A machine learning potential energy model construction method based on hierarchical active learning

Abstract

This invention relates to a method for constructing a machine learning potential energy model based on hierarchical active learning, comprising the following steps: constructing a database of background solvent molecules and lithium salt molecules; initializing a machine learning potential energy model committee; outer active learning automatically exploring the chemical species composition space of the electrolyte, generating electrolyte formulations and initial structures; determining whether a structure needs to be included in the labeling range based on uncertainty calculations; automatically constructing a liquid phase environment based on classical force field simulation methods; performing high-precision simulations based on first-principles calculations and labeling the energy and force values ​​of the structures; inner active learning driving the machine learning potential energy model to explore the configuration space and label the energy and force values ​​of structures with high uncertainty; obtaining a first-principles structure database of the real liquid phase environment through several iterations; and training to obtain the final high-performance machine learning potential energy model. This invention systematically and comprehensively samples the chemical species composition space of the electrolyte and the molecular simulation configuration space based on hierarchical active learning, thereby ensuring the predictive performance and simulation accuracy of the machine learning potential model on unknown and complex electrolyte systems.

Description

Technical Field

[0001] This invention relates to the fields of software engineering and development technology, and in particular to a method for constructing a machine learning potential model based on hierarchical active learning. Background Technology

[0002] Molecular dynamics simulations play a vital role in chemical reactions, materials design, and drug development, serving as a bridge between experiments and microscopic mechanisms. In electrolyte design and development, they function as a pre-screening method or provide insights into the mechanisms of microscopic molecular interactions after experiments.

[0003] Traditional molecular dynamics simulation techniques include classical molecular dynamics simulations and ab initio molecular dynamics simulations. Classical molecular dynamics simulations calculate energy and atomic forces using predefined atom types and force field parameters, and then use Newton's second law to solve for atomic velocities and positions at the next moment. However, their accuracy is insufficient, especially for weak interactions in electrolytes. Ab initio molecular dynamics simulations calculate energy and atomic forces using first-principles calculations of electronic structure wavefunctions, offering high accuracy, but the massive computational cost limits the scale and efficiency of the simulations. Molecular dynamics simulations based on machine learning potential energy models can extend the accuracy of first-principles calculations to larger time and spatial scales by learning the energy and force information from ab initio labels.

[0004] The key to obtaining high-performance machine learning potential models lies in the richness and breadth of sampled structures. Manual sampling introduces bias and significantly increases computational resource consumption. In contrast, active learning uses a model committee to decide whether a new structure should be included in the labeling scope. This reduces the number of labeled samples, thus reducing computational resource consumption, and allows for targeted learning of chemical spaces with high model prediction uncertainty, thereby improving model performance. However, for the complex chemical species composition and disordered molecular arrangement caused by the complex liquid phase environment in the electrolyte field, active learning sampling that only focuses on the configuration space while ignoring the chemical species composition space is far from sufficient.

[0005] The current research on machine learning potential energy models in the field of electrolytes has two main shortcomings: (1) Existing active learning algorithms sample the configuration space of a fixed composition under a predefined formulation, while ignoring the larger composition space of various solvent molecules and lithium salt molecules in various proportions. The single sampling dimension limits the application domain of machine learning potential energy models and makes it difficult to accurately calculate the actual complex electrolyte system; (2) Existing sampling methods require manual configuration of simulation conditions, which makes it difficult to apply to the sampling process of large-scale electrolyte formulations; (3) Existing sampling methods are limited to cluster data of a small number of atoms and cannot reflect the real solvated liquid phase environment. Due to the scale effect, there are systematic errors when calculating the energy and force of periodic systems. Summary of the Invention

[0006] In view of the above-mentioned problems in the existing technology, the technical problem to be solved by the present invention is: how to systematically sample the chemical composition space and simulated configuration space of large-scale electrolyte formulations in an active learning process, expand the application domain of machine learning potential energy model, and realize accurate calculation of actual complex electrolyte systems.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0008] A method for constructing a machine learning potential model based on hierarchical active learning includes the following steps:

