Material inter-atomic potential prediction method based on Bayesian flow network
By combining Bayesian flow network self-supervised learning with GNS backbone network, the problem of dependence on high-precision labeled data in existing technologies is solved, achieving more efficient prediction of interatomic potentials in materials, improving the accuracy and stability of the model, and making it applicable to a wider range of material types.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-19
- Publication Date
- 2026-04-03
AI Technical Summary
Existing machine learning methods for interatomic potentials rely on large amounts of high-precision labeled data, failing to fully utilize the structural and physical information of the data itself. This results in training objectives being limited to fitting a given label, making it difficult to meet the needs of complex material design.
A self-supervised learning strategy based on Bayesian flow networks is adopted to train unlabeled material structure data by adding and removing noise. Combined with the GNS backbone network, multimodal structure information is obtained through the self-supervised learning strategy, and fine-tuning is performed on limited labeled data to construct a material interatomic potential prediction model.
It significantly improves the accuracy and stability of interatomic potential prediction in materials, reduces the dependence on high-precision DFT labeled data, lowers computational costs, and enhances the model's generalization ability in cross-system material modeling.
Smart Images
Figure CN121789856A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the interdisciplinary field of computational materials science and artificial intelligence, and in particular relates to a method for predicting the interatomic potential of materials based on Bayesian flow networks. Background Technology
[0002] Interatomic potential is a mathematical model describing the interaction potential energy at the atomic scale. It determines the total energy, forces, and dynamic behavior of a system and forms the physical foundation for computational materials science research, including molecular dynamics simulations, structure optimization, and phase transition prediction. Traditionally, the construction of interatomic potential relies on first-principles calculations, such as methods based on density functional theory (DFT). DFT obtains the total energy and atomic forces by solving the electronic structure, offering high accuracy, but its computational complexity increases cubically with the number of electrons. It typically only handles small systems with a few hundred atoms and struggles to meet the million time steps required for nanosecond-level dynamics simulations, severely limiting its application in the design of complex materials.
[0003] In recent years, machine learning technology has made groundbreaking progress in the field of interatomic potentials. Researchers have used methods such as Gaussian process regression and graph neural networks, using first-principles calculations as training data, to construct machine learning interatomic potential functions (MLIP) that balance accuracy and efficiency. Compared with traditional methods, MLIP reduces computational complexity to the linear or sublinear level of the number of atoms, while maintaining near-DFT accuracy and improving speed by 3–5 orders of magnitude. This makes long-scale simulations of alloys, interfaces, and defect systems containing hundreds of thousands or even millions of atoms possible, greatly promoting the development of high-throughput material screening, dynamic process tracking, and cross-scale coupled simulations.
[0004] Currently, the mainstream technical approach for machine learning-based interatomic potential (MLIP) mostly adopts a supervised learning paradigm: large-scale labeled datasets (including physical quantities such as energy, force, and stress corresponding to atomic structures) generated by high-precision quantum chemical calculation methods such as DFT are used to approximate the real potential energy surface using deep neural network regression models. That is, by inputting atomic structure information, the corresponding energy, force, and stress values are output. In the process of realizing this invention, the inventors found that the existing technology has at least the following problems: such methods rely on a large amount of high-precision labeled data when modeling the real potential energy surface, and their training objective is limited to fitting a given label, failing to fully utilize the structural and physical information contained in the data itself. Summary of the Invention
[0005] Purpose of the invention: The technical problem to be solved by the present invention is to provide a method for predicting the interatomic potential of materials based on Bayesian flow networks, which addresses the shortcomings of the existing technology.
[0006] To address the aforementioned technical problems, this invention discloses a method for predicting interatomic potentials in materials based on Bayesian flow networks, comprising:
[0007] Step 1: Add noise to the unlabeled material structure data to obtain the noisy material structure data;
[0008] Step 2: Construct a first prediction model, which includes a GNS backbone network and a noise prediction head; use the first prediction model to denoise and train the noisy material structure data to obtain the trained first prediction model.
[0009] Step 3: Construct a second prediction model, which includes the GNS backbone network and the interatomic potential prediction head; use the second prediction model to train the material structure data with interatomic potential labels to obtain the trained second prediction model;
[0010] Step 4: Input the structural data of the material to be predicted into the trained second prediction model to obtain the prediction results of the interatomic potential of the material to be predicted.
