Crystal material generation method and system based on energy band condition and space group symmetry constraint

Abstract

The invention discloses a crystal material generation method and system based on an energy band condition and a space group symmetry constraint. The method comprises the following steps: obtaining crystal structure data; calculating energy band data corresponding to the crystal structure, performing standardization processing on the energy band data, and generating a training data set together with the crystal structure data; performing feature compression on the energy band data, and generating an energy band condition implicit vector by using an encoder; constructing a conditional autoregression generation model, and training an autoregression model by using the training data set to obtain a trained model; jointly inputting an embedded vector of a space group number and an energy band condition implicit vector into the trained model, generating an atom type, an atom fraction coordinate and a lattice parameter according to a Wyckoff alphabetic sequence, and generating a crystal material structure; and through position selection of Wyckoff letters in a symmetric mask constraint generation process, a lattice parameter consistency loss function optimization model is utilized, so that a predicted energy band of a crystal material structure is matched with a target energy band.

Description

Technical Field

[0001] The present invention belongs to the interdisciplinary field of computational material design and artificial intelligence, and specifically relates to a method and system for generating crystal materials based on energy band conditions and space group symmetry constraints. Background Art

[0002] Traditional methods for reverse material design rely on trial-and-error experiments or limited empirical rules, and are unable to effectively utilize electronic band data calculated using quantum mechanics. These electronic bands are key indicators of material performance and are crucial for their application. While theoretical physics allows for the precise electronic bands associated with specific material properties to be accurately derived using lattice models, previous methods have relied solely on trial-and-error experiments or limited empirical rules to generate crystalline materials with these theoretically superior electronic bands, preventing direct reverse material design from these bands.

[0003] Existing neural network models do not fully consider lattice symmetry (such as rotation and translation invariance) when processing crystal structures, resulting in prediction results that violate physical laws. Manual screening is required to obtain a reasonable crystal structure. Traditional methods usually simply map the three-dimensional crystal structure to a set of atomic coordinates in Euclidean space. This processing method essentially destroys the inherent translational symmetry and point group symmetry of the lattice. Taking convolutional neural networks as an example, their translation invariance is only designed for continuous space and cannot accurately match the discrete translational symmetry of crystal periodic points, resulting in the model showing non-physical sensitivity to basic operations such as unit cell supercell expansion and unit cell selection.

[0004] CN119007883A discloses a crystal structure prediction model. This model obtains the X-ray diffraction spectrum of a target crystal, performs crystal faceting processing using the crystal faceting module of the crystal structure prediction model, generates crystal face features, and inputs these features into an encoding module for processing to generate a target crystal structure prediction model for quickly and accurately predicting the crystal structure of a material. The main method of this patent is to design material crystals by linking X-ray diffraction spectra with crystal structure. However, these X-ray diffraction spectra themselves have no relevance to material applications. Actual material applications do not directly rely on the properties of X-ray diffraction spectra, which limits the application scenarios of this patent.

[0005] CN117423396A proposes a graph neural network crystal structure generation method based on a diffusion model, which introduces linear interpolation to improve the diversity and novelty of crystal generation. It uses a linear combination method to add Gaussian noise to two crystal structures according to their proportions, and then fuses them to obtain mixed Gaussian noise, and then gradually removes the noise. The two crystal structures can be smoothly and controllably fused, and in theory, tens of thousands of new crystal structures can be generated. The graph neural network used in this method cannot maintain spatial equivariance in atomic coordinate prediction, and requires a large amount of data enhancement to learn symmetry, which is inefficient. This architectural defect requires the model to perform a large amount of artificial enhancement on the training data (such as randomly rotating molecular conformations, mirror-flipping atomic arrangements, etc.) to force learning of the symmetry laws that should be endogenous to the physical system. This not only significantly increases the computational complexity (the amount of data amplification usually needs to reach 10-100 times the original data), but also introduces noise interference, making it difficult for the model to converge to a physically reasonable solution space. Summary of the Invention