[0009] S1: Construct a machine learning potential energy model M, where M is a general large atom model LAM;

[0010] S2: Select lithium battery electrolyte data from publicly available literature to construct a molecular database, which includes several original molecular configurations;

[0011] Define a time step B, and perform simulation operations on each original molecular configuration in the molecular database in chronological order to obtain several simulated molecular frame configurations corresponding to each original molecular configuration. The simulation operation on the simulated molecular frame configuration of the current time step is used to obtain the simulated molecular frame configuration of the next time step. Each original molecular configuration and its corresponding several simulated molecular frame configurations are sorted in chronological order and together constitute the molecular simulation trajectory of each original molecular configuration.

[0012] Then, calculate the energy information E of each molecular simulation frame configuration in each molecular simulation trajectory and the force information F of all atoms on each molecular simulation frame configuration;

[0013] Finally, all E and F are converted to the correct format and used as the initial training dataset D. The uncertainty threshold W is then calculated using D.

[0014] S3: Initialize M T times, each initialization yielding a corresponding machine learning potential model. A total of T machine learning potential models form a model committee, denoted as [model committee name missing]. ,in, This represents the t-th machine learning potential model;

[0015] Construct the loss function L, the expression for L is as follows:

[0016]

[0017] in, This represents the coefficient factor of the energy information loss function, used to adjust the degree of influence of energy information on the model. The loss function representing energy information, This represents the coefficient factor of the atomic force information loss function, used to adjust the degree of influence of atomic force information on the model. The loss function representing the force information of atoms. Indicates the number of atoms in the training system;

[0018] Let D be respectively The input to each machine learning potential model is used, and an optimizer is used to train each machine learning potential model separately. Training stops when L converges, resulting in the final trained model committee. ;

[0019] S4: Based on a hierarchical active learning framework, define the outer electrolyte chemical composition space and the inner electrolyte simulation configuration space; the outer electrolyte chemical composition space is used to obtain several electrolyte formulations; the inner electrolyte simulation configuration space is used to calculate the energy information and atomic force information of the molecular simulation frame configuration;

[0020] S5: Input D into the outer electrolyte chemical composition space, randomly combine all molecules in D, and obtain an electrolyte formula each time. The electrolyte formula contains solvent molecules and lithium salt molecules. After several iterations, N' electrolyte formulas are obtained.

[0021] S6: Let i=1, select the i-th electrolyte formula X from N' electrolyte formulas to construct the initial structure of the new electrolyte X' corresponding to X;

[0022] S7: Utilize Calculate the uncertainty value w of X'. If w is greater than W, then define X' as a high uncertainty structure. Then construct the automated liquid phase environment corresponding to X', obtain the pre-equilibrium structure of X', and proceed to the next step. If w is less than or equal to W, then ignore X, return to S6, and let i = i + 1.

[0023] S8: Using the pre-equilibrium structure of X' as the initial configuration for molecular simulation, the molecular simulation trajectory of X' is obtained by performing simulation operations on the initial configuration of molecular simulation using the method described in S2. Then, the energy information and atomic force information of each molecular simulation frame configuration in the molecular simulation trajectory of X' are calculated.

[0024] S9: Input the configuration of the last molecular simulation frame in the molecular simulation trajectory of X'. In the process, new molecular simulation trajectories are obtained, and then the energy information and atomic force information of each molecular simulation frame configuration in the new molecular simulation trajectory are calculated.