[0011] Further, step 1 includes:
[0012] Step 101: Initialize the distribution parameters related to the material structure data so that the distribution parameters all follow a standard normal distribution, so as to ensure that the model can cover a sufficient structural space in the early stage of training;
[0013] Step 102, sampling time step, the time step is used to control the intensity of the noise addition and denoising process of the material structure data. The introduction of the time step enables the GNS backbone network of the first prediction model to learn the multi-scale features of the material structure under different degrees of perturbation.
[0014] Step 103: Based on the sampled time step and the initialized distribution parameters, calculate the Bayesian flow distribution of the material structure at the time step;
[0015] Step 104: Sample the noisy structure to obtain the noisy material structure data.
[0016] Further, step 2 includes:
[0017] Step 201: Input the noisy material structure data into the first prediction model for denoising prediction to obtain denoised material structure data;
[0018] Step 202: Compare the denoised material structure data with the original material structure data and calculate the first loss function;
[0019] Step 203: Based on the first loss function, update the parameters of the first prediction model using the standard backpropagation algorithm, iterate the training, and finally obtain the trained first prediction model.
[0020] Furthermore, step 3 includes:
[0021] Step 301: The GNS backbone network in the second prediction model is initialized using the parameters of the GNS backbone network in the first prediction model after training.
[0022] Step 302: Input the material structure data with interatomic potential labels into the second prediction model;
[0023] Step 303: The second prediction model performs forward propagation and outputs the interatomic potential prediction results;
[0024] Step 304: Calculate the second loss function based on the inter-atomic potential prediction results and inter-atomic potential labels;
[0025] Step 305: Based on the second loss function, calculate the gradient of the second prediction model and perform backpropagation to update the parameters of the second prediction model in order to gradually optimize the model performance;
[0026] Step 306: Evaluate the performance of the second prediction model using an independent validation set;
[0027] Step 307: If the performance does not meet the preset convergence condition, return to step 303 to continue training;
[0028] Step 308: If the performance meets the preset convergence condition, the training ends and the trained second prediction model is obtained.
[0029] Step 1 introduces a Bayesian flow network. Step 2 uses a self-supervised learning strategy to jointly model multimodal structural information such as atomic coordinates and lattice matrices, obtaining high-quality potential characterizations of the material structure. Step 3 uses the trained first prediction model to initialize the network parameters of the second prediction model, and performs joint fine-tuning training on a small amount of high-precision labeled data for energy, force, and stress, enabling the second prediction model to converge quickly under limited labeling conditions and significantly improve prediction accuracy and generalization ability.
[0030] Further, in step 101, the material structure data is denoted as... Where A represents the atom type, X represents the atomic coordinates, and L represents the lattice matrix; the distribution parameters related to the material structure data include atomic coordinate distribution parameters and lattice matrix distribution parameters. Atomic coordinates and lattice matrices respectively characterize the local atomic degrees of freedom and the global unit cell deformation degrees of freedom of the material system. Through subsequent self-supervised learning stages, both types of variables are modeled simultaneously to obtain the first prediction model. This model learns the features related to local atomic forces and macroscopic unit cell stresses, thereby improving the modeling ability for force and stress prediction in the final supervised stage, resulting in the second prediction model.
[0031] Further, step 102 includes randomly sampling time steps from a uniform distribution U(0,1). This setting allows the first prediction model to access training samples evenly under different noise intensities, preventing it from being biased towards specific noise levels and thus improving its stability and generalization across the entire noise range.
[0032] Furthermore, in step 103, the Bayesian flow distribution of the material structure at the stated time step is calculated as follows:
[0033] ,
[0034] ,
[0035] in, and These are the distribution mean values of atomic coordinates and lattice matrix, respectively. The mean is Standard deviation is Gaussian distribution, and All of these are predefined hyperparameters. Indicates the atomic coordinate retention factor. The lattice matrix retention coefficient is represented by I, which represents the identity matrix. This distribution is used to characterize the probability changes of atomic coordinates and lattice matrix under perturbation, and is a key step in the subsequent generation of noisy structures.
[0036] Step 104 includes: sampling specific distribution parameters from the Bayesian flow distribution obtained in step 103, including the distribution mean of atomic coordinates and lattice matrix. and The structural data constituted This refers to the material structure data after adding noise at the current moment.
