Method for predicting new crystal structure based on generative adversarial model
By combining adversarial generative models with group theory and energy prediction proxy models, the complexity and high cost of crystal structure prediction are solved, enabling efficient and accurate generation and screening of new crystal structures. This approach is applicable to fields such as materials science, drug design, energy materials, and nanotechnology.
Patent Information
- Application Number
- CN202410907747.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-08
- Publication Date
- 2025-10-31
AI Technical Summary
Existing technologies for crystal structure prediction suffer from high computational costs, high complexity, global optimization problems, and complex handling of symmetry and periodicity, making traditional methods time-consuming, labor-intensive, and difficult to generate new crystal structures.
We employ an adversarial generative model-based approach that combines group theory, Wyckoff site analysis, and energy prediction proxy models to generate and screen crystal structures through machine learning, reducing reliance on experiments and achieving efficient prediction.
It reduces the time and cost of crystal structure prediction, improves the efficiency and accuracy of new crystal structure discovery, and is applicable to multiple scientific and engineering fields.
Smart Images
Figure CN120877970A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of new materials technology, and more specifically, to a method for predicting new crystal structures based on an adversarial generative model. Background Technology
[0002] Crystals play a crucial role in physics, chemistry, biology, materials science, and engineering. In crystallography, fully defining a crystal requires knowledge of its lattice structure, chemical composition, and the distribution of each element's sites, all while satisfying precise symmetry operations required by group theory. The possibilities and diversity of compositional combinations and symmetry site arrangements provide a near-infinite crystal composition space. In materials science research, finding new crystal structures is essential for developing new materials. However, traditional experimental methods are time-consuming and costly. While theoretical calculations, such as first-principles calculations, can predict crystal structures to some extent, they often require substantial computational resources and struggle to handle complex chemical and physical properties. In recent years, Generative Adversarial Networks (GANs) have shown great potential in areas such as image generation and natural language processing. This invention proposes an innovative GAN-based method aimed at generating and predicting three-dimensional crystal structures. This method consists of two core components: a generator and a discriminator. The generator is responsible for extracting the basic features and compositional rules of crystal structures from a database of known crystals, thereby generating new crystal structure candidates. It continuously explores novel and rational crystal structures during the generation process, pursuing innovation while ensuring compliance with fundamental chemical and physical principles. The discriminator acts as an evaluator, assessing the authenticity and quality of the generated crystals by comparing them with a real crystal library. This process not only guides the generator to optimize its generation strategy but also ensures the accuracy and reliability of the crystal structure. A significant advantage of this invention is its ability to automatically generate crystal structures through learning from large amounts of data, even in the absence of explicit prior knowledge. This method significantly reduces the time and cost required to find new crystal structures while increasing the probability of discovering rare or unknown crystal structures. Furthermore, this technology is highly adaptable, capable of being customized to generate crystal structures with specific physical, chemical, or functional properties according to specific application needs, providing a powerful impetus for advancements in materials science.
