A method, device, computer equipment and storage medium for predicting crystal properties

By setting the truncation radius and constructing the truncation function, the radial and angle matrix of the environment descriptor are obtained for embedding operations, and the multi-head self-attention mechanism is input, the problem of low prediction performance of crystal elements is solved, and more accurate and general Prediction effect with strong ability.

CN117747012BActive Publication Date: 2024-12-06JILIN UNIVERSITY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410157552.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-02-04
Publication Date
2024-12-06
Estimated Expiration
2044-02-04

AI Technical Summary

Technical Problem

In the prior art, the prediction performance of crystal elements is not high, resulting in insufficient prediction performance for unknown data.

Method used

By setting the truncation radius, the truncation function is constructed, the radial and angle matrix of the environment descriptor is obtained, and the embedding operation is performed to learn the feature representation with strong generalization ability, and input it into the multi-head self-attention mechanism for prediction.

Benefits of technology

Improve the accuracy and generalization ability of crystal element properties prediction, allowing more comprehensive consideration of various aspects of environmental descriptors to predict crystal elements more accurately.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117747012B_ABST
    Figure CN117747012B_ABST
Patent Text Reader

Abstract

The present invention provides a method, device, computer equipment and storage medium for predicting crystal properties. The method comprises: constructing a truncation function, obtaining an environmental descriptor of a crystal element to be predicted according to the truncation function; obtaining a radial matrix and an angular matrix of a central atom of the crystal element to be predicted according to the radial and angular values ​​of the environmental descriptor; embedding the radial matrix and the angular matrix, and inputting the result after the embedding operation into a multi-head self-attention mechanism for property prediction. The present invention obtains a better environmental descriptor according to the truncation function, captures the structural features of the crystal element by obtaining the radial matrix and the angular matrix of the environmental descriptor, embedding the radial and angular matrices helps to learn the advanced features of the crystal element, and inputting the embedding results into the multi-head self-attention mechanism can more comprehensively consider various aspects of the environmental descriptor, thereby more accurately predicting the properties of the chemical element.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of computational chemistry and physical technology, and in particular to a method, device, computer equipment and storage medium for predicting crystal properties. Background Art

[0002] Crystal elements usually refer to the elements that make up a crystal in a crystal structure. A crystal is a crystalline form of a substance in which atoms, molecules or ions are arranged in a highly ordered manner to form a solid with a periodic structure. This ordered structure gives crystals unique physical and chemical properties. In a crystal, atoms or molecules are arranged in a specific direction and spacing to form a lattice. The specific arrangement of the lattice determines the crystal structure of the crystal, and the elements present in the lattice are called crystal elements. The arrangement of these elements directly affects the properties of the crystal, such as optical, thermal and electrical properties. Each crystal element occupies a specific position in the crystal, and the arrangement and connection between them are determined by the structure of the crystal. Different elements and crystal structures lead to various types of crystals, including metallic crystals, covalent crystals, ionic crystals, etc. Predicting crystal elements has many functions and significances. Through the prediction of crystal elements, scientists and engineers can design and discover new materials more effectively. Understanding the properties and structure of crystal elements can help predict specific properties of materials, such as electrical conductivity, thermal conductivity, mechanical properties, etc., which is of great significance for the development of advanced materials with specific application requirements, such as electronic devices, optical materials, energy storage materials, etc. It can be used for high-throughput material screening. This means that a large number of potential materials can be evaluated quickly, thereby reducing the time and cost of laboratory experiments. This is of great significance for the rapid discovery of new materials with specific properties.

