A method for constructing a machine learning potential function model for solid-solid interface systems

Abstract

The present invention relates to the fields of materials and chemical computing, and in particular to a method for constructing a machine learning potential function model of a solid-solid interface system. The method comprises constructing a solid-solid interface system; calculating the solid-solid interface system based on first principles, recording the output data of each configuration in the solid-solid interface system at each step of its simulation process, and forming a high-precision data set; fine-tuning a pre-trained MACE model using the high-precision data set, and optimizing it using a physical constraint loss function to obtain a preliminary machine learning potential function model; performing hyperparameter tuning on the preliminary machine learning potential function model to obtain a fine-tuned machine learning potential function model; and the present invention adopts a machine learning potential function model based on a high-order equivariant graph neural network architecture to effectively learn the characteristics of the solid-solid interface system.

Description

Technical Field

[0001] The present invention relates to the fields of materials and chemical computing, and in particular to a method for constructing a machine learning potential function model of a solid-solid interface system. Background Art

[0002] The Machine Learning Interatomic Potential (MLIP) is an interatomic interaction model built using machine learning algorithms. It is used to describe the forces and energies between atoms in materials. Compared to traditional empirical potentials (such as the Morse potential and the Lennard-Jones potential), MLIP can capture more complex interatomic interactions by learning from quantum mechanical calculation data (such as DFT results). This model offers significant advantages in accuracy and generalization, and is widely used in molecular dynamics simulations, materials design, and reaction path analysis.

[0003] Solid-solid interfaces (hereafter referred to as SSIs) are ubiquitous in various materials, and their structure and properties significantly influence the overall performance of the material. Only with a deep understanding of interfaces can we more effectively construct interfaces with desired properties. The application of existing MLIPs in SSI systems faces multiple challenges, primarily due to the following reasons:

[0004] 1) Insufficient training data diversity: Solid-solid interfaces often involve the bonding of dissimilar materials, such as metals and ceramics, or semiconductors and insulators. These interfaces can exhibit complex features such as lattice mismatch, defects, and amorphous structures. Because these characteristics are often not adequately represented in existing training data, MLIPs suffer from insufficient accuracy in predicting the properties of these interfaces.

[0005] 2) Limitations in the treatment of long-range interactions: Many MLIPs focus primarily on short-range interactions, neglecting long-range interactions such as electrostatics and van der Waals forces. However, at solid-solid interfaces, especially those involving polar materials or charged defects, long-range interactions can significantly influence interfacial properties. The lack of accurate modeling of these interactions limits the applicability of MLIPs in such systems.

[0006] 3) Limited transferability and generalization: MLIPs are typically trained on specific material systems and structures, so their predictive power can decline when faced with unseen interface structures or novel material combinations. This lack of generalization is particularly prominent in solid-solid interface simulations.

[0007] 4) Complexity of interface structure: Solid-solid interfaces may contain complex structural features such as amorphous regions, lattice mismatch, and stress concentration. These features are difficult to fully capture in traditional MLIPs training, resulting in poor model performance when dealing with real interfaces. Summary of the Invention

[0008] To solve the above problems, the present invention provides a method for constructing a machine learning potential function model of a solid-solid interface system. By introducing a high-order equivariant graph neural network architecture, it effectively describes the characteristics of the solid-solid interface system and completes fine-tuning on a smaller data set and training rounds, reducing the training error by one order of magnitude compared to mainstream architectures.

[0009] The specific plan includes the following steps:

[0010] S1. Construct a solid-solid interface system, including the bulk and surface configurations of the two solid materials, as well as the solid-solid interface configuration of the two materials;

[0011] S2. Simulate the solid-solid interface system based on first principles, recording the output data of each configuration in the solid-solid interface system at each step of the simulation process to form a high-precision data set;

[0012] S3. Fine-tune the pre-trained MACE model using a high-precision dataset and optimize it using a physical constraint loss function to obtain a preliminary machine learning potential function model.

[0013] S4. performing hyperparameter tuning on the preliminary machine learning potential function model to obtain a fine-tuned machine learning potential function model;

[0014] S5. Dynamic simulation using fine-tuned machine learning potential function model.