[0003] Current technologies face several challenges in crystal structure prediction, primarily due to their high complexity and computational cost, the complexity of physical and chemical rules, global optimization issues, and constraints related to symmetry and periodicity. Crystal structures are formed by the ordered arrangement of atoms in three-dimensional space, involving numerous variables such as the types, numbers, arrangement, and interaction forces of atoms. Predicting an unknown crystal structure often requires enormous computational resources, especially for material systems with complex chemical and physical properties, making the process computationally expensive. Acquiring crystal structure data typically relies on experimental measurements, which are not only time-consuming and labor-intensive, but also often result in limited available experimental data for predicting new materials. The stability of crystals is influenced not only by the types and numbers of atoms but also by factors such as the electronic structure between atoms, chemical bonding, and macroscopic conditions (e.g., temperature, pressure). These complex physicochemical interactions make predicting crystal structures solely through computational methods extremely challenging. Crystal structure prediction is essentially a process of finding the globally lowest energy configuration. However, existing algorithms, especially local optimization algorithms, easily get trapped in local minima, missing more stable global optimal solutions. Therefore, developing algorithms that can effectively avoid local minima and quickly converge to the global minimum is an important direction for current research. The symmetry and periodicity of crystals are their fundamental characteristics, and these constraints impose extremely high requirements on numerical accuracy. Therefore, correctly handling these symmetries and periodicities during the prediction process is extremely complex. Summary of the Invention
[0004] To address the shortcomings of existing technologies, this invention aims to provide a method for predicting new crystal structures based on an adversarial generative model. This method combines group theory, Wyckoff site analysis, machine learning coding techniques, and an energy prediction surrogate model to achieve efficient structure prediction and screening. By predicting crystal structures through computational simulation, the dependence on experimental conditions and the number of trial and error attempts are reduced. The combination of adversarial generative models and surrogate models enables efficient prediction and rapid screening of crystal structures. Users can customize the generation of crystal structures by specifying or not specifying conditions such as composition, space group, and total number of atoms in the unit cell.
[0005] To achieve the above objectives, the present invention employs the following technical solution and steps:
[0006] Step S1: Collect structure files from open-source crystallography databases (such as Materials Project, Jarvis, The OpenQuantum Materials Database), and convert the structure files into standard crystallography information files with the .cif extension;
[0007] Step S2: Extract lattice parameter information from the crystal structure file collected in step S1, analyze Wyckoff sites by combining the coordinate information of atoms with the symmetry of the crystal, and encode this information into vectors that can be recognized by machine learning, and organize them into a training dataset.
[0008] Step S3: Input the training dataset provided in step S2 into the generative adversarial neural network to train the two components of the neural network: the generator and the discriminator;
[0009] Step S4: Input Gaussian random noise into the model trained in step S3. In this step, you can set the configuration file for the initial generated structure's spatial group and element ratio, and collect the generated structure into a database based on atomic environment simulation (ASE).
[0010] Step S5: Perform preliminary screening of the generated structures and apply a surrogate model with crystal structure energy prediction capabilities to evaluate thermodynamic and mechanical stability;
[0011] Step S6: Perform first-principles high-precision relaxation on the structure predicted by the surrogate model, and evaluate the stability of the generated structure by calculating its phonon pairs using the DFPT method.
[0012] Preferably, in step one, the crystallography database should have reasonable theoretical calculations and evaluations (such as Materials Project, Jarvis, The Open Quantum Materials Database) to ensure the quality of the training data.
[0013] Preferably, step S1 is as follows:
[0014] Step S1.1: Use Materials Project as the database for collecting crystal structures and use the official structure download tool;
[0015] Step S1.2: Use the Pymatgen software package provided by the Materials Project to convert the extracted crystal structure and save it as a file with the .cif extension;
[0016] Preferably, step S2 is as follows:
[0017] Step S2.1: Read the processed crystallographic information file, for any crystal sample The six parameters a, b, c, α, β, γ of the corresponding lattice and the corresponding space group information can be read directly.
[0018] Step S2.2: Based on its space group g and the International Crystallographic Group Symmetry Reference Table, reconstruct the crystal sample. At symmetric sites in space group g, obtain sets of asymmetric units that cannot be obtained from each other by symmetry operations. fractional coordinates
[0019] Step S2.3: Encode the local information of the crystal (encoder) L ) and global information encoding (encoder) G Encode them separately. For local area information encoding:
[0020]
[0021] The site fractional coordinates of all asymmetric units in space group g. More specifically, for different atoms in the crystal, they are represented as:
[0022]
[0023] Key j This indicates that different elements are in asymmetric units. The embedding in the code involves removing the element type during the Wyckoff-based encoding extraction process and not preserving the element's symbol information. For global information encoding, the main components are space group information and lattice parameter information.
[0024]
[0025] Obtain the encoding vector for a single crystal.
