A world-model-based molecular dynamics simulation method, system, terminal, and storage medium

By constructing an initial atomic system and training a world model, the problem that existing molecular dynamics simulation methods struggle to describe bonding changes and charge transfer is solved, improving the accuracy of the simulation, particularly in applications such as lithium-ion batteries, catalytic materials, and biomedicine.

CN123120072APending Publication Date: 2026-11-17PENG CHENG LAB
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611387415.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-08
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing molecular dynamics simulation methods struggle to accurately describe bond changes, electronic polarization, and charge transfer, leading to reduced simulation accuracy.

Method used

A world-model-based molecular dynamics simulation method is adopted to improve accuracy by constructing an initial atomic system, training the world model, and performing trajectory simulations.

Benefits of technology

It improves the accuracy of molecular dynamics simulations, especially in the fields of lithium-ion batteries, catalytic materials, and biomedicine.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123120072A_ABST
    Figure CN123120072A_ABST
Patent Text Reader

Abstract

This invention discloses a molecular dynamics simulation method, system, terminal, and storage medium based on a world model. The method includes: constructing atomic systems to obtain atomic element information and multiple initial atomic systems; constructing world model training samples based on the atomic element information and multiple initial atomic systems, and performing batch processing to obtain training data and test data; constructing a loss function based on the loss information of the training data to obtain a target loss function; iteratively training the initial world model based on the training data, the target loss function, and simulation conditions to obtain a target world model; and performing trajectory simulation using the target world model based on the test data to obtain a target dynamic trajectory. This invention constructs a world model, trains the world model based on model training samples to obtain a target world model, and performs trajectory simulation using the target world model to obtain a target dynamic trajectory, thereby improving the accuracy of molecular dynamics simulation.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of molecular dynamics simulation technology, and in particular to a molecular dynamics simulation method, system, terminal, and computer-readable storage medium based on a world model. Background Technology

[0002] Molecular dynamics simulations can reveal the structure, interactions, and dynamic evolution of matter at the atomic scale, serving as a crucial computational tool bridging microscopic mechanisms and macroscopic properties. In the materials industry, molecular dynamics simulations can be used to study crystal defects, ion diffusion, and interfacial reactions, aiding in the design and optimization of materials such as batteries, catalysts, and alloys. In the biopharmaceutical industry, they are widely used for protein conformation analysis, drug-target binding mechanism research, and lead compound screening, thereby reducing experimental trial-and-error costs, shortening R&D cycles, and improving product performance.

[0003] Existing molecular dynamics simulation methods calculate the forces acting on atoms using empirical force fields or first-principles methods and integrate atomic motion step by step in femtosecond time steps. However, this molecular dynamics simulation method is difficult to accurately describe bonding changes, electronic polarization, charge transfer, and complex reactions, resulting in reduced accuracy of molecular dynamics simulations.

[0004] Therefore, existing technologies still need to be improved and developed. Summary of the Invention

[0005] The main objective of this invention is to provide a molecular dynamics simulation method, system, terminal, and computer-readable storage medium based on a world model, aiming to solve the problem that existing molecular dynamics simulation methods are unable to accurately describe bonding changes, electronic polarization, charge transfer, and complex reactions, leading to a decrease in the accuracy of molecular dynamics simulations.

[0006] To achieve the above objectives, this invention provides a molecular dynamics simulation method based on a world model, comprising the following steps: The crystal information and simulation conditions of the target molecule are obtained, and the atomic system is constructed based on the crystal information and simulation conditions to obtain atomic element information and multiple initial atomic systems. Based on the atomic element information and multiple initial atomic systems, world model training samples are constructed to obtain multiple world model training samples. The multiple world model training samples are then processed in batches to obtain training data and test data. The loss information of the training data is configured to obtain sample loss information, and the loss function is constructed based on the sample loss information to obtain the target loss function. Determine an initial world model, and iteratively train the initial world model based on the training data, the target loss function, and the simulation conditions to obtain a target world model; The target world model is used to perform trajectory simulation processing based on the test data to obtain the target dynamic trajectory, and the simulation results are evaluated based on the target dynamic trajectory to obtain the simulation evaluation results.

[0007] Optionally, the molecular dynamics simulation method based on the world model, wherein acquiring the crystal information and simulation conditions of the target molecule, and performing atomic system construction processing based on the crystal information and simulation conditions to obtain atomic element information and multiple initial atomic systems, specifically includes: The initial crystal structure of the target molecule is obtained, and the initial crystal structure is normalized to obtain atomic element information; The unit cell of the target molecule is replicated to obtain a crystal supercell, and the crystal supercell is subjected to molecular system construction processing to obtain multiple molecular systems; By adding strain tensors to multiple molecular systems, multiple strain simulation systems are obtained. The simulation conditions are obtained, and the simulation conditions are configured for multiple strain simulation systems according to the simulation conditions to obtain multiple initial atomic systems.

[0008] Optionally, the molecular dynamics simulation method based on the world model, wherein the step of constructing world model training samples based on the atomic element information and multiple initial atomic systems to obtain multiple world model training samples, and processing the multiple world model training samples in batches to obtain training data and test data, specifically includes: A density functional theory calculation algorithm is determined, and the first parameters of multiple initial atomic systems are calculated using the density functional theory calculation algorithm to obtain multiple first parameters; Multiple initial atomic systems were subjected to structural optimization and single-point energy calculations to obtain multiple second parameters. The research objective is obtained, and trajectory generation processing is performed based on the research objective, the atomic element information, multiple first parameters and multiple second parameters to obtain multiple first-principles molecular dynamics trajectories; Training samples were constructed and processed based on multiple first-principles molecular dynamics trajectories to obtain multiple world model training samples; Multiple training samples of the world model are processed in batches to obtain training data and test data.

[0009] Optionally, in the world model-based molecular dynamics simulation method, the sample loss information includes average atomic coordinate loss, velocity prediction loss, average energy loss, force loss, stress loss, time continuity loss, and isotropic consistency loss. The process of configuring the loss information of the training data to obtain sample loss information, and constructing a loss function based on the sample loss information to obtain the target loss function, specifically includes: The training data is processed by defining the average atomic coordinate loss and the velocity prediction loss to obtain the average atomic coordinate loss and the velocity prediction loss. The training data is processed by defining average energy loss, force loss, and stress loss to obtain the average energy loss, the force loss, and the stress loss. The training data is processed using the temporal continuity loss definition to obtain the temporal continuity loss; Obtain the rotation matrix and translation vector, and perform equivariant consistency loss construction on the training samples based on the rotation matrix and translation vector to obtain the equivariant consistency loss. Obtain the weights of the loss terms, and perform function construction based on the weights of the loss terms, the average atomic coordinate loss, the velocity prediction loss, the average energy loss, the force loss, the stress loss, the time continuity loss, and the isovariance consistency loss to obtain the target loss function.