[0003] For decades, the first-principles method has been widely recognized as an effective means of simulating material properties. Since the first-principles method usually requires an approximate solution to the Schrödinger equation, its computational complexity is proportional to the cube of the number of electrons. Therefore, the huge computational consumption limits its use in the simulation of large-scale systems and complex physical processes. For example, nucleation, defects and phase transitions are usually affected by scale effects, and dynamic simulations of unit cells of different scales are required. In addition, the ion diffusion process in solid electrolytes has an important influence on the conductivity of electrolyte materials, which has great research value in industrial applications. However, the simulation of this process may need to be extended to the nanosecond level. The first-principles calculation method usually cannot cover this time and space scale. At the beginning of the last century, people developed many empirical potential or semi-empirical potential methods based on physical experience. The advantages are high computational efficiency and compliance with physical experience. For example, the Lennard-Jones potential and ESMSV potential used to study the properties of rare gases, the Morse potential used to calculate the vibration energy levels of diatomic molecules, and the EAM potential used to study metal or alloy systems. These potentials are usually constructed based on physical experience or experimental parameters, so the construction process is also relatively complicated. At the same time, they can usually only obtain relatively accurate analytical solutions in a certain type of specific material system, that is, they are less portable. As a data-driven mode, machine learning methods can well balance computational accuracy and computational efficiency to meet the requirements of large-scale system simulation. If the machine learning model is built with first-principles calculation data as the driver, the first-principles calculation accuracy can be obtained with high computational efficiency. At the same time, its good computational scalability also makes it easy to design parallel algorithms on heterogeneous supercomputers in the future to achieve large-scale parallel material simulation. The machine learning potential method was proposed in the 1990s and was first used to study the diffusion of molecules on the surface. However, due to the lack of suitable descriptors, it was limited to the study of simple low-dimensional systems. In 2007, Belher and Parrinello proposed the symmetry function (SF) method to maintain the rotation, translation and permutation symmetry of the material, and extended the neural network to the high-dimensional potential energy surface (HDNNP). Replacing the first-principles force field with a trained machine learning potential greatly accelerates the speed of molecular dynamics simulation. In 2010, Bartók et al. proposed the Gaussian approximation potential (GAP) based on the smooth overlap of atomic positions (SOAP) descriptor. This method can calculate the uncertainty of each configuration in the molecular dynamics simulation through the training set (prior), effectively improving the reliability of the machine learning potential atomic simulation. However, since the covariance matrix of the entire training set needs to be constructed during the training process of GAP, it may cause the risk of computer memory overflow when applied to the simulation of complex systems. In 2016, Alexander V. Shapeev et al. developed the moment tensor potential (MTP) based on the moment tensor descriptor. The MTP potential was the most accurate linear regression neural network potential model at that time.In 2018, Zhang et al. developed and promoted the DeepPMD-kit potential based on the previous work. Xie Tian et al., KTSchütt et al. applied graph neural network models (CGCNN, SchNet) based on the graph structure of crystals or molecules to the construction of machine learning potentials, further improving the accuracy of neural network models. Simon Batzner et al. and KTSchütt et al. established equivariant neural network models such as NequIP based on graph neural networks to directly learn vector properties such as atomic forces. Although graph neural networks and equivariant neural networks have high accuracy, tests have shown that the HDNNP method, which was developed by Belher et al. in the early days, has the disadvantages of being difficult to parallelize on a large scale and having slightly lower computational efficiency.

[0004] In the prior art, application publication number CN 108536998 A, application publication date: 2018.09.14, discloses a material structure description method suitable for machine learning potential energy surface construction. The method constructs a series of characteristic functions as input information. These characteristic functions use the interatomic bond length and bond angle as basic variables, and construct the atomic environment by combining power functions, truncation functions, spherical harmonic functions and trigonometric functions, including structural information such as bonding and coordination. The global potential energy surface data of the material system is trained using a machine learning method to obtain a machine learning potential energy surface.

[0005] However, when constructing the potential energy surface using the above method, directly using the feature function as input information may cause the machine learning model to only learn primary features and fail to learn more representative advanced features, resulting in low prediction performance for unknown data. Summary of the invention

[0006] The present invention provides a crystal property prediction method, device, computer equipment and storage medium, which can solve the technical problem of low performance in predicting crystal element properties in the prior art.