[0026] Step S2.4: Repeat steps S2.1, S2.2 and S2.3 for other crystal structures in the database to obtain all the crystal-encoded data, thus obtaining the training dataset for training the neural network model.
[0027] Preferably, step S3 is as follows:
[0028] Step S3.1: Construct the generator part of the generative adversarial neural network, which consists of a module composed of three fully connected layers and a ReLU activation function connected in series, and then connected to a module with a number of neurons equal to the number of neurons in the network. It consists of fully connected layers of equal length and Tanh activation functions.
[0029] Step S3.2: Construct the discriminator part of the generative adversarial neural network, which consists of a module composed of three fully connected layers and a LeakRelu activation function connected in series, then a module based on one-dimensional convolutional self-attention connected by residuals, and finally a fully connected layer without an activation function.
[0030] The working principle of the residual connection module is as follows:
[0031] Algorithm: One-dimensional convolutional self-attention module Input: -Input tensor×with dimensions(batch_size, channels, length) Parameters: -in_dim:Input dimension of the tensor -query_conv,key_conv,value_conv:Convolutionlayers -gamma: Learnablescaleparameter -softmax:Softmax function for attention 1.Initializequery_conv, key_conv, value_convwith 1x1 convolution layers 2.Initialize gamma as a learnable parameter function FORWARD(x) proj_query<-query_conv(x)reshaped to(batch_size,-1,length) proj_key<-key_conv(x)reshaped to(batch_size,-1,length) proj_value<-value_conv(x)reshaped to(batch_size,-1,length) e<-batch matrix multiplication of proj_query(permuted)and proj_key attention<-softmax(e)along the last dimension out<-batch matrix multiplication of proj_value and attention(permuted) out<-reshape out to the dimensions of x out<-gamma*out+x
[0032] Where batch_size, channels, and length are the batch size, number of channels, and crystal encoding length during model training, respectively. Other process parameters required for execution are explained in the algorithm table.
[0033] Step S3.3: The loss function is trained using the WGAN (Wasserstein Generative Adversarial Network) framework, and the loss function is designed as follows: Where c represents the sampling result from the true distribution p(c), z represents the sampling result from the Gaussian distribution p(z), and G θ and f w These represent the generator and discriminator, respectively, which contain parameters. The learning rate is fixed at 0.001, and the optimization method is set to Adam.
[0034] Step S3.4: Observe the change of the loss function during the training process. Stop training when the loss value is less than 0.001 after 50 optimization iterations and save the model.
[0035] Preferably, step S4 is as follows:
[0036] Step S4.1: Set the composition of the constraints, the target space group, and the total number of atoms in the unit cell;
[0037] Step S4.2: Input random noise based on Gaussian distribution into the adversarial generative model that has been trained in step S3.4. The model generates and automatically stores the structure into a preset database based on atomic environment simulation (ASE).
[0038] Preferably, step S5 is as follows:
[0039] Step S5.1: Convert the crystal structure stored in the ASE database in step S4.2 into a structure accepted by the surrogate model;
[0040] Step S5.2: Input the structure into the surrogate model for predicting crystal energy to obtain a basic estimate of the formation energy. The calculation formula for the formation energy is: E form =E tota1 -∑n i E i E form It is the energy of formation, E tota1 It is the total energy of the crystal, n iE represents the number of different atoms in the unit cell of a crystal. i It is the reference energy of different atoms in the crystal;
[0041] Step S5.3: Based on the crystal structure obtained after screening in S5.2, the phonon spectrum in the crystal is further calculated using a surrogate model combined with the Density Functional Perturbation Theory (DFPT) method.
[0042] Preferably, step S6 is as follows:
[0043] Step S6.1: Based on the structures screened in step S5.3, the candidate structures are finely optimized using first-principles calculation methods, the formation energy is calculated, and the crystal structures in the dataset are sorted according to the lowest value of the formation energy, thereby obtaining the priority of the predicted structures;
[0044] Step S6.2: Based on the priority crystal structure obtained in step S6.1, perform phonon calculations using a first-principles method to obtain the verification results.