[0010] Optionally, in the aforementioned molecular dynamics simulation method based on a world model, the initial world model includes a scientific foundation model, a molecular system encoding module, a simulation condition encoding module, and an isovariant dynamics expert model. The process of determining the initial world model, and iteratively training the initial world model based on the training data, the target loss function, and the simulation conditions to obtain the target world model, specifically includes: A scientific foundation model is determined, and the user instructions in the training data are processed by text conversion using the scientific foundation model to obtain instruction representation vectors. A simulation condition encoding module is determined, and the simulation condition and the instruction representation vector are processed by the simulation condition encoding module to obtain a unified condition vector. A molecular system encoding module is determined, and the atomic element information is processed by the molecular system encoding module to obtain the molecular system characterization; A variable dynamics expert model is determined, and the future trajectory of the molecule is obtained by performing future trajectory prediction processing based on the molecular system characterization and the unified condition vector through the variable dynamics expert model. The target world model is obtained by iteratively generating future trajectories based on the molecular future trajectory and the target loss function using the isovariant dynamics expert model.

[0011] Optionally, the molecular dynamics simulation method based on the world model, wherein the step of obtaining the target world model by iteratively generating future trajectories using the isovariant dynamics expert model based on the future molecular trajectory and the target loss function, specifically includes: Based on the future molecular trajectory and the molecular system characterization, the molecular system characterization is updated to obtain the updated molecular system characterization. The target world model is obtained by iteratively generating future trajectories using the isovariant dynamics expert model, based on the unified condition vector, the updated molecular system characterization, and the target loss function.

[0012] Optionally, the molecular dynamics simulation method based on the world model, wherein the step of performing trajectory simulation processing based on the test data using the target world model to obtain the target dynamic trajectory, and performing simulation result evaluation processing based on the target dynamic trajectory to obtain the simulation evaluation result, specifically includes: The target dynamic trajectory is obtained by performing a first trajectory simulation on the test data using the target world model. A classical force field molecular dynamics algorithm is determined, and the test data is processed by the classical force field molecular dynamics algorithm to obtain a comparative dynamic trajectory. Based on the target dynamic trajectory and the comparison dynamic trajectory, a Pearson correlation coefficient comparison is performed to obtain a first comparison result; A second comparison result is obtained by performing root mean square error comparison processing on the target dynamic trajectory and the comparison dynamic trajectory; A third comparison result is obtained by comparing the acceleration ratios of the target dynamic trajectory and the comparison dynamic trajectory. The simulation evaluation results are obtained based on the first comparison result, the second comparison result, and the third comparison result.

[0013] Furthermore, to achieve the above objectives, the present invention also provides a world-model-based molecular dynamics simulation system, wherein the world-model-based molecular dynamics simulation system comprises: An atomic system construction module is used to acquire the crystal information and simulation conditions of the target molecule, and to perform atomic system construction processing based on the crystal information and simulation conditions to obtain atomic element information and multiple initial atomic systems. The training sample construction module is used to construct world model training samples based on the atomic element information and multiple initial atomic systems, to obtain multiple world model training samples, and to process the multiple world model training samples in batches to obtain training data and test data. The loss function construction module is used to configure and process the loss information of the training data to obtain sample loss information, and to construct the loss function based on the sample loss information to obtain the target loss function. The world model training module is used to determine an initial world model, and to iteratively train the initial world model based on the training data, the target loss function, and the simulation conditions to obtain a target world model. The simulation evaluation module is used to perform trajectory simulation processing based on the test data using the target world model to obtain the target dynamic trajectory, and to evaluate the simulation results based on the target dynamic trajectory to obtain the simulation evaluation results.

[0014] Furthermore, to achieve the above objectives, the present invention also provides a terminal, wherein the terminal includes: a memory, a processor, and a world-model-based molecular dynamics simulation program stored in the memory and executable on the processor, wherein when the world-model-based molecular dynamics simulation program is executed by the processor, it implements the steps of the world-model-based molecular dynamics simulation method as described above.

[0015] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a world-model-based molecular dynamics simulation program, which, when executed by a processor, implements the steps of the world-model-based molecular dynamics simulation method as described above.

[0016] In this invention, crystal information and simulation conditions of the target molecule are acquired. Based on the crystal information and simulation conditions, atomic system construction is performed to obtain atomic element information and multiple initial atomic systems. Based on the atomic element information and the multiple initial atomic systems, world model training samples are constructed to obtain multiple world model training samples. These multiple world model training samples are processed in batches to obtain training data and test data. Loss information of the training data is configured to obtain sample loss information, and a loss function is constructed based on the sample loss information to obtain a target loss function. An initial world model is determined, and iterative training is performed on the initial world model based on the training data, the target loss function, and the simulation conditions to obtain a target world model. Trajectory simulation is performed using the target world model based on the test data to obtain a target dynamic trajectory, and simulation results are evaluated based on the target dynamic trajectory to obtain a simulation evaluation result. This invention constructs initial atomic systems, builds model training samples based on the initial atomic systems, constructs a world model, trains the world model based on the model training samples to obtain a target world model, and performs trajectory simulation using the target world model to obtain the target dynamic trajectory, thus improving the accuracy of molecular dynamics simulation. Attached Figure Description

[0017] Figure 1 This is a flowchart of a preferred embodiment of the molecular dynamics simulation method based on the world model of the present invention; Figure 2 This is a schematic diagram of the LiFePO4 crystal structure based on the world model-based molecular dynamics simulation method of this invention; Figure 3 This is a world model architecture diagram of the molecular dynamics simulation method based on the world model of this invention; Figure 4 This is a structural diagram of a preferred embodiment of the molecular dynamics simulation system based on the world model of the present invention; Figure 5 This is a structural diagram of a preferred embodiment of the terminal of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0019] Molecular dynamics simulations can reveal the structure, interactions, and dynamic evolution of matter at the atomic scale, serving as a crucial computational tool bridging microscopic mechanisms and macroscopic properties. In the materials industry, molecular dynamics simulations can be used to study crystal defects, ion diffusion, and interfacial reactions, aiding in the design and optimization of materials such as batteries, catalysts, and alloys. In the biopharmaceutical industry, they are widely used for protein conformation analysis, drug-target binding mechanism research, and lead compound screening, thereby reducing experimental trial-and-error costs, shortening R&D cycles, and improving product performance.

[0020] Existing molecular dynamics simulation methods calculate the forces acting on atoms using empirical force fields or first-principles methods and integrate atomic motion step by step in femtosecond time steps. However, this molecular dynamics simulation method is difficult to accurately describe bonding changes, electronic polarization, charge transfer, and complex reactions, resulting in reduced accuracy of molecular dynamics simulations.

[0021] To address the aforementioned issues, this invention proposes a molecular dynamics simulation method based on a world model. This method involves constructing an initial atomic system, generating model training samples based on the initial atomic system, constructing a world model, training the world model using the model training samples, obtaining a target world model, and then performing trajectory simulation using the target world model to obtain the target dynamic trajectory. This improves the accuracy of molecular dynamics simulations.