[0007] The present invention provides a method for predicting crystal element properties, comprising:

[0008] A cutoff radius is set, and a cutoff function is constructed according to the cutoff radius to focus on atoms in the crystal element to be predicted that are located within the cutoff radius, and an environment descriptor of the crystal element to be predicted is obtained according to the cutoff function;

[0009] According to the radial and angular dimensions of the environmental descriptors, the structural features of the central atom of the crystal element to be predicted are obtained: radial matrix and angular matrix;

[0010] By embedding the radial matrix and the angular matrix to learn the feature representation with strong generalization ability of the crystal elements to be predicted, the result of the embedding operation is input into the multi-head self-attention mechanism that can express various aspects of the environmental descriptor to predict the properties of the crystal elements.

[0011] Furthermore, the truncation function includes:

[0012]

[0013] Among them, f c (·) is the cosine truncation function, r ij is the distance between central atoms within the cutoff radius, r c is the cutoff radius that is set.

[0014] Furthermore, the step of obtaining the environment descriptor includes:

[0015] Get the three-body term for the central atom:

[0016]

[0017]

[0018] in, is the three-body term of the central atom, l refers to the order of the angle descriptor function Legendre polynomial, where l max is the maximum order allowed by the Legendre polynomial, n is the order of the descriptor Chebyshev polynomial, nAmax is the maximum order allowed by the descriptor Chebyshev polynomial, r ij is the distance between central atoms within the cutoff radius, r ik is the distance between neighboring atoms, P l (cosθ ijk ) is the l-th order Legendre polynomial, θ ijk Yes ij and r ik The angle between

[0019] g n (r ij ) is the distance r ij Function, T n is a Chebyshev polynomial of the first kind of order n.

[0020] Furthermore, the step of obtaining the radial matrix and the angle matrix includes:

[0021] The environment descriptor of the central atom is divided into two parts according to the radial direction and angle of the environment descriptor;

[0022] The two parts are organized into two parts of size according to the element type of the neighboring atoms. and size The matrix, where N ele is the number of element types.

[0023] Furthermore, the embedding operation on the radial matrix and the angle matrix includes:

[0024] Use two multi-layer fully connected neural networks with activation functions to combine the row vectors of the two matrices to form two vectors with lengths of and A vector of

[0025] Concatenate these two vectors together to get the embedding result.

[0026] Furthermore, the result after the embedding operation is input into the multi-head self-attention mechanism for property prediction, including:

[0027] Use a linear combination layer to transform the embedding result into a feature with a length of d dimensions;

[0028] The feature of length d is evenly divided into h parts, each of which is fed into a head, and the feature dimension of each head is d h =d / h;

[0029] Each head passes through three matrices Perform projection and construct query, key and value;

[0030] The query and key are used to calculate the attention score of the central atom to the neighboring atoms through the scaled cosine scoring function:

[0031] score=cos(Q i , K i ) / τ

[0032] Among them, score is the attention score of the central atom to the neighboring atoms, Q i is the query for the ith atom, K i is the bond of the ith atom, τ is the learned scalar;

[0033] Get the output of each head based on the attention score:

[0034] Head i =score×V i

[0035] Among them, Head i is the output of the ith head, V i is the value of the ith atom;

[0036] Concatenate the output of each header to get the final output:

[0037] MultiHead(x i )=Concatenate(head1,…,head h )W O

[0038] Among them, Concatenate(·) is a concatenation operation, x i is the atomic position of the structure, W O are the parameters to be trained in the neural network.

[0039] Furthermore, the connection method in the multi-head self-attention mechanism is layer normalization and residual connection.

[0040] A crystal property prediction device, comprising:

[0041] An environmental descriptor acquisition module is used to set a cutoff radius, construct a cutoff function for focusing on atoms in the crystal element to be predicted that are located within the cutoff radius according to the cutoff radius, and obtain an environmental descriptor of the crystal element to be predicted according to the cutoff function;

[0042] A radial matrix and an angle matrix acquisition module is used to acquire the structural features of the central atom of the crystal element to be predicted according to the radial direction and angle of the environmental descriptor: a radial matrix and an angle matrix;

[0043] The crystal property prediction module is used to learn the feature representation with strong generalization ability of the crystal elements to be predicted by embedding the radial matrix and the angular matrix, and input the result of the embedding operation into the multi-head self-attention mechanism that can express various aspects of the environmental descriptor to predict the properties of the crystal elements.

