A material performance analysis and system prediction method

Abstract

The present application relates to the technical field of computational chemistry and computational electrochemistry, and particularly relates to a material performance analysis and system prediction method.The material performance analysis and prediction method provided by the present application adopts a constant potential message passing model to analyze atomic configurations; a multi-head attention mechanism layer combines atomic features and initial features of over-electron generated atomic configuration nodes, and multiple atomic features are linearly fused to form fused features carrying over-electron information; a message passing layer utilizes a geometric tensor product to perform information interaction between each central atom and neighbor atoms, reconstructs atomic features and over-electrons, and finally obtains a giant potential energy of the atomic configuration, atomic force and work function by performing a permutation invariant summation operation on messages of all neighbor nodes.The present application realizes an electrochemical calculation method with potential control capability, physical consistency and cross-scale simulation efficiency to support interface process research under real electrochemical conditions.The present application overcomes the problems in the prior art that electrochemical reaction research cannot consider scale and precision, lacks precision of potential control and is insufficient in physical consistency.

Description

Technical Field

[0001] This invention relates to the fields of computational chemistry and computational electrochemistry, and in particular to a method and system for predicting material properties. Background Technology

[0002] Electrochemical reactions typically occur at the electrode-electrolyte interface, which exhibits complex kinetics modulated by potential, including key processes such as intermediate adsorption, charge transfer, and the dynamic evolution of the interfacial solvation layer and electrical double layer (EDL). Current mainstream research methods for elucidating these phenomena include:

[0003] 1. Density Functional Theory (DFT) calculations: can accurately predict electronic structure, adsorption energy, reaction barriers, etc., but have high computational costs and cannot be extended to nanometer / nanosecond-level interface systems; scale and accuracy are mutually exclusive. 2. Grand-Canonical DFT: achieves constant potential modeling by fixing the chemical potential, but still has limitations in simulating charge redistribution and solvent response; it is difficult to directly correlate the applied potential with the electronic structure response, and lacks precision in potential control.

[0004] 3. Machine Learning Potential Functions (MLPs): They excel in improving computational efficiency, but they struggle to describe potential dependence and maintain physical symmetry, resulting in insufficient physical consistency. Furthermore, they do not explicitly model the potential-driven dynamic evolution of charges. Summary of the Invention

[0005] To overcome the problems in existing technologies, such as the inability to balance scale and precision in electrochemical reaction research, lack of potential control precision, and insufficient physical consistency, this invention proposes a material performance analysis and prediction method. This method possesses constant potential control capability, symmetry preservation, machine learning acceleration, and a modeling framework that can be extended to interface systems, in order to accurately capture electrochemical interface behavior and promote the intelligent design of electrocatalysts.

[0006] The present invention proposes a material property analysis and prediction method. First, a constant potential message passing model is constructed to predict the giant potential energy, work function and atomic force under constant potential based on atomic configuration and potential information. The constant potential message passing model is used to analyze the molecular properties under the target potential. The potential information is characterized by the electrons passing through the molecule under the target potential.

[0007] The constant potential message passing model includes a multi-head attention mechanism layer and a message passing layer; the multi-head attention mechanism layer is used to fuse atomic features and overelectrons of atomic configurations and output fused features;

[0008] The message passing layer is used to aggregate and merge features and edge features of atomic configurations to obtain the giant potential energy, atomic force and work function of the atomic configuration.

[0009] Preferably, the message passing layer includes a first fully connected layer, a second fully connected layer, a matrix fusion module, a third fully connected layer, and an aggregation module;

[0010] The first fully connected layer is used to transform the fused features output by the multi-head attention mechanism layer into a node matrix; the second fully connected layer is used to transform the edge features of the atomic configuration into an edge matrix; the node matrix and the edge matrix are fused by the matrix fusion module to obtain a fused matrix;

[0011] The third fully connected layer processes the fusion matrix, reconstructs the node features of the atomic configuration, and rearranges the overelectrons;

[0012] The aggregation module obtains the giant potential energy, atomic force, and work function based on the node characteristics and overelectrons output by the third weighted connection layer.

[0013] Preferably, the aggregation module includes a giant potential energy prediction branch, an atomic force prediction branch, and a work function prediction branch; the giant potential energy prediction branch performs global summation pooling on the node features output by the third fully connected layer to output the giant potential energy; the atomic force prediction branch performs automatic differential calculation on the giant potential energy to obtain consistent atomic forces; the work function prediction branch performs global average pooling on the node features output by the third fully connected layer to output the system work function.

[0014] Preferably, the edge features are obtained by encoding the edges of the atomic configuration using equivariant convolution to obtain edge features that include radial and angular components.