[0045] Compared with the prior art, the present invention has the following obvious and prominent substantive features and significant advantages:
[0046] 1. This invention reduces the trial-and-error costs of experimental screening by generating models to fully utilize existing data and predict new crystal structures. This invention can guide experiments, avoiding high costs and blind spots;
[0047] 2. The generative model proposed in this invention, which combines crystallographic knowledge, can be applied not only to materials science, but also to other engineering fields, such as drug design, energy materials, and nanotechnology, and has broad application prospects.
[0048] 3. The generation method proposed in this invention retains a limited low-degree structure during the generation process, which can further support other scientific research such as machine learning force fields.
[0049] 4. This invention employs group theory to construct a reversible code for crystals starting from Wyckoff sites. It then combines this code with a pre-trained surrogate model (M3GNet is used as the surrogate model in this patent) with energy prediction capabilities to achieve rapid screening of generated structures. This reduces experimental trial-and-error costs; it implements a method for structure prediction based on an adversarial generative model; and it provides a method for rapid screening of generated structures. Customized crystal structure generation can be performed with or without specifying composition, space group, and the total number of atoms within the unit cell.
[0050] The attached figures illustrate a, b, c, α, β, and γ.
[0051] Figure 1 This is a statistical distribution diagram of the space group extracted from the crystal structure in Materials Prokect in the example.
[0052] Figure 2 This is a schematic diagram of the generator (a) and discriminator (b) in the embodiment.
[0053] Figure 3 The above is a scatter plot showing the lattice parameters of the generated structure before and after optimization in the embodiment. The horizontal axis represents the values before optimization, and the vertical axis represents the values after optimization. Wherein (a) is the length of the lattice parameter on the principal axis a; (b) is the length of the lattice parameter on the principal axis b; (c) is the length of the lattice parameter on the principal axis c; (d) is the angle between axis b and axis c; (e) is the angle between axis a and axis c; and (f) is the angle between axis b and axis a.
[0054] Figure 4 The examples show the phonon spectra of the new structures generated in different space groups and their stability calculated according to first principles, where (a) is the structure in space group 1; (b) is the structure in space group 25; (c) is the structure in space group 35; (d) is the structure in space group 71; (e) is the structure in space group 99; (f) is the structure in space group 123; (g) is the structure in space group 166; (h) is the structure in space group 194; and (i) is the structure in space group 221. Detailed Implementation
[0055] The present invention will now be described in detail with reference to the accompanying drawings and specific preferred embodiments, but this should not be construed as limiting the scope of protection of the present invention.
[0056] Example 1
[0057] In this embodiment, the ternary oxide Ba-R u The new structure of the -O system was verified by prediction and theoretical calculation, and the steps are as follows:
[0058] Step S1: Download all crystal structures from the Materials Project (MP) database and save the crystal structures as .cif files using the Pymatgen algorithm package provided by MP;
[0059] Step S2.1: Analyze and read the acquired crystal structure. The lattice parameters and space group information can be directly read from the .cif file.
[0060] Step S2.2: Consult the International Crystallographic Group Symmetry Reference Table based on the space group information to obtain the equivalent site information and the composition of asymmetric units for all crystal structures. In the ternary oxide Ba-Ru-O system, select only structural samples with 3 and 4 sites in the asymmetric units. Mark the samples that meet the requirement of having 3 or 4 sites in the asymmetric units and add them to the structure dataset for training.
[0061] Step S2.3: Encode the training dataset obtained after filtering in S2.2, with each crystal being encoded into a vector. In this vector, the fractional coordinates of the asymmetric units are placed first. The space group number g and the crystal lattice parameters a, b, c, α, β, γ were then spliced together.