[0025] The energy information and atomic force information of each molecular simulation frame configuration in the molecular simulation trajectory of X', and the energy information and atomic force information of all molecular simulation frame configurations in the new molecular simulation trajectory are converted into formats and merged into D to obtain the updated dataset D'; at this time, if i≥N', proceed to the next step, otherwise return to S6, and let i=i+1;

[0026] S10: Divide D' into the current training set and the current test set according to the proportion. Use the current training set as the input to M', and then train M' using L. Update the parameters of M' through gradient descent and backpropagation. Stop training when the maximum number of iterations is reached or the loss function no longer changes, and obtain the current model committee. ;

[0027] Input the current test set in batches Calculate the prediction accuracy metric for each model. ;like If convergence occurs, then the one with the lowest... The model definition of the indicator is The optimal machine learning potential model Save the high-uncertainty structures corresponding to the current N' electrolyte formulations and proceed to the next step; otherwise, return to S5 and regenerate the N' electrolyte formulations. The formula for calculating the indicator is as follows:

[0028]

[0029] in, Let D be the number of all structures in dataset D. For the predicted energy information of the i-th structure, For the true energy information of the i-th structure, Let be the number of atoms in the i-th structure. For the predicted force information of the j-th atom in the i-th structure, This represents the actual force information of the j-th atom in the i-th structure;

[0030] S11: Calculate the proportion v of the initial structure of the new electrolyte with high uncertainty structure corresponding to the current N' electrolyte formulations in all new electrolyte initial structures;

[0031] S12: Preset percentage threshold V. If v is less than V, then... The final trained machine learning potential model; otherwise, return to S5 and regenerate N' electrolyte formulations.

[0032] Preferably, the step in S2 to calculate the uncertainty threshold W using D is as follows:

[0033] First, the computational model committee calculated the variance of the predictions for any original molecular configuration. The calculation formula is as follows:

[0034]

[0035] in, This represents the energy prediction value of the i-th machine learning potential energy model for the original molecular configuration. This represents the value calculated using density functional theory. The number of models on the model committee;

[0036] Secondly, the prediction uncertainty of calculating the original molecular configuration. The calculation formula is as follows:

[0037]

[0038] in, Indicates the number of atoms in the original molecular configuration;

[0039] Next, calculate the difference between the predicted energy and the reference energy for each model corresponding to the original molecular configuration, and select the largest difference as the energy difference corresponding to the original molecular configuration. The calculation formula is as follows:

[0040]

[0041] in, This indicates that the maximum value is selected. This represents the maximum difference between the predicted energy of the i-th model and the reference energy calculated by density functional theory.

[0042] Finally, select Values ​​greater than 1 kcal / mol are obtained The model committee will select The uncertainty threshold W is defined as the uncertainty value corresponding to at least 90% of the original molecular configuration.

[0043] Preferably, the step of obtaining the pre-equilibrium structure of X' in S5 is as follows:

[0044] X' is input into Gromacs software for structural relaxation. Then, X' after structural relaxation is subjected to liquid phase structural pre-equilibrium operation in NPT ensemble to finally obtain the pre-equilibrium structure of X'.

[0045] Compared with the prior art, the present invention has at least the following advantages:

[0046] This technical solution improves the coverage of training samples through systematic sampling in two spaces, enhances the efficiency of large-scale sampling and labeling through automated simulation workflow, and reduces the impact of scale effect on prediction accuracy through automatic adjustment of electrolyte structure, thereby expanding the application domain of machine learning potential energy model and realizing accurate calculation and simulation of complex electrolyte systems.

[0047] (1) The designed hierarchical active learning method extends the single-dimensional active learning algorithm to a hierarchical active learning algorithm, thereby achieving a breakthrough in sampling dimensions. At the same time, it systematically samples the chemical composition space and the simulated configuration space of the electrolyte, avoiding the limitation of sample diversity on the predefined electrolyte formulation composition. This greatly improves the generalization performance of the machine learning potential model in real complex electrolyte systems and its potential for application in the actual electrolyte development and design process, thus expanding the application domain of the machine learning potential model.