[0044] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the crystal property prediction method is implemented.

[0045] A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned crystal property prediction method when executing the program.

[0046] The present invention provides a crystal property prediction method, device, computer equipment and storage medium. Compared with the prior art, the present invention has the following beneficial effects:

[0047] The present invention focuses on atoms located within a truncation radius in a crystal element to be predicted according to a truncation function, and obtains a better environment descriptor. The structural characteristics of the crystal element are captured by obtaining the radial matrix and the angular matrix of the environment descriptor. Embedding the radial and angular matrices helps to learn a feature representation with strong generalization ability of the crystal element to be predicted. The embedding result is input into a multi-head self-attention mechanism. The multi-head self-attention mechanism can consider various aspects of the environment descriptor more comprehensively, thereby more accurately predicting the properties of the crystal element. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0049] In the attached picture:

[0050] Figure 1 It is a flow chart provided in this manual;

[0051] Figure 2 It is a schematic diagram of the framework provided in this manual;

[0052] Figure 3 It is a schematic diagram of the embedding operation provided in this manual;

[0053] Figure 4 is a schematic diagram of a fully connected neural network provided in this manual;

[0054] Figure 5 This is a schematic diagram of the attention mechanism provided in this manual;

[0055] Figure 6 This is a schematic diagram of the multi-head attention mechanism provided in this manual;

[0056] Figure 7 This is a comparison chart of the accuracy of two machine learning potentials provided in this manual. Figure 6 The left side shows the energy root mean square error comparison of ARES-FF and ARES-ATTN for different elements. Figure 6 On the right is the comparison of the root mean square error of forces for different elements between ARES-FF and ARES-ATTN;

[0057] Figure 8 This is a schematic diagram of GPU training acceleration for different elements in different systems provided in this manual. The left side of the histogram is the Xeon system, and the right side is the NVDIA system;

[0058] Fig. 9 This is a parallel efficiency test of three different scale systems with 30,000, 240,000, and 1,920,000 atoms provided in this manual;

[0059] Fig.10 It is a schematic diagram of the computer equipment provided in this manual. DETAILED DESCRIPTION

[0060] The technical solutions in the embodiments of the present invention are described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention. However, it should be understood that the protection scope of the present invention is not limited by the specific implementation methods.

[0061] Example

[0062] The structure of a crystal can be described by the Bravais lattice and the unit cell. The Bravais lattice is an infinite, discrete lattice structure that is used to describe the periodic arrangement of atoms, molecules or ions in a crystal. The unit cell is the smallest repeating structural unit in a crystal and contains information about all atoms, molecules or ions in the crystal. In order to model the crystal structure, we must first describe the geometric environment of the atoms in space, and then embed the element information based on the location of the atoms. Figure 1-2 As shown in the figure, a method for predicting the properties of multi-component complex crystal elements is firstly input into the atomic position Obtain the atomic environment descriptor; then combine it with the element C in the crystal to be simulated to obtain an embedded representation; input the embedded representation into the multi-head self-attention mechanism to output the potential energy and force of the crystal structure to be predicted. Specifically, it includes the following steps:

[0063] Step 1: Get the atomic local environment descriptor

[0064] According to the principle of locality, the energy of the system can be approximately decomposed into the sum of the energy contributions of each atom. i and the atomic energy contribution E i The mapping relationship is represented by a neural network, where w i are the parameters to be trained:

[0065]

[0066] Among them, E total is the total energy of the system, q(·) is the descriptor calculation algorithm, r ij is the distance between atoms i and j, N i refers to the number of neighbor atoms in the local environment centered on atom i. Since different central atoms usually have different numbers of neighbor atoms, and the set R i It also cannot maintain the permutation symmetry of neighboring atoms, so it cannot be directly input into the neural network for training. i is mapped to an intermediate variable q of fixed dimension that preserves the symmetry, the so-called descriptor. The atomic forces and the virial tensor can be calculated by the following derivatives

[0067]

[0068]

[0069] Among them, F ij is the force exerted on atom i by atom j, U i is the atomic energy of atom i, W i Refers to the potential tensor of atom i, that is, the contribution of atom i to the unit cell pressure, which is independent of the trainable parameters above.

