Training method and device of crystal property prediction model and computer device

Abstract

The application relates to a training method and device of a crystal property prediction model and a computer device. The method comprises the following steps: obtaining a sample crystal and a property label corresponding to the sample crystal; the sample crystal comprises a plurality of sample atoms; determining respective atomic coordinates of each sample atom, and determining orbital characteristics of the sample atoms according to the atomic coordinates; determining respective atomic potential energies of the sample atoms according to the orbital characteristics, and determining a predicted property corresponding to the sample crystal according to the atomic potential energies; training the crystal property prediction model according to the difference between the predicted property and the property label, stopping when a first training stop condition is reached, and obtaining a trained crystal property prediction model; and the trained crystal property prediction model is used to determine the property of the crystal. The method can improve the accuracy of predicting the property of the crystal.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a training method, apparatus, and computer equipment for a crystal property prediction model. Background Technology

[0002] With the development of artificial intelligence technology, a technology has emerged to intelligently predict the properties of chemical substances. By measuring the performance of chemical substances, in-depth research can be conducted on materials such as batteries or pharmaceuticals.

[0003] Currently, existing methods for predicting the properties of chemical substances are typically based on graph neural networks. These methods construct a graph structure of a chemical substance and then use graph convolution to predict the properties corresponding to that graph structure. However, graph neural network-based prediction methods do not incorporate fundamental physical principles and are difficult to generalize to large crystal or macromolecular structures, meaning they lack transferability. Therefore, accurately predicting the properties of crystal structures is the problem this proposal aims to solve. Summary of the Invention

[0004] Therefore, it is necessary to provide a training method, apparatus, computer equipment, computer-readable storage medium, and computer program product for a crystal property prediction model that can improve the accuracy of crystal property prediction, in order to address the above-mentioned technical problems.

[0005] Firstly, this application provides a method for training a crystal property prediction model. The method includes:

[0006] Obtain the sample crystal and its corresponding property label; the sample crystal comprises multiple sample atoms;

[0007] Determine the atomic coordinates corresponding to each of the sample atoms, and determine the orbital characteristics of the sample atoms based on the atomic coordinates;

[0008] Based on the orbital characteristics, the atomic potential energy corresponding to each of the sample atoms is determined, and the predicted properties corresponding to the sample crystal are determined based on the atomic potential energy.

[0009] Based on the difference between the predicted properties and the property labels, the crystal property prediction model is trained until the first training stopping condition is met, thus obtaining a trained crystal property prediction model; the trained crystal property prediction model is used to determine the properties of the crystal.

[0010] In one embodiment, determining the orbital characteristics of the sample atom based on the atomic coordinates includes: determining a target rotation matrix of the sample atom based on the atomic coordinates corresponding to the sample atom; determining target orbital coordinates of the sample atom based on the target rotation matrix; determining a target wavefunction of the sample atom based on the target orbital coordinates; and combining the target orbital coordinates and the target wavefunction to obtain the orbital characteristics of the sample atom.

[0011] In one embodiment, determining the target rotation matrix of the sample atom based on the atomic coordinates corresponding to the sample atom includes: for each of the plurality of sample atoms, determining the neighboring sample atoms at the adjacent positions around the current sample atom; determining the first distance between the current sample atom and each of the neighboring sample atoms based on the atomic coordinates corresponding to the current sample atom and the neighboring sample atoms respectively; and combining each of the first distances to obtain the target rotation matrix corresponding to the current sample atom.

[0012] In one embodiment, determining the target orbit coordinates corresponding to the sample atom based on the target rotation matrix includes: for each sample atom among multiple sample atoms, taking the current sample atom as the origin of the atomic coordinate system and determining the hybrid orbit corresponding to the current sample atom in the atomic coordinate system; selecting at least one orbit point on the hybrid orbit at a preset interval and determining the candidate orbit coordinates corresponding to each orbit point; and performing coordinate transformation on each candidate orbit coordinate according to the target rotation matrix corresponding to the current sample atom to obtain the target orbit coordinates corresponding to each orbit point.

[0013] In one embodiment, determining the target wavefunction corresponding to the sample atom based on the target orbit coordinates includes: for each sample atom among multiple sample atoms, determining a second distance between the target orbit coordinates of the orbit point corresponding to the current sample atom and the current sample atom, and obtaining a first candidate wavefunction of the current sample atom based on the second distance; determining the angles between the target orbit coordinates of the orbit point corresponding to the current sample atom and each axis in the atomic coordinate system, and obtaining angle data; obtaining a second candidate wavefunction of the current sample atom based on the second distance and the angle data, and combining the first candidate wavefunction and the second candidate wavefunction to obtain the target wavefunction of the current sample atom.

[0014] In one embodiment, the crystal property prediction model includes a first network structure and a second network structure; determining the atomic potential energy corresponding to each of the sample atoms based on the orbital characteristics includes: obtaining the atomic potential energy corresponding to each sample atom through the first network structure based on the orbital characteristics of the sample atoms and a pre-acquired atomic potential energy model; determining the predicted properties corresponding to the sample crystal based on the atomic potential energy includes: determining the Hamiltonian matrix corresponding to the sample crystal through the first network structure based on the atomic potential energy corresponding to each of the sample atoms; and determining the predicted properties of the sample crystal through the second network structure based on the Hamiltonian matrix.

[0015] In one embodiment, before determining the Hamiltonian matrix corresponding to the sample crystal based on the atomic potential energy corresponding to each of the sample atoms through the first network structure, the method further includes: determining the spatial properties of the sample crystal, and when the spatial properties are momentum spatial properties, determining the lattice vector and wave vector of the sample crystal; and determining the Hamiltonian matrix corresponding to the sample crystal based on the atomic potential energy, the lattice vector and the wave vector.

[0016] In one embodiment, the method further includes: obtaining a special crystal; the special crystal comprising a plurality of special atoms; the special crystal being a crystal in which the bond length between each of the special atoms conforms to a preset length; performing initial pre-training on the crystal property prediction model using the special crystal until a second training stopping condition is met, thereby obtaining an initial crystal property prediction model; and training the crystal property prediction model based on the difference between the predicted property and the property label until a first training stopping condition is met, thereby obtaining a trained crystal property prediction model, comprising: training the initial crystal property prediction model based on the difference between the predicted property and the property label until a first training stopping condition is met, thereby obtaining a trained crystal property prediction model.