[0048] (2) The designed automated simulation workflow can automatically generate molecular structures based on a series of electrolyte formulations, automatically combine the molecular structures into electrolyte configurations, and achieve parallel simulation of classical molecular dynamics to obtain the pre-equilibrium structure of each electrolyte formulation, which accelerates the operation efficiency of hierarchical active learning and makes it possible to generate and predict tens of thousands of novel electrolyte formulations.

[0049] (3) The designed automated simulation workflow achieves electrolyte dynamics simulation that takes into account everything from low temperature to high temperature, from low concentration to high concentration, and from small space size to large space size by automatically adjusting temperature, concentration and space size. Compared with the isolated molecular cluster structure generated by existing methods, this method can generate real periodic structures on a large scale, minimizing the impact of scale effect on structural accuracy. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of a hierarchical active learning algorithm.

[0051] Figure 2 This is a flowchart illustrating the automated simulation process for a solvated liquid phase environment.

[0052] Figure 3Scatter plot showing the correlation between the predicted performance of the optimal lithium battery electrolyte using a machine learning potential energy model. Detailed Implementation

[0053] The present invention will now be described in further detail.

[0054] See Figures 1-3 A method for constructing a machine learning potential model based on hierarchical active learning includes the following steps:

[0055] S1: Construct a machine learning potential energy model M, where M is a general large atomic model (LAM). The general large atomic model LAM is existing technology; this invention uses the DPA-2 model. When using it, the hyperparameters of the machine learning potential energy model structure are first set. This model includes a three-layer embedding network that generates descriptors (i.e., feature vectors of the structure), containing 20, 40, and 80 neurons respectively. The fitting network for predicting atomic energy and force consists of four layers, each with 240 neurons. The cutoff radius of the embedding process is 6 angstroms. The loss function during training uses a weighted error value of energy and force errors. The model structure originates from: DPA-2: a large atomic model as a multi-task learner.

[0056] S2: A molecular database is constructed by selecting lithium battery electrolyte data from publicly available literature. The molecular database includes several original molecular configurations. Specifically, the molecular database used in this invention contains solvent molecules and lithium salt molecules.

[0057] Define a time step B, and perform simulation operations on each original molecular configuration in the molecular database in chronological order to obtain several simulated molecular frame configurations corresponding to each original molecular configuration. The simulation operation on the simulated molecular frame configuration of the current time step is used to obtain the simulated molecular frame configuration of the next time step. Each original molecular configuration and its corresponding several simulated molecular frame configurations are sorted in chronological order and together constitute the molecular simulation trajectory of each original molecular configuration.

[0058] Then, calculate the energy information E of each molecular simulation frame configuration in each molecular simulation trajectory and the force information F of all atoms on each molecular simulation frame configuration;

[0059] Finally, all E and F values ​​were converted to a new format and used as the initial training dataset D. The uncertainty threshold W was calculated using D. The molecular simulation trajectories were obtained by performing first-principles molecular dynamics simulations of each molecule in the database using the CP2K software. The energy information and the force information of the atoms were further obtained by calculating the simulation trajectory of each molecule using density functional theory. Both the CP2K software and density functional theory are existing technologies. The dpdata toolkit was used to convert the information into training data, which is also an existing technology.

[0060] The steps in S2 to calculate the uncertainty threshold W using D are as follows:

[0061] First, the computational model committee calculated the variance of the predictions for any original molecular configuration. The calculation formula is as follows:

[0062]

[0063] in, This represents the energy prediction value of the i-th machine learning potential energy model for the original molecular configuration. This represents the value calculated using density functional theory. The number of models on the model committee;

[0064] Secondly, the prediction uncertainty of calculating the original molecular configuration. , Standard deviation versus number of atoms and the number of committee models The normalized value is calculated using the following formula:

[0065]

[0066] in, Indicates the number of atoms in the original molecular configuration;

[0067] Next, calculate the difference between the predicted energy and the reference energy for each model corresponding to the original molecular configuration, and select the largest difference as the energy difference corresponding to the original molecular configuration. The calculation formula is as follows:

[0068]

[0069] in, This indicates that the maximum value is selected. This represents the maximum difference between the predicted energy of the i-th model and the reference energy calculated by density functional theory.