[0015] Furthermore, step S2 uses ab initio molecular dynamics simulation of the solid-solid interface system based on first principles to form a high-precision data set, including:

[0016] S21. Setting simulation parameters, including functional selection, molecular dynamics parameter setting, and computing resource configuration;

[0017] S22. simulating the solid-solid interface system based on the configuration of step S21, comprising:

[0018] The bulk and surface configurations of the two materials are simulated separately. Each step of the simulation process outputs a single-material sample. A single-material sample includes the total potential energy, atomic forces, and atomic coordinate data of a bulk or surface configuration.

[0019] Perform simulation on the solid-solid interface configuration of the two materials. Each step of the simulation process outputs a frame of heterogeneous interface sample, which includes the energy, atomic force and structural coordinate data of the solid-solid interface configuration;

[0020] S23. Single material samples and heterogeneous interface samples were screened and processed to obtain a high-precision dataset, of which heterogeneous interface samples accounted for 80%.

[0021] Furthermore, step S21 includes:

[0022] S211. Select PBE-PAW density functional;

[0023] S212. Set the molecular dynamics parameters: set the time step to 1 fs, the sampling step to 10,000 steps, select the NVT ensemble, and maintain the temperature at each target temperature using an Andersen thermal bath.

[0024] S213. Configure computing resources: Set the plane wave cutoff energy to 500 eV and use Γ-point sampling.

[0025] Furthermore, the pre-trained MACE model in step S3 adopts MACE-MP-0-medium, and the physical constraint loss function is composed of a weighted combination of an energy loss function and an atomic force loss function; the fine-tuning process of the pre-trained MACE model includes:

[0026] S31. Split the high-precision dataset into a training set and a validation set in a ratio of 9:1;

[0027] S32. The MACE-MP-0-medium is trained once using the training set. The prediction accuracy of the trained MACE-MP-0-medium in terms of energy and atomic forces is evaluated using the validation set to determine whether the convergence condition is met. If so, a preliminary machine learning potential function model is obtained. Otherwise, step S32 is executed again.

[0028] Furthermore, the fine-tuning of the pre-trained MACE model was run on a single GPU with 64-bit precision. When training MACE-MP-0-medium using the training set, the optimizer used was AdamW+AMSGrad, and the initial learning rate was set to 5 × 10⁻ 4 , batch size is 8, weight decay is 1 × 10⁻ 6 , combining cosine annealing with random weight averaging to achieve robust convergence within 150 epochs.

[0029] Furthermore, step S4 performs hyperparameter tuning on the preliminary machine learning potential function model, specifically including:

[0030] In terms of learning strategy, the initial learning rate is set and warm-up is introduced to adjust the initial learning rate; the random weight average learning rate and the start timing of random weight average are reset;

[0031] Adjust the physical constraint loss function and use the loss function L that combines energy, atomic force and stress total , redistribute the weights, expressed as L total =w E ×L E +w F ×L F +w S ×L S

[0032] Among them, L E represents the energy loss function, L F Represents the atomic force loss function, L S represents the stress loss function, w E represents the energy loss weight, w F Indicates the atomic force loss weight, w S represents the stress loss weight;

[0033] Increase batch_size and strengthen regularization of weight_decay and ema_decay;

[0034] Freeze the underlying parameters of the preliminary machine learning potential function model for training.

[0035] Beneficial effects of the present invention:

[0036] The present invention adopts a machine learning potential function model based on a high-order equivariant graph neural network architecture to effectively learn the characteristics of the solid-solid interface system, showing higher accuracy and faster convergence speed in complex systems such as solid-solid interfaces.

[0037] The method of the present invention has low requirements on the number of data sets for first-principles construction, and only requires a small number of data sets of a specific system to fine-tune the model. Specifically, compared with the mainstream high-dimensional neural network potential function model architecture, only 1000-2000 frames of data are required for fine-tuning. It can achieve high precision in fewer training rounds, with fast convergence speed. The training error is reduced by one order of magnitude compared with the mainstream machine learning potential function architecture, and has wide applicability to solid-solid interface systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 Flow chart of the method of the present invention;

[0039] Figure 2 Flowchart of the kinetic simulation for the present invention;

[0040] Figure 3 Schematic diagram of the data set structure of the embodiment of the present invention, (a) Na bulk phase and Na(110) surface; (b) Na3SbS4 bulk phase and Na3SbS4(110) surface; (c) Na(110)|Na3SbS4(110) heterogeneous interface;

[0041] Figure 4 A diagram showing the relationship between the energy and atomic force loss functions and the number of training rounds according to an embodiment of the present invention;

[0042] Figure 5 This is a graph showing the relationship between the RMSE of energy and atomic force and the number of training rounds;

[0043] Figure 6 The relationship between the energy and atomic force loss functions and the training rounds of the present invention is shown in FIG.

