CN-120473021-B - Method and device for predicting physical properties of electrolyte based on attention mechanism

CN120473021BCN 120473021 BCN120473021 BCN 120473021BCN-120473021-B

Abstract

The application relates to a method and a device for predicting physical properties of electrolyte based on an attention mechanism, which are applied to the field of chemical property prediction, wherein the method comprises the steps of obtaining electrolyte to be trained in a preset electrolyte formula database; the electrolyte formula database comprises formula component information of the electrolyte to be trained, wherein the formula component information comprises chemical group characteristics and site data, mask training is carried out on a preset autoregressive pretraining attention model according to the electrolyte to be trained to obtain a target electrolyte prediction model, the formula component information of the electrolyte to be predicted is obtained, and the formula component information of the electrolyte to be predicted is input into the target electrolyte prediction model to obtain target physical property prediction information of the electrolyte to be predicted. By the method, the application range of electrolyte physical property prediction is improved, and the accuracy of the prediction result is higher.

Inventors

CHEN ZHONGWEI
Liao Chenyi
MAO ZHIYU
ZHAO LEI

Assignees

中国科学院大连化学物理研究所

Dates

Publication Date: 20260512
Application Date: 20250408

Claims (8)

1. A method for predicting physical properties of an electrolyte based on an attention mechanism, the method comprising: The method comprises the steps of obtaining electrolyte to be trained in a preset electrolyte formula database, wherein the electrolyte formula database comprises formula component information of the electrolyte to be trained, and the formula component information comprises chemical group characteristics and site data; Performing mask training on a preset autoregressive pretraining attention model according to the electrolyte to be trained to obtain a target electrolyte prediction model, wherein the target electrolyte prediction model comprises at least one target coding module and at least one target decoding module, and the target coding module and the target decoding module are respectively provided with a target self-attention layer; The method comprises the steps of obtaining formula component information of electrolyte to be predicted, inputting the formula component information of the electrolyte to be predicted into a target electrolyte prediction model, masking the formula component information of the electrolyte to be predicted to obtain prediction mask information, inputting the prediction mask information into a target coding module, and obtaining an intermediate arbitrary feature vector of the electrolyte to be predicted based on a target self-attention layer; inputting an intermediate arbitrary feature vector of the electrolyte to be predicted to the target decoding module, and obtaining target physical property prediction information of the electrolyte to be predicted based on the self-attention layer; The target physical property prediction information comprises target solubility, the method further comprises the steps of obtaining attention weight of any intermediate feature vector, inputting prediction mask information to the target electrolyte prediction model in combination with the attention weight and a preset solubility improvement strategy to obtain target proportion of chemical groups and site data of the electrolyte to be predicted, and improving the target solubility of the electrolyte to be predicted based on the target proportion to obtain improved solubility.
2. The method for predicting the physical properties of an electrolyte based on an attention mechanism according to claim 1, wherein the preset autoregressive pre-training attention model comprises at least one coding module and at least one decoding module, wherein the coding module and the decoding module comprise self-attention layers; mask training is carried out on a preset autoregressive pretraining attention model according to the electrolyte to be trained to obtain a target electrolyte prediction model, and the method comprises the following steps: masking the formula component information of the electrolyte to be trained to obtain formula masking component information; inputting the formula mask component information to the coding module, and obtaining the intermediate prediction characteristic of the electrolyte to be trained based on the self-attention layer; inputting the formula mask component information and the intermediate prediction characteristics of the electrolyte to be trained to the decoding module, and obtaining a physical property prediction value of the electrolyte to be trained based on the self-attention layer; calculating a predicted loss value based on the physical property predicted value and chemical group characteristics and site data in the formula component information of the electrolyte to be trained; And performing iterative training on the preset autoregressive pre-training attention model based on the calculated predicted loss value until the predicted loss value is lower than a preset error threshold value, so as to obtain a target electrolyte predicted model.
3. The method for predicting physical properties of an electrolyte based on an attention mechanism according to claim 1, further comprising, after obtaining the target proportioning of the chemical group and site data on the electrolyte to be predicted: According to the target ratio, a preset molecular simulation method is combined, and the correlation and the bonding strength between the chemical groups and the site data are determined; and determining target groups and target sites in the chemical groups and site data based on the correlation and the bonding strength, and reconstructing the electrolyte to be predicted.
4. A method of predicting the physical properties of an electrolyte based on an attention mechanism as claimed in any one of claims 1 to 3, further comprising: and updating the electrolyte formula database according to the electrolyte to be predicted and the target physical property prediction information of the electrolyte to be predicted.
5. The method for predicting physical properties of an electrolyte based on an attention mechanism as recited in claim 4, wherein the predetermined electrolyte formulation database includes known electrolytes, and wherein the step of obtaining the electrolyte to be trained in the predetermined electrolyte formulation database includes: Obtaining structural information in formula component information of known electrolyte; dividing the structural information to determine label data contained in the known electrolyte; and processing the known electrolyte based on the label data to obtain the chemical group characteristics of the known electrolyte, and taking the known electrolyte as the electrolyte to be trained.
6. The device for predicting the physical property of the electrolyte based on the attention mechanism is characterized by comprising a data acquisition module, a model training module and a prediction module; The data acquisition module is used for acquiring electrolyte to be trained in a preset electrolyte formula database, wherein the electrolyte formula database comprises formula component information of the electrolyte to be trained, and the formula component information comprises chemical group characteristics and site data; The model training module is used for carrying out mask training on a preset autoregressive pretraining attention model according to the electrolyte to be trained to obtain a target electrolyte prediction model, wherein the target electrolyte prediction model comprises at least one target coding module and at least one target decoding module; The prediction module is used for obtaining recipe component information of the electrolyte to be predicted, inputting the recipe component information into a target electrolyte prediction model, carrying out mask processing on the recipe component information of the electrolyte to be predicted to obtain prediction mask information, inputting the prediction mask information into the target coding module, obtaining an intermediate arbitrary feature vector of the electrolyte to be predicted based on the target self-attention layer, inputting the intermediate arbitrary feature vector of the electrolyte to be predicted into the target decoding module, obtaining target physical property prediction information of the electrolyte to be predicted based on the self-attention layer, wherein the target physical property prediction information comprises target solubility, and is further used for obtaining attention weight of the intermediate arbitrary feature vector, combining the attention weight and a preset solubility improvement strategy, inputting the prediction mask information into the target electrolyte prediction model to obtain a target ratio of chemical groups and site data of the electrolyte to be predicted, and improving the target solubility of the electrolyte to be predicted based on the target ratio to obtain improved solubility.
7. An electronic device comprising a memory and a processor, wherein the memory has stored therein a computer program, the processor being arranged to run the computer program to perform the method of predicting the physical properties of an electrolyte based on an attention mechanism as claimed in any one of claims 1 to 5.
8. A computer-readable storage medium, on which a computer program is stored, characterized in that the computer program, when being executed by a processor, implements the steps of the method for predicting the physical properties of an electrolyte based on an attention mechanism as claimed in any one of claims 1 to 5.

