Deep learning based energy prediction method for metal-organic compounds

Abstract

The application discloses a kind of metal organic compound energy prediction methods based on deep learning, belong to compound property prediction technical field.It includes the following steps: obtaining the molecular structure information of metal organic compound and calculating electronic structure characteristics, constructs molecular graph and encodes generation geometry feature vector;Global fusion features are output through multimodal feature fusion model;Total energy value and molecular structure are predicted through multi-task prediction network;The mean and variance of the total energy value are obtained by statistical prediction, and the energy prediction value and confidence interval are obtained.The application constructs molecular graph and extracts geometric features, and simultaneously utilizes cross attention and gate fusion network to realize the deep interactive fusion of geometric and electronic structure features, solving the problem of inaccurate feature representation of metal organic compounds and poor multimodal feature fusion effect of traditional methods.

Description

Technical Field

[0001] This invention relates to the field of compound property prediction technology, and more specifically, to a method for predicting the energy of organometallic compounds based on deep learning. Background Technology

[0002] Organometallic compounds have extremely high application value in energy and new materials fields such as energy storage materials, catalytic materials, and optoelectronic materials due to their tunable molecular structure and rich electronic properties. Their energy characteristics are the core indicators for evaluating material performance and guiding material design and screening.

[0003] Currently, energy calculations for organometallic compounds mainly rely on quantum chemical calculation methods based on density functional theory. Although this method can provide high-precision energy and structure data, the calculation process requires complex quantum mechanical solutions to the electronic structure, resulting in high computational costs and long processing times.

[0004] Deep learning technology has provided new ideas for predicting material properties, but there are still many problems in predicting the energy of organometallic compounds: First, the molecular structure of organometallic compounds is complex, and conventional graph neural networks are insufficient in characterizing long-range interactions or specific coordination environments when modeling interatomic interactions; Second, the energy characteristics of materials are determined by both geometric and electronic structures, and existing methods have failed to achieve deep interactive integration of the two, easily losing structure-electronic coupling information; Third, most methods only predict a single property, ignoring the intrinsic relationship between energy and molecular structure, reducing the physical consistency of the model.

[0005] Therefore, there is an urgent need to develop a deep learning method that can accurately model molecular structures, achieve deep fusion of multimodal features, and complete joint prediction of multiple properties, so as to improve computational efficiency while ensuring accuracy and provide technical support for the research and development of new metal-organic compound materials. Summary of the Invention

[0006] The purpose of this invention is to provide a deep learning-based method for predicting the energy of organometallic compounds, which can solve the problems of insufficient characterization of long-range interactions or specific coordination environments, easy loss of structure-electron coupling information, and neglect of the intrinsic correlation of various physical properties in the energy prediction of organometallic compounds in the prior art.

[0007] To achieve the above objectives, the present invention aims to provide a deep learning-based method for predicting the energy of organometallic compounds, comprising the following steps: S1. Obtain the molecular structure information of organometallic compounds, and obtain the electronic structure characteristics through quantum chemical calculations and coordinate alignment; among which, the molecular structure information includes atomic coordinates and atomic numbers; S2. Identify atoms and edges in the molecular structure information to construct a molecular graph; S3. Encode the nodes and edges of the molecular graph, and update the node representation through a multi-layer graph convolutional network to generate geometric feature vectors; S4. Input the geometric feature vector and electronic structure feature into the multimodal feature fusion model for feature fusion, output the fused features, and further extract features through global average pooling to obtain the global fused features; the multimodal feature fusion model includes a cross-attention mechanism module and a gated fusion network module; S5. Input the global fusion features into the multi-task prediction network and output the predicted total energy value and predicted molecular structure of the organometallic compound; wherein, the multi-task prediction network includes a shared coding layer and a task-specific decoding layer, and the task-specific decoding layer includes two independent branches: a total energy decoding layer and a molecular structure decoding layer. S6. By statistically predicting the mean and variance of the total energy value, the predicted energy value and its confidence interval are obtained.