[0070] The improved atomic cluster expansion (ACE) method is selected. This form inherits the excellent characteristics of the ACE method, including a complete description of the local environment of the atom, linear computational complexity, and efficient convergence. In order to strike a balance between computational efficiency and accuracy, at most three-body terms are used. Specifically, for the central atom i, the two-body term The definition is as follows:

[0071]

[0072] Three-body term The definition is as follows:

[0073]

[0074] Among them, r ij and r ik is the distance between the central atom and its neighbor atoms within the cutoff radius, P l (cosθ ijk ) is the l-th order Legendre polynomial, θ ijk Yes ij and r ik The angle between them. g n (r ij ) is the distance r ij The function has the following form:

[0075]

[0076] Among them, T n is a Chebyshev polynomial of the first kind nth order, f c (r ij ) is the cosine truncation function, defined as follows:

[0077]

[0078] The cosine truncated function makes the function g n At the cutoff radius r cThe function value and derivative value at can smoothly approach zero, thus ensuring that when the atom passes through the cutoff radius, the atomic energy and atomic force change smoothly. Set the specific radial cutoff radius Small, which is computationally beneficial to reduce the computational cost of the angle descriptor. For the three-body term, the addition theorem of spherical harmonics can be used to calculate it in linear time complexity:

[0079]

[0080]

[0081] The above formula is the expanded expression of the l-th order Legendre polynomial. The simplified equation (2) can be rewritten as:

[0082]

[0083]

[0084] Among them, δ 0l It is the Kronecker symbol, which means that when l = 0, the value is 1, and when l ≠ 0, the value is 0. lm are spherical harmonics, where the subscripts 1 and m are indices of the order of the spherical harmonics.

[0085] Step 2: Embedding operation

[0086] Embedding crystal element information helps capture the chemical differences and physical properties between different elements. The type of element is directly related to the chemical bond properties and lattice structure of molecules or crystals, which helps the model learn these associations and improve the understanding of crystallization and molecular structure. In addition, for multi-element complex systems, the differences in electronic structure of different elements lead to different behaviors in materials. For example, for properties such as lattice dynamics, thermal conductivity, and electrical conductivity, the presence of different elements will affect the performance of these properties. By embedding the information of crystal elements, the changes in electronic structure can be better captured, and the electronic state and band structure can be more accurately modeled. For example, in the surface and interface problems of batteries, the performance of batteries is directly affected by multiple factors such as electrode materials, electrolytes, and interface interactions, and the simulation of battery problems usually involves multiple elements. Developing an embedding method that can accurately characterize the type of crystal elements and their position in the structure, the model can better predict the energy, structure, and other related properties of the system. Therefore, in general, crystal element information can improve the accuracy of machine learning models in predicting material properties. It is expected to bring new breakthroughs in fields such as materials science and energy storage, and provide a new perspective for in-depth understanding of the properties of multi-element condensed matter systems.

[0087] Since descriptors can only encode the spatial information of atomic environments, for complex compounds, explicitly encoding crystal element information can improve the accuracy of machine learning models. Therefore, an element embedding module is introduced to further embed element information on top of the spatial descriptors. The workflow of the embedding module is as follows: Figure 3 The module first divides the descriptor of the central atom into two parts according to the descriptor type (radial and angular), and then organizes each part into a matrix according to the element type of the neighboring atoms. ele The shapes of the two matrices obtained by the radial part and the angular part are and Next, two multi-layer fully connected neural networks with tanh activation function are used, as Figure 3 As shown, the two matrices and row vectors are combined to form two matrices with lengths of and Finally, concatenate these two vectors together.