[0015] Preferably, the radial component uses the Bessel function as the radial basis function; the angular component uses the spherical harmonic function.

[0016] Preferably, the multi-head attention mechanism layer includes multiple front-end fully connected layers, attention fusion modules, and back-end fully connected layers configured in parallel;

[0017] Each front-end fully connected layer corresponds one-to-one with the set atomic feature category. The atomic feature is the node information after the atomic configuration is decomposed, including the number of atoms and the types of atoms.

[0018] Each front-end fully connected layer processes the overelectron and corresponding atomic features. The attention fusion module uses a multi-head attention mechanism to fuse the feature data output by each front-end fully connected layer, and then splices the output multi-head attention features before inputting them into the back-end fully connected layer for processing to generate and output fused features.

[0019] Preferably, the specific steps are as follows:

[0020] The atomic configuration to be analyzed is obtained, atomic features are extracted, and the edges of the atomic configuration are encoded using isovariant convolution to obtain edge features; the overelectrons of the molecule to be analyzed under the target potential are obtained.

[0021] A multi-head attention layer processes atomic features and overelectrons to obtain fused features; a message passing layer processes the fused features and edge features to reconstruct atomic features and overelectrons.

[0022] The multi-head attention layer and message passing layer process the reconstructed atomic features and overelectrons repeatedly until the number of reconstructions of atomic features and overelectrons reaches a set value; the message passing layer processes the last reconstructed atomic features and overelectrons to obtain the giant potential energy, atomic force and work function of the atomic configuration.

[0023] The present invention proposes a material property analysis and prediction system, comprising:

[0024] The human-computer interaction module is used to set the object to be analyzed and the target potential;

[0025] The feature extraction module, connected to the human-computer interaction module, is used to call the atomic configuration of the object to be analyzed and generate atomic features, edge features, and overelectrons at the target potential.

[0026] The model prediction module, with its built-in constant potential message passing model, processes the atomic features, edge features, and overelectrons output by the feature extraction module to obtain the giant potential energy, atomic force, and work function of the object to be analyzed.

[0027] The present invention proposes a material property analysis and prediction device, which includes a memory and a processor. The memory stores a computer program, and the processor is connected to the memory. The processor is used to execute the computer program to realize the material property analysis and prediction method.

[0028] The present invention proposes a storage medium storing a computer program, which, when executed, is used to implement the material property analysis and prediction method.

[0029] The advantages of this invention are:

[0030] (1) The material performance analysis and prediction method proposed in this invention uses a constant potential message passing model to analyze atomic configuration. The multi-head attention mechanism layer combines atomic features (e.g., position information and element type) and overelectron generation to form initial features of atomic configuration nodes. Multiple atomic features are linearly fused to form fused features carrying overelectron information. The message passing layer uses the geometric tensor product to perform information interaction between each central atom and its neighboring atoms, reconstructing atomic features and overelectrons. Finally, the giant potential energy, atomic force, and work function of the atomic configuration are obtained by performing a substitution-invariant summation operation on the messages of all neighboring nodes. This invention realizes an electrochemical calculation method that combines potential control capability, physical consistency, and cross-scale simulation efficiency to support the study of interfacial processes under real electrochemical conditions.

[0031] (2) The third fully connected layer in this invention updates the node features through two MLP layers (including ReLU activation), and the updated features continue to maintain symmetry properties.

[0032] (3) The message passing layer works in conjunction with the multi-head attention mechanism layer to achieve iterative rearrangement of atomic features and overelectrons, realizing cross-layer connections and preserving the influence of charge features.

[0033] (4) This invention relates to a constant potential atomic simulation method for electrochemical interface catalysis research, covering solid-liquid interface modeling, reaction path prediction under potential control, charge redistribution and double layer structure evolution in electrocatalytic reactions (such as CO2 reduction and hydrogen evolution reaction). Attached Figure Description

[0034] Figure 1 This is a structural diagram of a constant potential message passing model proposed in this invention;

[0035] Figure 2 This is a flowchart of a material property analysis and prediction method proposed in this invention. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0037] Reference Figure 1 This embodiment proposes a constant potential message passing model that predicts the giant potential energy, work function, and atomic force under constant potential based on atomic configuration and potential information. The atomic configuration consists of nodes representing molecules and edges connecting the nodes. The potential information is characterized by the number of electrons in the molecule under the target potential. The number of electrons is the product of the difference between the number of electrons under the target potential and the number of electrons in the molecule under neutral conditions and the chemical potential of electrons in the power source. Neutral conditions refer to voltage neutrality, i.e., zero-point voltage.