[0008] As a further improvement to this technical solution, the method for obtaining the electronic structure features includes: Molecular structure files of organometallic compounds are obtained from public databases, and molecular structure information, including atomic coordinates and atomic numbers, is obtained through high-throughput modeling. Quantum chemical calculation software based on density functional theory was used to calculate and extract the force situation, electrostatic potential and orbital occupancy number of each atom, and then spliced ​​them in a fixed order to form the initial electronic structure features. The initial electronic structure features are aligned with atomic coordinates, and all electronic structure features are normalized to obtain the electronic structure features. .

[0009] As a further improvement to this technical solution, the method for constructing the molecular map includes: Based on atomic coordinates, the spatial distance between all atomic pairs is calculated using the Euclidean distance method. ; If spatial distance Not exceeding the preset distance threshold Then in the atom and Establishing edges Constructing a molecular map ,in, For a set of nodes, Let it be the set of edges.

[0010] As a further improvement to this technical solution, the encoding of nodes and edges of the molecular graph includes: Atom type embedding is obtained by one-hot encoding based on atomic number, combined with its corresponding atomic coordinates and electronic structure features. , splicing together to form nodes in the molecular diagram initial feature vector ; For each connection edge, based on the element types of the atoms at both ends, the corresponding atom type combination code is generated by looking up the preset atom type pair encoding mapping table. Using the spatial distance between opposite sides of the radial basis function set Encode to obtain distance feature vector The edges are then combined with the corresponding atom type codes to obtain the edges in the molecular graph. initial feature vector .

[0011] As a further improvement to this technical solution, the method for obtaining the geometric feature vector includes: The initial eigenvectors of the nodes in the molecular graph The initial eigenvectors of the edges Input a multi-layer graph convolutional network, and iteratively aggregate the features of each node itself and the features of its neighboring nodes and edges to update the node representation layer by layer; go through After passing the layer graph through convolution, we obtain the geometric feature vector corresponding to each node. .

[0012] As a further improvement to this technical solution, the method for obtaining the fusion feature includes: The geometric feature vector and electronic structure feature are input into the cross-attention mechanism module of the multimodal feature fusion model, and the attention feature vector is output. Attention feature vectors and geometric feature vectors are input into a gated fusion network module for dynamic weight fusion, outputting fused features. .

[0013] As a further improvement to this technical solution, the method for obtaining the global fusion feature includes: The fused features are input into a two-layer fully connected neural network, and after dimensionality expansion and compression, the refined features are obtained. ; Global average pooling is performed on the refined features of all atoms to obtain the globally fused features of the organometallic compound. .

[0014] As a further improvement to this technical solution, the method for obtaining the predicted total energy value includes: The globally fused features are input into the shared coding layer, and a deep shared representation is obtained through layer-by-layer nonlinear transformation. The shared coding layer consists of three fully connected layers. Deep shared representation Input the total energy decoding layer and output the predicted total energy value of the organometallic compound. The total energy decoding layer consists of two fully connected networks.

[0015] As a further improvement to this technical solution, the methods for obtaining predicted molecular structures include: For each atom Share its geometric feature vectors with the deep layer representation. The data is spliced ​​together, input to the molecular structure decoding layer, and outputted as the predicted coordinate offset of the atom. The molecular structure decoding layer consists of two fully connected networks. For each atom Offset its atomic coordinates from the corresponding predicted coordinates Adding them together gives the predicted steady-state structure coordinates of the atom; The final output is the predicted steady-state structure coordinates of all atoms, i.e., the predicted molecular structure.

[0016] As a further improvement to this technical solution, the method for obtaining the energy prediction value and its confidence interval includes: After training the multi-task prediction network, multiple forward propagations are performed on the same organometallic compound sample during the prediction phase. Random perturbations are introduced in each forward propagation to obtain a set of total energy prediction values ​​with random differences. ;in, Forward propagation count; Based on the set of total energy prediction values, calculate their mean as the energy prediction value. And calculate its variance, thereby obtaining the prediction standard deviation. ; Based on energy forecasts and the predicted standard deviation Construct a confidence level of confidence interval .

