A method for predicting the performance of MXenes materials based on a graph neural network

Abstract

The application belongs to the technical field of material performance prediction, and discloses a MXenes material performance prediction method based on a graph neural network, which comprises the following steps: obtaining structural characteristic parameters of a to-be-predicted MXenes material; respectively matching node derivative information, edge derivative information and global derivative information based on the structural characteristic parameters; constructing an initial feature vector set based on the node derivative information, the edge derivative information and the global derivative information, and based on an edge feature vector and a global feature vector; and performing hierarchical processing on each feature vector in the initial feature vector set through each layer of the graph neural network to obtain a target performance index of the to-be-predicted MXenes material. The scheme effectively solves the key technical problems in the prior art, such as insufficient coding of geometric information and inability to effectively distinguish intra-layer and inter-layer interactions.

Description

Technical Field

[0001] This invention belongs to the field of material property prediction technology, specifically relating to a method for predicting material properties based on graph neural networks (MXenes). Background Technology

[0002] Traditional machine learning methods are only applicable to MXenes with single functional groups, and their effectiveness is limited for complex features; they cannot accurately predict MXenes with mixed surface terminations. Existing graph neural network models suffer from insufficient geometric information encoding and inability to effectively distinguish intra-layer and inter-layer interactions when dealing with the unique layered structure and complex interactions of MXenes, resulting in limited prediction accuracy and performance saturation.

[0003] In the prior art, there is a method of predicting the dynamic viscosity of MXene-soybean oil nanofluid by using a long short-term memory network (LSTM) combined with a Bayesian optimization technique. First, different concentrations of MXene nanofluid are prepared through experiments, and a rheometer is used to measure the dynamic viscosity in the temperature range of 25-85°C to obtain 200 sample data. Then, an LSTM deep learning model is constructed, with temperature and nanoplatelet concentration as input parameters, and the long-term dependence relationship in the sequence data is processed through the memory unit. The Bayesian optimization method is used to automatically adjust the hyperparameters of the LSTM, and the network is trained through the back propagation algorithm, finally realizing the prediction of the dynamic viscosity of the MXene nanofluid. However, this method only predicts a single performance parameter and cannot simultaneously predict multiple key performance parameters of the material, so its application range is limited. The traditional LSTM network structure is mainly used to process time series data, which cannot effectively represent the atomic-level spatial structure of MXene materials and the complex interaction between atoms, and lacks a deep understanding of the geometric structure information of the material; it relies heavily on a large amount of experimental data, and different ratio nanofluid samples need to be prepared and subjected to tedious rheological tests, which is high in experimental cost and long in cycle, and cannot realize high-throughput material screening; the input features are limited to macroscopic temperature and concentration parameters, ignoring the microstructure characteristics of MXene materials, which leads to insufficient understanding of the nature of the material by the model and limits the prediction accuracy. Further, in the prior art, there is also a method of using a VICReg graph contrast learning framework. The core technology is to generate enhanced graph pairs through six random graph enhancement technologies, and use the VICReg loss function for contrast learning training. After training, linear regression, random forest, LightGBM, XGBoost, and other downstream models are used to predict the performance of embedded features, which can predict seven MXenes performance parameters including work function, Fermi level density of states, Poisson ratio, bulk modulus, shear modulus, Young's modulus, and magnetism. However, this method relies on the random graph enhancement strategy to change the molecular graph structure, which may destroy the key structural information of MXenes materials, especially for MXenes materials with fine layered structure, random deletion of nodes or edges will lose important intra-layer-inter-layer interaction information; it needs to be pre-trained unsupervised before using traditional machine learning models for downstream prediction, the training process is complex and the calculation efficiency is low; it is not designed specifically for the unique two-dimensional layered structure of MXenes, and lacks the ability to distinguish different interaction modes of covalent bonds and van der Waals forces, which limits the further improvement of prediction accuracy.