[0062] Step S2.4: Traverse all crystal structures of the training preparation dataset obtained after filtering in S2.2, and perform max-min normalization on the dataset.
[0063] Step S3.1: Construct the generator part of the generative adversarial neural network, which consists of a module composed of three fully connected layers and a ReLU activation function connected in series, and then connected to a module with a number of neurons equal to the number of neurons in the network. It consists of fully connected components of equal length and Tanh activation functions;
[0064] Step S3.2: Construct the discriminator part of the generative adversarial neural network, which consists of a module composed of three fully connected layers and a LeakRelu activation function connected in series, then a module based on one-dimensional convolutional self-attention connected by residuals, and finally a fully connected layer without an activation function.
[0065] Step S3.3: The loss function is trained using the WGAN (Wasserstein Generative Adversarial Network) framework, and the loss function is designed as follows: Where c represents the sampling result from the true distribution p(c), z represents the sampling result from the Gaussian distribution p(z), and G θ and f w These represent the generator and discriminator, respectively, which contain parameters. The learning rate is fixed at 0.001, and the optimization method is set to Adam.
[0066] Step S3.4: Observe the change of the loss function during the training process. Stop training when the loss value is less than 0.001 after 50 optimization iterations, and save the model as a .pth file.
[0067] Step S4.1: Set the constraint composition to Ba, Ru and 0 elements, set no restrictions on the target space group, and set the total number of atoms in the unit cell to be less than or equal to 20;
[0068] Step S4.2: Load the saved model and input random noise based on Gaussian distribution into the adversarial generative model that has been trained in step S3.4. The model generates and automatically stores the structure into a preset database based on atomic environment simulation (ASE).
[0069] Step S5.1: Convert the structure in the ASE library into an object that can be read by the proxy model M3GNet using Pymatgen;
[0070] Step S5.2: Predict the formation energy of the structure processed in S5.1;
[0071] Step S5.3: Identify structures with formation energy less than 0 in the structure formation energy prediction of S5.2, calculate the phonon spectrum based on the surrogate model M3GNet, and retain structures that do not have imaginary frequencies in the reciprocal space of the crystal;
[0072] Step S6.1: Based on the structures selected in Step S5.3, and using VASP software, a plane wave cutoff energy of 600 eV is designed using first-principles calculations. Fine-tuning of the structure is performed using GGA-PBE functionals combined with van der Waals corrections, with a convergence criterion of 10 constrained to the energy level. After convergence, the formation energy is calculated, and the crystal structures in the dataset are ranked according to their lowest formation energy values to obtain the priority of the predicted structures.
[0073] Step S6.2: Calculate the phonon spectrum of the structure obtained in S6.1 one by one according to its priority to obtain the final rigorously selected new structure.
Claims
1. A method for predicting new crystal structures based on an adversarial generative model, characterized in that it includes: Step 1. Data Acquisition and Preprocessing: 1.1 Crystal structure files are collected from open-source crystallography databases and converted into standard crystallography information files with the .cif extension; 1.2 Extract lattice parameters and atomic coordinate information from the crystal structure file, and perform Wyckoff site analysis using the crystal symmetry to encode the crystal structure information and form a training dataset in vector form; Step 2. Construction and training of the adversarial generative model: An adversarial generative neural network is trained using the aforementioned training dataset. This adversarial generative neural network comprises a generator and a discriminator. The generator is responsible for generating new crystal structures, and the discriminator is responsible for evaluating the authenticity of the generated structures. The two compete against each other during the training process to optimize the quality of the generated crystal structures. Step 3. Structure Generation and Preliminary Screening: 3.1 Input Gaussian random noise into the trained adversarial generative model to generate a new crystal structure; 3.2 A surrogate model with energy prediction capability is used to preliminarily screen the generated crystal structures and evaluate their thermodynamic and mechanical stability; Step 4. Validation and Stability Assessment: 4.1 First-principles high-precision relaxation is performed on the crystal structures that have passed the initial screening to further optimize the crystal structure parameters; 4.2 Phonon spectra were calculated using density functional perturbation theory to further evaluate the stability of the crystal structure.
