CN-121999948-A - Two-dimensional Janus material and energy level parameter prediction method and device thereof

CN121999948ACN 121999948 ACN121999948 ACN 121999948ACN-121999948-A

Abstract

The invention provides a two-dimensional Janus material and an energy level parameter prediction method and device thereof, which are characterized in that a matrix and a Janus material crystal structure are constructed into a graph structure, local information is aggregated through a graph attention network with edge embedding after node/edge characteristics are improved, a global attention mechanism is used for obtaining graph level embedding, differential characteristics are obtained by difference between the two embedding, and a conduction band bottom offset, a valence band top offset and a vacuum energy level difference are predicted through multitasking of a multi-layer perceptron, and in addition, a model is optimized through multitasking weighting loss. According to the invention, by utilizing differential learning of parent-Janus pairing, the calculation cost is reduced, the prediction precision and generalization capability are improved, high-throughput screening can be realized, and the engineering design of Janus material energy bands is supported.

Inventors

HU CHAOYANG
ZHU XIAOJUN
Tao Maoxiang
LI QUAN
CAO WENTAO
HU YUQIANG

Assignees

江西理工大学南昌校区

Dates

Publication Date: 20260508
Application Date: 20260331

Claims (10)

1. A two-dimensional Janus material and an energy level parameter prediction method thereof, which are characterized in that the method comprises the following steps: Respectively converting the crystal structure diagrams of the parent material and the Janus material into corresponding diagram structures; converting node characteristics and edge characteristics in each graph structure into high-dimensional vectors to obtain lifted node embedding and edge embedding; According to the lifted node embedding and edge embedding, attention weight of each edge in the corresponding graph structure is calculated, and neighborhood node characteristics are aggregated according to the attention weight to update node representation, so that corresponding contextualization node embedding is obtained; Calculating global attention weights according to the corresponding contextualization node embedding, and carrying out weighted summation on all contextualization node embedding of the corresponding graph structure according to the global attention weights to respectively obtain parent graph embedding and Janus graph embedding; Performing characteristic difference on the parent diagram embedding and the Janus diagram embedding to obtain differential characteristics, inputting the differential characteristics into a multi-layer perceptron to predict, and outputting 3 labels, wherein the labels comprise conduction band bottom offset, valence band top offset and vacuum energy level difference; And carrying out loss function evaluation on the predicted label and the real label, and then carrying out back propagation to obtain the optimal parameters of the multi-layer perceptron.
2. The two-dimensional Janus material and energy level parameter prediction method according to claim 1, wherein in the step of converting crystal structure diagrams of a parent material and a Janus material into corresponding diagram structures, respectively, crystal structure information of the parent material and the Janus material is obtained, wherein the crystal structure information comprises lattice parameters, atomic fraction coordinates, element types, layer sequence/upper and lower surface identifications, and then a unified neighborhood construction rule is adopted to generate a crystal diagram under a periodic boundary condition Wherein the node Is an atom, an edge Representing an adjacency.
3. The two-dimensional Janus material and energy level parameter prediction method according to claim 2, wherein in the step of converting node features and edge features in each graph structure into high-dimensional vectors to obtain lifted node embedding and edge embedding, the node features at least comprise atomic numbers, electronegativity, atomic radii, valence electron related descriptions and layer indexes, the edge features at least comprise interatomic distances or bond lengths, and the node features are input into a node multi-layer perceptron to obtain node embedding, wherein the node embedding is represented as: ; Wherein h i is node embedding of central atom i, i is central node index, x i is node characteristic, MLP v is full-connected neural network special for node characteristic, The vector element is a real number, and d is the dimension of the vector; Inputting the edge characteristics into the edge multi-layer perceptron to obtain edge embedding, wherein the edge embedding is expressed as: ; Wherein h ij is the edge embedding of edge (i, j), j is the neighbor node index, MLP e is the fully connected neural network dedicated to edge features, e ij is the edge feature, The vector element is a real number, and d is the dimension of the vector.
4. The two-dimensional Janus material and the energy level parameter prediction method thereof according to claim 3, wherein the step of calculating the attention weight of each side in the corresponding graph structure according to the lifted node embedding and the side embedding, and aggregating the neighborhood node characteristics according to the attention weight to update the node representation, and obtaining the attention weight of each side in the step of embedding the corresponding contextualized node comprises the following calculation formula: ; where a ij is the attention weight of the edge (i, j), For the neighborhood Softmax normalization operator, leakyReLU is a nonlinear activation function, For the transposition of the attention vector, W is the node feature linear transformation matrix, W e is the edge feature linear transformation matrix, h i is the node embedding of the central atom i, h j is the node embedding of the neighbor atom j, h ij is the edge embedding of the edge (i, j), Vector splicing operators; the calculation formula of embedding the contextualization nodes is as follows: ; Wherein, the For contextualized node embedding, σ is a nonlinear activation function, A summation operation is performed for all neighbors j of the central atom i.