[0004] Further, the patent CN120340696A discloses a high polymer material performance prediction method based on deep learning, which is mainly for predicting the aging performance of high polymer materials. The method first prepares a group of high polymer material samples with different variable parameter combinations, respectively performs initial tensile test and tensile test after accelerated aging test, and obtains tensile performance parameters such as breaking strength and elongation. Then two independent multilayer perceptron deep learning network models are constructed: the initial performance prediction model adopts the structure of 5 feature inputs→128 neurons in the first hidden layer→64 neurons in the second hidden layer→32 neurons in the third hidden layer→1 neuron output layer, to predict the initial tensile performance of the material; the aging performance prediction model adopts the structure of 7 feature inputs→128 neurons in the first hidden layer→64 neurons in the second hidden layer→32 neurons in the third hidden layer→1 neuron output layer, to predict the final tensile performance of the material. All hidden layers use ReLU activation function, and mean square error is used as the loss function for training. Finally, the comprehensive performance of the material is evaluated through the environmental harshness coefficient calculation and the performance evaluation index. This method is only applicable to the specific performance prediction of high polymer materials and cannot be extended to the multi-performance prediction of other types of materials such as two-dimensional materials MXenes. Using traditional multilayer perceptron network, the effective representation of atomic-level structure information of materials is lacking, and the complex interactions and geometric spatial relationships between atoms cannot be captured. The environmental harshness coefficient calculation uses a simple linear weighting method and does not consider the nonlinear coupling effects between different environmental factors and the intrinsic characteristics of the material structure. This method is mainly based on experimental data and relies on a large number of physical test sample preparations, which is time-consuming and cannot meet the demand for high-throughput material screening and rapid performance prediction. SUMMARY

[0005] In view of the deficiencies of the prior art, the present application provides a MXenes material performance prediction method based on a graph neural network.

[0006] The method comprises:

[0007] Obtaining structure characteristic parameters of the MXenes material to be predicted, the structure characteristic parameters comprising atomic coordinates, atomic types, lattice parameters and edge characteristic information, the edge characteristic information comprising interatomic distances, angle information and chemical bond types;

[0008] Based on the structure characteristic parameters, respectively matching node-derived information, edge-derived information and global-derived information;

[0009] Based on the node-derived information, the edge-derived information and the global-derived information, constructing an initial feature vector set, wherein the initial feature vector set comprises a node feature vector constructed based on the node-derived information, an edge feature vector constructed based on the edge-derived information and a global feature vector constructed based on the global-derived information.

[0010] perform hierarchical processing on each feature vector in the initial feature vector set through each layer of the graph neural network to obtain the target performance index of the MXenes material to be predicted.

[0011] In some embodiments, the obtaining of the structural characteristic parameter of the MXenes material to be predicted comprises:

[0012] obtaining the MXenes material to be predicted from a database;

[0013] performing structural file analysis on the MXenes material to be predicted through a density functional theory, calculating performance data, and performing secondary verification on the performance data measured by a model to obtain the structural characteristic parameter.

[0014] In some embodiments, the matching of the node-derived information, the edge-derived information and the global-derived information based on the structural characteristic parameter respectively comprises:

[0015] matching atomic attribute information of each atom based on the atomic type to obtain the node-derived information, the atomic attribute information at least including atomic number, valence and atomic radius;

[0016] encoding interaction features between atoms based on the edge feature information to obtain edge-derived information for describing atom pair interaction information;

[0017] determining a target extraction dimension of the global-derived information;

[0018] extracting each dimension information matching the target extraction dimension based on the atomic coordinates and the atomic type to obtain the global-derived information.

[0019] In some embodiments, the constructing of the initial node feature vector based on the node-derived information comprises:

[0020] inputting the node-derived information into a multi-layer perception machine based on a mapping function to map the node-derived information into a high-dimensional hidden space vector to obtain the initial node feature vector represented based on the high-dimensional hidden space vector;

[0021] wherein the mapping function is:

[0022] (1);

[0023] wherein, is a node feature vector of an atomic node i, ELU is an exponential linear unit activation function, is an original feature vector of the atomic node i, is an embedding weight matrix, is a bias term.