[0017] In one embodiment, the crystal property prediction model includes a first network structure and a second network structure; the first network structure is a deep neural network structure; the second network structure is a convolutional neural network structure; the deep neural network structure includes two hidden layers, each hidden layer includes multiple neurons, and the activation function of each hidden layer is a linear rectified function; the convolutional neural network structure includes convolutional layers, max pooling layers, and multilayer perceptron layers.

[0018] Secondly, this application also provides a training device for a crystal property prediction model. The device includes:

[0019] The sample acquisition module is used to acquire a sample crystal and the property label corresponding to the sample crystal; the sample crystal includes multiple sample atoms;

[0020] The property prediction module is used to determine the atomic coordinates corresponding to each of the sample atoms, and to determine the orbital characteristics of the sample atoms based on the atomic coordinates; based on the orbital characteristics, to determine the atomic potential energy corresponding to each of the sample atoms, and to determine the predicted properties corresponding to the sample crystal based on the atomic potential energy;

[0021] The first training module is used to train the crystal property prediction model based on the difference between the predicted property and the property label until the first training stopping condition is met, thus obtaining a trained crystal property prediction model; the trained crystal property prediction model is used to determine the properties of the crystal.

[0022] Thirdly, this application also provides a computer device. The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to perform the following steps:

[0023] Obtain the sample crystal and its corresponding property label; the sample crystal comprises multiple sample atoms;

[0024] Determine the atomic coordinates corresponding to each of the sample atoms, and determine the orbital characteristics of the sample atoms based on the atomic coordinates;

[0025] Based on the orbital characteristics, the atomic potential energy corresponding to each of the sample atoms is determined, and the predicted properties corresponding to the sample crystal are determined based on the atomic potential energy.

[0026] Based on the difference between the predicted properties and the property labels, the crystal property prediction model is trained until the first training stopping condition is met, thus obtaining a trained crystal property prediction model; the trained crystal property prediction model is used to determine the properties of the crystal.

[0027] Fourthly, this application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program, which, when executed by a processor, performs the following steps:

[0028] Obtain the sample crystal and its corresponding property label; the sample crystal comprises multiple sample atoms;

[0029] Determine the atomic coordinates corresponding to each of the sample atoms, and determine the orbital characteristics of the sample atoms based on the atomic coordinates;

[0030] Based on the orbital characteristics, the atomic potential energy corresponding to each of the sample atoms is determined, and the predicted properties corresponding to the sample crystal are determined based on the atomic potential energy.

[0031] Based on the difference between the predicted properties and the property labels, the crystal property prediction model is trained until the first training stopping condition is met, thus obtaining a trained crystal property prediction model; the trained crystal property prediction model is used to determine the properties of the crystal.

[0032] Fifthly, this application also provides a computer program product. The computer program product includes a computer program that, when executed by a processor, performs the following steps:

[0033] Obtain the sample crystal and its corresponding property label; the sample crystal comprises multiple sample atoms;

[0034] Determine the atomic coordinates corresponding to each of the sample atoms, and determine the orbital characteristics of the sample atoms based on the atomic coordinates;

[0035] Based on the orbital characteristics, the atomic potential energy corresponding to each of the sample atoms is determined, and the predicted properties corresponding to the sample crystal are determined based on the atomic potential energy.

[0036] Based on the difference between the predicted properties and the property labels, the crystal property prediction model is trained until the first training stopping condition is met, thus obtaining a trained crystal property prediction model; the trained crystal property prediction model is used to determine the properties of the crystal.

[0037] The training method, apparatus, computer equipment, storage medium, and computer program product of the aforementioned crystal property prediction model, by acquiring a sample crystal, can determine the atomic coordinates corresponding to each sample atom in the sample crystal, and then determine the orbital characteristics corresponding to the sample atom based on the atomic coordinates. By determining the orbital characteristics corresponding to the sample atom, the atomic potential energy corresponding to the sample atom can be determined based on the orbital characteristics, and then the predicted property corresponding to the sample crystal can be determined based on the atomic potential energy corresponding to each sample atom. By determining the predicted property and obtaining property labels, the crystal property prediction model can be trained based on the difference between the predicted property and the property label to obtain a trained crystal property prediction model. Since this application predicts the orbital characteristics corresponding to each sample atom in the sample crystal, and then predicts the atomic potential energy of the sample atom based on the orbital characteristics, and then determines the crystal property based on the atomic potential energy, compared with the traditional graph neural network method that does not introduce basic physical principles, this application uses the orbital characteristics determined by physical principles as the basis for predicting crystal properties, thereby greatly improving the accuracy of crystal property prediction. At the same time, since the atomic potential energy corresponding to each sample atom is determined, the properties of the sample crystal can be determined directly based on the atomic potential energy, thereby improving the efficiency of predicting the properties of the crystal. Attached Figure Description

[0038] Figure 1 This is a diagram illustrating the application environment of a training method for a crystal property prediction model in one embodiment.

[0039] Figure 2 This is a flowchart illustrating the training method for a crystal property prediction model in one embodiment;

[0040] Figure 3 This is a schematic diagram of the process for preprocessing sample crystals in one embodiment;

[0041] Figure 4 This is a flowchart illustrating the process of determining the predicted properties of a sample crystal in one embodiment;

[0042] Figure 5 This is a flowchart illustrating the process of determining the target wave function of a sample crystal in one embodiment;

[0043] Figure 6 This is a structural block diagram of a training device for a crystal property prediction model in one embodiment;

[0044] Figure 7 This is a structural block diagram of the track characteristic module in one embodiment;

[0045] Figure 8 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0047] The training method for the crystal property prediction model provided in this application embodiment can be applied to, for example... Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. Terminal 102 sends the acquired sample crystal and its property labels to server 104, which then predicts the properties of the received sample crystal. Server 104 determines the difference between the predicted properties and the property labels, trains a crystal property prediction model based on this difference, and returns the trained model to terminal 102. Terminal 102 can receive the trained crystal property prediction model and determine the properties of the crystal using it. Terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, and smart in-vehicle devices. Portable wearable devices can include smartwatches, smart bracelets, and head-mounted devices. Server 104 can be implemented using a standalone server or a server cluster consisting of multiple servers.