5. The two-dimensional Janus material and energy level parameter prediction method according to claim 4, wherein in the steps of obtaining a parent map embedding and a Janus map embedding respectively, a global attention weight is calculated according to the corresponding contextualization node embedding, and according to the global attention weight, all contextualization node embedding weights of the corresponding map structure are summed up, and the calculation formula of the global attention weight is: ; Wherein, the For global attention weights, softmax is a global softmax normalization operator, Is the transpose of the global query vector, tanh is a nonlinear activation function, W r is a learnable node feature projection parameter, Embedding a contextualized node, b r being a learnable additive bias parameter; The calculation formula of the graph embedding is as follows: ; Wherein h G is the graph embedding, As a global attention weight, For the embedding of the contextualized nodes, For the vector elements to be real numbers, Is the dimension of the vector.
6. The two-dimensional Janus material and the energy level parameter prediction method thereof according to claim 5, wherein the method is characterized in that a parent diagram is embedded and the Janus diagram is embedded to perform characteristic difference to obtain differential characteristics, the differential characteristics are input into a multi-layer perceptron to perform prediction, 3 labels are output, the parent diagram is embedded and spliced with the differential characteristics, and then the obtained result is input into the multi-layer perceptron to perform prediction, and the splicing result is expressed as: ; ; In order to achieve the result of the stitching, As a characteristic of the difference, For the embedding of the Janus diagram, For the embedding of the parent image(s), For the vector concatenation operator, For the vector elements to be real numbers, Is the dimension of the vector.
7. The two-dimensional Janus material and energy level parameter prediction method according to claim 6, wherein in the step of performing loss function evaluation on the predicted label and the real label and then back-propagating to obtain the optimal parameters of the multi-layer perceptron, the total loss is expressed as: ; ; Wherein, the In order to account for the total loss, For the kth subtask loss, MSE is the mean square error, For the predictive label of the kth subtask, For the k-th subtask real label, lambda 1 、λ 2 、λ 3 is the weight coefficient, L 1 is the conduction band bottom offset predictive loss, L 2 is the valence band top offset predictive loss, and L 3 is the vacuum level difference predictive loss, respectively.
8. A two-dimensional Janus material and energy level parameter prediction device thereof, which is used for realizing the two-dimensional Janus material and energy level parameter prediction method thereof according to any one of claims 1-7, wherein the device comprises: The first conversion module is used for respectively converting crystal structure diagrams of the parent material and the Janus material into corresponding diagram structures; The second conversion module is used for converting the node characteristics and the edge characteristics in each graph structure into high-dimensional vectors to obtain lifted node embedding and edge embedding; The first calculation module is used for calculating the attention weight of each edge in the corresponding graph structure according to the lifted node embedding and edge embedding, and aggregating the neighborhood node characteristics according to the attention weight to update the node representation so as to obtain the corresponding contextualization node embedding; The second calculation module is used for respectively calculating global attention weights according to the corresponding contextualization node embedding, and respectively obtaining parent diagram embedding and Janus diagram embedding according to the global attention weights and the weighted summation of all contextualization node embedding of the corresponding diagram structure; The prediction module is used for performing characteristic difference on the parent diagram embedding and the Janus diagram embedding to obtain differential characteristics, inputting the differential characteristics into the multi-layer perceptron to predict, and outputting 3 labels, wherein the labels comprise conduction band bottom offset, valence band top offset and vacuum level difference; And carrying out loss function evaluation on the predicted label and the real label, and then carrying out back propagation to obtain the optimal parameters of the multi-layer perceptron.
9. A computer readable storage medium having stored thereon a computer program, which when executed by a processor implements the two-dimensional Janus material and the energy level parameter prediction method of any of claims 1-7.
10. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, the processor implementing the two-dimensional Janus material and energy level parameter prediction method of any of claims 1-7 when the program is executed.