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. In this deep learning-based method for predicting the energy of metal-organic compounds, a molecular graph is constructed and geometric features are extracted using a graph convolutional network. At the same time, a deep interactive fusion of geometric and electronic structural features is achieved through a cross-attention and gating fusion network. Compared with traditional methods that can only extract shallow artificial features or simply splice multimodal features, this method fully explores the coupling relationship between structural and electronic features, solves the problems of inaccurate feature representation and poor fusion effect, and improves the quality of basic features for energy prediction.

[0018] 2. In this deep learning-based method for predicting the energy of organometallic compounds, a multi-task prediction network is designed to simultaneously predict the total energy and molecular structure, and to quantify the uncertainty of the prediction results. Compared with the traditional single-task prediction method without confidence assessment, this method utilizes the intrinsic physical relationship between energy and molecular structure to improve prediction accuracy, and at the same time provides a confidence interval for the energy prediction value, solving the problem of lack of reliable reference for the prediction results, and providing a more valuable decision basis for material screening and design. Attached Figure Description

[0019] Figure 1 This is a flowchart of the overall method of the present invention. Detailed Implementation

[0020] 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 skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Example: Please see Figure 1 As shown, this embodiment provides a deep learning-based method for predicting the energy of organometallic compounds, including the following steps: S1. Obtain the molecular structure information of organometallic compounds, and obtain the electronic structure characteristics through quantum chemical calculations and coordinate alignment; among which, the molecular structure information includes atomic coordinates and atomic numbers; In this embodiment, the molecular structure file of the organometallic compound is obtained from a public database, and high-throughput modeling is used to obtain the atomic coordinates. and atomic number Molecular structure information in, The first in the representative molecular structure The three-dimensional spatial coordinates of an atom (usually in Å or nm). This represents the set of coordinates of all atoms, encompassing the spatial position information of every atom in the organometallic compound; For the first Atomic number of each atom, This represents the set of atomic numbers of all atoms in a molecule; Calculations were performed using quantum chemical computational software based on density functional theory (such as VASP or Quantum ESPRESSO) to extract the force situation of each atom. electrostatic potential and the number of orbital occupancy They are then assembled in a fixed order to form the initial electronic structure features. ; in, For the force situation, indicate the first The force vector of an atom in three-dimensional space, with units of eV / Å; For the first Electrostatic potential at each atom; Indicates correspondence The number of electrons filling the atomic orbitals of a quantum number. Angular quantum number representing atomic orbitals (e.g.) =0 corresponds to the s orbital. =1 corresponds to the p orbital. =2 corresponds to the d orbital). Represents the magnetic quantum number (ranging from -1 to +1). It is the set of all atomic orbital occupancy numbers; For the first The initial electronic structure characteristics of an atom integrate electronic structure information including the force conditions, electrostatic potential, and orbital occupancy number of that atom. For feature dimensions (e.g.) ); The initial electronic structure features are aligned with atomic coordinates, and all electronic structure features are Z-score normalized to obtain the electronic structure features. ;

[0022] in, The mean of the corresponding feature. The standard deviation of the corresponding feature. These are the standardized electronic structure features.

[0023] S2. Identify atoms and edges in the molecular structure information to construct a molecular graph; In this embodiment, the spatial distance between all atomic pairs is calculated using the Euclidean distance method based on atomic coordinates. ;

[0024] in, The first in an organometallic compound molecule The atom and the first Spatial distance between atoms (in Å); Let be the Euclidean norm of the vector; If spatial distance Not exceeding the preset distance threshold (like (Based on the typical metal-ligand bond length range), then in the atom and Establishing edges Constructing a molecular map ,in, For a set of nodes (atoms). For a set of edges (bonds or interactions); in, Molecular diagrams constructed for organometallic compounds. For the set of nodes in the molecular graph, Let be the set of edges of the molecular graph; Connecting the first in the molecular diagram The atomic node and the first The edges of each atomic node are an abstract graph structure representing the bonding / interaction between atoms.