[0037] Further, step 201 includes: the first prediction model outputs predicted structural noise, including atomic coordinate noise. and lattice matrix noise ; Calculate the atomic coordinates for denoising With lattice matrix :
[0038] ,
[0039] ,
[0040] Structural data This refers to the denoised material structure data.
[0041] Furthermore, in step 202, the first loss function adopts a weighted L2 loss, where the weights are determined by predefined hyperparameters. , The first loss function L1 is defined jointly with the time step t as follows:
[0042] .
[0043] This definition can balance the training objective of the first prediction model under different noise intensities, making it gradually approach the steady-state distribution of the material.
[0044] Furthermore, in step 3, the interatomic potential of the material includes the material's energy, force, and stress. The interatomic potential labels include an energy label E, a force label F, and a stress label S. The interatomic potential prediction results include predicted energy. Predictive power and predicted stress ;
[0045] Step 304 includes: calculating the mean square error between the predicted energy, predicted force, predicted stress, and the true label, and then weighting and summing them according to predefined weights to obtain the second loss function L2.
[0046] ,
[0047] in Represents predefined energy weights. Indicates the defined force weights. This represents a predefined stress weight.
[0048] The second loss function performs weighted joint optimization of energy, atomic force, and stress errors, enabling the second prediction model to simultaneously meet the requirements of energy consistency and mechanical response consistency during training, thereby improving the overall accuracy and stability of the second prediction model in multi-physical quantity prediction tasks.
[0049] Beneficial effects: From a technical perspective, this application (1) introduces a self-supervised pre-training mechanism based on Bayesian flow, which is no longer limited to the traditional supervised learning paradigm. It can make full use of unlabeled material structure data to learn more comprehensive potential representations. (2) During the pre-training process, it simultaneously models multimodal information such as atom type, atom coordinates, and lattice matrix. Compared with methods that only consider a single coordinate mode, it can capture more complete material structure features. (3) In the fine-tuning stage, the pre-trained model parameters are used for initialization, which enables the model to converge faster under limited labeled data and significantly improves the prediction accuracy and stability of energy, force, and stress. (4) The overall process is compatible with existing mainstream network architectures (such as graph neural networks like GNS), which is convenient to combine with existing machine learning inter-atomic potential model frameworks and has good versatility and scalability.
[0050] From an application perspective, this application (1) can effectively reduce the dependence on large-scale, high-precision DFT labeled data, thereby significantly reducing the computational cost and time required for model training. (2) It has better generalization in cross-system material modeling tasks and can be applied to a wider range of chemical spaces and material types. (3) By more accurately modeling the relationship between energy and force fields, it improves the reliability and efficiency of downstream applications such as molecular dynamics simulation and material property prediction, providing strong support for the rapid screening and design of new materials. Attached Figure Description
[0051] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.
[0052] Figure 1 A flowchart illustrating a method for predicting interatomic potentials in materials based on Bayesian flow networks, provided in this application embodiment.
[0053] Figure 2 A flowchart illustrating the process of obtaining noisy material structure data in a material interatomic potential prediction method based on Bayesian flow networks, as provided in this application embodiment.
[0054] Figure 3 The flowchart of obtaining the first prediction model after training is provided in a material interatomic potential prediction method based on Bayesian flow networks for embodiments of this application.
[0055] Figure 4 The flowchart illustrates the process of obtaining a trained second prediction model in a material interatomic potential prediction method based on Bayesian flow networks, as provided in this application embodiment. Detailed Implementation
[0056] The embodiments of the present invention will now be described with reference to the accompanying drawings.
[0057] This application offers a more effective self-supervised learning strategy, beyond existing supervised learning paradigms, to learn more complete structural representations from a large amount of unlabeled material structure data. Specifically, by simultaneously modeling multimodal information of atomic coordinates and lattice matrices, a self-supervised pre-training method based on Bayesian flow networks is designed. This enables the interatomic potential prediction model to fully explore the physical laws within the data, obtaining potential representations that are robust to different material systems. Furthermore, supervised fine-tuning is performed using limited labeled energy, force, and stress data, thereby significantly improving the accuracy and generalization ability of interatomic potential modeling.