[0044] Figure 7 Schematic diagram of the simulation operation of the Na(110)|Na3SbS4(110) heterointerface in an embodiment of the present invention. DETAILED DESCRIPTION

[0045] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0046] The present invention provides a method for constructing a machine learning potential function model of a solid-solid interface system, such as Figure 1 As shown, the following steps are included:

[0047] S1. Construct a solid-solid interface system, including the bulk and surface configurations of the two solid materials, as well as the solid-solid interface configuration of the two materials.

[0048] Specifically, molecular modeling is used to construct the solid-solid interface system to be studied, which should include the bulk configuration of material A and the bulk configuration of material B. Furthermore, the stable surface structure (i.e., surface configuration) of material A and material B, as well as the solid-solid interface configuration of materials A and B, are obtained. Materials A and B are two different materials, which can be metal and ceramic, semiconductor and insulator, or other different types of materials.

[0049] S2. Simulate the solid-solid interface system based on first principles, and record the output data of each configuration in the solid-solid interface system at each step of the simulation process to form a high-precision dataset.

[0050] Specifically, first principles is a method for studying the properties of matter by directly solving quantum mechanical equations, starting from the most fundamental laws of physics without relying on empirical parameters or simplifying assumptions. It can be used to investigate the electronic structure and related physical and chemical properties of materials and molecular systems. Ab initio molecular dynamics (AIMD) is an extension of first principles methods to the study of dynamic processes. It combines the high precision of quantum mechanics with the time-evolution capabilities of molecular dynamics, providing an "atomic-scale camera" for studying the dynamic behavior of materials.

[0051] The present invention is based on first principles and uses AIMD to simulate the solid-solid interface system, including the following process:

[0052] S21. Set simulation parameters, including the selection of functionals, the setting of molecular dynamics parameters, and the configuration of computing resources.

[0053] S22. simulating the solid-solid interface system based on the configuration of step S21, comprising:

[0054] The bulk and surface configurations of the two materials are simulated separately. Each step of the simulation process outputs a single-material sample. A single-material sample includes the total potential energy, atomic forces, and atomic coordinate data of a surface configuration or a bulk configuration.

[0055] Perform simulation on the solid-solid interface configuration of the two materials. Each step of the simulation process outputs a frame of heterogeneous interface sample, which includes the energy, atomic force and structural coordinate data of the solid-solid interface configuration;

[0056] S23. Single material samples and heterogeneous interface samples were screened and processed to obtain a high-precision dataset, of which heterogeneous interface samples accounted for 80%.

[0057] S3. Fine-tune the pre-trained MACE model using a high-precision dataset and optimize it using a physical constraint loss function to obtain a preliminary machine learning potential function model.

[0058] Specifically, with the development of large-scale machine learning models (such as neural networks, Transformer, etc.), pre-trained large models have been gradually introduced into atomistic simulations. Pre-trained large models use public material databases (such as Materials Project, Open Catalyst Project, QM9, etc.) for large-scale training to obtain a potential energy function representation with versatility and high expressiveness. However, in order to apply the pre-trained large model to a specific system (such as a specific reaction interface, catalyst surface, etc.), fine-tuning training is still required to further improve the prediction accuracy of the pre-trained large model in a specific physical system.

[0059] MACE (Matrix Completion for Atomic-scale Energy) is a machine learning-based interatomic potential (MLIP) framework. Based on a high-order equivariant graph neural network architecture, MACE leverages the message-passing concept from graph neural networks (GNNs) to treat each atom as a node in a graph. By performing a spherical harmonic expansion of the local atomic environment and coupling it with the Clebsch-Gordan tensor, MACE achieves high-order equivariant modeling of the system's energy. This approach enables MACE to naturally handle the complexity of molecular structures and the interactions between atoms, as GNNs can effectively capture information propagation and dependencies within graph-structured data.

[0060] The present invention adopts a pre-trained MACE model and a transfer learning strategy, using the existing weights of the pre-trained MACE model as initialization parameters, and further optimizes it on a small sample set of the target system (such as the high-precision dataset constructed by the present invention).

[0061] Furthermore, during the fine-tuning process, a strategy of freezing some common layer parameters can be adopted, that is, freezing the parameters of shallow layers (such as feature extraction layers) and only training high-level parameters to prevent overfitting; a full parameter fine-tuning strategy can also be adopted to obtain higher accuracy.