Description

Two-dimensional Janus material and energy level parameter prediction method and device thereof Technical Field The invention belongs to the technical field of energy level parameter prediction, and particularly relates to a two-dimensional Janus material, an energy level parameter prediction method and an energy level parameter prediction device thereof. Background Two-dimensional Janus materials (such as MXY monolayer, M is transition metal, X/Y is different chalcogen or halogen element) break mirror symmetry through upper and lower surface asymmetric substitution, can introduce built-in dipole and obviously regulate and control energy band structure. For applications such as photocatalysis, electronic device contact engineering and heterojunction energy band alignment, the energy level parameters (conduction band bottom CBM, valence band top VBM, vacuum energy level difference delta phi) are core evaluation indexes. The prior art generally determines energy level parameters based on first sexual principle calculations (e.g., DFT) that require geometric optimization, self-consistent calculations, energy band/state density analysis, and vacuum level alignment by planar average electrostatic potential alignment to determine conduction band bottom (CBM), valence band top (VBM), and vacuum level difference (ΔΦ). The process mainly has the following problems: (1) The calculation cost is high, the steps of geometric optimization, self-consistent alignment, vacuum potential alignment and the like are required to be completed for each candidate material, and large-scale high-throughput screening is difficult to support; (2) The data multiplexing efficiency is low, janus materials are usually constructed by parent symmetric materials (such as MX 2), and the existing flow does not fully utilize the parent-disturbance relation, so that repeated calculation is caused; (3) Alignment is inconsistent with the fiducials-different computational settings (vacuum layer thickness, dipole corrections, pseudopotentials and functional, etc.) may introduce systematic deviations that make it difficult for band-edge labels across the data sources to agree. Disclosure of Invention Based on this, the embodiment of the invention provides a two-dimensional Janus material and an energy level parameter prediction method and device thereof, which aim to support high-throughput screening and energy band engineering design by using a parent-Janus pairing relation and a method for rapidly predicting energy level parameters (conduction band bottom CBM, valence band top VBM and vacuum energy level difference delta phi) of the Janus material under a low-cost condition. A first aspect of an embodiment of the present invention provides a two-dimensional Janus material and an energy level parameter prediction method thereof, where the method includes: Respectively converting the crystal structure diagrams of the parent material and the Janus material into corresponding diagram structures; converting node characteristics and edge characteristics in each graph structure into high-dimensional vectors to obtain lifted node embedding and edge embedding; According to the lifted node embedding and edge embedding, attention weight of each edge in the corresponding graph structure is calculated, and neighborhood node characteristics are aggregated according to the attention weight to update node representation, so that corresponding contextualization node embedding is obtained; Calculating global attention weights according to the corresponding contextualization node embedding, and carrying out weighted summation on all contextualization node embedding of the corresponding graph structure according to the global attention weights to respectively obtain parent graph embedding and Janus graph embedding; Performing characteristic difference on the parent diagram embedding and the Janus diagram embedding to obtain differential characteristics, inputting the differential characteristics into a multi-layer perceptron to predict, and outputting 3 labels, wherein the labels comprise conduction band bottom offset, valence band top offset and vacuum energy level difference; And carrying out loss function evaluation on the predicted label and the real label, and then carrying out back propagation to obtain the optimal parameters of the multi-layer perceptron. Further, in the step of converting the crystal structure diagrams of the parent material and the Janus material into the corresponding diagram structures, crystal structure information of the parent material and the Janus material is obtained, wherein the crystal structure information comprises lattice parameters, atomic fraction coordinates, element types and layer sequence/upper and lower surface identifiers, and then a unified neighborhood construction rule is adopted to generate a crystal diagram under a periodic boundary conditionWherein the nodeIs an atom, an edgeRepresenting an adjacency. Further, in the