[0022] The preferred embodiment of the molecular dynamics simulation method based on the world model described in this invention, such as... Figure 1 As shown, the world model-based molecular dynamics simulation method includes the following steps: Step S10: Obtain the crystal information and simulation conditions of the target molecule, and perform atomic system construction processing based on the crystal information and simulation conditions to obtain atomic element information and multiple initial atomic systems.

[0023] Specifically, the initial crystal structure of the target molecule is obtained, the initial crystal structure is standardized to obtain atomic element information, the unit cell of the target molecule is replicated to obtain a crystal supercell, and the crystal supercell is subjected to molecular system construction processing to obtain multiple molecular systems.

[0024] The world-model-based molecular dynamics simulation method proposed in this invention can be applied to multiple fields. For example, in the field of lithium-ion batteries, it can be used to simulate ion migration and structural phase transitions in positive and negative electrode materials, solid electrolytes, and electrode-electrolyte interfaces, evaluate the ion diffusion capability of materials, and assist in the screening of high-rate electrode materials. In the field of catalytic materials, it can be used to simulate the adsorption, diffusion, reaction, and desorption processes of reactants on catalyst surfaces, and analyze catalytic reaction pathways and active sites. In the field of biomedicine, it can be used to simulate the conformational changes and long-term dynamic behavior of proteins, nucleic acids, small drug molecules, and their complexes. In specific implementation, this embodiment selects the long-range bulk diffusion simulation of lithium (Li) ions in lithium iron phosphate (LiFePO4) materials as a specific application scenario for introduction.

[0025] For the long-range bulk diffusion of Li ions in LiFePO4, the research objective is defined as: to simulate the long-term migration process of Li ions under different delithiation degrees, Li vacancy concentrations, temperatures, and strain conditions inside the LiFePO4 crystal, and to obtain the migration trajectory, mean square displacement, diffusion coefficient, and other results of Li ions.

[0026] The initial crystal structure of LiFePO4 was obtained from a publicly available crystal structure database. This example uses an orthorhombic olivine structure, and the initial structure includes the atomic types, atomic coordinates, lattice vectors, space group, and periodic boundary information for Li, Fe, P, and O elements. The obtained LiFePO4 crystal structure was normalized, converting the atomic coordinates to Cartesian coordinates, and the lattice matrix was represented as follows: ,in, , and These represent three lattice vectors, respectively. For the first... Calculate Cartesian coordinates for each atom. ,in, Represents the lattice matrix. Represents fractional coordinates.

[0027] The processed information, such as atom types and atom coordinates, is used as atomic element information for subsequent training sample construction.

[0028] Next, a LiFePO4 supercell system was constructed. Replication was performed along the three lattice directions of the LiFePO4 unit cell. , and Next, the size is obtained as The supercell. The total number of atoms in the supercell. Represented as ,in, This indicates the number of atoms in a single unit cell. Each unit cell of LiFePO4 contains 4 LiFePO4 units, therefore =4x(1+1+1+4)=28, in this embodiment, we choose... =4、 =12 and =4, which enables Li ions to undergo continuous transitions between multiple lattice sites, thus forming migration trajectories that can be used to calculate long-range diffusion properties.

[0029] Molecular systems with different degrees of delithiation and Li vacancy concentrations were constructed. The LiFePO4 system was represented as follows: ,in, This indicates the proportion of Li occupied, and .like Figure 2 As shown, based on a preset delithiation degree, a certain number of Li atoms are removed from the Li lattice sites of the intact LiFePO4 supercell, resulting in Li vacancy concentrations of varying degrees. System. For example, for a supercell containing 768 Li lattice sites, when removing... When there are 10 Li atoms, the Li vacancy concentration is expressed as: ,in, This indicates the number of lattice sites in the supercell.

[0030] The corresponding Li occupancy ratio is expressed as .

[0031] In the specific implementation, they are constructed separately. =0.95, 0.75, 0.50, 0.25 and 0.05 For the same vacancy concentration, multiple different Li vacancy configurations can be generated using random sampling, uniform distribution sampling, clustered distribution sampling, or energy-based screening methods to cover different local chemical environments.

[0032] Furthermore, strain tensor addition processing is performed on multiple molecular systems to obtain multiple strain simulation systems; simulation conditions are obtained, and simulation condition configuration processing is performed on multiple strain simulation systems according to the simulation conditions to obtain multiple initial atomic systems.

[0033] Construct a simulation system incorporating lattice strain. (This is based on the previously constructed...) Based on this, the lattice matrix Applying strain tensor (This can represent uniaxial tension, uniaxial compression, shear strain, or isotropic volumetric strain), yielding the strained lattice matrix. .

[0034] Define the conditions for the molecular dynamics simulation. Set the temperature for the atomic system after applying the strain tensor. ,pressure Simulation ensemble, time step and target simulation duration Temperatures can be set to 300K, 500K, 800K, and 1000K; the simulation ensemble can be a microcanonical ensemble, a canonical ensemble, or an isothermal-isobaric ensemble.

[0035] The initial atomic system with configured simulation conditions undergoes structural checks and energy optimization. First, the system is checked for atomic overlap, abnormal bond lengths, excessively short interatomic distances, or unreasonable local structures. Then, density functional theory is used to geometrically optimize the atomic positions until the total energy change, maximum atomic force, and stress satisfy the preset convergence conditions. The optimized atomic structure serves as the initial state for subsequent high-precision data generation and world model simulation.

[0036] Step S20: Based on the atomic element information and multiple initial atomic systems, construct world model training samples to obtain multiple world model training samples, and process the multiple world model training samples in batches to obtain training data and test data.

[0037] Specifically, a density functional theory calculation algorithm is determined, and the first parameters of multiple initial atomic systems are calculated using the density functional theory calculation algorithm to obtain multiple first parameters; structural optimization processing and single-point energy calculation processing are performed on multiple initial atomic systems to obtain multiple second parameters.

[0038] Defining density functional theory (DFT) calculation parameters. For different initial atomic systems, DFT calculations based on plane waves and pseudopotentials are used to obtain the system energy, atomic forces, and stresses. The exchange-correlation functional is approximated using the generalized gradient method. The plane wave cutoff energy, k-point sampling density, electronic self-consistent convergence threshold, and atomic force convergence threshold are set according to the target accuracy. For example, the plane wave cutoff energy can be set to 400–700 eV, and the electronic energy convergence threshold can be set to... eV, and the maximum stress threshold for structural optimization is set to 0.01–0.05 eV / Å.

[0039] For atomic systems with different supercell sizes, delithiation degrees, Li vacancy arrangements, and strain states, structural optimization and single-point energy calculations were performed to obtain the total system energy E, atomic forces F, and stress tensor for each configuration. And the optimized atomic coordinates. Force on each atom The following is obtained by differentiating the total energy of the system with respect to its atomic coordinates: The data obtained in this step is used to enable the world model to learn the mapping relationship between different atomic structures and energy and forces, covering equilibrium structures, metastable structures, and high-energy non-equilibrium structures.