[0070] Finally, select Values ​​greater than 1 kcal / mol are obtained The model committee will select The uncertainty threshold W is defined as the value corresponding to at least 90% of the original molecular configuration; the defined threshold needs to enable the model committee to capture 90% of the energy error. Structures with an energy error greater than 1 kcal / mol, i.e., 90% energy error. Uncertainty value of structures greater than 1 kcal / mol It needs to be above the threshold;

[0071] S3: Initialize M T times, each initialization yielding a corresponding machine learning potential model. A total of T machine learning potential models form a model committee, denoted as [model committee name missing]. ,in, This represents the t-th machine learning potential model;

[0072] Construct the loss function L, the expression for L is as follows:

[0073]

[0074] in, This represents the coefficient factor of the energy information loss function, used to adjust the degree of influence of energy information on the model. The loss function representing energy information, This represents the coefficient factor of the atomic force information loss function, used to adjust the degree of influence of atomic force information on the model. The loss function representing the force information of atoms. Indicates the number of atoms in the training system;

[0075] Let D be respectively The input to each machine learning potential model is used, and an optimizer is used to train each machine learning potential model separately. Training stops when L converges, resulting in the final trained model committee. The optimizer uses Adam, which is existing technology.

[0076] S4: Based on a hierarchical active learning framework, define the outer electrolyte chemical composition space and the inner electrolyte simulation configuration space; the outer electrolyte chemical composition space is used to obtain several electrolyte formulations; the inner electrolyte simulation configuration space is used to calculate the energy information and atomic force information of the molecular simulation frame configuration;

[0077] The hierarchical active learning method used in the designed hierarchical active learning framework is as follows: the active learning of the outer electrolyte chemical composition space is responsible for generating novel electrolyte formulations in batches and judging the predictive ability of these formulations by the machine learning potential model. Electrolyte formulations with high uncertainty are included in the sample pool that needs to be labeled. The active learning of the inner electrolyte simulation configuration space is responsible for generating electrolyte simulation configurations in batches under a certain formulation. Electrolyte simulation configurations with high uncertainty are included in the sample pool that needs to be labeled.

[0078] S5: Input D into the outer electrolyte chemical composition space, and randomly combine all molecules in D. Generally, random combination is performed according to the commonly used or obtained electrolyte concentration. Each random combination yields an electrolyte formulation, and the electrolyte formulation contains solvent molecules and lithium salt molecules. After several iterative combinations, N' electrolyte formulations are obtained. Each electrolyte formulation contains solvent molecules and lithium salt molecules, but the types of these two molecules are different, and the proportions of these two molecules in the electrolyte formulation are also different. These two molecules include many types, for example: lithium salt molecules include lithium hexafluorophosphate LiPF6, lithium bis(oxalato)borate, Li[B(C2O4)2], lithium bis(trifluoromethanesulfonyl)imide Li[N(SO2CF3)2], etc.; solvent molecules include ethylene carbonate C3H4O3, dimethyl carbonate C3H6O3, ethylene glycol dimethyl ether C4H 10 O2, etc.;

[0079] The steps for obtaining the pre-equilibrium structure of X' in S5 are as follows:

[0080] X' is input into Gromacs software for structural relaxation. This operation is based on the molecular structure optimization and charge distribution of the quantum chemistry Gaussian software and aims to eliminate unreasonable atomic contacts. Then, X' after structural relaxation is subjected to liquid phase structure pre-equilibrium operation under the NPT ensemble to finally obtain the pre-equilibrium structure of X'. The liquid phase structure pre-equilibrium operation is based on the atomic type identification and force field parameter matching of the GAFF classical force field under the NPT ensemble, with the aim of making the density of the electrolyte structure close to the actual density.