[0048] In one embodiment, such as Figure 2 As shown, a training method for a crystal property prediction model is provided. The method is illustrated using a computer device as an example, which can provide... Figure 1 The training method for the crystal property prediction model in the terminal or server includes the following steps:

[0049] Step 202: Obtain the sample crystal and its corresponding property label; the sample crystal includes multiple sample atoms.

[0050] A crystal is a structural material composed of a large number of microscopic substances, such as atoms or molecules, arranged in an orderly manner according to certain rules. By studying the arrangement rules or characteristics of these microscopic substances, the relevant properties of the crystal can be determined. These properties may include the total energy of the crystal, electronic band structure, phonon band structure, atomization energy, and dielectric constant. The sample crystal can be a carbon crystal with a relatively complex atomic arrangement, or a carbon crystal with uniform bond lengths between atoms, such as diamond. The property labels corresponding to the sample crystal are training labels that correspond to the relevant properties of the crystal, such as electronic band structure labels and atomization energy labels.

[0051] In one embodiment, a user can obtain sample crystals from a cloud platform, a publicly available dataset, or a dataset provided by an enterprise, and input the sample crystals into a computer device until a trained crystal property prediction model is obtained. The computer device then stores the trained crystal property prediction model.

[0052] In one embodiment, the user can also use a dedicated chemical structure detection instrument to detect and acquire the sample crystals acquired in real time. Therefore, this application does not limit the method of acquiring sample crystals.

[0053] In one embodiment, when a user needs to predict the properties of a crystal, a computer device can provide crystal property prediction services to the user through a trained crystal property prediction model.

[0054] Step 204: Determine the atomic coordinates corresponding to each sample atom, and determine the orbital characteristics of the sample atom based on the atomic coordinates.

[0055] The entire sample crystal can be viewed as a spatial lattice, in which sample atoms are arranged according to certain rules. Therefore, each sample atom can be simplified as a coordinate point.

[0056] Specifically, the computer device uses any sample atom in the spatial lattice as the central origin of the lattice coordinate system, and then determines the atomic coordinates corresponding to each sample atom based on this central origin. The computer device calculates the atomic distances between all sample atoms based on the atomic coordinates corresponding to each sample atom, and simultaneously determines and stores the atomic environment corresponding to each sample atom within a preset radius. The atomic environment can represent the neighboring sample atoms within the preset radius centered on each sample atom, and the corresponding atomic coordinates of the neighboring sample atoms.

[0057] Furthermore, once the atomic coordinates are obtained, the computer device can determine the target orbital coordinates corresponding to the sample atom based on the atomic coordinates, and determine the orbital characteristics of the sample atom based on the target orbital coordinates. Here, the target orbital coordinates refer to the coordinates obtained by modifying the orbital coordinates of multiple orbital points on the sample atom's orbit based on a preset rotation matrix. An atomic orbital is a mathematical function describing the wave-like behavior of electrons in an atom, indicating the possible positions of electrons in the space outside the atomic nucleus and the probability of them appearing at those positions. Orbital coordinates refer to the coordinates of the orbital points corresponding to the sample atom.

[0058] In one embodiment, determining the orbital characteristics of a sample atom based on its atomic coordinates includes: determining the target rotation matrix of the sample atom based on its atomic coordinates; determining the target orbital coordinates of the sample atom based on the target rotation matrix; determining the target wavefunction of the sample atom based on its target orbital coordinates; and combining the target orbital coordinates and the target wavefunction to obtain the orbital characteristics of the sample atom.

[0059] Among them, the wave function is a function that can quantitatively describe the properties of microscopic particles. It can be used to calculate the possible positions of electrons in the space outside the atomic nucleus and their corresponding probabilities, and can also be used to determine other properties between atoms and electrons.

[0060] Specifically, the computer device retrieves pre-stored atomic environments from a database and determines the atomic coordinates included in those environments. Then, based on the atomic coordinates corresponding to the sample atoms, it determines the target rotation matrix of the sample atoms. The computer device selects multiple orbital points on the sample atoms' orbits and determines the candidate orbital coordinates for these points. Based on the target rotation matrix and the candidate orbital coordinates, the computer device determines the target orbital coordinates corresponding to the sample atoms. The computer device acquires a wavefunction model and inputs the target orbital coordinates of the sample atoms into the wavefunction model to obtain the target wavefunction for the sample atoms. Finally, by combining the target orbital coordinates and the target wavefunction, the orbital characteristics of the sample atoms are obtained.

[0061] In one embodiment, before determining the orbital properties of the sample atoms, preprocessing of the sample atoms is also involved, such as... Figure 3 As shown, a method for preprocessing a sample crystal is provided, including the following steps: S302: Read the lattice vector of the sample crystal and determine the corresponding spatial properties; S304: Calculate and store the distance between all sample atoms; S306: Store the atomic environment within the vicinity of each sample atom; S308: Uniformly select orbital points within a preset radius of each sample atom and save the position information of the orbital points; S310: Calculate and store the relevant positions of the orbital points relative to neighboring sample atoms.

[0062] In this embodiment, after preprocessing the sample crystal, the calculated relevant information is stored in the database, so that the atomic environment corresponding to each sample atom can be directly extracted from the database, simplifying the data calculation process for model training.

[0063] Step 206: Based on the orbital characteristics, determine the atomic potential energy corresponding to each sample atom, and determine the predicted properties corresponding to the sample crystal based on the atomic potential energy.

[0064] The orbital characteristics include the target orbital coordinates and target wavefunctions corresponding to the sample atoms. Atomic potential energy refers to the potential energy of electrons in the electric field of the sample atoms, such as carbon potential energy. Predicted properties may include the atomization energy of the crystal, total energy, electronic band structure, phonon band structure, etc.

[0065] In one embodiment, the crystal property prediction model includes a first network structure and a second network structure; such as Figure 4 As shown, a method for determining the predicted properties of a sample crystal is provided, comprising the following steps:

[0066] Step 402: Using the first network structure, the atomic potential energy corresponding to the sample atom is obtained based on the orbital characteristics of the sample atom and the pre-acquired atomic potential energy model.

[0067] Step 404: Using the first network structure, determine the Hamiltonian matrix corresponding to the sample crystal based on the atomic potential energy corresponding to each sample atom.