[0024] In some embodiments, constructing an initial edge feature vector based on the edge-derived information comprises:

[0025] Combining encoding of the edge-derived information based on Gaussian radial basis functions and spherical harmonic functions to generate an edge edge feature vector representing the interaction of the atom pair, wherein the edge feature vector is represented as:

[0026] (2).

[0027] In the formula, is the feature vector of edge (i, j), RBF is Gaussian radial basis function, SBF is spherical harmonic function, is distance encoding, which refers to the Euclidean distance between atom node i and atom node j:

[0028] (3).

[0029] is angle encoding, which is a key to the MXenes layered structure:

[0030] (4).

[0031] In some embodiments, constructing a global feature vector based on the global-derived information comprises:

[0032] Aggregating mapping of the global-derived information through a multi-layer perception mechanism or a nonlinear change structure to obtain the global feature vector, wherein the global feature vector is represented as:

[0033] (5).

[0034] In the formula, represents the global feature vector of the MXenes material to be predicted, is an aggregation function, is the node feature vector of the i-th atom node.

[0035] In some embodiments, the layered processing of each feature vector in the initial feature vector set through each layer of the graph neural network to obtain the target performance indicator of the MXenes material to be predicted comprises:

[0036] Processing the input node feature vector and edge feature vector through the multi-head attention graph convolution layer of the graph neural network combined with attention weights to obtain the graph convolution feature vector of the MXenes material to be predicted, and then inputting the graph convolution feature vector to the global attention layer of the graph neural network;

[0037] The global attention layer of the graph neural network receives the graph convolution feature vector and performs correlation analysis to generate a global attention enhanced feature. The global attention enhanced feature is processed through a feature reweighting mechanism to obtain a global correlation feature vector of the to-be-predicted MXenes material, which is input to a fusion prediction layer of the graph neural network.

[0038] The fusion prediction layer of the graph neural network obtains the global feature vector and the global correlation feature vector input by the global attention layer, performs fusion prediction on the global feature vector and the global correlation feature vector, and outputs the predicted MXenes material target performance index.

[0039] In some embodiments, the attention weight is determined based on the following formula:

[0040] (6) ;

[0041] wherein, is the attention weight of atom node i to atom node j in the kth attention head, , are query and key vectors respectively, is the attention head dimension, is the attention weight matrix, is the neighbor set of atom node i.

[0042] In some embodiments, the processing of the global attention enhanced feature through the feature reweighting mechanism to obtain the global correlation feature vector of the to-be-predicted MXenes material includes:

[0043] In combination with the multi-head attention function, the covalent bond interaction of atoms in the same layer is reweighted;

[0044] In combination with the cross-layer attention function, the van der Waals force interaction of atoms across layers is reweighted;

[0045] wherein, the multi-head attention function is:

[0046] (7) ;

[0047] The cross-layer attention function is:

[0048] (8) ;

[0049] wherein, is the intra-layer attention feature of the lth layer, is the inter-layer attention feature of the lth layer, is the multi-head attention function, is the cross-layer attention function.

[0050] In some embodiments, the global feature vector and the global correlation feature vector are fused for prediction, and outputting the predicted target performance index of the MXenes material comprises:

[0051] The global correlation feature vector is transformed by a multi-layer MLP after being weighted and pooled by attention, and a global transformed feature vector is obtained;

[0052] The global transformed feature vector and the global feature vector are fused for prediction to output the target performance index;

[0053] Wherein, the fully connected network layer is according to the following formula:

[0054] (9) ;

[0055] (10) ;

[0056] (11) ;

[0057] Wherein, is the output of the first layer, r is the fusion feature vector, is the output of the lth layer, y is the final prediction value, and are the weight matrix and bias term of the 0th layer respectively, and are the weight matrix and bias term of the lth layer respectively, and are the weight matrix and bias term of the output layer.