[0081] S6: Let i=1, select the i-th electrolyte formulation X from N' electrolyte formulations to construct the initial structure of the new electrolyte X' corresponding to X; the initial structure construction uses the Packmol tool, which is an existing technology.

[0082] S7: Utilize Calculate the uncertainty value w of X'. If w is greater than W, then define X' as a high uncertainty structure. Then construct the automated liquid phase environment corresponding to X', obtain the pre-equilibrium structure of X', and proceed to the next step. If w is less than or equal to W, then ignore X, return to S6, and let i = i + 1.

[0083] S8: Using the pre-equilibrium structure of X' as the initial configuration for molecular simulation, the molecular simulation trajectory of X' is obtained by performing simulation operations on the initial configuration of molecular simulation using the method described in S2. Then, the energy information and atomic force information of each molecular simulation frame configuration in the molecular simulation trajectory of X' are calculated.

[0084] S9: Input the configuration of the last molecular simulation frame in the molecular simulation trajectory of X'. In the process, new molecular simulation trajectories are obtained, and then the energy information and atomic force information of each molecular simulation frame configuration in the new molecular simulation trajectory are calculated. Starting from the last frame structure of the first-principles molecular dynamics simulation, the model in the model committee is called to perform a longer machine learning potential energy simulation in the inner electrolyte simulation configuration space to explore the inner electrolyte simulation configuration space. The machine learning model committee evaluates the uncertainty of all the generated new molecular simulation frame configurations.

[0085] The energy information and atomic force information of each molecular simulation frame configuration in the molecular simulation trajectory of X', and the energy information and atomic force information of all molecular simulation frame configurations in the new molecular simulation trajectory are converted into formats and merged into D to obtain the updated dataset D'; at this time, if i≥N', proceed to the next step, otherwise return to S6, and let i=i+1;

[0086] S10: Divide D' into the current training set and the current test set proportionally, typically in an 8:2 ratio. Use the current training set as input to M', then train M' using L, updating the parameters of M' through gradient descent and backpropagation. Stop training when the maximum number of iterations is reached or the loss function no longer changes, thus obtaining the current model committee. ;

[0087] Input the current test set in batches Calculate the prediction accuracy metric for each model. ;like If convergence occurs, then the one with the lowest... The model definition of the indicator is The optimal machine learning potential model The system saves the high-uncertainty structures corresponding to the current N' electrolyte formulations and proceeds to the next step; otherwise, it returns to S5 and regenerates N' electrolyte formulations. Different models differ only in their initialization parameters and will always converge synchronously or not converge. The value difference will not be significant. The formula for calculating the indicator is as follows:

[0088]

[0089] in, Let D be the number of all structures in dataset D. For the predicted energy information of the i-th structure, For the true energy information of the i-th structure, Let be the number of atoms in the i-th structure. For the predicted force information of the j-th atom in the i-th structure, This represents the actual force information of the j-th atom in the i-th structure;

[0090] S11: Calculate the proportion v of the initial structure of the new electrolyte with high uncertainty structure corresponding to the current N' electrolyte formulations in all new electrolyte initial structures;

[0091] S12: Preset percentage threshold V. If v is less than V, then... The final trained machine learning potential model; otherwise, return to S5 and regenerate N' electrolyte formulations; after testing, The proportion of high uncertainty structures always decreases in tandem, and after 6 iterations, both can simultaneously satisfy the convergence condition.

