme to be screened by using a high-entropy electrode material design model, and screening the high-entropy doping scheme meeting the preset performance requirement.
2. The method for designing an electrode material based on high-entropy doping screening and attention as claimed in claim 1, wherein, The matrix structure is Li x Fe y Cl z , wherein x is more than 0 and less than or equal to 2 y is more than 0 and less than or equal to 2 0<z is less than or equal to 5; the physicochemical rules include ionic radius rules, electronegativity rules, valence rules, earth abundance rules, and toxicity rules; The screening based on the preset physicochemical rules to form a high-entropy doping element library comprises the following steps: and screening out elements with ion radius, electronegativity, valence state, earth abundance and toxicity reaching standards from the existing doped element library according to preset screening standards of each physicochemical rule to form the high-entropy doped element library.
3. The method for designing the electrode material based on the high-entropy doping screening and attention according to claim 1, wherein the hierarchical clustering based on the high-entropy doping element library is characterized by comprising the following steps of: Extracting characteristics of each element in the high-entropy doping element library to form a characteristic vector, wherein the characteristics comprise ion radius, electronegativity and valence state; Dividing elements corresponding to the feature vectors into a plurality of non-overlapping clusters based on Euclidean distance of the feature vectors in a feature space by adopting a hierarchical clustering algorithm, wherein the Euclidean distance from the feature vectors of the elements in each cluster to the center of the cluster is smaller than a preset distance threshold; And selecting elements in each cluster according to the preset quantity, and combining all selected elements to obtain a high-entropy doping scheme, wherein the high-entropy doping scheme comprises the composition of the elements and the doping proportion of each element.
4. The method for designing an electrode material based on high-entropy doping screening and attention according to claim 1, wherein the high-entropy doping, structure optimization and energy barrier calculation are performed on a base structure based on the high-entropy doping scheme to obtain a high-entropy electrode material data set, and the method comprises the following steps: Performing structural optimization on the matrix structure to obtain a matrix structure with the lowest energy as an optimized matrix structure; Performing high-entropy doping on the optimized substrate structure based on a high-entropy doping scheme to generate a high-entropy doping structure, wherein each high-entropy doping scheme correspondingly generates a high-entropy doping structure; Performing structural optimization on each high-entropy doping structure to obtain a high-entropy doping structure with the lowest energy as an optimized doping structure; performing energy barrier calculation on the optimized doping structure to obtain a lithium ion diffusion energy barrier corresponding to the optimized doping structure; and constructing a high-entropy electrode material data set, wherein the high-entropy electrode material data set comprises the atomic three-dimensional coordinates, unit cell parameters, a stress matrix, total energy and lithium ion diffusion energy barriers of the optimized doping structure.
5. The method for designing an electrode material based on high-entropy doping screening and attention according to claim 1, wherein the step of acquiring the graph structure input data based on the open source material database comprises the steps of: extracting original data of electrode materials from an open source material database; Performing feature engineering on the original data to extract multi-scale material features of the electrode material, wherein the multi-scale material features comprise chemical features, atomic features and crystallographic features; converting the multi-scale material characteristics of the electrode material into a graph structure, wherein atoms are graph nodes, interatomic interactions or chemical bonds are graph edges, the node characteristics are chemical characteristics and atomic characteristics, and the edge characteristics are interatomic bond lengths and bond angles; And forming graph structure input data by utilizing graph structures corresponding to all electrode material samples.
6. The electrode material design method based on high-entropy doping screening and attention according to claim 1, wherein the graph neural network model comprises an input layer, a feature extraction layer, a relation capturing layer and an output layer, wherein the input layer is used for receiving graph structure input data, the feature extraction layer comprises at least two stacked convolution layers used for aggregation and learning of local atomic environment information, the relation capturing layer comprises at least one stacked attention layer and is used for capturing long-range dependency relations among atoms in a crystal structure of an electrode material based on a self-attention mechanism, and the output layer is used for outputting material performance predicted values.
7. The electrode material design method based on high entropy doping screening and attention according to claim 1, wherein in the pre-training process of the graph neural network model, an average absolute error is selected as a loss function, a gradient descent algorithm is used for updating graph neural network model parameters so as to minimize the loss function, and in the continuous iteration process, iteration is stopped when the loss descent amplitude of the graph neural network model is smaller than a preset value or reaches a preset iteration round, so that the pre-trained graph neural network model is obtained.
8. The method for designing an electrode material based on high entropy doping screening and attention according to claim 1, wherein the optimizing training of the pre-trained graph neural network model based on the high entropy electrode material data set and hierarchical self-distillation comprises: Taking a pre-trained graph neural network model as a training frame of hierarchical self-distillation; In the pre-trained graphic neural network model, the high-entropy electrode material dataset is used as input, the output of an attention layer is used as a teacher layer, the output of a convolution layer is used as a student layer, learning loss is selected as a learning loss function, a gradient descent algorithm is used for updating parameters of the pre-trained graphic neural network model, local features output by the student layer are gradually aligned with structural features of the teacher layer, the minimum learning loss function is achieved, training is stopped until the loss converges or the maximum iteration number is reached, and therefore a high-entropy electrode material design model is obtained, and the high-entropy electrode material design model is used for outputting the total energy and diffusion energy barrier prediction result of a high-entropy doping scheme to be screened.