[0068] The atomic potential energy model includes a first atomic potential energy model corresponding to real space properties and a second atomic potential energy model corresponding to momentum space properties. The Hamiltonian matrix is ​​an N×N matrix, where N is the total number of orbitals included in the sample crystal. For the diagonal elements of the matrix, H... μμ =α μμ H μμ This represents the orbital energy corresponding to orbital μ, and parameter α is a trainable constant in the first network structure. The off-diagonal elements of the matrix can be obtained through the orbital characteristics of each sample atom.

[0069] Specifically, the computer device inputs the orbital characteristics of the sample atoms into the first network structure, determines the cross terms between orbitals, and combines this with a pre-acquired atomic potential energy model to obtain the orbital energies corresponding to the off-diagonal elements of the sample atoms in the matrix. By combining the orbital energies on the diagonal and off-diagonal sides of the matrix, the atomic potential energy of the sample atom is obtained. The computer device then combines the atomic potential energies of each sample atom to determine the Hamiltonian matrix corresponding to the sample crystal. Specifically, when the orbital characteristics form a 4×4 matrix, the atomic potential energy of each sample atom includes the orbital energies of 4×4 elements, and this 4×4 atomic potential energy is used as the matrix element in the Hamiltonian matrix.

[0070] For example, when the spatial properties of the sample crystal are real space properties, also known as R space, the first atomic potential energy model is obtained, that is, the off-diagonal elements of the Hamiltonian matrix are:

[0071]

[0072] Where μ is the orbital pointed to by the row vectors in the Hamiltonian matrix, v is the orbital pointed to by the column vectors in the Hamiltonian matrix, and H... μv This is the orbital overlap energy between the μ row vector orbit and the v column vector orbit, S μv For the intersections between tracks.

[0073] In one embodiment, when the orbital characteristics are not a 4×4 matrix, the atomic potential energy corresponding to the sample atom can be obtained by combining the orbital energies on the diagonal of the matrix and the corresponding orbital energies on the off-diagonal of the matrix.

[0074] In one embodiment, before determining the Hamiltonian matrix corresponding to the sample crystal based on the atomic potential energy of each sample atom through the first network structure, the method further includes: determining the spatial properties of the sample crystal, and when the spatial properties are momentum spatial properties, determining the lattice vector and wave vector of the sample crystal; and determining the Hamiltonian matrix corresponding to the sample crystal based on the atomic potential energy, lattice vector, and wave vector.

[0075] Among them, spatial properties can include spatiotemporal properties and momentum-space properties. Real space properties are visible characteristics of electrons, such as electron density; momentum-space properties are invisible energy characteristics, such as the energy of different energy levels.

[0076] Specifically, the computer equipment determines the spatial properties of the sample atoms. When the spatial property is a real space property, the intersection terms between orbitals are directly input into the acquired first atomic potential energy model to obtain the orbital energies corresponding to the elements of the sample atoms on the off-diagonal side of the matrix. When the spatial property is a momentum space property, the computer equipment determines the lattice vector and wave vector of the sample crystal, and directly inputs the intersection terms between orbitals, the lattice vector, and the wave vector into the acquired second atomic potential energy model to obtain the orbital energies corresponding to the elements of the sample atoms on the off-diagonal side of the matrix.

[0077] For example, when the spatial properties of the sample crystal are momentum-space properties, also known as K-space, the second atomic potential energy model is obtained, that is, the off-diagonal elements corresponding to the Hamiltonian matrix are:

[0078]

[0079] Here, R is the lattice vector, k is the wave vector, and parameter K is a trainable constant in the first network structure. Therefore, the Hamiltonian matrix can be periodically constructed for the atomic potential energy in K space, and the properties of K space require a pre-defined wave vector k.

[0080] In one embodiment, before determining the Hamiltonian matrix corresponding to the sample crystal based on the atomic potential energy of each sample atom using the first network structure, the method further includes: determining the spatial properties of the sample crystal, and when the spatial properties are real space properties, determining the lattice vector of the sample crystal; and determining the Hamiltonian matrix corresponding to the sample crystal based on the atomic potential energy and the lattice vector. The lattice vector is used to determine the distance between atoms. For example, in a square lattice with a side length of 10, when the atomic coordinates of two sample atoms are (1, 0, 0) and (9, 0, 0) respectively, the distance between these two sample atoms is determined to be 2, not 8, based on the lattice vector.

[0081] In this embodiment, by pre-determining the spatial properties of the sample crystal, different atomic potential energy models can be selected to train the sample crystal, thereby improving the accuracy of subsequent training of the crystal property prediction model.

[0082] Step 406: Using the second network structure, determine the predicted properties of the sample crystal based on the Hamiltonian matrix.

[0083] Specifically, the computer device inputs the Hamiltonian matrix corresponding to the sample crystal into the second network structure to determine the predicted properties of the sample crystal. When the predicted property of the sample crystal is atomization energy, the second network structure can be adjusted to sum the atomic potential energies in the Hamiltonian matrix; when the predicted property of the sample crystal is electronic band structure, the second network structure can be adjusted to diagonalize the atomic potential energies in the Hamiltonian matrix. Therefore, this application does not limit the specific method of adjusting the Hamiltonian matrix.

[0084] In one embodiment, the target rotation matrix of the sample atom is determined based on the atomic coordinates corresponding to the sample atom using a first network structure.

[0085] Step 208: Based on the difference between the predicted properties and the property labels, train the crystal property prediction model until the first training stopping condition is met, and obtain the trained crystal property prediction model; the trained crystal property prediction model is used to determine the properties of the crystal.

[0086] The computer equipment can determine the loss value of the loss function through the backpropagation algorithm, and update the model parameters in the crystal property prediction model along the gradient direction of the loss value, so that the loss function converges, that is, the first training stopping condition is met, and the crystal property prediction model is obtained.

[0087] In one embodiment, the first training stopping condition can be the number of iterations in the training process. For example, when the number of iterations reaches 1000, the training of the crystal property prediction model can be stopped.