[0088] The element embedding operation has the following benefits in this method: 1. The embedding operation learns a higher-level representation of the chemical environment descriptor. By embedding the environmental descriptor into a more semantically informative space, the model can better capture the key features that describe the chemical environment around the element, which helps improve the system's understanding and modeling capabilities of complex chemical structures. 2. The embedding operation of crystal element information can help reduce the dimensionality of the data, thereby reducing the complexity of the calculation and improving the generalization ability of the model. By mapping the environmental descriptor to a more compact representation space, some noise or redundant information can be removed, allowing the model to focus more on key structural features. 3. Improve the learning efficiency of the model: Inputting the results of the crystal element embedding operation into the multi-head self-attention mechanism can better capture the relationship between the local and the global. The self-attention mechanism can dynamically adjust the weights according to the importance of different positions, allowing the model to more flexibly handle environmental information of different scales and complexities. 4. Enhance the generalization ability of the model: The embedding operation can help the model better generalize to unseen chemical environments. By learning more abstract and expressive features, the model can better adapt to different chemical structures instead of just memorizing specific instances in the training data.

[0089] Step 3: Attention module prediction

[0090] The attention module takes descriptors as input, such as Figure 5 As shown. First, a linear combination layer is used to change the dimension to d. A multi-head self-attention mechanism similar to that in Transformer is used. The feature of length d is evenly divided into h parts, each of which is fed into a head. In this case, the feature dimension of each head is d. h =d / h. Figure 6As shown, each head first passes through three matrices Projection is performed to construct query, key and value. The query and key are used to calculate the attention score of the central atom to the neighboring atoms through the scoring function. The scaled cosine scoring function is used, and because of its natural normalization, a milder attention value can be obtained. The output of each head can be interpreted as a weighted sum of values,

[0091] score=cos(Q i ,K i ) / τ

[0092] Head i =score×V i

[0093] Where τ is a learnable scalar. Each head in the multi-head mechanism focuses on a different part of the input features, thereby achieving more comprehensive information capture. The multi-head output is concatenated by each head and can be written as,

[0094] MultiHead(x i )=Concatenate(head1,...,head h )W O

[0095] The number of atoms in the physical system to be simulated is often large. Therefore, similar to the treatment of descriptors, an artificial cutoff radius is set. The central atom only needs to calculate the attention score within the truncation radius, thus reducing the amount of calculation. By using the attention mechanism, the important information contributed by the neighboring atoms to the central atom is learned. In addition, since the output of the descriptor module already contains the truncation radius The attention layer reconstructs the truncation radius based on the local information within The weighted representation within , so the farthest (where N attn represents the number of attention modules) is associated with the representation of the central atom, such as Figure 4 To ensure that the model containing multiple layers of multi-attention modules can be trained effectively, layer normalization and residual connections are used.

[0096] ACNN (Attention Coupled Neural Network) is a software for building machine learning atomic interaction potential models. ACNN couples the multi-head self-attention mechanism to the traditional short-range machine learning potential. The advantages of this design are: first, the multi-head attention mechanism can perform focused calculations on the surrounding neighboring atoms, and automatically assign greater weights to the neighboring atoms that contribute more to the central atom, which helps to obtain a better description of the local environment of the atom. Second, while coupling the attention mechanism to the short-range neural network, the message passing mechanism is introduced, which allows ACNN to include more distant interactions into the scope of the machine learning potential, so ACNN has better accuracy than the traditional short-range machine learning potential. On the other hand, the modular design of coupling the attention layer on the short-range neural network potential allows users to freely decide whether to use the attention mechanism according to specific requirements. In this way, machine learning potential models with different accuracies can be trained according to the actual situation and task requirements of the system to be simulated.

[0097] The algorithm of the non-attention neural network model has high parallel scalability and can realize large-scale system simulation of millions of atoms on a variety of computing hardware (CPU, GPU). The parallel scalability of the attention neural network is slightly lower than the former, but it can capture more comprehensive atomic environment information, thereby obtaining a more accurate machine learning potential model.