[0038] The constant potential message passing model includes a multi-head attention mechanism layer and a message passing layer;

[0039] Multi-head attention mechanism layer: includes multiple parallel front-end fully connected layers, attention fusion module and back-end fully connected layer.

[0040] Each front-end fully connected layer corresponds one-to-one with the set atomic feature category. The atomic feature is the node information after the atomic configuration is decomposed, including the number of atoms and the types of atoms. The number of atoms is the number of nodes on the atomic configuration, and the types of atoms are the types of elements on each node of the atomic configuration.

[0041] Each front-end fully connected layer processes the overelectrons and corresponding atomic features. The attention fusion module uses a multi-head attention mechanism to fuse the feature data output by each front-end fully connected layer, and then splices the output multi-head attention features before inputting them into the back-end fully connected layer for processing. The back-end fully connected layer outputs the fused features of atomic features and overelectrons at the target potential, which are the new features with overelectrons.

[0042] The message passing layer includes a first fully connected layer, a second fully connected layer, a matrix fusion module, a third fully connected layer, and an aggregation module;

[0043] The first fully connected layer is used to access the fused features output by the multi-head attention mechanism layer and transform them into a node matrix;

[0044] The second fully connected layer is used to access the edge features of the atomic configuration and transform them into an edge matrix. The edge features are obtained by encoding the edges of the atomic configuration using isovariant convolution to obtain the initial node feature representation h. v (a,c,q) and edge features including radial and angular components; the radial components use Bessel functions as radial basis functions; the angular components use spherical harmonic functions Y. lm , where l = 0, 1, 2.

[0045] In practice, when encoding the edges of atomic configurations, Clebsch–Gordan coefficients are used to ensure SO(3) isovariability; [The remaining text appears to be incomplete and requires further context.] The cutoff radius is used to determine the adjacency graph to ensure the effectiveness of edge feature representation.

[0046] The matrix fusion module merges the node matrix and the edge matrix to obtain a fused matrix. The matrix fusion module can employ matrix fusion algorithms such as fully connected layers.

[0047] The third fully connected layer processes the fusion matrix, reconstructs the node features of the atomic configuration, and rearranges the overelectrons;

[0048] The aggregation module includes a giant potential energy prediction branch, an atomic force prediction branch, and a work function prediction branch;

[0049] The giant potential energy prediction branch performs global summation pooling on the node features output by the third fully connected layer to output the giant potential energy;

[0050] The atomic force prediction branch performs automatic differential calculations on the giant potential energy to obtain consistent atomic forces;

[0051] The work function prediction branch performs global average pooling on the node features output by the third fully connected layer, and outputs the system work function.

[0052] That is, the giant potential energy prediction branch adopts a global summation pooling layer, the atomic force prediction branch adopts a differential calculation function model, and the work function prediction branch adopts a global average pooling layer.

[0053] In practice, the constant potential message passing model is obtained through machine learning on a labeled dataset.

[0054] Reference Figure 2 In this embodiment, a method for analyzing molecular properties using a constant potential message passing model is employed, and the steps are as follows:

[0055] S1. Obtain the atomic configuration of the molecule to be analyzed, extract atomic features, and encode the edges of the atomic configuration using isovariant convolution to obtain edge features; set the target potential and the overelectrons at the target potential;

[0056] S2. The multi-head attention layer processes atomic features and overelectrons to obtain fusion features;

[0057] S3. The message passing layer processes the fused features and edge features to obtain the reconstructed node features and rearranged overelectrons output by the third fully connected layer.

[0058] S4. Determine whether the message passing count t has reached the set value T; the message passing count is the number of node feature reconstructions of the atomic configuration, and also the number of transelectron rearrangements; the initial value of t is 0;

[0059] If not, update the message passing count t to t+1, then update the atomic configuration with the reconstructed node features, re-extract the atomic features and rearranged overelectrons of the atomic configuration as input to the multi-head attention layer, and return to step S2;

[0060] If so, the aggregation module processes the node features output by the third fully connected layer to obtain the giant potential energy, atomic force, and work function of the atomic configuration.

[0061] This constant potential message passing model was obtained through machine learning using a known dataset.

[0062] This embodiment also proposes a material property analysis and prediction system, including:

[0063] The human-computer interaction module is used to set the object to be analyzed and the target potential; the object to be analyzed is the molecule to be analyzed.

[0064] The feature extraction module, connected to the human-computer interaction module, is used to call the atomic configuration of the object to be analyzed and generate atomic features, edge features, and overelectrons at the target potential.