[0058] The present application includes but is not limited to the following benefits: (1) The scheme combines the graph neural network technology with the unique layered structure characteristics of MXenes materials, effectively solves the key technical problems such as insufficient geometric information coding and inability to effectively distinguish intra-layer and inter-layer interactions in the prior art through the combination of Gaussian radial basis function and spherical harmonic function coding, multi-level feature fusion mechanism and other technical means; (2) The scheme realizes the collaborative modeling of local interaction and global structure attribute through the fusion layer, which not only retains the fineness of local features, but also supplements the integrity of global information, finally realizes the collaborative modeling of local interaction and global structure attribute through the fusion layer; (3) The present application is specially designed for the unique two-dimensional layered structure characteristics of MXenes materials, and the geometric information of the layered structure is specially coded through the angle coding formula, which can accurately capture the difference between the intra-layer covalent bond interaction and the inter-layer van der Waals force interaction of MXenes materials, and solve the technical problem that the traditional method cannot effectively represent the layered structure of MXenes; (4) Compared with the traditional density functional theory calculation method which requires a large amount of computing resources and time cost, the present application can realize fast prediction based on the trained graph neural network model, greatly reducing the computing cost and improving the material screening efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0059] 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 prior art description will be briefly introduced as follows.

[0060] Figure 1 is a flowchart of the MXenes material performance prediction method based on the graph neural network of the present application embodiment;

[0061] Figure 2 is a flowchart of obtaining the structure characteristic parameters of the MXenes material to be predicted of the present application embodiment;

[0062] Figure 3 is a flowchart of matching and deriving information based on the structure characteristic parameters of the present application embodiment;

[0063] Figure 4 is a flowchart of determining the target performance index of the predicted MXenes material of the present application embodiment;

[0064] Figure 5 is a prediction result verification diagram of the MXenes material performance prediction method based on the graph neural network of the present application embodiment. DETAILED DESCRIPTION

[0065] The embodiment of the application provides a MXenes material performance prediction method based on a graph neural network, and the method comprises the following steps: acquiring structure characteristic parameters of a to-be-predicted MXenes material, wherein the structure characteristic parameters comprise atomic coordinates, atomic types, lattice parameters and edge characteristic information, and the edge characteristic information comprises atomic distance, angle information and chemical bond types; based on the structure characteristic parameters, node derivative information, edge derivative information and global derivative information are matched respectively; based on the node derivative information, the edge derivative information and the global derivative information, an initial feature vector set is constructed, wherein the initial feature vector set comprises a node feature vector constructed based on the node derivative information, an edge feature vector constructed based on the edge derivative information and a global feature vector constructed based on the global derivative information; each feature vector in the initial feature vector set is processed in layers through each layer of the graph neural network, so as to obtain a target performance index of the to-be-predicted MXenes material. The scheme combines the graph neural network technology and the specific layered structure characteristics of the MXenes material, and effectively solves the key technical problems such as insufficient geometric information coding and inability to effectively distinguish intra-layer and inter-layer interactions in the prior art by means such as combined coding of Gaussian radial basis functions and spherical harmonic functions and a multi-level feature fusion mechanism.

[0066] The terms "first", "second", "third", "fourth" and the like in the description, claims, and drawings of the application, and those above and below (if any) are used for distinguishing between similar objects and not necessarily for describing a specific sequential or chronological order. It is to be understood that the use of the terms so termed herein is to be interpreted to allow that combinations of the features of the various embodiments (and of the various claims) can provide particularly desired or expeditious results, either as combined in one application, or in two or more applications simultaneously, without necessarily excluding other embodiments or applications from the technology that can provide similar benefits.