[0092] Experimental content and results

[0093] Based on the above steps, the first-principles label database obtained by the hierarchical active learning algorithm of this invention contains 375,781 novel electrolyte structures, covering 11 electrolyte-related elements including hydrogen, lithium, boron, carbon, nitrogen, oxygen, fluorine, silicon, phosphorus, sulfur, and chlorine. The number of atoms with diverse chemical local environments exceeds 45 million. The performance of the optimal lithium-ion battery electrolyte machine learning potential energy model obtained through training is as follows: Figure 3As shown, the mean absolute error of energy for the training set is 0.2785 eV, and the mean absolute error of force is 0.0584 eV / Å. For the test set, the mean absolute error of energy is 0.1874 eV, and the mean absolute error of force is 0.0532 eV / Å. Furthermore, to further verify the effectiveness of this invention, it was compared with existing electrolyte machine learning potential energy models. The model trained by this invention demonstrates the strongest application capability in real, complex electrolyte systems. This experiment tested and compared the prediction accuracy of energy and force using the method of this invention with several other newly developed complex electrolyte structures. It can be seen that the model of this method has the highest prediction accuracy for energy and force, and also achieves the most stable large-scale molecular dynamics simulation (see Table 1).

[0094] Table 1. Simulation Comparison Table of the Invention Model and Existing Models

[0095]

[0096] In contrast, existing models published in the literature cannot complete the simulation tasks of these electrolyte systems. The results show that the hierarchical active learning-based machine learning potential energy model development method for lithium battery electrolytes proposed in this invention demonstrates excellent capabilities in sampling electrolyte formulations and simulation configurations with complex chemical compositions. This enables the machine learning potential energy model to achieve optimal generalization performance, which will contribute to the development of multifunctional electrolytes for various needs.

[0097] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for constructing a machine learning potential model based on hierarchical active learning, characterized in that: Includes the following steps: S1: Construct a machine learning potential energy model M, where M is a general large atom model LAM; S2: Select lithium battery electrolyte data from publicly available literature to construct a molecular database, which includes several original molecular configurations; Define a time step B, and perform simulation operations on each original molecular configuration in the molecular database in chronological order to obtain several simulated molecular frame configurations corresponding to each original molecular configuration. The simulation operation on the simulated molecular frame configuration of the current time step is used to obtain the simulated molecular frame configuration of the next time step. Each original molecular configuration and its corresponding several simulated molecular frame configurations are sorted in chronological order and together constitute the molecular simulation trajectory of each original molecular configuration. Then, calculate the energy information E of each molecular simulation frame configuration in each molecular simulation trajectory and the force information F of all atoms on each molecular simulation frame configuration; Finally, all E and F are converted to the correct format and used as the initial training dataset D. The uncertainty threshold W is then calculated using D. S3: Initialize M T times, each initialization yielding a corresponding machine learning potential model. A total of T machine learning potential models form a model committee, denoted as [model committee name missing]. ,in, This represents the t-th machine learning potential model; Construct the loss function L, the expression for L is as follows: in, This represents the coefficient factor of the energy information loss function, used to adjust the degree of influence of energy information on the model. The loss function representing energy information, This represents the coefficient factor of the atomic force information loss function, used to adjust the degree of influence of atomic force information on the model. The loss function representing the force information of atoms. Indicates the number of atoms in the training system; Let D be respectively The input to each machine learning potential model is used, and an optimizer is used to train each machine learning potential model separately. Training stops when L converges, resulting in the final trained model committee. ; S4: Based on a hierarchical active learning framework, define the outer electrolyte chemical composition space and the inner electrolyte simulation configuration space; the outer electrolyte chemical composition space is used to obtain several electrolyte formulations; the inner electrolyte simulation configuration space is used to calculate the energy information and atomic force information of the molecular simulation frame configuration; S5: Input D into the outer electrolyte chemical composition space, randomly combine all molecules in D, and obtain an electrolyte formula each time. The electrolyte formula contains solvent molecules and lithium salt molecules. After several iterations, N' electrolyte formulas are obtained. S6: Let i=1, select the i-th electrolyte formula X from N' electrolyte formulas to construct the initial structure of the new electrolyte X' corresponding to X; S7: Utilize Calculate the uncertainty value w of X'. If w is greater than W, then define X' as a high uncertainty structure. Then construct the automated liquid phase environment corresponding to X', obtain the pre-equilibrium structure of X', and proceed to the next step. If w is less than or equal to W, then ignore X, return to S6, and let i = i + 1. S8: Using the pre-equilibrium structure of X' as the initial configuration for molecular simulation, the molecular simulation trajectory of X' is obtained by performing simulation operations on the initial configuration of molecular simulation using the method described in S2. Then, the energy information and atomic force information of each molecular simulation frame configuration in the molecular simulation trajectory of X' are calculated. S9: Input the configuration of the last molecular simulation frame in the molecular simulation trajectory of X'. In the process, new molecular simulation trajectories are obtained, and then the energy information and atomic force information of each molecular simulation frame configuration in the new molecular simulation trajectory are calculated. The energy information and atomic force information of each molecular simulation frame configuration in the molecular simulation trajectory of X', and the energy information and atomic force information of all molecular simulation frame configurations in the new molecular simulation trajectory are converted into formats and merged into D to obtain the updated dataset D'; at this time, if i≥N', proceed to the next step, otherwise return to S6, and let i=i+1; S10: Divide D' into the current training set and the current test set according to the proportion. Use the current training set as the input to M', and then train M' using L. Update the parameters of M' through gradient descent and backpropagation. Stop training when the maximum number of iterations is reached or the loss function no longer changes, and obtain the current model committee. ; Input the current test set in batches Calculate the prediction accuracy metric for each model. ;like If convergence occurs, then the one with the lowest... The model definition of the indicator is The optimal machine learning potential model Save the high-uncertainty structures corresponding to the current N' electrolyte formulations and proceed to the next step; otherwise, return to S5 and regenerate the N' electrolyte formulations. The formula for calculating the indicator is as follows: in, Let D be the number of all structures in dataset D. For the predicted energy information of the i-th structure, For the true energy information of the i-th structure, Let be the number of atoms in the i-th structure. For the predicted force information of the j-th atom in the i-th structure, This represents the actual force information of the j-th atom in the i-th structure; S11: Calculate the proportion v of the initial structure of the new electrolyte with high uncertainty structure corresponding to the current N' electrolyte formulations in all new electrolyte initial structures; S12: Preset percentage threshold V. If v is less than V, then... The final trained machine learning potential model; otherwise, return to S5 and regenerate N' electrolyte formulations.