[0062] Specifically, during fine-tuning training, the model is optimized using physical constraint loss functions (such as energy error, force error, and multi-body constraints). To prevent overfitting, fine-tuning training incorporates training mechanisms such as early stopping and regularization.

[0063] In the embodiment of the present invention, the process of obtaining a preliminary machine learning potential function model can be described as follows:

[0064] A1. Split the high-precision dataset into a training set and a validation set in a ratio of 9:1. Note that the ratio of single-material samples to heterogeneous interface samples is consistent in both the training and validation sets.

[0065] A2. Fine-tune the pre-trained MACE model using the training set. A physical constraint loss function, consisting of a weighted energy loss function and an atomic force loss function, is used.

[0066] A3. During training, the physical constraint loss and root mean squared error (RMSE) are output at each epoch, and the model parameters are optimized through backpropagation.

[0067] A4. Obtain the currently fine-tuned MACE model and evaluate the prediction accuracy of energy and atomic forces on the validation set. Compare the predictions with the DFT results to assess the model's accuracy and generalization capabilities.

[0068] A5. If the convergence condition is met (the maximum number of iterations is reached or the physical constraint loss obtained through the validation set increases for multiple consecutive rounds), a preliminary machine learning potential function model is obtained. Otherwise, return to step A2.

[0069] Specifically, when using a high-precision dataset to fine-tune the pre-trained MACE model, the material configuration (i.e., the coordinate file in three-dimensional space) is used as the model input, and the energy and atomic forces corresponding to the material configuration are used as the model output.

[0070] S4. Perform hyperparameter tuning on the preliminary machine learning potential function model to obtain a fine-tuned machine learning potential function model.

[0071] Specifically, hyperparameter tuning is a key step in improving model accuracy and generalization. When migrating a general, preliminary machine learning potential function model to a specific system, achieving high-precision predictions of physical quantities such as energy and atomic forces requires systematic tuning of the training hyperparameters of the preliminary machine learning potential function model. Well-optimized hyperparameter configuration can achieve the following: 1) accelerate convergence and stabilize the training process; 2) improve predictive power, especially for small samples; and 3) reduce overfitting and enhance the model's generalization ability to new structures.

[0072] Tuning strategies for key hyperparameters include:

[0073] 1) Learning rate (lr): The learning rate in the fine-tuning stage is generally a small value (such as 10 -4) to avoid excessive perturbations to the original pre-trained parameters. You can also use a learning rate decay strategy (such as cosine annealing, exponential decay) or an adaptive optimizer (such as AdamW, Radam).

[0074] 2) Number of frozen layers: This controls which layers retain their parameters during transfer learning and which layers are fine-tuned. This method freezes the bottom layers (low-level feature extraction layers) and only updates the higher layers.

[0075] 3) Adjust the physical constraint loss function and redistribute the loss function weights. In this embodiment of the present invention, when tuning key hyperparameters, a stress loss function is added to the original physical constraint loss function. Based on the focus of the task, the weights of energy, atomic force, or stress in the physical constraint loss function are adjusted accordingly; the calculation formula is:

[0076] L total =w E ×L E +w F ×L F +w S ×L S

[0077] Among them, L E represents the energy loss function, L F Represents the atomic force loss function, L S represents the stress loss function, w E represents the energy loss weight, w F Indicates the atomic force loss weight, w S Represents the stress loss weight.

[0078] S5. Dynamic simulation using fine-tuned machine learning potential function model.

[0079] Specifically, after obtaining a fine-tuned machine learning potential function model, its generalization ability can be used to construct a larger structural model and perform dynamic simulations to obtain the interface system properties of interest, including thermodynamic properties (radial distribution function, coordination number, etc.) and kinetic properties (diffusion coefficient, self-diffusion coefficient, conductivity, etc.).

[0080] like Figure 2 As shown in Figure 2, the specific process of dynamic simulation includes:

[0081] Build a molecular model to obtain simulation parameters such as atomic speed, position, temperature, stress, etc. as simulation input.

[0082] Call the fine-tuned machine learning potential function model to calculate the atomic forces.

[0083] The atomic force and position at the next moment are obtained through time integration.

[0084] Determine whether the convergence conditions are met. Otherwise, re-call the fine-tuning machine learning potential function model and repeat the above steps.