[0088] In the training method of the aforementioned crystal property prediction model, by acquiring a sample crystal, the atomic coordinates corresponding to each sample atom in the sample crystal can be determined, and then the orbital characteristics corresponding to the sample atom can be determined based on the atomic coordinates. By determining the orbital characteristics corresponding to the sample atom, the atomic potential energy corresponding to the sample atom can be determined based on the orbital characteristics, and then the predicted property corresponding to the sample crystal can be determined based on the atomic potential energy corresponding to each sample atom. By determining the predicted property and obtaining property labels, the crystal property prediction model can be trained based on the difference between the predicted property and the property label to obtain a trained crystal property prediction model. Since this application predicts the orbital characteristics corresponding to each sample atom in the sample crystal, and then predicts the atomic potential energy of the sample atom based on the orbital characteristics, and then determines the crystal property based on the atomic potential energy, compared with the traditional graph neural network method that does not introduce basic physical principles, this application uses the orbital characteristics determined based on physical principles as the basis for predicting crystal properties, thereby greatly improving the accuracy of crystal property prediction. At the same time, since the atomic potential energy corresponding to each sample atom is determined, the property of the sample crystal can be directly determined based on the atomic potential energy, thereby improving the efficiency of crystal property prediction.

[0089] In one embodiment, determining the target rotation matrix of a sample atom based on the atomic coordinates corresponding to the sample atom includes: for each of the multiple sample atoms, determining the neighboring sample atoms at the adjacent positions around the current sample atom; determining the first distance between the current sample atom and each neighboring sample atom based on the atomic coordinates corresponding to the current sample atom and the neighboring sample atoms respectively; and combining each first distance to obtain the target rotation matrix corresponding to the current sample atom.

[0090] Here, the first distance represents the distance between a sample atom and its neighboring sample atoms.

[0091] Specifically, the computer device directly uses the atomic coordinates of the current sample atom and the atomic coordinates of neighboring sample atoms as input to the deep neural network, calculates the first distance between the current sample atom and each neighboring sample atom, and then outputs a unit vector through the deep neural network to represent the rotation axis parameter and rotation angle parameter corresponding to the current sample atom. Finally, based on the rotation axis parameter and rotation angle parameter, the target rotation matrix corresponding to the current sample atom is calculated.

[0092] For example, when sample atom 'o' is taken as the origin of the lattice coordinate system, the atomic coordinates of each sample atom can be determined. When sample atom 'a' is taken as the current sample atom, if the neighboring sample atoms around sample a are determined to include sample atom b and sample atom c, then the atomic coordinates of sample atom are (xa, ya, za), the atomic coordinates of sample atom b are (xb, yb, zb), and the atomic coordinates of sample atom c are (xc, yc, zc). The computer calculates the first distance between sample atom a and sample atom b and sample atom c respectively, and outputs a unit vector through a deep neural network. The rotation axis parameters can be represented by this unit vector. The rotation angle parameter is θ. The target rotation matrix A corresponding to sample atom a is calculated using these two parameters.

[0093] The expression for the target rotation matrix can be:

[0094]

[0095] In this embodiment, the target rotation matrix corresponding to the current sample atom can be determined by the atomic coordinates of the current sample atom and the neighboring sample atoms, so that the orbital characteristics of the sample atom can be accurately obtained based on the target rotation matrix.

[0096] In one embodiment, determining the target orbit coordinates corresponding to the sample atom based on the target rotation matrix includes: for each sample atom among multiple sample atoms, taking the current sample atom as the origin of the atomic coordinate system and determining the hybrid orbit corresponding to the current sample atom in the atomic coordinate system; selecting at least one orbit point on the hybrid orbit at a preset interval and determining the candidate orbit coordinates corresponding to each orbit point; and performing coordinate transformation on each candidate orbit coordinate according to the target rotation matrix corresponding to the current sample atom to obtain the target orbit coordinates corresponding to each orbit point.

[0097] During atomic bonding, due to the mutual influence between atoms, several atomic orbitals of similar energy and different types can linearly combine, redistribute energy, and determine spatial orientation, thus forming a new number of atomic orbitals. This orbital recombination is called hybridization, and the new orbitals formed after hybridization are called hybrid orbitals, such as sp orbitals. 3 Hybrid orbitals, sp 2 Hybrid orbitals, etc.

[0098] Each sample atom typically forms a hybrid orbital, and each hybrid orbital includes multiple sub-orbitals, all centered around the sample atom. For sp... 3 Hybrid orbitals include one s orbital and three p orbitals.

[0099] Specifically, for each sample atom among multiple sample atoms, the computer device uses the current sample atom as the origin of the atomic coordinate system and determines the hybrid orbital corresponding to the current sample atom within the atomic coordinate system. For example, using sample atom 'a' as the origin of the atomic coordinate system. The computer device selects at least one orbital point on the hybrid orbital at preset intervals and determines the candidate orbital coordinates corresponding to each orbital point. The selection of orbital points can be within a sphere with a radius of 3 angstroms, using the current sample atom as the origin, and the preset interval can be 1 angstrom, etc. For example, when selecting orbital point 1 and orbital point 2 on the hybrid orbital of sample atom 'a', the candidate orbital coordinates of orbital point 1 are determined to be (x1, y1, z1), and the candidate orbital coordinates of orbital point 2 are determined to be (x2, y2, z2). The computer device performs coordinate transformation on each candidate orbital coordinate according to the target rotation matrix corresponding to the current sample atom to obtain the target orbital coordinates corresponding to each orbital point. For example, referring to the example above, the target rotation matrix A corresponding to sample atom a has been determined. Based on the target rotation matrix A, the target orbit coordinates of orbit point 1 are (X1, Y1, Z1), and the target orbit coordinates of orbit point 2 are (X2, Y2, Z2).

[0100] In this embodiment, the target orbit coordinates corresponding to each orbit point are obtained according to the target rotation matrix. Thus, the orbital characteristics of sample atoms can be determined more accurately based on the target orbit coordinates, thereby improving the prediction accuracy of the crystal property prediction model.

[0101] In one embodiment, such as Figure 5 As shown, a method for determining the target wavefunction corresponding to a sample atom is provided, including the following steps:

[0102] Step 502: For each sample atom among multiple sample atoms, determine the second distance between the target orbit coordinates of the orbit point corresponding to the current sample atom and the current sample atom, and obtain the first candidate wavefunction of the current sample atom based on the second distance.

[0103] Wherein, the second distance represents the distance between the target orbit coordinates of the orbit point corresponding to the sample atom and the sample atom; the first candidate wavefunction represents the wavefunction corresponding to the s orbit.

[0104] Specifically, when the computer device determines the target orbit coordinates for each orbital point, it can also determine the second distance between the target orbit coordinates of that orbital point and the current sample atom. The computer device then acquires the wavefunction model and inputs the second distance into the wavefunction model to obtain the first candidate wavefunction corresponding to the current sample atom. The wavefunction model includes both s-orbital and p-orbital wavefunction models.