2. The method for constructing a machine learning potential model based on hierarchical active learning as described in claim 1, characterized in that: The steps in S2 to calculate the uncertainty threshold W using D are as follows: First, the computational model committee calculated the variance of the predictions for any original molecular configuration. The calculation formula is as follows: in, This represents the energy prediction value of the i-th machine learning potential energy model for the original molecular configuration. This represents the value calculated using density functional theory. The number of models on the model committee; Secondly, the prediction uncertainty of calculating the original molecular configuration. The calculation formula is as follows: in, Indicates the number of atoms in the original molecular configuration; Next, calculate the difference between the predicted energy and the reference energy for each model corresponding to the original molecular configuration, and select the largest difference as the energy difference corresponding to the original molecular configuration. The calculation formula is as follows: in, This indicates that the maximum value is selected. This represents the maximum difference between the predicted energy of the i-th model and the reference energy calculated by density functional theory. Finally, select Values ​​greater than 1 kcal / mol are obtained The model committee will select The uncertainty threshold W is defined as the uncertainty value corresponding to at least 90% of the original molecular configuration.

3. The method for constructing a machine learning potential model based on hierarchical active learning as described in claim 2, characterized in that: The steps for obtaining the pre-equilibrium structure of X' in S5 are as follows: X' is input into Gromacs software for structural relaxation. Then, X' after structural relaxation is subjected to liquid phase structural pre-equilibrium operation in NPT ensemble to finally obtain the pre-equilibrium structure of X'.

Images