[0098] The method provided in this specification obtains a descriptor that only contains information within the truncation radius by setting a cutoff radius, and inputs the descriptor into a multi-head self-attention mechanism. The multi-head self-attention mechanism only needs to calculate the attention score within the truncation radius, which reduces the amount of calculation. In addition, the multi-head self-attention mechanism constructs a weighted representation within the cutoff radius, which can associate the farthest atomic information with the representation of the central atom, thereby improving the recognition accuracy.

[0099] The trained machine learning model can be used for molecular dynamics simulation of large-scale systems or long-term physical processes, predicting the properties of materials such as hardness, strength, thermal conductivity, etc., and can also optimize the structure of materials to achieve better performance and stability. ACNN is written in C++, which provides complete object-oriented programming support, including encapsulation, inheritance, and polymorphism, making the program easier to understand, expand, and maintain while having high operating efficiency. In addition, ACNN training implements CUDA C++ GPU acceleration, which greatly improves the computing efficiency during the training process.

[0100] This specification also provides experimental validation of the method for predicting the chemical properties of elements:

[0101] like Figure 7As shown, the accuracy test of ACNN includes single substances, complex multi-element systems, etc. The energy error is 10-3~10-4eV / atom, and the force error is less than May be used for simulation of actual systems.

[0102] like Figure 8 As shown in the figure, ACNN uses CUDA C++ to test computing efficiency. The test includes single elements, complex multi-element systems, etc. For complex systems with a mass of more than 100 atoms, GPU can provide up to 8.5 times acceleration.

[0103] like Fig. 9 As shown in the figure, the parallel scalability test of ACNN on CPU shows that the parallel efficiency of 384 cores is higher than 70%. For a system of about 2 million atoms, the molecular dynamics learning simulation speed is about 1.5s / step.

[0104] In addition, ACNN has realized interfaces with material design software such as ARES, LAMMPS, and CALYPSO, forming a unified material research and development and design platform that can be used to accelerate the structure search of large-scale complex systems, material stability, and the simulation of various mechanical and thermal properties of materials.

[0105] The specific definition of the crystal element property prediction device can be found in the definition of the crystal element property prediction method mentioned above, which will not be repeated here. Each module in the above-mentioned crystal element property prediction device can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory in the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0106] This specification also provides a computer-readable storage medium, which stores a computer program. The computer program can be used to execute the above-mentioned crystal element property prediction method.

[0107] This manual also provides Fig.10 The structural diagram of the computer device shown in FIG. Fig.10 As shown, at the hardware level, the computer device includes a processor, an internal bus, a network interface, a memory, and a non-volatile memory, and may also include hardware required for other services. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it to implement the above-mentioned crystal element property prediction method.

[0108] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database 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 or optical memory, etc. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0109] The above-mentioned embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A method for predicting crystal properties, characterized in that: include: A cutoff radius is set, and a cutoff function is constructed according to the cutoff radius to focus on atoms in the crystal element to be predicted that are located within the cutoff radius, and an environment descriptor of the crystal element to be predicted is obtained according to the cutoff function; The truncation function comprises: Among them, f c (·) is the cosine truncation function, r ij is the distance between central atoms within the cutoff radius, r c is the set cutoff radius; The step of obtaining the environment descriptor comprises: Get the three-body term for the central atom: in, is the three-body term of the central atom, l refers to the order of the angle descriptor function Legendre polynomial, where l max is the maximum allowed order of Legendre polynomials, n is the order of the descriptor Chebyshev polynomials, is the maximum order allowed for the Chebyshev polynomial in the descriptor, r ij is the distance between the central atoms within the cutoff radius, r ik is the distance between neighboring atoms, P l (cosθ ijk ) is the lth-order Legendre polynomial, θ ijk Yes ij and r ik The angle between n (r ij is the distance r ij Function, T n is a Chebyshev polynomial of the first kind and order n; According to the radial direction and angle of the environmental descriptor, the structural features of the central atom of the crystal element to be predicted are obtained: a radial matrix and an angle matrix; the steps of obtaining the radial matrix and the angle matrix include: The environment descriptor of the central atom is divided into two parts according to the radial direction and angle of the environment descriptor; The two parts are organized into two parts of size according to the element type of the neighboring atoms. and size The matrix, where N ele is the number of element types; By embedding the radial matrix and the angular matrix to learn the feature representation with strong generalization ability of the crystal elements to be predicted, the result of the embedding operation is input into the multi-head self-attention mechanism that can express various aspects of the environmental descriptor to predict the properties of the crystal elements.