[0105] For example, when the wave function model includes:

[0106]

[0107] Where r is the second distance between the orbital point and the sample atom, a0 is the Bohr radius constant, and the parameter ζ is used to ensure normalization. Therefore, for orbital point 1 of sample atom a, the second distance r1 is... The first candidate wavefunction is |s(r1)).

[0108] In one embodiment, the orbital expression for a carbon atom can be approximated by the orbital expression for a hydrogen atom.

[0109] Step 504: Determine the angles between the target orbit coordinates of the orbit point corresponding to the current sample atom and each axis in the atomic coordinate system to obtain angle data.

[0110] The angle data includes first angle data, second angle data, and third angle data.

[0111] Specifically, the computer equipment determines the angle between the target orbit coordinates corresponding to the orbit point and the horizontal axis of the coordinate system, obtaining the first angle data; it determines the angle between the target orbit coordinates corresponding to the orbit point and the vertical axis of the coordinate system, obtaining the second angle data; and it determines the angle between the target orbit coordinates corresponding to the orbit point and the vertical axis of the coordinate system, obtaining the third angle data. For example, determining the angles between the target orbit coordinates (X1, Y1, Z1) of orbit point 1 and each axis yields the first angle data θ1, the second angle data θ2, and the third angle data θ3.

[0112] Step 506: Based on the second distance and angle data, obtain the second candidate wavefunction of the current sample atom, and combine the first candidate wavefunction and the second candidate wavefunction to obtain the target wavefunction of the current sample atom.

[0113] The second candidate wavefunction includes a third, a fourth, and a fifth candidate wavefunction. The third candidate wavefunction represents the wavefunction corresponding to the px orbital, the fourth candidate wavefunction represents the wavefunction corresponding to the py orbital, and the fifth candidate wavefunction represents the wavefunction corresponding to the pz orbital.

[0114] Specifically, the computer device inputs the first angle data and the second distance into the wave function model to obtain the third candidate wave function corresponding to the sample atom. The computer device inputs the second angle data and the second distance into the wave function model to obtain the fourth candidate wave function corresponding to the sample atom. At the same time, the computer device inputs the third angle data and the second distance into the wave function model to obtain the fifth candidate wave function corresponding to the sample atom.

[0115] For example, when the wave function model includes:

[0116]

[0117] Where r is the second distance between the orbital point and the sample atom, and θ is the angle between the target orbital coordinates and each axis in the atomic coordinate system. Therefore, taking orbital point 1 of sample atom a as an example, the second candidate wavefunction includes the third candidate wavefunction |p x (r1)), the fourth candidate wavefunction |p y (r1)), the fifth candidate wavefunction |p z (r1)>.

[0118] Furthermore, the computer device synthesizes the first, third, fourth, and fifth candidate wavefunctions to obtain the target wavefunction of the current sample atom. For example, the computer device can obtain the target wavefunction of the current sample atom using the following formula:

[0119]

[0120]

[0121] In this embodiment, the target wave function corresponding to the current sample atom is constructed by using the coordinate features of the orbital point corresponding to the current sample atom. This allows the orbital characteristics of the sample atom to be quickly determined based on the target wave function, thereby improving the efficiency of subsequent model training.

[0122] In one embodiment, the method further includes: obtaining a special crystal; the special crystal includes multiple special atoms; the special crystal is a crystal in which the bond length between each special atom conforms to a preset length; performing initial pre-training on the crystal property prediction model using the special crystal until a second training stopping condition is met, thereby obtaining an initial crystal property prediction model; and training the crystal property prediction model based on the difference between the predicted properties and the property labels until a first training stopping condition is met, thereby obtaining a trained crystal property prediction model, including: training the initial crystal property prediction model based on the difference between the predicted properties and the property labels until a first training stopping condition is met, thereby obtaining a trained crystal property prediction model.

[0123] The special crystal can be a crystal with a relatively stable interatomic structure, such as diamond. Furthermore, the special crystal can be obtained through data augmentation, which diversifies the data samples and thus enhances the generalization ability of the trained three-crystal property prediction model. The specific process of initial pre-training the crystal property prediction model using special crystals can be referenced from the process of training the crystal property prediction model using sample crystals described above; therefore, it will not be repeated here in the embodiments of this application.

[0124] In one embodiment, the second training stopping condition can be the number of iterations in the training process. For example, when the number of iterations reaches 1000, the training of the initial crystal property prediction model can be stopped.

[0125] In one embodiment, multilayer graphene can be used as a special crystal to perform initial pre-training of the crystal property prediction model.

[0126] In this embodiment, unlike the backpropagation algorithm that directly applies to sample crystals and their properties, it is necessary to first pre-train the crystal property prediction model using special crystals with different bond lengths and their properties. This can significantly reduce problems such as gradient vanishing and gradient exploding during the training process.

[0127] In one embodiment, the crystal property prediction model includes a first network structure and a second network structure; the first network structure is a deep neural network structure; the second network structure is a convolutional neural network structure; the deep neural network structure includes two hidden layers, each hidden layer includes multiple neurons, and the activation function of each hidden layer is a linear rectified function; the convolutional neural network structure includes convolutional layers, max pooling layers, and multilayer perceptron layers.

[0128] In this configuration, when the first network structure is a deep neural network, it includes two hidden layers, each containing multiple neurons (e.g., 100 neurons). The activation function for each hidden layer is a linear rectified function, such as ReLU. When the second network structure is a convolutional neural network (CNN), the first layer can be a convolutional layer with 130 channels, a kernel size of 3, a stride of 1, and padding of 2. The corresponding activation function can be ReLU. Alternatively, the first layer can also be a convolutional layer with 32 channels, a kernel size of 3, a stride of 2, and padding of 1. The third layer is an adaptive max-pooling layer with an output shape of 128×128. The fourth layer can be a convolutional layer with 64 channels, a kernel size of 3, a stride of 2, and padding of 1. The corresponding activation function is ReLU. The fifth layer of a convolutional neural network can be a convolutional layer with 64 channels, a kernel size of 3, a stride of 2, padding of 1, and the corresponding activation function is ReLU. The sixth layer of the convolutional neural network is a multilayer perceptron layer containing 100 hidden neurons, producing the desired output.