[0040] Furthermore, the research objective is obtained, and trajectory generation processing is performed based on the research objective, the atomic element information, multiple first parameters, and multiple second parameters to obtain multiple first-principles molecular dynamics trajectories.

[0041] Based on the research objectives of the Li ion long-range bulk diffusion mission, and targeting different... Within a pre-defined parameter range, the system generates continuous dynamic trajectories using first-principles molecular dynamics. For each system, at each time step, the forces acting on atoms are calculated using density functional theory, and the atomic positions and velocities are updated using an integral method.

[0042] For time step The state of a single-frame atomic system is represented as ; in, Indicates the type of atom, Represents atomic coordinates, Represents atomic velocity. Represents the lattice matrix. Represents the system's energy. This indicates that the atom is subjected to force. Represents the stress tensor. The state transition between adjacent time steps is represented as... By continuously preserving the system states at different time steps, first-principles molecular dynamics trajectories with time sequence can be obtained.

[0043] Furthermore, training samples are constructed based on multiple first-principles molecular dynamics trajectories to obtain multiple world model training samples; these multiple world model training samples are then processed in batches to obtain training data and test data.

[0044] The constructed first-principles molecular dynamics trajectories were screened and cleaned. First, configurations that failed to converge in electronic self-consistency calculations, exhibited abnormal atomic forces, had excessively short interatomic distances, or suffered non-target structural disruptions were removed. Then, the continuity of timestamps within the same system was checked. For repetitive or highly similar configurations, deduplication was performed based on atomic environment descriptors, root mean square deviation of structure, or energy differences to avoid excessive concentration of training data near equilibrium states. Samples with energy, force, or stress exceeding reasonable ranges were manually checked and removed.

[0045] Construct training samples for the world model. For each system, at time... Previous historical states are used as model input, and future states are used as input. Trajectory blocks at each time step are used as supervisory labels. A single training sample is represented as... ,in, Indicates the length of the history status window. Indicates the simulation conditions. Indicates the future The atomic coordinates, displacements, velocities, energies, and forces at each time step were recorded. All cleaned data were grouped according to atomic system, trajectory, and simulation conditions to obtain training, validation, and test sets. The training and validation sets were used as training data, and the test set as test data. Further independent test sets were constructed, including temperatures, delithiation levels, vacancy arrangements, and strain ranges not present during the training phase, to verify the model's generalization ability under out-of-distribution conditions.

[0046] Step S30: Configure the loss information of the training data to obtain sample loss information, and construct the loss function based on the sample loss information to obtain the target loss function.

[0047] Specifically, the sample loss information includes average atomic coordinate loss, velocity prediction loss, average energy loss, force loss, stress loss, time continuity loss, and isotropic consistency loss; the training data is processed by defining the average atomic coordinate loss and the velocity prediction loss to obtain the average atomic coordinate loss and the velocity prediction loss.

[0048] The world model in this invention includes a scientific foundation model, a molecular system encoding module, a simulation condition encoding module, and an isovariant dynamics expert model. A loss function is defined during the world model training process based on the world model's architecture.

[0049] First, we define the average atomic coordinates and velocity prediction loss.

[0050] The average atomic coordinate loss is expressed as: ; in, This represents the total number of negative samples in the entire batch. Indicates the first Atoms in Coordinates of time, This represents the prediction of the first world model. Atoms in When calculating the coordinate error, the predicted coordinates at each time point are expanded according to the periodic boundary conditions to avoid inaccurate error calculations caused by periodic boundary jumps.

[0051] The velocity prediction loss is expressed as: ; in, Indicates the first Atoms in The speed of time This represents the prediction of the first world model. Atoms in Predicting speed at any given moment.

[0052] Furthermore, the training data is processed by defining average energy loss, force loss, and stress loss to obtain the average energy loss, the force loss, and the stress loss.

[0053] Define average energy, force, and stress loss.

[0054] The average energy loss is expressed as: ; in, Indicates the first Atoms in The energy of moments This represents the predicted energy obtained through the world model.

[0055] The loss of force is expressed as: ; in, Indicates the first Atoms in The force at any moment, This represents the predicted forces obtained through the world model.

[0056] Stress loss is expressed as: ; in, Indicates the first Atoms in Stress at any moment This represents the predicted stress obtained through the world model.

[0057] By introducing energy, force, and stress losses, the future trajectory generated by the model is made consistent with the potential energy surface of density functional theory.

[0058] Furthermore, the training data is processed by defining a temporal continuity loss to obtain the temporal continuity loss.

[0059] Define temporal continuity and kinematic consistency loss.

[0060] Construct a time continuity loss based on the relationship between predicted coordinates and velocity: .

[0061] By constructing a time continuity loss, the velocity changes of adjacent time steps are made consistent with the direction of the force on the atom.

[0062] Further, the rotation matrix and translation vector are obtained, and the training samples are processed by the rotation matrix and translation vector to construct the equivariant consistency loss, thereby obtaining the equivariant consistency loss.

[0063] Define the equivariant consistency loss. Randomly apply a rotation matrix to the training samples. Translation vector The model predictions were performed on the original system and the transformed system respectively. The original prediction results were then compared with the prediction results of the transformed system after undergoing the same rotation and translation transformations to obtain the isovariant consistency loss.

[0064] Equivariant consistency loss is expressed as: ; in, This indicates predictions made using an isovariant dynamics expert model within the world model. This represents the training samples.

[0065] The robustness of the model to different coordinate systems, cell orientations, and supercell construction methods is improved by using equivariant consistency loss.

[0066] Furthermore, the weights of the loss terms are obtained, and the target loss function is obtained by constructing a function based on the weights of the loss terms, the average atomic coordinate loss, the velocity prediction loss, the average energy loss, the force loss, the stress loss, the time continuity loss, and the isovariance consistency loss.

[0067] The total training loss of the world model is expressed as: ; ; ; in, Indicates the total loss. Indicates data loss. Indicates constraint loss, This represents the average atomic coordinate loss weight. Indicates the weight of the speed prediction loss. Indicates the average energy loss weight. Indicates the weight of force loss. Indicates the stress loss weight. Indicates the weight of time continuity loss. This represents the weight of the equivariant consistency loss.

[0068] Each loss weight can be dynamically adjusted according to different training stages. For example, in the early stage of training, the weights of coordinate, energy, and force losses can be increased to enable the model to learn local potential energy surfaces; in the later stage of training, the weights of temporal continuity and long-range rolling losses can be increased to enable the model to learn stable long-term dynamic evolution laws.

[0069] Step S40: Determine the initial world model, and iteratively train the initial world model according to the training data, the target loss function, and the simulation conditions to obtain the target world model.