[0085] In some embodiments, the present invention selects two materials, metal Na and electrolyte Na3SbS4, to construct relevant fine-tuning machine learning potential function models.

[0086] The first step is to build a solid-solid interface system:

[0087] For metallic Na, a body-centered cubic (bcc) unit cell structure with a measured lattice constant a=4.20 Å at 300K was selected. Based on this, a 4 × 4 × 4 regular cell (2 atoms per cell) was constructed, totaling 128 atoms. The (110) surface configuration of metallic Na is based on a 3 × 3 repeating of the bcc-(110) plane within the plane, with 10 atomic layers (≈20 Å thickness) stacked along the normal direction, totaling 180 atoms. A 20 Å vacuum layer was added in the normal direction to eliminate periodic mirror coupling.

[0088] For the electrolyte Na3SbS4, the I4̅3m crystal system (a = b = c = 7.24 Å) was used to construct a bulk configuration of 2 × 2 × 2 supercells (16 atoms per unit cell), totaling 128 atoms. The (110) surface configuration of the electrolyte Na3SbS4 consists of 2 × 2 repeating tetragonal (110) surfaces in the plane, stacked in 10 layers (≈25 Å thickness) along the normal direction, with 224 atoms. An 18 Å vacuum layer is also provided along the normal direction.

[0089] Constructing the Na(110)|Na3SbS4(110) heterojunction: The aforementioned Na(110) layer and the Na3SbS4(110) layer were used for ZSL crystal plane matching, and the in-plane strain was controlled within ±5%. An initial gap of 3 Å was reserved between the two layers along the z direction and then they were bonded together. A 20 Å vacuum layer was then added to the whole. The total number of atoms in the interface model was 272.

[0090] The three sets of detailed models (Na bulk and Na(110) surface, Na3SbS4 bulk and Na3SbS4(110) surface, and Na(110)|Na3SbS4(110) heterointerface) can cover the multi-scale features of bulk, surface, and heterointerface, laying an accurate structural foundation for subsequent ab initio molecular dynamics (AIMD) data acquisition and machine learning potential function training. The data structure is as follows Figure 3 As shown in Figure 2, the bulk-surface-interface hierarchical structure of metal Na and electrolyte Na3SbS4 on the (110) crystal plane clearly shows the arrangement and bonding of the two-phase atoms.

[0091] The second step is to prepare a high-precision dataset:

[0092] All systems were simulated using PBE-PAW density functional theory. Rapid structural warm-up was performed to a maximum ion force of < 0.02 eV Å⁻¹ at a 500 eV plane wave cutoff and Γ-point sampling. The system was then transferred to the NVT ensemble with a time step of 1 fs and 10,000 sampling steps (10 ps). The temperature was maintained stable at each target temperature using an Andersen thermal bath. The temperature scheme was as follows: ① Na (bulk metal and (110) surface) was run independently at 300 K and 400 K; ② Na3SbS4 (bulk phase and (110) surface) was run at 300 K, 500 K, 700 K, and 900 K; and ③ Na(110)|Na3SbS4(110) heterointerface was run at 300 K, 500 K, 700 K, and 900 K. The electronic iteration convergence threshold was set to 1 × 10⁻. 5 eV; each step outputs energy, atomic forces, and structural coordinate data, resulting in a cumulative total of approximately 165,000 frames of high-quality data (i.e., 165,000 frames of DFT-level AIMD configurations). The metallic Na and electrolyte Na3SbS4 were simulated for their bulk and (110) surfaces, yielding approximately 115,000 frames. The Na(110)|Na3SbS4(110) heterointerface simulation yielded approximately 50,000 frames, laying the foundation for subsequent machine learning potential function training.

[0093] After removing 2% of outlier frames with non-convergent electrons or abnormal stresses from the 165,000 frames of high-quality data, and to highlight the potential function's ability to predict interface behavior, only approximately 400 frames (≈0.35%) were randomly retained from the bulk / surface portion, and approximately 1600 frames (≈3%) from the interface portion. This resulted in a reduced dataset of 2000 frames, 80% of which were interface samples. This reduced dataset was then randomly divided into an 1800-frame training set and a 200-frame validation set at a 9:1 ratio, ensuring a consistent ratio between the two types of samples in training and validation. All frames were uniformly written in the extxyz format, including energy and atomic forces, for direct input into the MACE network for potential function training.