[0025] S3. Encode the nodes and edges of the molecular graph, and update the node representation through a multi-layer graph convolutional network to generate geometric feature vectors; In this embodiment, atom type embedding is obtained by one-hot encoding based on atomic number, combined with its corresponding atomic coordinates and electronic structure features. , splicing together to form nodes in the molecular diagram initial feature vector ; in, Molecular diagram The Middle The initial node features of each atomic node are determined by atomic type embedding and atomic coordinates. and electronic structure characteristics It is pieced together; The dimension of the initial node feature vector, such as ; For each connection edge, based on the element types of the atoms at both ends, the corresponding atom type combination code is generated by looking up the preset atom type pair encoding mapping table. Among them, the atomic type combination encoding is the first The atom and the first The atomic number of each atom is used to encode features (such as one-hot encoding or embedding encoding) to characterize the combination of atom types connected by the edge (such as metal-carbon, metal-nitrogen, etc.); the atom type pair encoding mapping table is a pre-defined lookup table that maps any two atomic element types (such as Fe-O, CH) to a fixed-length unique integer or vector, used to distinguish different chemical bond types or interaction categories in graph neural networks.

[0026] Using the spatial distance between opposite sides of the radial basis function set Encode to obtain distance feature vector The edges are then combined with the corresponding atom type codes to obtain the edges in the molecular graph. initial feature vector ; For each edge Extend distance information using radial basis functions:

[0027] in, The first one after radial basis function expansion - The distance feature vector of each edge; It is a natural exponential function; For the preset distance from the center point (e.g.) ), The width hyperparameter of the radial basis functions. This represents the total distance from the center point; The distance from the center point set is the first One center value (in Å). The index of the center point; Edges in the molecular graph The assigned initial feature vector includes spatial distance Atomic type combination encoding; The initial eigenvectors of the nodes in the molecular graph The initial eigenvectors of the edges Input a multi-layer graph convolutional network, and iteratively aggregate the features of each node itself and the features of its neighboring nodes and edges to update the node representation layer by layer; go through After passing the convolutional vector through a layered graph (e.g., 3 layers), the geometric feature vector corresponding to each node is obtained. ; Specifically, a 3-layer graph convolutional network is used to update the node representation:

[0028] in, For graph convolutional networks The first layer Feature vectors of atomic nodes This represents the feature vector of the nodes in the current layer. The SiLU activation function; For nodes A set of neighbors, where the elements in the set satisfy a distance threshold. 、 and the All atomic nodes connected by an edge exist at a given node. ; This represents element-wise multiplication; For graph convolutional networks The node features of the layer undergo a linear transformation weight matrix. For graph convolutional networks The neighbor node feature linear transformation weight matrix of the layer is a learnable parameter matrix; For the output of the first Geometric feature vectors of atomic nodes, This represents the total number of convolutional layers in the graph convolutional network. denoted as the dimension of the geometric eigenvectors.

[0029] S4. Input the geometric feature vector and electronic structure feature into the multimodal feature fusion model for feature fusion, output the fused features, and further extract features through global average pooling to obtain the global fused features; the multimodal feature fusion model includes a cross-attention mechanism module and a gated fusion network module; In this embodiment, the geometric feature vector With electronic structure characteristics It serves as the cross-attention mechanism module for the input multimodal feature fusion model and outputs attention feature vectors.

[0030] in, For geometric modal query vectors, For electronic mode bond vectors, This is the electronic mode value vector; For querying the matrix, The key matrix, The matrix consists of value matrices, all of which are learnable linear transformation matrices; For the key attention dimension, For the attention dimension, This is the scaling factor; The softmax function normalizes the attention score into a probability distribution, ensuring that the sum of the weights is 1. For attention mechanism functions, This is the attention feature vector; Attention feature vectors and geometric feature vectors are input into a gated fusion network module for dynamic weight fusion, outputting fused features. ;

[0031]

[0032] in, For the gated weight vector, It is the Sigmoid activation function. The learnable gated weight matrix, This represents a vector concatenation operation; To integrate features, it combines geometric and electronic structural information.

[0033] Fusion features Inputting into a two-layer fully connected neural network, after dimensionality expansion and compression, yields the refined features. ;

[0034] in, For the first layer of weights, For the first layer bias; Second layer weights, For the second layer bias; ReLU activation function, used to introduce non-linear expressive power; These are the characteristics after refinement; Global average pooling is performed on the refined features of all atoms to obtain the global fused features of the entire organometallic compound. ;

[0035] in, This represents the total number of atoms in the compound. For the first The refined characteristics of each atomic node; This is a global fusion feature, representing the overall structure-electronic characteristics of the entire compound.