[0070] Specifically, the initial world model includes a scientific foundation model, a molecular system encoding module, a simulation condition encoding module, and an isovariant dynamics expert model; the scientific foundation model is determined, and the user instructions in the training data are processed by text conversion using the scientific foundation model to obtain an instruction representation vector; the simulation condition encoding module is determined, and the simulation conditions and the instruction representation vector are processed by simulation condition encoding to obtain a unified condition vector; the molecular system encoding module is determined, and the atomic element information is processed by molecular system encoding to obtain a molecular system representation.

[0071] like Figure 3 As shown, the world model includes a scientific foundation model, a molecular system encoding module, a simulation condition encoding module, and an isovariant dynamics expert model. Human instructions are input into the scientific foundation model, molecular systems into the molecular system encoding module, and simulation conditions into the simulation condition encoding module. The outputs of the scientific foundation model, the molecular system encoding module, and the simulation condition encoding module share the representation of the world model and are input into the output predictive future trajectory block of the isovariant dynamics expert model. The predictive future trajectory block is used for autoregressive world state updates to iteratively predict future trajectories.

[0072] The input to the World Model (WM) is represented as follows: ; in, This indicates human instructions or research objectives entered by the user. Indicates the type of atom, Indicates the current atom coordinates. Indicates the current atomic velocity. Represents the lattice matrix. This indicates simulation conditions such as temperature, pressure, time step, and target simulation duration. Represents the historical dynamic state.

[0073] World Model Outputs the Future Trajectory blocks at each time step .

[0074] Scientific Basic Model This is used to understand human input commands, material composition, and research objectives, and to transform natural language into command representation vectors. First, the user commands are... Convert it into a text sequence, and then obtain the instruction representation vector through the scientific foundation model. For example, if the input user instruction is "simulate long-range bulk diffusion of Li ions under 500K conditions", the scientific foundation model will encode the above instruction into a corresponding instruction representation vector.

[0075] Analog Conditional Encoding Module Complete simulation conditions, such as temperature, pressure, time step, degree of delithiation, Li vacancy concentration, strain tensor, and target simulation duration, are fully encoded. Continuous variables are mapped to continuous condition vectors through a normalization layer and a multilayer perceptron, while discrete variables are mapped to discrete condition vectors through an embedding layer. The simulation condition representations obtained in this step are then fused with the instruction representation vectors output by the scientific foundation model to obtain a unified condition vector. .

[0076] Molecular system coding module Representing an atomic system as a three-dimensional graph with periodic boundary conditions can be achieved using an isovariant graph neural network. In this network, each atom corresponds to a node. Atom node features include at least atomic number, element type, mass, charge state, current coordinates, and velocity; when the periodic shortest distance between two atoms is less than a preset cutoff radius... At that time, a graph edge is constructed between two atoms. .

[0077] After considering periodic boundary conditions, atoms With atoms The relative displacement between them is expressed as: ; in, , Each represents an atom With atoms Location coordinates, This represents the periodic translation vector that minimizes the distance between atoms.

[0078] atomic distance Expanded into radial basis functions, relative directions Represented as directional features, and local atomic environment features are extracted using an isomorphic graph neural network to obtain atomic features. initial moment Hidden state ,in, Represents atoms kind of.

[0079] By fusing the hidden states of all atoms, we can obtain the time step. Characterization of corresponding molecular systems .

[0080] Furthermore, an isovariant dynamics expert model is determined, and the future trajectory of the molecule is obtained by performing future trajectory prediction processing based on the molecular system characterization and the unified condition vector through the isovariant dynamics expert model.

[0081] Equivariant dynamics expert model An isofluid matching model is adopted, using the current molecular system characterization and unified condition vector as input, for each future time step The model predicts the displacement of each atom. speed change And predicting energy Predicting forces and predicted stress .

[0082] The predicted future trajectory output by the isotropic dynamics expert model is: .

[0083] The predicted coordinates and predicted velocities are expressed as follows: .

[0084] When the input atomic system undergoes rotation or translation, the model output should satisfy the corresponding isovariability, that is, for the rotation matrix... Translation vector and training samples The model satisfies .

[0085] Furthermore, based on the molecular future trajectory and the molecular system characterization, the molecular system characterization is updated to obtain an updated molecular system characterization; through the isovariant dynamics expert model, based on the unified condition vector, the updated molecular system characterization, and the target loss function, an iterative future trajectory generation process is performed to obtain the target world model.

[0086] Construct a trajectory block generation and autoregressive update mechanism. The model generates trajectory blocks during a single inference process. Instead of just generating the next time step, the system generates future trajectories for each consecutive time step. The final state of the generated trajectory block is used as the initial state for the next inference, resulting in an updated characterization of the molecular system. ,in, This indicates a state update. The next trajectory block is then generated. By repeatedly performing trajectory block generation and world state updates, a complete world state representation and dynamic trajectory over the target simulation duration are obtained.

[0087] The model parameters are continuously adjusted during iterative training based on the target loss function. A phased training strategy is employed to train the world model. The first phase trains the atomic system encoding module and the energy and force prediction module, enabling the model to learn the mapping relationship between the local atomic environment and the potential energy surface. The second phase trains the trajectory block generation module, enabling the model to learn short-term dynamic changes. The third phase jointly fine-tunes the scientific foundation model, the simulation condition encoding module, the molecular system encoding module, and the isovariant dynamics expert model, enabling the model to perform corresponding tasks according to different user instructions and simulation conditions. When the model converges or reaches the preset number of iterations, training is complete, and the target world model is obtained.

[0088] Step S50: Perform trajectory simulation processing based on the test data using the target world model to obtain the target dynamic trajectory, and perform simulation result evaluation processing based on the target dynamic trajectory to obtain the simulation evaluation result.

[0089] Specifically, the target dynamic trajectory is obtained by performing a first trajectory simulation process on the test data using the target world model.

[0090] Based on the target world model, respectively in the test set and independent test sets The test is conducted by inputting the initial atomic structure, simulation conditions and research instructions into the world model, continuously executing the trajectory block generation process until the cumulative simulation time reaches the target simulation time. During the simulation, atomic coordinates, velocity, energy, force, stress and the total running time of the simulation process are saved.

[0091] The simulation results of Li ion diffusion trajectories are processed. For Li ions crossing periodic boundaries, the unfolded coordinates are recovered based on their positional changes in adjacent time steps and the lattice vector. Let the encapsulation coordinates of the Li ion be... The periodic translation count is Then the expanded coordinate representation is The true long-range migration distance of Li ions is calculated using unfolded coordinates, thus avoiding the misidentification of Li ions as reverse jumps when they enter from one side of the supercell to the other.

[0092] Calculate the mean square displacement of Li ions. For those containing For a system with 1 Li ion, the average square displacement is expressed as: Considering the directionality of Li ion diffusion, calculate the average square displacement along the a, b, and c directions of the lattice respectively: , where α represents the direction. By comparing the mean squared displacement (MSD) curves in different directions, the main diffusion directions and diffusion channels of Li ions are identified.