[0094] The third step is to fine-tune the large model:

[0095] This embodiment of the present invention uses the officially released foundation model (mace-mpa-0-medium) as the starting point, uses mace_train.xyz (1800 frames) and mace_test.xyz (200 frames) as the training set and validation set, respectively, and runs on a single GPU with 64-bit precision; the optimizer uses AdamW+AMSGrad, with an initial learning rate of 5 × 10⁻4 , batch size 8, weight decay 1 × 10⁻ 6 , and robust convergence was achieved within 150 epochs using a combination of cosine annealing and stochastic weight averaging (SWA). The loss function weights energy and atomic forces 1.0:10.0 (interface-dominated, thus ignoring stresses), and scales are unified using RMS-forces scaling. EMA (0.99) and gradient clipping (‖∇‖ ≤ 10) are enabled during training to ensure generalization and numerical stability. The random seed is fixed at 123. It is worth noting that the elemental benchmark energies E0 = {Na -0.258 eV, S -0.889 eV, Sb -1.427 eV} are derived from isolated atom DFT benchmark energies calculated using the Vienna ab initioSimulation Package (VASP) using the same computational settings as AIMD, ensuring consistency between the fine-tuned model and the original data source at the energy zero point.

[0096] Step 4: Model evaluation:

[0097] When fine-tuning the large model using the training set, obtain the physical constraint loss (i.e., the weighted loss of energy and atomic forces) for each epoch, as well as the RMSE based on energy and atomic forces in each epoch. Use the validation set to perform multi-epoch validation on the fine-tuned large model and obtain the physical constraint loss (i.e., the weighted loss of energy and atomic forces) for each epoch, as well as the RMSE based on energy and atomic forces in each epoch. Plot the variation of the physical constraint loss over epochs during training and validation to verify convergence and overfitting, such as Figure 4 As shown in the figure, we plot the changes of energy-based RMSE and atomic force-based RMSE over epochs during training to visually compare their convergence speed and stability as training progresses. Figure 5 To further examine the absolute error distribution, diagonal scatter plots of energy and atomic force are drawn on the training set and validation set, as shown in Figure 6 As shown, black dots represent the training set, gray dots represent the validation set, the horizontal axis represents the DFT reference value, and the vertical axis represents the model prediction value. A dashed y = x reference line is also included to help identify systematic deviations, outliers, and linear correlations. This evaluation process, consisting of a loss curve, RMSE evolution, and two diagonal scatter plots, ensures that the model achieves consistent accuracy with the reference DFT data in both energy and force, providing a reliable potential function for subsequent large-scale interface dynamics simulations.

[0098] Step 5: Hyperparameter tuning:

[0099] Targeted fine-tuning of several key hyperparameters was performed: Regarding loss weights, the energy loss, atomic force loss, and stress loss weights were reallocated to ensure the model simultaneously focuses on the accuracy of interface energy, atomic forces, and stresses. Regarding the learning strategy, the initial lr was adjusted, warm-up was introduced, and the swa_lr (the learning rate used by SWA) and the start timing of stochastic weight averaging were reset to smooth out early oscillations and improve later generalization. Regarding stability optimization, the batch_size was increased, and regularization was strengthened on the decay rate (ema_decay) in weight decay (weight_decay) and the exponential moving average (EMA) to reduce the risk of overfitting. These parameter-level optimizations resulted in faster convergence, smaller training-validation error gaps, and significantly improved the overall fit quality of energy, atomic forces, and stresses.

[0100] Step 6, molecular dynamics simulation:

[0101] During the molecular dynamics phase, we integrated the system using metal units and three-dimensional periodic boundary conditions with a time step of 1 fs. After reading the interface configuration, we fixed the upper and lower atoms within approximately 15 Å as rigid walls, and calculated the interactions of the remaining atoms using the MACE potential function. The simulation was performed under 298 K NVT conditions for 2 ns (≈ 2,000,000 steps). During the simulation, the trajectory was output every 100 steps, periodically written to a restart file, and the final state was saved using write_data for subsequent analysis and long-term expansion. Figure 7 As shown in Figure 3, we used machine learning potential functions to run a 2ns simulation of the Na(110)|Na3SbS4(110) heterointerface, demonstrating the dynamic evolution of Na⁺ as it continues to diffuse from the metal side to the electrolyte side.