[0067] For the sake of understanding, the specific flow of the embodiment of the application is described below, in particular, Figure 1 The flowchart of the MXenes material performance prediction method based on the graph neural network of the embodiment of the application is shown in Figure 1 , comprising the following steps:

[0068] S100, acquiring structure characteristic parameters of a to-be-predicted MXenes material.

[0069] The structure characteristic parameters comprise atomic coordinates, atomic types, lattice parameters and edge characteristic information, and the edge characteristic information comprises atomic distance, angle information and chemical bond types.

[0070] In particular, Figure 2As another flowchart of the MXenes material performance prediction method based on the graph neural network of the embodiments of the present application, the flowchart is an exemplary illustration of step S100, see 2, and includes the following steps:

[0071] S200, obtaining a to-be-predicted MXenes material from a database.

[0072] S202, performing structural file analysis on the to-be-predicted MXenes material by density functional theory, calculating performance data, and performing secondary verification with the performance data measured by the model to obtain structural characteristic parameters.

[0073] In an example, 23,857 MXenes structure samples can be obtained from the MXenes Database database (website https: / / anant.mrc.iisc.ac.in / apps / mxene), each sample containing structural parameters and performance parameters calculated by density functional theory (DFT).

[0074] The specific operation is as follows:

[0075] a) Analyze the structural file to obtain atomic coordinates, atomic types, and lattice parameters;

[0076] b) Represent each MXenes crystal structure as G = (V, E), where node V represents an atom and edge E represents an interaction relationship between atoms with a distance less than 8 Å;

[0077] c) Obtain performance parameters: magnetic moment (μB), band gap (eV), and lattice constant (Å);

[0078] d) Divide the dataset into a training set of 19,085, a validation set of 2,386, and a test set of 2,386 in a ratio of 8:1:1.

[0079] S102, based on the structural characteristic parameters, respectively matching node derived information, edge derived information and global derived information.

[0080] Specifically, Figure 3 As another flowchart of the MXenes material performance prediction method based on the graph neural network of the embodiments of the present application, the flowchart is an exemplary illustration of step S102, see 3, and includes the following steps:

[0081] S300, based on the atomic type, matching the atomic attribute information of each atom to obtain node derived information, the atomic attribute information at least including atomic number, valence and atomic radius.

[0082] S302, encode the interaction feature between atoms based on the edge feature information to obtain edge derived information for describing atom pair interaction information.

[0083] S304, determine the target extraction dimension of the global derived information.

[0084] S306, based on the atomic coordinates and atomic types, extract each dimension information matched with the target extraction dimension to obtain the global derived information.

[0085] S104, based on the node derived information, the edge derived information and the global derived information, construct an initial feature vector set.

[0086] The initial feature vector set includes a node feature vector constructed based on the node derived information, an edge feature vector constructed based on the edge derived information, and a global feature vector constructed based on the global derived information.

[0087] Specifically, constructing the initial node feature vector based on the node derived information includes:

[0088] Based on the mapping function, the node derived information is input into the multi-layer perception machine to map the node derived information into a high-dimensional hidden space vector, and an initial node feature vector represented based on the high-dimensional hidden space vector is obtained;

[0089] The mapping function is:

[0090] (1);

[0091] In the formula, is the node feature vector of the atomic node i, ELU is the exponential linear unit activation function, is the original feature vector of the atomic node i, is the embedding weight matrix, is the bias term.

[0092] In an example, the atomic number, valence, atomic radius and other 14-dimensional original features of the atom are mapped to a 300-dimensional hidden space through the multi-layer perception machine, and the node feature vector is obtained by calculating formula (1).

[0093] Constructing the initial edge feature vector based on the edge derived information includes:

[0094] The edge derived information is combined and encoded based on the Gaussian radial basis function and the spherical harmonic function to generate an edge feature vector representing the interaction of the atom pair, wherein the edge feature vector is represented as:

[0095] (2);

[0096] In the formula, is the eigenvector of edge (i, j), RBF is Gaussian radial basis function, SBF is spherical harmonic function, is distance encoding, which refers to the Euclidean distance between atom node i and atom node j:

[0097] (3).