[0129] In one embodiment, the initial learning rate of both the crystal property prediction model and the initial crystal property prediction model can be set to 0.0001, and the learning rate is halved after every 100 iterations. The entire training process can include 1000 iterations.

[0130] In one embodiment, an open-source deep learning framework, such as the PyTorch framework, can be used to build a crystal property prediction model, and the crystal property prediction model can be optimized using the Adam algorithm (Adaptive Momentum Estimation) based on a batch stochastic gradient descent optimizer.

[0131] In this embodiment, by employing a combination of a first network structure and a second network structure, and providing specific implementation structures for deep neural network structure and convolutional neural network structure respectively, the crystal property prediction model can be trained accurately and efficiently, thereby realizing the prediction of crystal properties.

[0132] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0133] Based on the same inventive concept, this application also provides a training device for a crystal property prediction model to implement the training method for the crystal property prediction model described above. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more crystal property prediction model training device embodiments provided below can be found in the limitations of the crystal property prediction model training method described above, and will not be repeated here.

[0134] In one embodiment, such as Figure 6 As shown, a training device 600 for a crystal property prediction model is provided, comprising: a sample acquisition module 602, a property prediction module 604, and a first training module 606, wherein:

[0135] The sample acquisition module 602 is used to acquire the sample crystal and the corresponding property label of the sample crystal; the sample crystal includes multiple sample atoms.

[0136] The property prediction module 604 is used to determine the atomic coordinates corresponding to each sample atom, and to determine the orbital characteristics of the sample atom based on the atomic coordinates; based on the orbital characteristics, to determine the atomic potential energy corresponding to each sample atom, and to determine the predicted properties corresponding to the sample crystal based on the atomic potential energy.

[0137] The first training module 606 is used to train the crystal property prediction model based on the difference between the predicted properties and the property labels until the first training stopping condition is met, thus obtaining a trained crystal property prediction model; the trained crystal property prediction model is used to determine the properties of the crystal.

[0138] In one embodiment, the property prediction module 604 further includes an orbital characteristic module 6041, which is used to determine the target rotation matrix of the sample atom based on the atomic coordinates corresponding to the sample atom; determine the target orbital coordinates corresponding to the sample atom based on the target rotation matrix; determine the target wave function corresponding to the sample atom based on the target orbital coordinates; and obtain the orbital characteristics of the sample atom by combining the target orbital coordinates and the target wave function.

[0139] In one embodiment, the crystal property prediction model includes a first network structure and a second network structure; the property prediction module 604 is further configured to obtain the atomic potential energy corresponding to the sample atom through the first network structure based on the orbital characteristics of the sample atom and the pre-acquired atomic potential energy model; and to determine the predicted properties corresponding to the sample crystal based on the atomic potential energy, including: determining the Hamiltonian matrix corresponding to the sample crystal through the first network structure based on the atomic potential energy corresponding to each sample atom; and determining the predicted properties of the sample crystal based on the Hamiltonian matrix through the second network structure.

[0140] In one embodiment, the property prediction module 604 further includes a spatial attribute module 6042, used to determine the spatial attributes of the sample crystal, and when the spatial attribute is a momentum spatial attribute, to determine the lattice vector and wave vector of the sample crystal; and to determine the Hamiltonian matrix corresponding to the sample crystal based on the atomic potential energy, lattice vector and wave vector.

[0141] In one embodiment, the training device 600 for the crystal property prediction model further includes a second training module 608 for acquiring a special crystal; the special crystal includes multiple special atoms; the special crystal is a crystal in which the bond length between each special atom conforms to a preset length; the crystal property prediction model is initially pre-trained using the special crystal until a second training stopping condition is met, thus obtaining an initial crystal property prediction model; the crystal property prediction model is trained according to the difference between the predicted properties and the property labels until a first training stopping condition is met, thus obtaining a trained crystal property prediction model, including: training the initial crystal property prediction model according to the difference between the predicted properties and the property labels until a first training stopping condition is met, thus obtaining a trained crystal property prediction model.

[0142] In one embodiment, the training device 600 for the crystal property prediction model further includes a network structure module 610 for determining a first network structure and a second network structure in the crystal property prediction model; the first network structure is a deep neural network structure; the second network structure is a convolutional neural network structure; the deep neural network structure includes two hidden layers, each hidden layer includes multiple neurons, and the activation function of each hidden layer is a linear rectified function; the convolutional neural network structure includes convolutional layers, max pooling layers, and multilayer perceptron layers.

[0143] In one embodiment, such as Figure 7 As shown, a trajectory characteristic module 6041 is provided, including: a rotation matrix determination module 6041a, a coordinate determination module 6041b, and a wave function determination module 6041c, wherein:

[0144] The rotation matrix determination module 6041a is used to determine the neighboring sample atoms at the adjacent positions around the current sample atom for each sample atom among multiple sample atoms; determine the first distance between the current sample atom and each neighboring sample atom according to the atomic coordinates corresponding to the current sample atom and the neighboring sample atoms respectively; and obtain the target rotation matrix corresponding to the current sample atom by combining each first distance.

[0145] The coordinate determination module 6041b is used to determine the hybrid orbit corresponding to the current sample atom in the atomic coordinate system as the origin of the atomic coordinate system for each sample atom among multiple sample atoms; select at least one orbital point on the hybrid orbital at a preset interval and determine the candidate orbital coordinates corresponding to each orbital point; and perform coordinate transformation on each candidate orbital coordinate according to the target rotation matrix corresponding to the current sample atom to obtain the target orbital coordinates corresponding to each orbital point.

[0146] The wavefunction determination module 6041c is used to determine, for each sample atom among multiple sample atoms, the second distance between the target orbit coordinates of the orbit point corresponding to the current sample atom and the current sample atom, and obtain the first candidate wavefunction of the current sample atom based on the second distance; determine the angles between the target orbit coordinates of the orbit point corresponding to the current sample atom and each axis in the atomic coordinate system, and obtain angle data; obtain the second candidate wavefunction of the current sample atom based on the second distance and angle data, and combine the first candidate wavefunction and the second candidate wavefunction to obtain the target wavefunction of the current sample atom.

[0147] Each module in the training device for the aforementioned crystal property prediction model can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0148] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 8 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores training data for crystal property prediction models. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a training method for a crystal property prediction model.

[0149] Those skilled in the art will understand that Figure 8 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0150] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0151] In one embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0152] In one embodiment, a computer program product or computer program is provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and executes the computer instructions, causing the computer device to perform the steps in the above method embodiments.