[0102] In the present invention, unless otherwise clearly stipulated and limited, the terms "installation", "setting", "connection", "fixation", "rotation" and the like should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can be the internal connection of two elements or the interaction relationship between two elements. Unless otherwise clearly defined, ordinary technicians in this field can understand the specific meanings of the above terms in the present invention according to the specific circumstances.

[0103] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for constructing a machine learning potential function model of a solid-solid interface system, characterized in that: The following steps are involved: S1. Construct a solid-solid interface system, including the bulk and surface configurations of the two solid materials, as well as the solid-solid interface configuration of the two materials; S2. Simulate the solid-solid interface system based on first principles, recording the output data of each configuration in the solid-solid interface system at each step of the simulation process to form a high-precision data set; Step S2 uses ab initio molecular dynamics simulation based on first principles to simulate the solid-solid interface system to form a high-precision data set. The process includes: S21. Setting simulation parameters, including functional selection, molecular dynamics parameter setting, and computing resource configuration; S22. simulating the solid-solid interface system based on the configuration of step S21, comprising: The bulk and surface configurations of the two materials are simulated separately. Each step of the simulation process outputs a single-material sample. A single-material sample includes the total potential energy, atomic forces, and atomic coordinate data of a bulk or surface configuration. Perform simulation on the solid-solid interface configuration of the two materials. Each step of the simulation process outputs a frame of heterogeneous interface sample, which includes the energy, atomic force and structural coordinate data of the solid-solid interface configuration; S23. Single-material samples and heterogeneous interface samples were screened to obtain a high-precision dataset, of which heterogeneous interface samples accounted for 80%; S3. Fine-tune the pre-trained MACE model using a high-precision dataset and optimize it using a physical constraint loss function to obtain a preliminary machine learning potential function model. S4. performing hyperparameter tuning on the preliminary machine learning potential function model to obtain a fine-tuned machine learning potential function model; S5. Dynamic simulation using fine-tuned machine learning potential function model.

2. The method for constructing a machine learning potential function model of a solid-solid interface system according to claim 1, characterized in that: Step S21 includes: S211. Select PBE-PAW density functional; S212. Set the molecular dynamics parameters: set the time step to 1 fs, the sampling step to 10,000 steps, select the NVT ensemble, and maintain the temperature at each target temperature using an Andersen thermal bath. S213. Configure computing resources: Set the plane wave cutoff energy to 500 eV and use Γ-point sampling.

3. The method for constructing a machine learning potential function model of a solid-solid interface system according to claim 1, characterized in that: The pre-trained MACE model in step S3 uses MACE-MP-0-medium, and the physical constraint loss function is composed of a weighted combination of an energy loss function and an atomic force loss function. The fine-tuning process of the pre-trained MACE model includes: S31. Split the high-precision dataset into a training set and a validation set in a ratio of 9:1; S32. The MACE-MP-0-medium is trained once using the training set. The prediction accuracy of the trained MACE-MP-0-medium in terms of energy and atomic forces is evaluated using the validation set to determine whether the convergence condition is met. If so, a preliminary machine learning potential function model is obtained. Otherwise, step S32 is executed again.

4. The method for constructing a machine learning potential function model of a solid-solid interface system according to claim 3, characterized in that: Fine-tuning of the pre-trained MACE model was performed on a single GPU with 64-bit precision. The MACE-MP-0-medium model was trained on the training set using AdamW+AMSGrad as the optimizer and an initial learning rate of 5 × 10⁻. 4 , batch size is 8, weight decay is 1 × 10⁻ 6 , combining cosine annealing with random weight averaging to achieve robust convergence within 150 epochs.

5. The method for constructing a machine learning potential function model of a solid-solid interface system according to claim 1, characterized in that: Step S4 performs hyperparameter tuning on the preliminary machine learning potential function model, specifically including: In terms of learning strategy, the initial learning rate is set and warm-up is introduced to adjust the initial learning rate; the random weight average learning rate and the start timing of random weight average are reset; Adjust the physical constraint loss function and use the loss function L that combines energy, atomic force and stress total , redistribute weights, loss function L total Expressed as: L total =w E ×L E +w F ×L F +w S ×L S , Among them, L E represents the energy loss function, L F Represents the atomic force loss function, L S represents the stress loss function, w E represents the energy loss weight, w F Indicates the atomic force loss weight, w S represents the stress loss weight; Increase batch_size and strengthen regularization of weight_decay and ema_decay; Freeze the underlying parameters of the preliminary machine learning potential function model for training.

Images