[0098] is angle encoding, which is a key for MXenes layered structure:

[0099] (4).

[0100] In an example, 64 Gaussian radial basis functions are adopted; angle encoding: spherical harmonic function is adopted to encode the angle information between atoms, and is specially designed for MXenes layered structure, and edge eigenvectors are obtained by using formulas (2), (3) and (4) for calculation.

[0101] The global eigenvector is constructed based on the global derived information, which includes:

[0102] The global derived information is aggregated and mapped by a multi-layer perception mechanism or a nonlinear change structure to obtain a global eigenvector, wherein the global eigenvector is represented as follows:

[0103] (5).

[0104] In the formula, represents the global eigenvector of the MXenes material to be predicted, is an aggregation function, is the node eigenvector of the i-th atom node.

[0105] In an example, c) extracts 6-dimensional global features such as total number of atoms, average atomic mass, total number of electrons, etc. of the material, and the global eigenvector is obtained by using formula (5) for calculation.

[0106] S106, each eigenvector in the initial eigenvector set is processed by layers of the graph neural network to obtain a target performance index of the MXenes material to be predicted.

[0107] Specifically, Figure 4 is another flowchart of the MXenes material performance prediction method based on the graph neural network of the embodiments of the present application, which is a specific description of step S106, and refer to Figure 4 , including the following steps:

[0108] S400, the node feature vector and the edge feature vector input are processed by the multi-head attention graph convolution layer of the graph neural network combined with the attention weight, and the graph convolution feature vector of the to-be-predicted MXenes material is obtained and input to the global attention layer of the graph neural network.

[0109] In an example, a multi-head attention graph convolution network containing 20 graph convolution layers is constructed herein, each layer using 8 attention heads, and the attention head dimension is 37.5 (300 / 8):

[0110] a) learn the interatomic interaction through the multi-head attention mechanism, and calculate the attention weight by formula (6);

[0111] b) add residual connection and layer normalization after each graph convolution layer to prevent gradient vanishing. Specifically, the attention weight is determined based on the following formula:

[0112] (6) ;

[0113] In the formula, is the attention weight of atom node i to atom node j in the kth attention head, , are the query and key vectors respectively, is the attention head dimension, is the attention weight matrix, is the neighbor set of atom node i.

[0114] S402, the global attention layer of the graph neural network receives the graph convolution feature vector and performs correlation analysis to generate a global attention enhanced feature, and processes the global attention enhanced feature through a feature reweighting mechanism to obtain a global correlation feature vector of the to-be-predicted MXenes material, and inputs the global correlation feature vector to the fusion prediction layer of the graph neural network.

[0115] Specifically, the covalent bond interaction of atoms in the same layer is reweighted in combination with the multi-head attention function;

[0116] In combination with the cross-layer attention function, the van der Waals force interaction of atoms across layers is reweighted;

[0117] Wherein, the multi-head attention function is:

[0118] (7) ;

[0119] The cross-layer attention function is:

[0120] (8) ;

[0121] In the formula, is the intra-layer attention feature of the lth layer, inter-layer attention feature of the first layer, multi-head attention function, cross-layer attention function.

[0122] S404, the fusion prediction layer of the graph neural network obtains the global feature vector and the global correlation feature vector input by the global attention layer, fuses and predicts the global feature vector and the global correlation feature vector, and outputs the predicted MXenes material target performance index.

[0123] Specifically, the fusion prediction of the global feature vector and the global correlation feature vector includes:

[0124] The global correlation feature vector is transformed by the multi-layer MLP to obtain the global transformed feature vector;

[0125] The global transformed feature vector and the global feature vector are fused and predicted to output the target performance index;

[0126] In an example, the MLP includes three layers:

[0127] a) the first layer: 512-dimensional input is mapped to 256-dimensional, using ELU activation function and 0.1 dropout;

[0128] b) the second layer: 256-dimensional is mapped to 128-dimensional, using ELU activation function and 0.1 dropout;

[0129] c) the output layer: 128-dimensional is mapped to 1-dimensional prediction value, and is calculated by formula (9)-(11).