[0058] This application proposes a method for predicting interatomic potentials in materials based on Bayesian flow networks. The overall process is as follows: Figure 1 As shown, it includes:
[0059] Step 1: Add noise to the unlabeled material structure data to obtain the noisy material structure data;
[0060] like Figure 2 As shown, the specific steps include the following:
[0061] Step 101: Input the 3D structure of the material and initialize the distribution parameters. Input the 3D structure information of the material. including atomic types Atomic coordinates and lattice matrix Simultaneously, the distribution parameters related to the material structure are initialized, specifically including atomic coordinate distribution parameters and lattice matrix distribution parameters. Initially, these distribution parameters all follow a standard normal distribution to ensure that the model can cover sufficient structure space in the early stages of training.
[0062] Step 102: Sampling Time Step. Randomly sample a time step from the uniform distribution U(0,1). This is used to control the intensity of the noise addition and denoising processes. The introduction of time steps enables the model to learn the multi-scale features of the material structure under different degrees of perturbation.
[0063] Step 103: Calculate the Bayesian flow distribution. Based on the sampling time step and the initialized distribution parameters, calculate the Bayesian flow distribution of the material structure at that time step.
[0064]
[0065]
[0066] in, and These are the distribution mean values of atomic coordinates and lattice matrix, respectively. The mean is Standard deviation is Gaussian distribution, and Both are predefined hyperparameters, and in the actual implementation, both hyperparameters can be set to 0.02. Indicates the atomic coordinate retention factor. The lattice matrix retention coefficient is represented by I, which represents the identity matrix. This distribution is used to characterize the probability changes of atomic coordinates and lattice matrix under perturbation, and is a key step in the subsequent generation of noisy structures.
[0067] Step 104: Sampling the noisy structure. Specific distribution parameters, including the atomic coordinates and the mean distribution of the lattice matrix, are sampled from the Bayesian flow distribution obtained in Step 103. and The structure formed This refers to the material structure after noise has been added at the current moment.
[0068] Step 2: Construct a first prediction model, which includes a GNS backbone network and a noise prediction head; use the first prediction model to denoise and train the noisy material structure data to obtain the trained first prediction model; such as Figure 3 As shown, the specific steps include the following:
[0069] Step 201: Add noise-increasing material structure The data is input into the first prediction model for prediction. In this embodiment, the first prediction model uses the GNS (Graph Network Simulator) backbone network as the base network to model inter-atomic interactions and local environment information. A noise prediction head is then connected to the GNS backbone network to output the predicted structural noise, including atomic coordinate noise. and lattice matrix noise In the specific implementation process, the noise prediction head can use a linear layer. Specifically, the output dimension of the atomic coordinate noise prediction head is 3, and the output dimension of the lattice matrix noise prediction head is 9.
[0070] Calculate the atomic coordinates for denoising With lattice matrix Specifically, it is done through the following formula:
[0071]
[0072]
[0073] Obtain the denoised atomic coordinates With lattice matrix The structure formed That is, the material structure after noise reduction.
[0074] Step 202: Calculate the training loss. Compare the predicted structure obtained in Step 201 with the original true structure and calculate the first loss function. This embodiment uses a weighted L2 loss, where the weights are determined by predefined hyperparameters. , The time step t is determined in conjunction with the time step t, and is specifically defined as follows:
[0075]
[0076] This design can balance the training objective of the model under different noise intensities, making it gradually approach the steady-state distribution of the material.
[0077] Step 203: Based on the first loss function, update the parameters of the first prediction model using the standard backpropagation algorithm, iterate the training, and finally obtain the trained first prediction model.
[0078] Steps 1 and 2 involve training the GNS backbone network on unlabeled material structure data using a noise-adding and denoising approach, enabling it to learn the steady-state distribution representation of the material structure and thus laying the foundation for subsequent supervised fine-tuning.
[0079] Step 3: Construct a second prediction model, which includes the GNS backbone network and the interatomic potential prediction head; use the second prediction model to train the material structure data with interatomic potential labels to obtain the trained second prediction model;
[0080] After completing the pre-training, this embodiment further performs supervised fine-tuning on limited labeled data to achieve high-precision modeling of material energy, force, and stress. The fine-tuning process is as follows: Figure 4 As shown, the specific steps include:
[0081] Step 301: The GNS backbone network in the second prediction model is initialized using the parameters of the GNS backbone network in the trained first prediction model. In the specific implementation, the inter-atomic potential prediction head can use a randomly initialized multilayer perceptron.