[0006] The present invention aims to solve the deficiencies of the prior art and provides the following solutions:

[0007] A method for generating a crystal material based on energy band conditions and space group symmetry constraints comprises the following steps:

[0008] Acquiring crystal structure data, the crystal structure data including: a space group number, a first atomic species, a Wyckoff letter position sequence, a first atomic fractional coordinate, and a first lattice parameter;

[0009] Using high-throughput first-principles calculations to calculate energy band data corresponding to the crystal structure, normalizing the energy band data, and generating a training dataset together with the crystal structure data;

[0010] Performing feature compression on the energy band data and generating an energy band conditional latent vector using an encoder;

[0011] Constructing a conditional autoregressive generative model, and training the autoregressive model using the training data set to obtain a trained model;

[0012] Inputting the embedding vector of the space group number and the energy band conditional latent vector into a decoder in the trained model, generating a second atomic species, a second atomic fractional coordinate, and a second lattice parameter in Wyckoff alphabetical order, and generating a crystal material structure;

[0013] By selecting the positions of Wyckoff letters during the generation process using symmetry mask constraints, and utilizing the lattice parameter consistency loss function to optimize the model, the predicted energy bands of the generated crystal material structure are matched with the target energy bands.

[0014] Preferably, the workflow of the encoder includes:

[0015] Extracting local and global features of the energy band data using a 3D convolutional neural network, wherein the local and global features include: energy band flatness, band gap width, and Dirac cone morphology;

[0016] A graph convolutional network is used to map the k points in the energy band data into graph structure nodes, and the energy band dispersion relation is captured through edge connections to obtain a topology-sensitive energy band descriptor;

[0017] The energy band descriptor is fused with the local and global features through a cross attention mechanism to obtain the energy band conditional latent vector.

[0018] Preferably, the conditional autoregressive generative model is an improved Transformer structure, comprising: a splicing module and a decoder;

[0019] The splicing module is used to splice the embedding vector of the space group number and the band conditional latent vector;

[0020] The decoder is used to decode the spliced ​​data, output the atomic species, the atomic fraction coordinates and the lattice parameters, and generate the crystal material structure.

[0021] Preferably, the symmetry mask constraint includes:

[0022] Dynamically shield invalid Wyckoff letters according to the current space group number;

[0023] The lattice parameters are forced to satisfy the geometric constraints of the corresponding crystal system.

[0024] Preferably, the lattice parameter consistency loss function includes: Wyckoff letter classification cross entropy loss, element type cross entropy loss, fractional coordinate log-likelihood loss and lattice parameter MSE loss.

[0025] The present invention also provides a crystal material generation system based on energy band conditions and space group symmetry constraints, wherein the system applies the above-described method and includes: a parameter acquisition module, a training set construction module, a latent vector generation module, a model construction module, a structure generation module, and an optimization module;

[0026] The parameter acquisition module is used to acquire crystal structure data, wherein the crystal structure data includes: space group number, first atomic species, Wyckoff letter position sequence, first atomic fractional coordinates and first lattice parameters;

[0027] The training set construction module uses high-throughput first-principles calculations to calculate the energy band data corresponding to the crystal structure, standardizes the energy band data, and generates a training data set together with the crystal structure data;

[0028] The latent vector generation module is used to perform feature compression on the energy band data and generate an energy band conditional latent vector using an encoder;

[0029] The model building module is used to build a conditional autoregressive generation model, and use the training data set to train the autoregressive model to obtain a trained model;

[0030] The structure generation module is used to jointly input the embedding vector of the space group number and the energy band condition latent vector into the decoder in the trained model, generate a second atomic species, a second atomic fractional coordinate and a second lattice parameter in Wyckoff alphabetical order, and generate a crystal material structure;