Description

Method and device for predicting physical properties of electrolyte based on attention mechanism Technical Field The application relates to the field of chemical property prediction, in particular to a method and a device for predicting physical properties of electrolyte based on an attention mechanism. Background The common electrolyte comprises lithium salt, solvent, additives and the like, and meanwhile, the physical properties of the electrolyte are closely related to chemical composition and proportion, and the electrolyte comprises various salt groups, organic groups, active coordination elements and the like from the chemical composition. Because of the variety of electrolytes and their component differences, the exploration requirements for the chemical space of electrolytes are high, so that the performance of electrolytes needs to be optimized. Current physical prediction methods of electrolyte properties involve a number of aspects including theoretical calculations, model construction, experimental verification, etc., such as simulating the motion trajectories of all particles in a system by solving newton's equations of motion. In electrolyte research, molecular dynamics simulation can be used for predicting diffusion coefficient, conductivity, viscosity, structural characteristics and the like of ions in a solvent, or predicting important information such as energy change, interaction energy between ions and the solvent, reaction path and the like in the dissolution process through quantum chemical calculation. Different prediction methods may be suitable for different types of electrolyte systems. For example, some classical force field based molecular dynamics simulations may not accurately describe complex ion-solvent interactions. Also, quantum chemistry is generally only capable of handling smaller scale systems (e.g., a few to tens of atoms) due to computational resource limitations, which limits its direct applicability to real electrolyte environments, while the results may deviate to some extent due to model simplification, parameter selection, and approximation. Therefore, the use range of predicting the physical properties of the electrolyte is narrow in the existing method, and the obtained accuracy has errors. Aiming at the problems of narrow application range and low accuracy of the electrolyte physical property prediction method in the related art, no effective solution is proposed at present. Disclosure of Invention The embodiment provides a prediction method and a prediction device for physical properties of electrolyte based on an attention mechanism, so as to solve the problems of narrow application range and low accuracy of the prediction method for physical properties of electrolyte in the related technology. In a first aspect, in this embodiment, there is provided a method for predicting physical properties of an electrolyte based on an attention mechanism, the method comprising: The method comprises the steps of obtaining electrolyte to be trained in a preset electrolyte formula database, wherein the electrolyte formula database comprises formula component information of the electrolyte to be trained, and the formula component information comprises chemical group characteristics and site data; performing mask training on a preset autoregressive pre-training attention model according to the electrolyte to be trained to obtain a target electrolyte prediction model; And acquiring formula component information of the electrolyte to be predicted, and inputting the formula component information of the electrolyte to be predicted into the target electrolyte prediction model to obtain target physical property prediction information of the electrolyte to be predicted. In some embodiments, the preset autoregressive pre-training attention model comprises at least one coding module and at least one decoding module, wherein the coding module and the decoding module comprise self-attention layers; mask training is carried out on a preset autoregressive pretraining attention model according to the electrolyte to be trained to obtain a target electrolyte prediction model, and the method comprises the following steps: masking the formula component information of the electrolyte to be trained to obtain formula masking component information; inputting the formula mask component information to the coding module, and obtaining the intermediate prediction characteristic of the electrolyte to be trained based on the self-attention layer; inputting the formula mask component information and the intermediate prediction characteristics of the electrolyte to be trained to the decoding module, and obtaining a physical property prediction value of the electrolyte to be trained based on the self-attention layer; calculating a predicted loss value based on the physical property predicted value and chemical group characteristics and site data in the formula component information of the electrolyte to b