[0082] Step 302: Input the material structure with interatomic potential tags. (The text then repeats the last part about energy (…)) ),force( ) and stress ( The material's three-dimensional structure labeled with ) is input into the second prediction model. The input structure includes atom type, atom coordinates, and lattice matrix.
[0083] Step 303: Forward computation of the second prediction model. Based on the input structure, the second prediction model performs forward propagation and outputs the prediction results, including the corresponding energy ( ),force( and stress () )value.
[0084] Step 304: Calculate the second loss function. Calculate the mean squared error between the predicted energy, predicted force, predicted stress, and the true label, and then sum them according to predefined weights to obtain the overall loss function:
[0085]
[0086] in, Represents predefined energy weights. Indicates the defined force weights. This represents the predefined stress weights. In the actual implementation, all three weights can be set to 1.
[0087] Step 305: Parameter Update. Based on the second loss function from step 304, the gradient is calculated and backpropagation is performed to update the parameters of the second prediction model, thereby gradually optimizing the performance of the second prediction model.
[0088] Step 306: Performance Evaluation. The current second prediction model is evaluated using an independent validation set, and metrics such as energy error, force error, and stress error are calculated to monitor generalization performance during training.
[0089] Step 307: Convergence Judgment. Determine whether the performance of the validation set meets the preset convergence criteria. If the convergence criteria are not met, return to step 303 to continue training; if the convergence criteria are met, proceed to step 208. The convergence criteria refer to training reaching a preset number of epochs, or continuous validation for multiple training epochs without a decrease in validation loss.
[0090] Step 308: End training. The fine-tuning phase ends, yielding the final second prediction model.
[0091] The table below shows a comparison between the method of this application and the existing directly supervised training GNS model method GNS-FT using the same dataset. As can be seen from the table, the method of this application has achieved significant advantages in material structure optimization indicators (MAE, R^2, RMSD) and stable material success discovery indicators (F1, Acc, DAF, Prec), demonstrating the powerful performance of the method of this application in modeling the interatomic potential energy of crystal materials.
[0092]
[0093] The explanations for each indicator are as follows:
[0094] F1: F1 score
[0095] Acc: Accuracy of stable material predictions
[0096] DAF: Acceleration rate for stable materials discovery
[0097] Prec: Accuracy of stable material discovery
[0098] MAE: Mean Absolute Error of Forming Energy Convex Hull Prediction
[0099] R^2: Coefficient of determination for predicting the convex hull formation energy.
[0100] RMSD: Root Mean Square Error in Material Structure Optimization
[0101] In its specific implementation, this application provides a computer storage medium and a corresponding data processing unit. The computer storage medium is capable of storing a computer program, which, when executed by the data processing unit, can run the invention's content regarding a method for predicting interatomic potentials of materials based on Bayesian flow networks, as well as some or all of the steps in various embodiments. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.
[0102] Those skilled in the art will clearly understand that the technical solutions in the embodiments of the present invention can be implemented using computer programs and their corresponding general-purpose hardware platforms. Based on this understanding, the technical solutions in the embodiments of the present invention, or the parts that contribute to the prior art, can be embodied in the form of computer programs, i.e., software products. These computer program software products can be stored in a storage medium and include several instructions to cause a device containing a data processing unit (which may be a personal computer, server, microcontroller, MUU, or network device, etc.) to execute the methods described in various embodiments or certain parts of the embodiments of the present invention.
[0103] This invention provides a method for predicting interatomic potentials in materials based on Bayesian flow networks. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment of the invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.
Claims
1. A method for predicting interatomic potentials in materials based on Bayesian flow networks, characterized in that, include: Step 1: Add noise to the unlabeled material structure data to obtain the noisy material structure data; Step 2: Construct a first prediction model, which includes a GNS backbone network and a noise prediction head; use the first prediction model to denoise and train the noisy material structure data to obtain the trained first prediction model. Step 3: Construct a second prediction model, which includes the GNS backbone network and the interatomic potential prediction head; use the second prediction model to train the material structure data with interatomic potential labels to obtain the trained second prediction model; Step 4: Input the structural data of the material to be predicted into the trained second prediction model to obtain the prediction results of the interatomic potential of the material to be predicted.