[0153] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0154] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0155] The above embodiments are merely illustrative of several implementation methods of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A training method for a crystal property prediction model, characterized in that, The method includes: Obtain the sample crystal and its corresponding property label; the sample crystal comprises multiple sample atoms; Determine the atomic coordinates corresponding to each of the sample atoms; Based on the atomic coordinates corresponding to the sample atoms, determine the target rotation matrix of the sample atoms; For each sample atom among multiple sample atoms, the current sample atom is taken as the origin of the atomic coordinate system, and the hybrid orbit corresponding to the current sample atom in the atomic coordinate system is determined. At least one orbital point on the hybrid orbit is selected at a preset interval, and the candidate orbital coordinates corresponding to each orbital point are determined. Based on the target rotation matrix corresponding to the current sample atom, coordinate transformation is performed on each candidate orbit coordinate to obtain the target orbit coordinates corresponding to each orbit point. Based on the target orbital coordinates, determine the target wavefunction corresponding to the sample atom; By combining the target orbital coordinates and the target wavefunction, the orbital characteristics of the sample atom are obtained; Based on the orbital characteristics, the atomic potential energy corresponding to each of the sample atoms is determined, and the predicted properties corresponding to the sample crystal are determined based on the atomic potential energy. Based on the difference between the predicted properties and the property labels, the crystal property prediction model is trained until the first training stopping condition is met, thus obtaining a trained crystal property prediction model; the trained crystal property prediction model is used to determine the properties of the crystal.

2. The method according to claim 1, characterized in that, The step of determining the target rotation matrix of the sample atom based on the atomic coordinates corresponding to the sample atom includes: For each sample atom in a plurality of sample atoms, the neighboring sample atoms at the nearest positions around the current sample atom are determined; Based on the atomic coordinates corresponding to the current sample atom and the neighboring sample atoms, a first distance is determined between the current sample atom and each of the neighboring sample atoms; By combining each of the first distances, the target rotation matrix corresponding to the current sample atom is obtained.

3. The method according to claim 1, characterized in that, The step of determining the target wavefunction corresponding to the sample atom based on the target orbital coordinates includes: For each sample atom among multiple sample atoms, determine the second distance between the target orbit coordinates of the orbit point corresponding to the current sample atom and the current sample atom, and obtain the first candidate wavefunction of the current sample atom based on the second distance; Determine the angles between the target orbit coordinates of the orbit point corresponding to the current sample atom and each axis in the atomic coordinate system to obtain angle data; Based on the second distance and the angle data, the second candidate wavefunction of the current sample atom is obtained, and the target wavefunction of the current sample atom is obtained by combining the first candidate wavefunction and the second candidate wavefunction.

4. The method according to claim 1, characterized in that, The crystal property prediction model includes a first network structure and a second network structure; determining the atomic potential energy corresponding to each sample atom based on the orbital characteristics includes: Using the first network structure, the atomic potential energy corresponding to the sample atom is obtained based on the orbital characteristics of the sample atom and the pre-acquired atomic potential energy model; The step of determining the predicted properties corresponding to the sample crystal based on the atomic potential energy includes: Using the first network structure, the Hamiltonian matrix corresponding to the sample crystal is determined based on the atomic potential energy corresponding to each sample atom. The predicted properties of the sample crystal are determined based on the Hamiltonian matrix using the second network structure.

5. The method according to claim 4, characterized in that, Before determining the Hamiltonian matrix corresponding to the sample crystal based on the atomic potential energy corresponding to each sample atom through the first network structure, the method further includes: Determine the spatial properties of the sample crystal, and when the spatial properties are momentum spatial properties, determine the lattice vector and wave vector of the sample crystal; The Hamiltonian matrix corresponding to the sample crystal is determined based on the atomic potential energy, the lattice vector, and the wave vector.

6. The method according to claim 1, characterized in that, The method further includes: Obtain a special crystal; the special crystal comprises multiple special atoms; the special crystal is a crystal in which the bond length between each of the special atoms conforms to a preset length; The crystal property prediction model is initially pre-trained using the special crystal until the second training stopping condition is met, thus obtaining the initial crystal property prediction model. The step of training the crystal property prediction model based on the difference between the predicted property and the property label until a first training stopping condition is met, thereby obtaining a trained crystal property prediction model, includes: Based on the difference between the predicted properties and the property labels, the initial crystal property prediction model is trained until the first training stopping condition is met, thus obtaining a trained crystal property prediction model.

7. The method according to any one of claims 1 to 6, characterized in that, The crystal property prediction model includes a first network structure and a second network structure; the first network structure is a deep neural network structure; the second network structure is a convolutional neural network structure; the deep neural network structure includes two hidden layers, each hidden layer includes multiple neurons, and the activation function of each hidden layer is a linear rectified function; the convolutional neural network structure includes convolutional layers, max pooling layers, and multilayer perceptron layers.

8. A training device for a crystal property prediction model, characterized in that, The device includes: The sample acquisition module is used to acquire a sample crystal and the property label corresponding to the sample crystal; the sample crystal includes multiple sample atoms; The property prediction module is used to determine the atomic coordinates corresponding to each of the sample atoms; determine the target rotation matrix of the sample atom based on the atomic coordinates; for each sample atom among multiple sample atoms, the current sample atom is taken as the origin of the atomic coordinate system, and the hybrid orbital corresponding to the current sample atom in the atomic coordinate system is determined; at least one orbital point on the hybrid orbital is selected at a preset interval, and the candidate orbital coordinates corresponding to each orbital point are determined; according to the target rotation matrix corresponding to the current sample atom, the coordinates of each candidate orbital coordinate are transformed to obtain the target orbital coordinates corresponding to each orbital point; the target wavefunction corresponding to the sample atom is determined based on the target orbital coordinates; the orbital characteristics of the sample atom are obtained by combining the target orbital coordinates and the target wavefunction; the atomic potential energy corresponding to each sample atom is determined based on the orbital characteristics, and the predicted properties corresponding to the sample crystal are determined based on the atomic potential energy. The first training module is used to train the crystal property prediction model based on the difference between the predicted property and the property label until the first training stopping condition is met, thereby obtaining a trained crystal property prediction model; the trained crystal property prediction model is used to determine the properties of the crystal.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

11. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images