[0031] The optimization module selects the position of Wyckoff letters in the symmetry mask constraint generation process and optimizes the model using the lattice parameter consistency loss function to match the predicted energy band of the generated crystal material structure with the target energy band.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] This invention uses Wyckoff sequence generation and dynamic masking to ensure 100% conformance to space group symmetry, avoiding the generation of invalid structures. A conditional latent vector guides the model to generate materials that simultaneously satisfy symmetry constraints and target electronic properties. An auxiliary energy band prediction network reduces the number of DFT verifications by over 70%, accelerating the materials R&D cycle. It supports cross-domain generation from semiconductors to topological materials, suitable for cutting-edge applications such as photovoltaics and quantum computing. This invention deeply integrates crystallographic symmetry principles with energy band property data, achieving efficient and controllable new material design through an innovative generative model architecture and optimization strategy. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0035] Figure 1 Schematic diagram of a method flow in an embodiment of the present invention;

[0036] Figure 2 Schematic diagram of the Wyckoff position mask mechanism according to an embodiment of the present invention;

[0037] Figure 3 The crystal structure and energy band diagram generated for Example 3 of the present invention;

[0038] Figure 4 The crystal structure and energy band diagram generated for Example 4 of the present invention. DETAILED DESCRIPTION

[0039] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0040] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0041] Example 1

[0042] In this embodiment, if Figure 1 As shown, a method for generating a crystal material based on energy band conditions and space group symmetry constraints includes the following steps:

[0043] S1. Acquire crystal structure data, where the crystal structure data includes: space group number, first atomic species, Wyckoff letter position sequence, first atomic fractional coordinate, and first lattice parameter.

[0044] In this example, crystal structure data is obtained and the space group number, first atomic species, Wyckoff position sequence, first atomic fractional coordinate, and first lattice parameter are extracted from the crystal structure data file (CIF). The Wyckoff position sequence representation encodes the crystal structure as a Wyckoff position sequence, which is generated autoregressively in descending symmetry order. This representation naturally embeds space group symmetry constraints, ensuring that the generated structure strictly adheres to the symmetry operation rules defined by the International Tables of Crystallography.

[0045] S2. Use high-throughput first-principles calculations to calculate the energy band data corresponding to the crystal structure, standardize the energy band data, and generate a training dataset together with the crystal structure data.

[0046] In this embodiment, the energy band data (including k-point grid energy distribution, band gap, and band topological characteristics) obtained by high-throughput first-principles calculations are combined to normalize the energy band data to eliminate energy dimension differences, and the lattice parameters are normalized by crystal system classification (such as only retaining the lattice length a parameter for the cubic crystal system). The crystal structure data and the energy band data are associated with the unique identifier material ID to ensure a one-to-one correspondence between structure and energy band during training, and to construct a standardized training data set.

[0047] S3. Perform feature compression on the band data and use the encoder to generate the band conditional latent vector.

[0048] In this embodiment, the encoder workflow includes: using a 3D convolutional neural network to extract local and global features of the band data, where the local and global features include band flatness, band gap width, and Dirac cone morphology; using a graph convolutional network to map k points in the band data into graph structure nodes, capturing the band dispersion relationship through edge connections, and obtaining a topology-sensitive band descriptor; and fusing the band descriptor with the local and global features through a cross-attention mechanism to obtain a 128-dimensional band conditional latent vector.

[0049] S4. Build a conditional autoregressive generative model and train the autoregressive model using the training dataset to obtain a trained model.

[0050] The conditional autoregressive generation model is an improved Transformer structure, including: a splicing module and a decoder; the splicing module is used to splice the embedding vector of the space group number and the band conditional latent vector; the decoder is used to decode the spliced ​​data, output the atomic species, atomic fractional coordinates and lattice parameters, and generate the crystal material structure.

[0051] S5. The embedded vector of the space group number and the band condition latent vector are jointly input into the decoder in the model, and the second atomic species, the second atomic fractional coordinates and the second lattice parameters are generated in Wyckoff alphabetical order, and the crystal material structure is generated.