[0036] S5. Input the global fusion features into the multi-task prediction network and output the predicted total energy value and predicted molecular structure of the organometallic compound; wherein, the multi-task prediction network includes a shared coding layer and a task-specific decoding layer, and the task-specific decoding layer includes two independent branches: a total energy decoding layer and a molecular structure decoding layer. In this embodiment, a multi-task prediction network is constructed, which includes a shared coding layer and a task-specific decoding layer. The shared coding layer consists of three fully connected layers, and the task-specific decoding layer includes two independent branches: a total energy decoding layer and a molecular structure decoding layer. Each independent branch consists of two fully connected layers. Global fusion features Input a shared coding layer, and extract deep shared representations related to energy and atomic forces through layer-by-layer nonlinear transformation. The shared coding layer consists of three fully connected layers.

[0037] in, To share the weight matrix of the first layer of the coding layer, To share the weight matrix of the second layer of the coding layer, The weight matrix of the third layer of the shared coding layer; To share the bias vector of the first layer of the coding layer, For the shared second layer bias vector of the coding layer, This is the bias vector for the third layer of the shared coding layer; ReLU activation function, used to introduce non-linear expressive power; It is a deep shared representation of the output of the shared coding layer, which serves both the total energy and atomic force prediction tasks. Deep shared representation Input the total energy decoding layer and output the predicted total energy value of the organometallic compound. ;

[0038] in, This is the weight matrix of the first layer of the total energy decoding layer. This is the weight matrix of the second layer of the total energy decoding layer; The bias of the first layer of the total energy decoding layer. The bias of the second layer of the total energy decoding layer; The total energy value is predicted in eV, corresponding to the ground state energy calculated by density functional theory in practical application scenarios. During the training phase, the mean square error between the predicted total energy value and the actual total energy value from the quantum chemistry computation database is calculated and used as the total energy prediction loss function:

[0039] in, This represents the number of training samples; For the first The true total energy value of each sample is derived from quantum chemical calculations in publicly available datasets (such as calculations using VASP). For multi-task prediction networks, the first The predicted total energy value for each sample; This represents the loss function value for the total energy prediction task.

[0040] In this embodiment, for each atom Its geometric feature vector Shared representation with deep layers The data is spliced ​​together, input to the molecular structure decoding layer, and outputted as the predicted coordinate offset of the atom. The molecular structure decoding layer consists of two fully connected networks.

[0041] in, This is the weight matrix of the first layer of the molecular structure decoding layer. This is the weight matrix of the second layer of the molecular structure decoding layer; The first layer bias is used for the molecular structure decoding layer. The second layer bias is used for the molecular structure decoding layer; For the first The predicted coordinate offset of each atom is a three-dimensional vector (corresponding to the three spatial dimensions of x, y, and z), which represents the magnitude and direction of the model's adjustment to the initial atom coordinates; For each atom Offset its atomic coordinates from the corresponding predicted coordinates Adding them together gives the predicted steady-state structure coordinates of the atom;

[0042] in, The atomic coordinates input in step S1 To predict steady-state structure coordinates; The final output is the predicted steady-state structure coordinates of all atoms, i.e., the predicted molecular structure; During the training phase, the mean square error between the predicted steady-state structure coordinates and the actual steady-state structure coordinates is calculated and used as the loss function for molecular structure prediction.

[0043] in, The total number of atoms in a single organometallic compound molecule; For the first In the nth sample The true steady-state structure coordinates of each atom (usually obtained by high-precision quantum chemical calculations); For the first In the nth sample Predicted steady-state structure coordinates of each atom; The Euclidean norm of a vector; The smaller the value, the closer the predicted molecular structure is to the true structure. Based on the total energy prediction loss function and the molecular structure prediction loss function, the multi-task joint loss is formed by adding them according to the set weights, and the parameters of the multi-task prediction network are optimized by backpropagation. Multi-task joint loss:

[0044] in, The weight hyperparameters for the total energy prediction task. The weight hyperparameters for molecular structure prediction tasks are typically set based on task importance or gradient magnitude (e.g., =0.5, =0.5); The total loss function for multi-task joint training is denoted as . The parameters of the multi-task prediction network include the weight matrix and biases in the fully connected layers that share the encoding layer and the task-specific decoding layer.