2. The method for predicting new crystal structures based on an adversarial generative model according to claim 1, characterized in that, Step 1.2 specifically includes: Step 1.2.1: Read the processed crystallography information file and select a sample of any crystal. Directly read the six parameters a, b, c, α, β, γ of the corresponding lattice and the corresponding space group. Step 1.2.2: Reconstruct the crystal sample based on space group g and the International Crystallographic Group Symmetry Reference Table. Obtain asymmetric unit groups at symmetric sites in space group g. fractional coordinates Step 1.2.3: Encode the crystal structure information into a machine learning-recognizable vector form, including encoding the local and global information of the crystal, as follows: Local information encoding of the crystal (ecoder) L The formula is as follows: In the formula, Let be the fractional coordinates of the positions of all asymmetric units in the space group g; The different atoms in a crystal are represented as: In the formula, Key j Indicates different elements in asymmetric units Embedded within; Encoder for global information in crystal G This includes information on space group and lattice parameters: Obtain the encoding vector for a single crystal. Step 1.2.4: Repeat steps S2.1, S2.2 and S2.3 for other crystal structures in the database until all crystal-encoded data are obtained, which is the training dataset for training the neural network model.
3. The method for predicting new crystal structures based on an adversarial generative model according to claim 2, characterized in that, The Wyckoff site analysis utilizes group theory to analyze the symmetry and Wyckoff sites in crystals, transforming complex crystal structures into asymmetric unit groups.
4. The method for predicting new crystal structures based on an adversarial generative model according to claim 1, characterized in that, Step 2, the construction and training of the adversarial generative model, specifically includes: Step S2.1: Construct the generator part of the generative adversarial neural network, which consists of a module composed of three fully connected layers and a ReLU activation function connected in series, and then connected to a module with a number of neurons equal to the number of neurons in the network. It consists of fully connected components of equal length and Tanh activation functions; Step S2.2: Construct the discriminator part of the adversarial generative neural network, which consists of a module composed of three fully connected layers and a LeakRelu activation function connected in series, then a module based on one-dimensional convolutional self-attention connected by residuals, and finally a fully connected layer without an activation function. Step S2.3: The loss function is trained using the WGAN framework. Where c represents the sampling result from the true distribution p(c), and z represents the sampling result from the Gaussian distribution p(z), G θ and f w These represent a generator and a discriminator with parameters, respectively; the learning rate is fixed at 0.001, and the optimization method is set to Adam. Step 2.4: Observe the changes in the loss function during the training process. Stop training when the loss value is less than 0.001 after 50 optimization iterations, and save the adversarial generative model.
5. The method for predicting new crystal structures based on an adversarial generative model according to claim 1, characterized in that, Step 3.1 is as follows: 3.1.1 Set the constraint composition, target space group, and total number of atoms in the unit cell; 3.1.2 Input random noise based on Gaussian distribution into the trained adversarial generative model. The model generates and automatically stores the structure in a pre-set database based on atomic environment simulation (ASE).
6. The method for predicting new crystal structures based on an adversarial generative model according to claim 1, characterized in that, Step 3.2 is as follows: 3.2.1 By inputting the crystal structure into the surrogate model for predicting crystal energy, a basic estimate of the formation energy is obtained. The calculation formula for the formation energy is: E form =E total -∑n i E i E form It is the energy of formation, E total It is the total energy of the crystal, n i E represents the number of different atoms in the unit cell of a crystal. i It is the reference energy of different atoms in the crystal; 3.2.2 Based on the crystal structure obtained after preliminary screening, the phonon spectrum in the crystal is further calculated using a surrogate model combined with the Density Functional Perturbation Theory (DFPT) method.