[0093] Calculate the Li ion diffusion coefficient. Perform linear fitting on the time period after the MSD curve enters the linear diffusion region. 3D diffusion, diffusion coefficient Represented as: We used multiple time points for averaging and calculated the confidence interval of the diffusion coefficient using a block statistical method to reduce the impact of random fluctuations in a single trajectory on the results.

[0094] Further, a classical force field molecular dynamics algorithm is determined, and the test data is processed by the classical force field molecular dynamics algorithm to obtain a comparative dynamic trajectory; the Pearson correlation coefficient is compared between the target dynamic trajectory and the comparative dynamic trajectory to obtain a first comparison result; the root mean square error is compared between the target dynamic trajectory and the comparative dynamic trajectory to obtain a second comparison result; the acceleration ratio is compared between the target dynamic trajectory and the comparative dynamic trajectory to obtain a third comparison result; and the simulation evaluation result is obtained based on the first comparison result, the second comparison result, and the third comparison result.

[0095] Using the same initial structure and simulation conditions, simulations were performed using classical force-field molecular dynamics. MSD indices were calculated on the same test configuration. Indicators and total simulation process runtime .

[0096] The simulation results of the molecular dynamics simulation method and the classical force field molecular dynamics method disclosed in this invention are compared with the true values ​​on the test set to verify the advantages of the method in terms of simulation accuracy, speed, and generalization. For the performance metrics, the Pearson correlation coefficient (PCC) with the baseline method (density functional theory precision simulation method) in the time dimension is calculated; for the performance metrics, the root mean square error (RMSE) between the method and the baseline method is calculated; for the speed performance metrics, the speedup ratio relative to the baseline method is calculated. Finally, the generalization ability of the method is verified by comparing the differences with the test results.

[0097] As shown in Table 1, the results show that the method of the present invention has significant advantages over the classical force field molecular dynamics method in terms of simulation accuracy, speed and generalization.

[0098] Table 1: Comparison of Molecular Dynamics Simulation Results of the Invention with Other Simulation Methods

[0099] The technical effects of this invention include: A molecular dynamics simulation framework based on a world model is proposed to predict the future dynamic evolution of a system, which can be applied to systems such as battery materials, catalytic materials, semiconductor materials, and biomolecules. A shared-world state characterization method for molecular dynamics simulation is proposed, which encodes atom type, atom coordinates, velocity, lattice parameters, periodic boundary conditions, and historical dynamic information into a unified characterization to describe the current state of a molecular system and its dynamic laws of evolution over time. An isovariant dynamics expert model and trajectory block generation method are proposed to satisfy the constraints of rotation, translation and periodic boundary conditions. The model generates future trajectory blocks containing multiple continuous time steps from the current world state and realizes long-term molecular dynamics simulation through trajectory block autoregression update. A multi-objective joint loss training method for molecular dynamics simulation models is proposed, which integrates atomic trajectory prediction accuracy, potential energy surface consistency, and dynamic continuity into the model optimization process, thereby improving the physical rationality, temporal continuity, and long-scale simulation stability of future trajectory predictions. A multi-level verification mechanism is proposed to comprehensively verify the accuracy, physical consistency, computational efficiency, and generalization ability of molecular dynamics simulation methods from aspects such as microscopic trajectory accuracy, macroscopic diffusion properties, computational efficiency, and generalization ability.

[0100] This invention constructs an initial atomic system, builds model training samples based on the initial atomic system, constructs a world model, trains the world model based on the model training samples, obtains a target world model, and performs trajectory simulation through the target world model to obtain the target dynamic trajectory, thereby improving the accuracy of molecular dynamics simulation.

[0101] Furthermore, such as Figure 4 As shown, based on the above-described world-model-based molecular dynamics simulation method, this invention also provides a world-model-based molecular dynamics simulation system, wherein the world-model-based molecular dynamics simulation system includes: The atomic system construction module 51 is used to acquire the crystal information and simulation conditions of the target molecule, and to perform atomic system construction processing based on the crystal information and simulation conditions to obtain atomic element information and multiple initial atomic systems. The training sample construction module 52 is used to construct world model training samples based on the atomic element information and multiple initial atomic systems to obtain multiple world model training samples, and to process the multiple world model training samples in batches to obtain training data and test data. The loss function construction module 53 is used to configure and process the loss information of the training data to obtain sample loss information, and to construct the loss function based on the sample loss information to obtain the target loss function. The world model training module 54 is used to determine an initial world model and iteratively train the initial world model based on the training data, the target loss function, and the simulation conditions to obtain a target world model. The simulation evaluation module 55 is used to perform trajectory simulation processing based on the test data using the target world model to obtain the target dynamic trajectory, and to perform simulation result evaluation processing based on the target dynamic trajectory to obtain the simulation evaluation result.

[0102] Furthermore, such as Figure 5 As shown, based on the above-mentioned molecular dynamics simulation method and system based on the world model, the present invention also provides a terminal, which includes a processor 10, a memory 20 and a display 30. Figure 5 Only some of the terminal components are shown; however, it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0103] In some embodiments, the memory 20 may be an internal storage unit of the terminal, such as a hard disk or memory. In other embodiments, the memory 20 may be an external storage device of the terminal, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc. Further, the memory 20 may include both internal and external storage devices. The memory 20 is used to store application software and various types of data installed on the terminal, such as the program code installed on the terminal. The memory 20 can also be used to temporarily store data that has been output or will be output. In one embodiment, the memory 20 stores a world-model-based molecular dynamics simulation program 40, which can be executed by the processor 10 to implement the world-model-based molecular dynamics simulation method of this application.

[0104] In some embodiments, the processor 10 may be a central processing unit (CPU), a microprocessor, or other data processing chip, used to run program code stored in the memory 20 or process data, such as executing the world model-based molecular dynamics simulation method.

[0105] In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. The display 30 is used to display information on the terminal and to display a visual user interface.

[0106] In one embodiment, when the processor 10 executes the world-model-based molecular dynamics simulation program 40 in the memory 20, the following steps are performed: The crystal information and simulation conditions of the target molecule are obtained, and the atomic system is constructed based on the crystal information and simulation conditions to obtain atomic element information and multiple initial atomic systems. Based on the atomic element information and multiple initial atomic systems, world model training samples are constructed to obtain multiple world model training samples. The multiple world model training samples are then processed in batches to obtain training data and test data. The loss information of the training data is configured to obtain sample loss information, and the loss function is constructed based on the sample loss information to obtain the target loss function. Determine an initial world model, and iteratively train the initial world model based on the training data, the target loss function, and the simulation conditions to obtain a target world model; The target world model is used to perform trajectory simulation processing based on the test data to obtain the target dynamic trajectory, and the simulation results are evaluated based on the target dynamic trajectory to obtain the simulation evaluation results.