[0045] S6. By statistically predicting the mean and variance of the total energy value, the predicted energy value and its confidence interval are obtained; In this embodiment, after training the multi-task prediction network, multiple forward propagations are performed on the same organometallic compound sample during the prediction phase. Random perturbations are introduced in each forward propagation to obtain a set of total energy prediction values ​​with random differences. ;in, Forward propagation number (e.g.) =50); Based on the set of total energy prediction values, calculate their mean as the energy prediction value. And calculate its variance, thereby obtaining the prediction standard deviation. ;

[0046]

[0047] in, Indicates the first The predicted total energy value (in eV) obtained from the sampling. To predict the mean energy (i.e., the predicted energy value). To predict variance; The prediction standard deviation reflects the confidence of the multi-task prediction network in predicting the current sample. The smaller the prediction standard deviation, the more concentrated the predicted values ​​from multiple samplings are, and the more reliable the prediction results are. Based on energy forecasts and the predicted standard deviation Construct a confidence level of (For example =0.05 (corresponding to a 95% confidence level) confidence interval ;

[0048] in, The upper part of the standard normal distribution Quantiles are determined by the confidence level; for example, a 95% confidence level corresponds to... ; Standard error; The confidence interval is a numerical range that represents the probability that the actual total energy value will fall within this range at a specified confidence level. This confidence interval quantifies the range of credibility of the model's prediction of total energy and can be used in practical applications to evaluate the reliability of the prediction results and assist in decision-making during the material design or screening process. It should be noted that uncertainty can be quantified in two different ways: one is by averaging multiple predictions. and its standard error The first is to construct the mean confidence interval, reflecting the model's accuracy in estimating the expected energy value; the second is to use the predicted standard deviation. Directly construct the fluctuation range of a single prediction This invention uses the mean confidence interval method, which reflects the confidence range of the model's prediction results for a single sample. The main reason for adopting this method is that in practical material screening applications, researchers typically focus on the model's stable estimates of the energy properties of a compound, rather than the random fluctuations of a single prediction. By taking the mean from multiple samples and providing its confidence interval, the overall confidence level of the model's predictions for that compound can be reflected more reliably, avoiding misjudgments caused by the randomness of a single prediction. If it is necessary to assess the fluctuation range of a single prediction, the prediction standard deviation can be used directly. As a reference indicator.

[0049] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A deep learning-based method for predicting the energy of organometallic compounds, characterized in that: Includes the following steps: The molecular structure information of organometallic compounds is obtained, and the electronic structure characteristics are obtained through quantum chemical calculations and coordinate alignment; the molecular structure information includes atomic coordinates and atomic numbers. Identify atoms and edges in the molecular structure information to construct a molecular graph; The nodes and edges of the molecular graph are encoded, and the node representations are updated through a multi-layer graph convolutional network to generate geometric feature vectors. Geometric feature vectors and electronic structure features are input into a multimodal feature fusion model for feature fusion, outputting fused features. These features are then further extracted using global average pooling to obtain global fused features. The multimodal feature fusion model includes a cross-attention mechanism module and a gated fusion network module. The global fusion features are input into a multi-task prediction network, which outputs the predicted total energy value and predicted molecular structure of the organometallic compound. The multi-task prediction network includes a shared coding layer and a task-specific decoding layer. The task-specific decoding layer includes two independent branches: a total energy decoding layer and a molecular structure decoding layer. By statistically predicting the mean and variance of the total energy value, the predicted energy value and its confidence interval are obtained.

2. The deep learning-based method for predicting the energy of organometallic compounds according to claim 1, characterized in that: The methods for obtaining the electronic structure features include: Molecular structure files of organometallic compounds are obtained from public databases, and molecular structure information, including atomic coordinates and atomic numbers, is obtained through high-throughput modeling. Quantum chemical calculation software based on density functional theory was used to calculate and extract the force situation, electrostatic potential and orbital occupancy number of each atom, and then spliced ​​them in a fixed order to form the initial electronic structure features. The initial electronic structure features are aligned with atomic coordinates, and all electronic structure features are normalized to obtain the electronic structure features. .