9. The method for designing an electrode material based on high-entropy doping screening and attention according to claim 8, wherein the expression of the learning loss is as follows: L loss =αL hard +(1-α)L soft Wherein L loss is learning loss, L hard is hard target loss, L soft is hierarchical knowledge distillation soft loss, and alpha is loss weighting coefficient.
10. An electrode material design system based on high entropy doping screening and attention, comprising: The scheme design module is used for determining the matrix structure of the electrode material, and screening the electrode material based on a preset physicochemical rule to form a high-entropy doping element library; the data set construction module is used for carrying out high-entropy doping, structure optimization and energy barrier calculation on the matrix structure based on the high-entropy doping scheme to obtain a high-entropy electrode material data set; the network pre-training module is used for acquiring graph structure input data based on an open source material database, constructing a graph neural network model, pre-training the graph neural network model based on the graph structure input data, and obtaining a pre-trained graph neural network model; The optimization training module is used for performing optimization training on the pre-trained graph neural network model based on the high-entropy electrode material data set and the hierarchical self-distillation to obtain a high-entropy electrode material design model; And the performance prediction module is used for predicting the performance of the high-entropy doping scheme to be screened by using the high-entropy electrode material design model to obtain the high-entropy doping scheme meeting the preset performance requirement.

Description

Electrode material design method and system based on high entropy doping screening and attention Technical Field The application relates to the fields of computing material science, artificial intelligence and new energy material intersection, in particular to an electrode material design method and system based on high-entropy doping screening and attention. Background Today's society is in the key period of energy conversion and industry upgrade. With the rapid development of electric vehicles, large-scale energy storage systems and wearable electronic device markets, the global demand for new battery materials with high performance, high safety, long life, and low cost has increased dramatically. In particular, all-solid-state batteries are considered to be the most promising direction for next-generation battery technology due to their inherent high safety and high energy density potential. Therefore, the rapid and efficient development of electrode materials having excellent ion transport properties and structural stability has become a core bottle problem restricting the breakthrough development of new energy industries. The introduction of the high-entropy material provides a huge component space and unprecedented degree of freedom for the design of novel high-performance electrode materials. However, this explosive growth in component space makes traditional first principles calculations and "trial and error" experiments very costly, inefficient, and lacking effective guidelines. Researchers cannot systematically explore tens of thousands of potential high entropy doping schemes, severely hampering the speed of material innovation in this field. Although the quantum mechanical method (first principle calculation) can provide high-precision calculation results (lithium ion diffusion energy barriers) of atomic scale, the calculation complexity is extremely high, the calculation amount is huge, and the quantum mechanical method cannot be directly applied to a large-scale high-entropy doping system for high-flux screening, so that sharp contradiction between calculation precision and calculation efficiency is caused. In recent years, the rapid development of artificial intelligence technology has become a non-negligible technical means in various fields. The deep learning method represented by the neural network can mine the physical and chemical information and the structure-property function mechanism hidden behind the neural network in big data, and greatly improves the calculation efficiency. Therefore, deep learning has become a key technique for solving the contradiction between accuracy and efficiency. However, deep learning approaches still face significant technical challenges in the application of high entropy systems. Firstly, the problems of data sparsity and cold start are that the high-quality first sex principle calculation data are extremely scarce, and the sufficient marked data are difficult to acquire to train a reliable model, so that model training faces the bottleneck of cold start. Secondly, the limitation of remote dependency capturing is that a local convolution mechanism adopted by a traditional graph neural network is difficult to effectively capture long-range lattice disorder and atomic association in a high-entropy system, and the long-range lattice disorder and atomic association are key factors influencing lithium ion migration. More importantly, the network level knowledge is inefficient in that in the training process of the graphic neural network, the shallow network of the model can only learn basic local low-order features, and the deep network can learn more abstract high-order feature knowledge and global association. However, due to the scarcity of high-entropy data, higher-order knowledge learned by deep networks is difficult to efficiently guide and multiplex to shallow networks, resulting in inefficient utilization of knowledge inside the model and suboptimal convergence of the model. How to solve the problems of sparse data and remote association capture of a high-entropy system based on an artificial intelligence technology, improve the transmission efficiency of internal knowledge of a model network, and realize high-flux and high-performance prediction under quantum mechanical precision so as to accelerate meeting of urgent requirements of society on novel high-performance electrode materials, and is a great problem to be further solved in the field of computing material science. The prior art lacks a method for predicting and designing the performance of the electrode material, which can simultaneously solve the problems of sparse data, low calculation efficiency and the like of a high-entropy system. Disclosure of Invention Aiming at the defects in the prior art, the application aims to provide an electrode material design method and system based on high-entropy doping screening and attention. According to one aspect of the present application, there is p