[0107] The process of acquiring the crystal information and simulation conditions of the target molecule, and then constructing an atomic system based on the crystal information and simulation conditions to obtain atomic element information and multiple initial atomic systems, specifically includes: The initial crystal structure of the target molecule is obtained, and the initial crystal structure is normalized to obtain atomic element information; The unit cell of the target molecule is replicated to obtain a crystal supercell, and the crystal supercell is subjected to molecular system construction processing to obtain multiple molecular systems; By adding strain tensors to multiple molecular systems, multiple strain simulation systems are obtained. The simulation conditions are obtained, and the simulation conditions are configured for multiple strain simulation systems according to the simulation conditions to obtain multiple initial atomic systems.

[0108] Specifically, the process of constructing world model training samples based on the atomic element information and multiple initial atomic systems to obtain multiple world model training samples, and then processing these multiple world model training samples in batches to obtain training data and test data, includes: A density functional theory calculation algorithm is determined, and the first parameters of multiple initial atomic systems are calculated using the density functional theory calculation algorithm to obtain multiple first parameters; Multiple initial atomic systems were subjected to structural optimization and single-point energy calculations to obtain multiple second parameters. The research objective is obtained, and trajectory generation processing is performed based on the research objective, the atomic element information, multiple first parameters and multiple second parameters to obtain multiple first-principles molecular dynamics trajectories; Training samples were constructed and processed based on multiple first-principles molecular dynamics trajectories to obtain multiple world model training samples; Multiple training samples of the world model are processed in batches to obtain training data and test data.

[0109] The sample loss information includes average atomic coordinate loss, velocity prediction loss, average energy loss, force loss, stress loss, time continuity loss, and isotropic consistency loss. The process of configuring the loss information of the training data to obtain sample loss information, and constructing a loss function based on the sample loss information to obtain the target loss function, specifically includes: The training data is processed by defining the average atomic coordinate loss and the velocity prediction loss to obtain the average atomic coordinate loss and the velocity prediction loss. The training data is processed by defining average energy loss, force loss, and stress loss to obtain the average energy loss, the force loss, and the stress loss. The training data is processed using the temporal continuity loss definition to obtain the temporal continuity loss; Obtain the rotation matrix and translation vector, and perform equivariant consistency loss construction on the training samples based on the rotation matrix and translation vector to obtain the equivariant consistency loss. Obtain the weights of the loss terms, and perform function construction based on the weights of the loss terms, the average atomic coordinate loss, the velocity prediction loss, the average energy loss, the force loss, the stress loss, the time continuity loss, and the isovariance consistency loss to obtain the target loss function.

[0110] The initial world model includes a scientific foundation model, a molecular system coding module, a simulation condition coding module, and an isovariant dynamics expert model. The process of determining the initial world model, and iteratively training the initial world model based on the training data, the target loss function, and the simulation conditions to obtain the target world model, specifically includes: A scientific foundation model is determined, and the user instructions in the training data are processed by text conversion using the scientific foundation model to obtain instruction representation vectors. A simulation condition encoding module is determined, and the simulation condition and the instruction representation vector are processed by the simulation condition encoding module to obtain a unified condition vector. A molecular system encoding module is determined, and the atomic element information is processed by the molecular system encoding module to obtain the molecular system characterization; A variable dynamics expert model is determined, and the future trajectory of the molecule is obtained by performing future trajectory prediction processing based on the molecular system characterization and the unified condition vector through the variable dynamics expert model. The target world model is obtained by iteratively generating future trajectories based on the molecular future trajectory and the target loss function using the isovariant dynamics expert model.

[0111] Specifically, the step of generating the target world model through iterative future trajectory generation using the isovariant dynamics expert model based on the molecular future trajectory and the target loss function includes: Based on the future molecular trajectory and the molecular system characterization, the molecular system characterization is updated to obtain the updated molecular system characterization. The target world model is obtained by iteratively generating future trajectories using the isovariant dynamics expert model, based on the unified condition vector, the updated molecular system characterization, and the target loss function.

[0112] Specifically, the process of simulating the target's dynamic trajectory using the target world model based on the test data to obtain the target's dynamic trajectory, and then evaluating the simulation results based on the target's dynamic trajectory to obtain the simulation evaluation results, includes: The target dynamic trajectory is obtained by performing a first trajectory simulation on the test data using the target world model. A classical force field molecular dynamics algorithm is determined, and the test data is processed by the classical force field molecular dynamics algorithm to obtain a comparative dynamic trajectory. Based on the target dynamic trajectory and the comparison dynamic trajectory, a Pearson correlation coefficient comparison is performed to obtain a first comparison result; A second comparison result is obtained by performing root mean square error comparison processing on the target dynamic trajectory and the comparison dynamic trajectory; A third comparison result is obtained by comparing the acceleration ratios of the target dynamic trajectory and the comparison dynamic trajectory. The simulation evaluation results are obtained based on the first comparison result, the second comparison result, and the third comparison result.

[0113] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a world-model-based molecular dynamics simulation program, which, when executed by a processor, implements the steps of the world-model-based molecular dynamics simulation method as described above.

[0114] In summary, this invention provides a molecular dynamics simulation method, system, and terminal based on a world model. The method includes: acquiring crystal information and simulation conditions of a target molecule; constructing atomic systems based on the crystal information and simulation conditions to obtain atomic element information and multiple initial atomic systems; constructing world model training samples based on the atomic element information and multiple initial atomic systems to obtain multiple world model training samples, and processing the multiple world model training samples in batches to obtain training data and test data; configuring loss information of the training data to obtain sample loss information, and constructing a loss function based on the sample loss information to obtain a target loss function; determining an initial world model; iteratively training the initial world model based on the training data, the target loss function, and the simulation conditions to obtain a target world model; performing trajectory simulation processing based on the target world model and the test data to obtain a target dynamic trajectory, and evaluating the simulation results based on the target dynamic trajectory to obtain a simulation evaluation result. This invention constructs an initial atomic system, builds model training samples based on the initial atomic system, constructs a world model, trains the world model based on the model training samples, obtains a target world model, and performs trajectory simulation through the target world model to obtain the target dynamic trajectory, thereby improving the accuracy of molecular dynamics simulation.

[0115] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal. Unless otherwise specified, an element defined by the phrase "comprising one" does not exclude the presence of other identical elements in the process, method, article, or terminal that includes that element.

[0116] Of course, those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware (such as a processor, controller, etc.). The program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The computer-readable storage medium can be a memory, magnetic disk, optical disk, etc.

[0117] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A molecular dynamics simulation method based on a world model, characterized in that, The world-model-based molecular dynamics simulation method includes: The crystal information and simulation conditions of the target molecule are obtained, and the atomic system is constructed based on the crystal information and simulation conditions to obtain atomic element information and multiple initial atomic systems. Based on the atomic element information and multiple initial atomic systems, world model training samples are constructed to obtain multiple world model training samples. The multiple world model training samples are then processed in batches to obtain training data and test data. The loss information of the training data is configured to obtain sample loss information, and the loss function is constructed based on the sample loss information to obtain the target loss function. Determine an initial world model, and iteratively train the initial world model based on the training data, the target loss function, and the simulation conditions to obtain a target world model; The target world model is used to perform trajectory simulation processing based on the test data to obtain the target dynamic trajectory, and the simulation results are evaluated based on the target dynamic trajectory to obtain the simulation evaluation results.