[0052] S6. By constraining the position selection of Wyckoff letters during the generation process using symmetry masks, the training model is optimized using the lattice parameter consistency loss function to ensure that the predicted energy bands of the generated crystalline material structure match the target energy bands.

[0053] In this embodiment, the symmetry mask constraint includes: dynamically masking invalid Wyckoff letters according to the current space group number (e.g., space group No. 62 only allows letters ae), filtering other letter options through the mask layer, such as Figure 2As shown, the generation of asymmetric structures is eliminated from the root, ensuring that the structure conforms to space group symmetry; the lattice parameters are forced to satisfy the geometric constraints of the corresponding crystal system (for example, the tetragonal system forces a=b≠c, α=β=γ=90°). The lattice parameter consistency loss functions include: Wyckoff letter classification cross entropy loss, element type cross entropy loss, fractional coordinate log-likelihood loss, and lattice parameter MSE loss. The weight ratio of the lattice parameter MSE loss can be adjusted to above 0.8 in the early stages of training, and other losses can be gradually increased in the later stages to balance convergence stability and performance orientation.

[0054] In this embodiment, after generating the crystal material structure, it is necessary to: (1) automatically correct it according to the geometric rules of the crystal system to which the space group belongs to ensure that the lattice parameters and symmetry are strictly consistent; (2) call the open source tool pymatgen to perform symmetry analysis on the structure and automatically detect the consistency of the space group. If the deviation exceeds the threshold, that is, the atomic position offset > Trigger the iterative correction mechanism and readjust the fractional coordinates; (3) Submit the generated structure to the first-principles calculation software for energy band calculation, and feed the results back to the model for fine-tuning, forming a "generation-verification-optimization" closed loop to continuously improve the generation accuracy.

[0055] Example 2

[0056] In this embodiment, a crystal material generation system based on energy band conditions and space group symmetry constraints includes: a parameter acquisition module, a training set construction module, a latent vector generation module, a model construction module, a structure generation module and an optimization module.

[0057] The parameter acquisition module is used to acquire crystal structure data, which includes: space group number, first atomic species, Wyckoff letter position sequence, first atomic fractional coordinates and first lattice parameters.

[0058] The training set construction module uses high-throughput first-principles calculations to calculate the energy band data corresponding to the crystal structure, standardizes the energy band data, and generates a training data set together with the crystal structure data.

[0059] The latent vector generation module is used to perform feature compression on the band data and use the encoder to generate the band conditional latent vector.

[0060] The model building module is used to build a conditional autoregressive generative model, and use the training data set to train the autoregressive model to obtain a trained model.

[0061] The structure generation module is used to jointly input the embedding vector of the space group number and the band condition latent vector into the decoder in the trained model, generate the second atomic species, the second atomic fractional coordinates and the second lattice parameters in Wyckoff alphabetical order, and generate the crystal material structure.

[0062] The optimization module constrains the position selection of Wyckoff letters in the generation process through symmetry mask, and optimizes the model using the lattice parameter consistency loss function to make the predicted energy band of the generated crystal material structure match the target energy band.

[0063] Example 3

[0064] In this embodiment, if Figure 3 As shown in the figure, a III-V semiconductor with a band gap of 0.20 eV is designed: (1) Input the band conditions (direct band gap = 0.2 eV ± 0.01 eV, and appears at the Γ point), space group No. 216 (cubic system); (2) The model generates the Wyckoff sequence "216-a-Ga-0-0-0-c-As-1 / 4-3 / 4-3 / 4", lattice parameters DFT verified the band gap to be 0.19 eV.

[0065] Example 4

[0066] In this embodiment, if Figure 4 As shown in the figure, a Bi2Te3 analogue with topological surface states is generated: (1) the input band condition (the Dirac point cannot be located at the Γ point) is space group No. 164 (trigonal crystal system); (2) the output structure is RhSeTe, and DFT calculations confirm that it has non-trivial topological order.

[0067] The embodiments described above are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by persons skilled in the art should fall within the scope of protection defined by the claims of the present invention.