[0130] The fully connected network layer is according to the following formula:

[0131] (9) ;

[0132] (10) ;

[0133] (11) ;

[0134] In the formula, is the output of the first layer, r is the fusion feature vector, is the output of the lth layer, y is the final prediction value, and are the weight matrix and the bias term of the 0th layer respectively, and are the weight matrix and the bias term of the lth layer respectively, and The weight matrix and bias term of the output layer are output.

[0135] Further, in the present example, the graph neural network is pre-trained, and the training parameters are set as follows:

[0136] • Optimizer: Adam, initial learning rate 1x10⁻³, weight decay 0;

[0137] • Learning rate scheduling: reduced to 1x10⁻ 4 after 50 epochs;

[0138] • Loss function: mean absolute error (MAE);

[0139] • Number of training rounds: 100 epochs;

[0140] • Early stopping strategy: stop if the validation set loss does not improve for 10 consecutive epochs.

[0141] The prediction results on the test set of 2,386 samples are shown in Table 1:

[0142] Table 1

[0143] Performance metrics MAE RMSE R² Magnetic moment 0.0807 μΒ 0.1123 μΒ 0.9513 Bandgap 0.0097 eV 0.0134 eV 0.9748 Lattice constant 0.0032 Å 0.0045 Å 0.9986

[0144] According to the evaluation criteria, the prediction is considered accurate when MAE ≤ 0.1 μB (magnetic moment), 0.02 eV (band gap), 0.005 Å (lattice constant) and RMSE ≤ 0.15 μB (magnetic moment), 0.03 eV (band gap), 0.008 Å (lattice constant). All indicators of the present embodiment meet the accurate prediction criteria, and the prediction result verification graph is shown in Figure 5 , where (a) is the lattice constant result graph, (b) is the band gap result graph, and (c) is the magnetic moment result graph.

[0145] The above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or make equivalent replacements for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for predicting the properties of MXenes materials based on graph neural networks, characterized in that, The method includes: Obtain the structural feature parameters of the MXenes material to be predicted. The structural feature parameters include atomic coordinates, atomic type, lattice parameters, and edge feature information. The edge feature information includes interatomic distance, angle information, and chemical bond type. Based on the structural feature parameters, node-derived information, edge-derived information, and global-derived information are matched respectively. Based on the node-derived information, edge-derived information, and global-derived information, an initial feature vector set is constructed, wherein the initial feature vector set includes node feature vectors constructed based on the node-derived information, edge feature vectors constructed based on the edge-derived information, and global feature vectors constructed based on the global-derived information. The graph neural network uses a multi-head attention graph convolutional layer with attention weights to process the input node feature vector and edge feature vector, and then inputs the graph convolutional feature vector of the MXenes material to be predicted into the global attention layer of the graph neural network. The global attention layer of the graph neural network receives the graph convolution feature vector and performs correlation analysis to generate global attention enhancement features. The global attention enhancement features are then processed through a feature reweighting mechanism to obtain the global correlation feature vector of the MXenes material to be predicted, which is then input into the fusion prediction layer of the graph neural network. The fusion prediction layer of the graph neural network acquires the global feature vector and receives the global associated feature vector input from the global attention layer, performs fusion prediction on the global feature vector and the global associated feature vector, and outputs the predicted target performance index of the MXenes material. The construction of edge feature vectors based on the edge-derived information includes: The edge-derived information is combined and encoded using Gaussian radial basis functions and spherical harmonic functions to generate edge feature vectors representing the interactions between atomic pairs. These edge feature vectors are characterized as follows: （2）； In the formula, Let be the eigenvector of edge (i,j), RBF be the Gaussian radial basis function, and SBF be the spherical harmonic function. It is distance encoding, referring to the Euclidean distance between atomic node i and atomic node j: （3）； It is angle encoding. （4）。 2. The method for predicting MXenes material properties based on graph neural networks according to claim 1, characterized in that, The process of obtaining the structural feature parameters of the MXenes material to be predicted includes: Retrieve the MXenes materials to be predicted from the database; The structural files of the MXenes material to be predicted are parsed using density functional theory, performance data is calculated, and secondary verification is performed with the performance data measured by the model to obtain the structural characteristic parameters.