2. The crystal property prediction method according to claim 1, characterized in that: The embedding operation of the radial matrix and the angle matrix comprises: Use two multi-layer fully connected neural networks with activation functions to combine the row vectors of the two matrices to form two vectors with lengths of and A vector of Concatenate these two vectors together to get the embedding result.

3. The crystal property prediction method according to claim 1, characterized in that: The result after the embedding operation is input into the multi-head self-attention mechanism for property prediction, including: Use a linear combination layer to transform the embedding result into a feature with a length of d dimensions; The feature of length d is evenly divided into h parts, each of which is fed into a head, and the feature dimension of each head is d h =d / h; Each head passes through three matrices Perform projection and construct query, key and value; The query and key are used to calculate the attention score of the central atom to the neighboring atoms through the scaled cosine scoring function: score=cos(Q i ,K i ) / τ Among them, score is the attention score of the central atom to the neighboring atoms, Q i is the query for the ith atom, K i is the bond of the ith atom, τ is the learned scalar; Get the output of each head based on the attention score: Head i =score×V i Among them, Head i is the output of the ith head, V i is the value of the ith atom; Concatenate the output of each header to get the final output: MultiHead(x i )=Concatenate(head1,…,head h )W O Among them, Concatenate(·) is a concatenation operation, x i is the atomic position of the structure, W O are the parameters to be trained in the neural network.

4. The method for predicting crystal properties according to claim 3, characterized in that: The connection method in the multi-head self-attention mechanism is layer normalization and residual connection.

5. A crystal property prediction device, characterized in that: include: An environmental descriptor acquisition module is used to set a cutoff radius, construct a cutoff function for focusing on atoms in the crystal element to be predicted that are located within the cutoff radius according to the cutoff radius, and obtain an environmental descriptor of the crystal element to be predicted according to the cutoff function; The truncation function comprises: Among them, f c (·) is the cosine truncation function, r ij is the distance between the central atoms within the cutoff radius, r c is the set cutoff radius; The step of obtaining the environment descriptor comprises: Get the three-body term for the central atom: in, is the three-body term of the central atom, l refers to the order of the angle descriptor function Legendre polynomial, where l max is the maximum allowed order of Legendre polynomials, n is the order of the descriptor Chebyshev polynomials, is the maximum order allowed for the Chebyshev polynomial in the descriptor, r ij is the distance between the central atoms within the cutoff radius, r ik is the distance between neighboring atoms, P l (cosθ ijk ) is the lth-order Legendre polynomial, θ ijk Yes ij and r ik The angle between n (r ij is the distance r ij Function, T n is a Chebyshev polynomial of the first kind and order n; The radial matrix and angle matrix acquisition module is used to acquire the structural features of the central atom of the crystal element to be predicted according to the radial direction and angle of the environmental descriptor: the radial matrix and the angle matrix; the acquisition steps of the radial matrix and the angle matrix include: The environment descriptor of the central atom is divided into two parts according to the radial direction and angle of the environment descriptor; The two parts are organized into two parts of size according to the element type of the neighboring atoms. and size The matrix, where N ele is the number of element types; The crystal property prediction module is used to learn the feature representation with strong generalization ability of the crystal elements to be predicted by embedding the radial matrix and the angular matrix, and input the result of the embedding operation into the multi-head self-attention mechanism that can express various aspects of the environmental descriptor to predict the properties of the crystal elements.

6. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 4 is implemented.

7. A computer device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method described in any one of claims 1 to 4 is implemented.