Claims

1. A method for generating crystal materials based on energy band conditions and space group symmetry constraints, characterized in that: The following steps are involved: Acquiring crystal structure data, the crystal structure data including: a space group number, a first atomic species, a Wyckoff letter position sequence, a first atomic fractional coordinate, and a first lattice parameter; Using high-throughput first-principles calculations to calculate energy band data corresponding to the crystal structure, normalizing the energy band data, and generating a training dataset together with the crystal structure data; Performing feature compression on the energy band data and generating an energy band conditional latent vector using an encoder; Constructing a conditional autoregressive generative model, and training the autoregressive model using the training data set to obtain a trained model; Inputting the embedding vector of the space group number and the energy band conditional latent vector into a decoder in the trained model, generating a second atomic species, a second atomic fractional coordinate, and a second lattice parameter in Wyckoff alphabetical order, and generating a crystal material structure; By selecting the positions of Wyckoff letters during the generation process using symmetry mask constraints, and utilizing the lattice parameter consistency loss function to optimize the model, the predicted energy bands of the generated crystal material structure are matched with the target energy bands.

2. The method for generating a crystal material based on energy band conditions and space group symmetry constraints according to claim 1, characterized in that: The encoder workflow includes: Extracting local and global features of the energy band data using a 3D convolutional neural network, wherein the local and global features include: energy band flatness, band gap width, and Dirac cone morphology; A graph convolutional network is used to map the k points in the energy band data into graph structure nodes, and the energy band dispersion relation is captured through edge connections to obtain a topology-sensitive energy band descriptor; The energy band descriptor is fused with the local and global features through a cross attention mechanism to obtain the energy band conditional latent vector.

3. The method for generating a crystal material based on energy band conditions and space group symmetry constraints according to claim 1, characterized in that: The conditional autoregressive generative model is an improved Transformer structure, including: a splicing module and a decoder; The splicing module is used to splice the embedding vector of the space group number and the band conditional latent vector; The decoder is used to decode the spliced ​​data, output the atomic species, the atomic fraction coordinates and the lattice parameters, and generate the crystal material structure.

4. The method for generating a crystal material based on energy band conditions and space group symmetry constraints according to claim 1, characterized in that: The symmetry mask constraints include: Dynamically shield invalid Wyckoff letters according to the current space group number; The lattice parameters are forced to satisfy the geometric constraints of the corresponding crystal system.

5. The method for generating crystal materials based on energy band conditions and space group symmetry constraints according to claim 1, characterized in that: The lattice parameter consistency loss function includes: Wyckoff letter classification cross entropy loss, element type cross entropy loss, fractional coordinate log-likelihood loss and lattice parameter MSE loss.

6. A system for generating crystal materials based on energy band conditions and space group symmetry constraints, the system applying the method according to any one of claims 1 to 5, characterized in that: include: Parameter acquisition module, training set construction module, latent vector generation module, model construction module, structure generation module and optimization module; The parameter acquisition module is used to acquire crystal structure data, wherein the crystal structure data includes: space group number, first atomic species, Wyckoff letter position sequence, first atomic fractional coordinates and first lattice parameters; The training set construction module uses high-throughput first-principles calculations to calculate the energy band data corresponding to the crystal structure, standardizes the energy band data, and generates a training data set together with the crystal structure data; The latent vector generation module is used to perform feature compression on the energy band data and generate an energy band conditional latent vector using an encoder; The model building module is used to build a conditional autoregressive generation model, and use the training data set to train the autoregressive model to obtain a trained model; The structure generation module is used to jointly input the embedding vector of the space group number and the energy band condition latent vector into the decoder in the trained model, generate a second atomic species, a second atomic fractional coordinate and a second lattice parameter in Wyckoff alphabetical order, and generate a crystal material structure; The optimization module selects the position of Wyckoff letters in the symmetry mask constraint generation process and optimizes the model using the lattice parameter consistency loss function to match the predicted energy band of the generated crystal material structure with the target energy band.

Images