3. The method for predicting MXenes material properties based on graph neural networks according to claim 1, characterized in that, The matching of node-derived information, edge-derived information, and global-derived information based on the structural feature parameters includes: Based on the atom type, the atomic attribute information of each atom is matched to obtain the node-derived information. The atomic attribute information includes at least atomic number, chemical valence, and atomic radius. Based on the edge feature information, the interaction features between atoms are encoded to obtain edge-derived information used to describe the interaction information of atom pairs; Determine the target extraction dimension of the global derived information; Based on the atomic coordinates and the atomic type, information of each dimension that matches the target extraction dimension is extracted to obtain the global derived information.

4. The method for predicting MXenes material properties based on graph neural networks according to claim 1, characterized in that, Constructing the initial node feature vector based on the node-derived information includes: Based on the mapping function, the node-derived information is input into a multilayer perceptron to map the node-derived information into a high-dimensional hidden space vector, thereby obtaining the initial node feature vector represented by the high-dimensional hidden space vector. The mapping function is: （1）； In the formula, Let be the node feature vector of atomic node i, and ELU be the exponential linear unit activation function. The original feature vector of atomic node i, To embed the weight matrix, This is a bias term.

5. The method for predicting MXenes material properties based on graph neural networks according to claim 1, characterized in that, Constructing a global feature vector based on the aforementioned global derived information includes: The global derived information is aggregated and mapped using a multi-layer sensing mechanism or a non-linear transformation structure to obtain the global feature vector, wherein the global feature vector is characterized as follows: （5）； In the formula, This represents the global feature vector of the MXenes material to be predicted. It is an aggregate function. V is the node feature vector of the i-th atomic node, where V is the atomic node.

6. The method for predicting MXenes material properties based on graph neural networks according to claim 1, characterized in that, Attention weights are determined based on the following formula: （6）； In the formula, Let be the attention weight of atomic node i to atomic node j in the k-th attention head. , These are the query and key vectors, respectively. For the attention head dimension, Here is the attention weight matrix. Let i be the set of neighbors of atomic node i.

7. The method for predicting MXenes material properties based on graph neural networks according to claim 1, characterized in that, The process of processing the global attention-enhanced features through a feature reweighting mechanism to obtain the global association feature vector of the MXenes material to be predicted includes: By incorporating a multi-head attention function, the covalent interactions of atoms in the same layer are reweighted; By incorporating a cross-layer attention function, the van der Waals interactions between cross-layer atoms are reweighted.

8. The method for predicting MXenes material properties based on graph neural networks according to claim 1, characterized in that, The global feature vector and the globally correlated feature vector are fused and predicted to output the predicted target performance index of the MXenes material, including: Attention-weighted pooling and fully connected network layers are used to nonlinearly transform the global correlation feature vector through a multi-layer MLP to obtain the global transformed feature vector. The global transformation feature vector and the global feature vector are fused and predicted to output the target performance index; The fully connected network layer is based on the following formula: （9）； （10）； （11）； In the formula, The first layer output is r, where r is the fused feature vector. Let y be the output of layer l, and y be the final predicted value. and These are the weight matrix and bias term for layer 0, respectively. and These are the weight matrix and bias term of the l-th layer, respectively. and Here, represents the weight matrix and bias term of the output layer, and ELU is the exponential linear unit activation function.

Images