[0065] The model prediction module, with its built-in constant potential message passing model, processes the atomic features, edge features, and overelectrons output by the feature extraction module to obtain the giant potential energy, atomic force, and work function of the object to be analyzed.

[0066] Of course, those skilled in the art will recognize that the present invention is not limited to the details of the exemplary embodiments described above, but also includes the same or similar structures that can be implemented in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0067] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

[0068] The technologies, shapes, and structures not described in detail in this invention are all known technologies.

Claims

1. A method of material performance analysis and prediction, characterized by, Firstly, a constant potential message passing model is constructed to predict the grand potential, work function and atomic force of the molecule under constant potential based on the atomic configuration and potential information; the constant potential message passing model is used to analyze the performance of the molecule under target potential; the potential information is characterized by the excess electron of the molecule under target potential; The constant potential message passing model includes a multi-head attention mechanism layer and a message passing layer; The multi-head attention mechanism layer is used to fuse the atomic features of the atomic configuration and the excess electron, and output the fused features; The message passing layer is used to aggregate the fused features and the edge features of the atomic configuration to obtain the grand potential, atomic force and work function of the atomic configuration; The message passing layer includes a first fully connected layer, a second fully connected layer, a matrix fusion module, a third fully connected layer and an aggregation module; The first fully connected layer is used to convert the fused features output by the multi-head attention mechanism layer into a node matrix; The second fully connected layer is used to convert the edge features of the atomic configuration into an edge matrix; the node matrix and the edge matrix are fused into a fused matrix through the matrix fusion module; The third fully connected layer processes the fused matrix to reconstruct the node features of the atomic configuration and rearrange the excess electron; The aggregation module obtains the grand potential, atomic force and work function based on the node features and excess electron output by the third fully connected layer; The aggregation module includes a grand potential prediction branch, an atomic force prediction branch and a work function prediction branch; the grand potential prediction branch performs global sum pooling on the node features output by the third fully connected layer to output the grand potential; the atomic force prediction branch performs automatic differentiation calculation on the grand potential to obtain consistent atomic force; The work function prediction branch performs global average pooling on the node features output by the third fully connected layer to output the system work function; The multi-head attention mechanism layer includes a plurality of front-end fully connected layers, an attention fusion module and a rear-end fully connected layer arranged in parallel; Each front-end fully connected layer corresponds to a set of atomic feature categories; the atomic features are node information after the atomic configuration is disassembled, including the number of atoms and the types of each atom; Each front-end fully connected layer processes the excess electron and the corresponding atomic features; the attention fusion module uses a multi-head attention mechanism to fuse the feature data output by each front-end fully connected layer, and inputs the output multi-head attention features into the rear-end fully connected layer for processing to generate and output the fused features.

2. The material performance analysis and prediction method of claim 1, wherein The edge features are obtained by using equivariant convolution to encode the edges of the atomic configuration to obtain edge features containing radial components and angular components.

3. The material performance analysis and prediction method according to claim 2, wherein The radial component uses Bessel function as the radial basis function; the angular component uses spherical harmonics.

4. The material performance analysis and prediction method of claim 1, wherein The specific steps are as follows: An atomic configuration to be analyzed is obtained, atomic features are extracted, and equivariant convolution is used to encode the edges of the atomic configuration to obtain edge features; the excess electron of the molecule to be analyzed under target potential is obtained; The multi-head attention layer is used to process the atomic features and the excess electron to obtain the fused features; The message passing layer is used to process the fused features and the edge features to reconstruct the atomic features and the excess electron; The multi-head attention layer and the message passing layer are used for reconstructing atomic features and over electrons, until the number of times of reconstruction of the atomic features and the over electrons reaches a set value; and the message passing layer is used for processing the last reconstructed atomic features and over electrons, so as to obtain the giant potential energy, atomic force and work function of the atomic configuration.

5. A material property analysis and prediction system employing the material property analysis and prediction method according to any one of claims 1 to 4, characterized by, The method comprises the following steps: An interactive module is configured to set an object to be analyzed and a target potential. An atomic configuration of the object to be analyzed is called by a feature extraction module connected to the interactive module, and atomic features, edge features and over electrons under the target potential are generated. A model prediction module with a constant potential message passing model is used to process the atomic features, edge features and over electrons output by the feature extraction module, so as to obtain the giant potential energy, atomic force and work function of the object to be analyzed.

6. A material performance analysis prediction device characterized by comprising: The computer program is stored in the memory and executed by the processor to realize the material performance analysis and prediction method according to any one of claims 1-4.

7. A storage medium, characterized by The computer program is stored in the memory and executed by the processor to realize the material performance analysis and prediction method according to any one of claims 1-4.

Images