3. The deep learning-based energy prediction method for organometallic compounds according to claim 2, characterized in that: The method for constructing the molecular map includes: Based on atomic coordinates, the spatial distance between all atomic pairs is calculated using the Euclidean distance method. ; If spatial distance Not exceeding the preset distance threshold Then in the atom and Establishing edges Constructing a molecular map ,in, For a set of nodes, Let it be the set of edges.

4. The deep learning-based method for predicting the energy of organometallic compounds according to claim 3, characterized in that: The encoding of nodes and edges in the molecular graph includes: Atom type embedding is obtained by one-hot encoding based on atomic number, combined with its corresponding atomic coordinates and electronic structure features. , splicing together to form nodes in the molecular diagram initial feature vector ; For each connection edge, based on the element types of the atoms at both ends, the corresponding atom type combination code is generated by looking up the preset atom type pair encoding mapping table. Using the spatial distance between opposite sides of the radial basis function set Encode to obtain distance feature vector The edges are then combined with the corresponding atom type codes to obtain the edges in the molecular graph. initial feature vector .

5. The deep learning-based energy prediction method for organometallic compounds according to claim 4, characterized in that: The method for obtaining the geometric feature vector includes: The initial eigenvectors of the nodes in the molecular graph The initial eigenvectors of the edges Input a multi-layer graph convolutional network, and iteratively aggregate the features of each node itself and the features of its neighboring nodes and edges to update the node representation layer by layer; go through After passing the layer graph through convolution, we obtain the geometric feature vector corresponding to each node. .

6. The deep learning-based method for predicting the energy of organometallic compounds according to claim 1, characterized in that: The methods for obtaining the fusion features include: The geometric feature vector and electronic structure feature are input into the cross-attention mechanism module of the multimodal feature fusion model, and the attention feature vector is output. Attention feature vectors and geometric feature vectors are input into a gated fusion network module for dynamic weight fusion, outputting fused features. .

7. The deep learning-based energy prediction method for organometallic compounds according to claim 6, characterized in that: The methods for obtaining the global fusion features include: The fused features are input into a two-layer fully connected neural network, and after dimensionality expansion and compression, the refined features are obtained. ; Global average pooling is performed on the refined features of all atoms to obtain the globally fused features of the organometallic compound. .

8. The deep learning-based method for predicting the energy of organometallic compounds according to claim 5, characterized in that: The method for obtaining the predicted total energy value includes: The globally fused features are input into the shared coding layer, and a deep shared representation is obtained through layer-by-layer nonlinear transformation. The shared coding layer consists of three fully connected layers. Deep shared representation Input the total energy decoding layer and output the predicted total energy value of the organometallic compound. The total energy decoding layer consists of two fully connected networks.

9. The deep learning-based energy prediction method for organometallic compounds according to claim 8, characterized in that: The methods for obtaining the predicted molecular structure include: For each atom Share its geometric feature vectors with the deep layer representation. The data is spliced ​​together, input to the molecular structure decoding layer, and outputted as the predicted coordinate offset of the atom. The molecular structure decoding layer consists of two fully connected networks. For each atom Offset its atomic coordinates from the corresponding predicted coordinates Adding them together gives the predicted steady-state structure coordinates of the atom; The final output is the predicted steady-state structure coordinates of all atoms, i.e., the predicted molecular structure.

10. The deep learning-based energy prediction method for organometallic compounds according to claim 9, characterized in that: The methods for obtaining the energy prediction value and its confidence interval include: After training the multi-task prediction network, multiple forward propagations are performed on the same organometallic compound sample during the prediction phase. Random perturbations are introduced in each forward propagation to obtain a set of total energy prediction values ​​with random differences. ;in, Forward propagation count; Based on the set of total energy prediction values, calculate their mean as the energy prediction value. And calculate its variance, thereby obtaining the prediction standard deviation. ; Based on energy forecasts and the predicted standard deviation Construct a confidence level of confidence interval .

Images