2. The molecular dynamics simulation method based on a world model according to claim 1, characterized in that, The process of acquiring the crystal information and simulation conditions of the target molecule, and then constructing an atomic system based on the crystal information and simulation conditions to obtain atomic element information and multiple initial atomic systems, specifically includes: The initial crystal structure of the target molecule is obtained, and the initial crystal structure is normalized to obtain atomic element information; The unit cell of the target molecule is replicated to obtain a crystal supercell, and the crystal supercell is subjected to molecular system construction processing to obtain multiple molecular systems; By adding strain tensors to multiple molecular systems, multiple strain simulation systems are obtained. The simulation conditions are obtained, and the simulation conditions are configured for multiple strain simulation systems according to the simulation conditions to obtain multiple initial atomic systems.

3. The molecular dynamics simulation method based on a world model according to claim 1, characterized in that, The process of constructing world model training samples based on the atomic element information and multiple initial atomic systems to obtain multiple world model training samples, and then processing these multiple world model training samples in batches to obtain training data and test data, specifically includes: A density functional theory calculation algorithm is determined, and the first parameters of multiple initial atomic systems are calculated using the density functional theory calculation algorithm to obtain multiple first parameters; Multiple initial atomic systems were subjected to structural optimization and single-point energy calculations to obtain multiple second parameters. The research objective is obtained, and trajectory generation processing is performed based on the research objective, the atomic element information, multiple first parameters and multiple second parameters to obtain multiple first-principles molecular dynamics trajectories; Training samples were constructed and processed based on multiple first-principles molecular dynamics trajectories to obtain multiple world model training samples; Multiple training samples of the world model are processed in batches to obtain training data and test data.

4. The molecular dynamics simulation method based on a world model according to claim 1, characterized in that, The sample loss information includes average atomic coordinate loss, velocity prediction loss, average energy loss, force loss, stress loss, time continuity loss, and isotropic consistency loss. The process of configuring the loss information of the training data to obtain sample loss information, and constructing a loss function based on the sample loss information to obtain the target loss function, specifically includes: The training data is processed by defining the average atomic coordinate loss and the velocity prediction loss to obtain the average atomic coordinate loss and the velocity prediction loss. The training data is processed by defining average energy loss, force loss, and stress loss to obtain the average energy loss, the force loss, and the stress loss. The training data is processed using the temporal continuity loss definition to obtain the temporal continuity loss; Obtain the rotation matrix and translation vector, and perform equivariant consistency loss construction on the training samples based on the rotation matrix and translation vector to obtain the equivariant consistency loss. Obtain the weights of the loss terms, and perform function construction based on the weights of the loss terms, the average atomic coordinate loss, the velocity prediction loss, the average energy loss, the force loss, the stress loss, the time continuity loss, and the isovariance consistency loss to obtain the target loss function.

5. The molecular dynamics simulation method based on a world model according to claim 1, characterized in that, The initial world model includes a scientific foundation model, a molecular system coding module, a simulation condition coding module, and an isotropic dynamics expert model; The process of determining the initial world model, and iteratively training the initial world model based on the training data, the target loss function, and the simulation conditions to obtain the target world model, specifically includes: A scientific foundation model is determined, and the user instructions in the training data are processed by text conversion using the scientific foundation model to obtain instruction representation vectors. A simulation condition encoding module is determined, and the simulation condition and the instruction representation vector are processed by the simulation condition encoding module to obtain a unified condition vector. A molecular system encoding module is determined, and the atomic element information is processed by the molecular system encoding module to obtain the molecular system characterization; A variable dynamics expert model is determined, and the future trajectory of the molecule is obtained by performing future trajectory prediction processing based on the molecular system characterization and the unified condition vector through the variable dynamics expert model. The target world model is obtained by iteratively generating future trajectories based on the molecular future trajectory and the target loss function using the isovariant dynamics expert model.

6. The molecular dynamics simulation method based on a world model according to claim 5, characterized in that, The step of generating the target world model through iterative future trajectory generation using the isovariant dynamics expert model based on the molecular future trajectory and the target loss function specifically includes: Based on the future molecular trajectory and the molecular system characterization, the molecular system characterization is updated to obtain the updated molecular system characterization. The target world model is obtained by iteratively generating future trajectories using the isovariant dynamics expert model, based on the unified condition vector, the updated molecular system characterization, and the target loss function.

7. The molecular dynamics simulation method based on a world model according to claim 1, characterized in that, The process of simulating the target's dynamic trajectory using the target world model based on the test data to obtain the target's dynamic trajectory, and then evaluating the simulation results based on the target's dynamic trajectory to obtain the simulation evaluation results, specifically includes: The target dynamic trajectory is obtained by performing a first trajectory simulation on the test data using the target world model. A classical force field molecular dynamics algorithm is determined, and the test data is processed by the classical force field molecular dynamics algorithm to obtain a comparative dynamic trajectory. Based on the target dynamic trajectory and the comparison dynamic trajectory, a Pearson correlation coefficient comparison is performed to obtain a first comparison result; A second comparison result is obtained by performing root mean square error comparison processing on the target dynamic trajectory and the comparison dynamic trajectory; A third comparison result is obtained by comparing the acceleration ratios of the target dynamic trajectory and the comparison dynamic trajectory. The simulation evaluation results are obtained based on the first comparison result, the second comparison result, and the third comparison result.

8. A molecular dynamics simulation system based on a world model, characterized in that, The world-model-based molecular dynamics simulation system includes: An atomic system construction module is used to acquire the crystal information and simulation conditions of the target molecule, and to perform atomic system construction processing based on the crystal information and simulation conditions to obtain atomic element information and multiple initial atomic systems. The training sample construction module is used to construct world model training samples based on the atomic element information and multiple initial atomic systems, to obtain multiple world model training samples, and to process the multiple world model training samples in batches to obtain training data and test data. The loss function construction module is used to configure and process the loss information of the training data to obtain sample loss information, and to construct the loss function based on the sample loss information to obtain the target loss function. The world model training module is used to determine an initial world model, and to iteratively train the initial world model based on the training data, the target loss function, and the simulation conditions to obtain a target world model. The simulation evaluation module is used to perform trajectory simulation processing based on the test data using the target world model to obtain the target dynamic trajectory, and to evaluate the simulation results based on the target dynamic trajectory to obtain the simulation evaluation results.

9. A terminal, characterized in that, The terminal includes a memory, a processor, and a world-model-based molecular dynamics simulation program stored in the memory and executable on the processor. When the processor executes the world-model-based molecular dynamics simulation program, it implements the steps of the world-model-based molecular dynamics simulation method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a world-model-based molecular dynamics simulation program, which, when executed by a processor, implements the steps of the world-model-based molecular dynamics simulation method as described in any one of claims 1-7.