2. The method for predicting interatomic potentials of materials based on Bayesian flow networks according to claim 1, characterized in that, Step 1 includes: Step 101: Initialize the distribution parameters related to the material structure data so that all distribution parameters follow a standard normal distribution; Step 102, sampling time step, the time step is used to control the intensity of the noise addition and denoising process of the material structure data; Step 103: Based on the sampled time step and the initialized distribution parameters, calculate the Bayesian flow distribution of the material structure at the time step; Step 104: Sample the noisy structure to obtain the noisy material structure data.
3. The method for predicting interatomic potentials of materials based on Bayesian flow networks according to claim 2, characterized in that, Step 2 uses the first prediction model to denoise and train the noisy material structure data to obtain the trained first prediction model, including: Step 201: Input the noisy material structure data into the first prediction model for denoising prediction to obtain denoised material structure data; Step 202: Compare the denoised material structure data with the original material structure data and calculate the first loss function; Step 203: Based on the first loss function, update the parameters of the first prediction model using the standard backpropagation algorithm, iterate the training, and finally obtain the trained first prediction model.
4. The method for predicting interatomic potentials of materials based on Bayesian flow networks according to claim 3, characterized in that, Step 3 uses the second prediction model to train the material structure data labeled with interatomic potentials, obtaining the trained second prediction model, including: Step 301: The GNS backbone network in the second prediction model is initialized using the parameters of the GNS backbone network in the first prediction model after training. Step 302: Input the material structure data with interatomic potential labels into the second prediction model; Step 303: The second prediction model performs forward propagation and outputs the interatomic potential prediction results; Step 304: Calculate the second loss function based on the inter-atomic potential prediction results and inter-atomic potential labels; Step 305: Based on the second loss function, calculate the gradient of the second prediction model and perform backpropagation to update the parameters of the second prediction model; Step 306: Evaluate the performance of the second prediction model using an independent validation set; Step 307: If the performance does not meet the preset convergence condition, return to step 303 and continue training; Step 308: If the performance meets the preset convergence condition, the training ends and the trained second prediction model is obtained.
5. The method for predicting interatomic potentials of materials based on Bayesian flow networks according to claim 4, characterized in that, In step 101, the material structure data is denoted as... Where A represents the atom type, X represents the atom coordinates, and L represents the lattice matrix; the distribution parameters related to the material structure data include atomic coordinate distribution parameters and lattice matrix distribution parameters.
6. The method for predicting interatomic potentials of materials based on Bayesian flow networks according to claim 5, characterized in that, Step 102 includes randomly sampling time steps from a uniform distribution U(0,1). .
7. The method for predicting interatomic potentials of materials based on Bayesian flow networks according to claim 6, characterized in that, In step 103, the Bayesian flow distribution of the material structure at the stated time step is calculated as follows: ， ， in, and These are the distribution mean values of atomic coordinates and lattice matrix, respectively. The mean is Standard deviation is Gaussian distribution, and All of these are predefined hyperparameters. Indicates the atomic coordinate retention factor. I represents the lattice matrix retention coefficient, and I represents the identity matrix. Step 104 includes: sampling specific distribution parameters from the Bayesian flow distribution obtained in step 103, including the distribution mean of atomic coordinates and lattice matrix. and The structural data constituted This refers to the material structure data after adding noise at the current moment.
8. The method for predicting interatomic potentials of materials based on Bayesian flow networks according to claim 7, characterized in that, Step 201 includes: the first prediction model outputs predicted structural noise, including atomic coordinate noise. and lattice matrix noise ; Calculate the atomic coordinates for denoising With lattice matrix : ， ， Structural data This refers to the denoised material structure data.
9. The method for predicting interatomic potentials of materials based on Bayesian flow networks according to claim 8, characterized in that, In step 202, the first loss function is a weighted L2 loss, where the weights are determined by predefined hyperparameters. , The first loss function L1 is defined jointly with the time step t as follows: 。 10. The method for predicting interatomic potentials of materials based on Bayesian flow networks according to claim 9, characterized in that, In step 3, the interatomic potential of the material includes its energy, force, and stress. The interatomic potential labels include an energy label E, a force label F, and a stress label S. The interatomic potential prediction results include the predicted energy. Predictive power and predicted stress ; Step 304 includes: calculating the mean square error between the predicted energy, predicted force, predicted stress, and the true label, and then weighting and summing them according to predefined weights to obtain the second loss function L2. ， in Represents predefined energy weights. Indicates the defined force weights. This represents a predefined stress weight.