Application scene-oriented machine learning force field model convenient development method and device

Abstract

The invention discloses an application scene-oriented machine learning force field model convenient development method and device, and particularly relates to the technical field of force field simulation. A modularized standard model training framework is provided, rapid loading, flexible configuration and efficient training can be achieved, and the threshold and engineering complexity of model development are greatly reduced; a semantic recognition and machine learning force field model screening mechanism based on a large language model is introduced, and automation and intellectualization of model selection are achieved; large model distillation and small model fine tuning strategies are combined, training efficiency and prediction precision are considered, and the method is suitable for resource limited scenes; integration with a mainstream molecular simulation platform is realized, and efficient operation of the model on heterogeneous computing platforms such as a GPU and a CPU is supported.

Description

Technical Field

[0001] This invention relates to the field of force field simulation technology, and more specifically to a convenient development method and apparatus for machine learning force field models for application scenarios. Background Technology

[0002] With the increasing demands for simulation accuracy and efficiency in fields such as new material design, fabrication process optimization, and application scenario strategies, traditional first-principles calculation methods (DFT), while offering advantages in accuracy, suffer from high computational costs that severely limit their application in large-scale systems, long-term dynamic processes, and high-throughput screening tasks. When dealing with problems involving complex reaction pathway exploration and interface behavior simulation, using DFT for molecular dynamics simulations or energy surface searches often struggles to balance efficiency and accuracy. In recent years, the development of Machine Learning Force Fields (MLFFs) has offered new possibilities for overcoming these bottlenecks. MLFFs learn the mapping relationship between atomic structure and physical quantities, achieving efficient approximation of potential energy surfaces while significantly reducing computational overhead and maintaining accuracy close to that of DFTs. This characteristic makes them promising for large-scale structure sampling, complex reaction process modeling, and molecular dynamics simulations, becoming a research hotspot in computational materials and computational catalysis.

[0003] Currently, various types of machine learning force field models have been proposed, such as those based on deep neural networks, graph neural networks (GNNs), and kernel regression models. These methods have achieved certain results on different material systems. However, existing technologies still face many challenges in practical applications. On the one hand, the expressive power and generalization performance of force field models are limited; on the other hand, the contradiction between model training efficiency and inference speed and accuracy makes it difficult to meet the needs of industrial-grade or high-throughput computing. Therefore, the field of MLFF still needs continuous exploration and development. However, the development of most current force field models relies on specific model frameworks and descriptor designs, resulting in a lack of unified development standards and systems among different models. The highly customized and reusable model building process makes it difficult to promote or iterate models quickly, restricting the application of MLFF in a wider range of fields. At the same time, the existing toolchain and development process have not yet formed a complete closed loop, making it difficult to cover the entire lifecycle from data acquisition, feature construction, model training, performance evaluation to actual deployment.

[0004] Therefore, there is an urgent need for a general, modular, and scalable method and apparatus for creating, screening, and optimizing machine learning force field models. This apparatus should encompass standardized input / output interfaces, modular training processes, rapid access and flexible combination of different model architectures, professional force field training and optimization, and external platform integration technologies and standards. These technologies and standards will help build a unified system for evaluating and applying machine learning force fields, improving the development efficiency, accuracy comparison, and widespread application of force field models. Summary of the Invention

[0005] To address these issues, this invention provides a convenient development method and apparatus for machine learning force field models tailored to specific application scenarios.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a convenient development method for machine learning force field models oriented towards application scenarios, comprising the following steps:

[0007] Step 1: Standardize the input and output interfaces for building unified machine learning force field models; the input interface includes standard formats for microstructure data, topological link information, and physical quantity label input data; the output interface includes standard structures and units for energy, force, and stress data; construct a standard model training framework and modularize it. This training framework supports the rapid configuration, training, validation, and inference of machine learning force field models for application scenarios. The module functions include a model definition module, a loss function module, an optimizer module, and an evaluation metric module.

[0008] Step 2: Based on the created application scenario description table, material system description table, and machine learning force field model performance evaluation table, a force field evaluation and matching mechanism is formed to build a force field model database and knowledge base;

[0009] By leveraging large language models to vectorize user needs, we can match machine learning force field models with user-submitted application scenarios and material systems.

[0010] Step 3: Use a general model to quickly obtain high-quality data of the target system, and then fine-tune it by combining a small amount of first-principles DFT data. This will improve the prediction accuracy while maintaining low computational cost, so as to realize the acquisition and application of professional models in specific scenarios.

[0011] By calling the standardized input / output interface and modular training framework described in step 1, the professional force field model is modified and optimized to support the user's specific application scenarios and material systems.

[0012] Step 4: Based on the standardized interface and modular training framework described in Steps 1 and 3, construct a code organization structure that supports unified management of multiple types of force field models, so as to improve the R&D efficiency and adaptability of machine learning force field models in specific systems and application scenarios.

[0013] Preferably, the application-scenario-oriented machine learning force field model described in step 1 includes multilayer perceptron (MLP), graph neural network (GNN), and kernel method.

[0014] Preferably, step 4 specifically includes:

[0015] (1) Decouple the functions of data processing, model definition, training process and evaluation index to form a pluggable and scalable modular system so as to realize the sharing and reuse of various force field models;

[0016] (2) Supports the development of multiple models coexisting and switching, and unifies the management of configuration files, training and application processes of various models, making it convenient for model comparison and deployment;

[0017] (3) Provide a collaborative development interface between models to support the collaborative calling and management of multiple models in model distillation, integration training and structure transfer tasks, so as to improve the R&D efficiency and adaptability of machine learning force field models in specific systems and application scenarios.

[0018] Preferably, a unified code organization structure based on multiple models is used to achieve the integration and interaction of force field models with mainstream molecular dynamics simulation software, including ASE and LAMMPS. Through interface encapsulation and data adaptation, the obtained model can be used for molecular dynamics simulation tasks, and supports automated structure transfer and calculation calls.

[0019] Based on the code organization structure that supports unified management of multiple types of force field models, the interface between force field models and heterogeneous computing hardware is realized. Through a unified backend calling mechanism, the distributed scheduling and efficient management of model training and inference tasks on different hardware resources are supported to ensure computing performance and scalability.

[0020] The present invention also discloses a convenient development device for machine learning force field models for application scenarios, which performs the above-described method.

[0021] This invention establishes a unified input / output interface standard; proposes a modular standard model training framework that enables rapid loading, flexible configuration, and efficient training, significantly reducing the threshold and engineering complexity of model development; introduces a semantic recognition and machine learning force field model selection mechanism based on a large language model to automate and intelligently select models; combines large model distillation and small model fine-tuning strategies to balance training efficiency and prediction accuracy, making it suitable for resource-constrained scenarios; and achieves integration with mainstream molecular simulation platforms, supporting efficient model operation on heterogeneous computing platforms such as GPUs and CPUs. This invention constructs a standardized, intelligent, and platform-based machine learning force field development technology and specification, significantly improving the development efficiency of force field models and providing strong technical support for complex material modeling and advanced force field research. Attached Figure Description

[0022] Figure 1 A flowchart illustrating the overall process of the solution provided by this invention;

[0023] Figure 2 This is a schematic diagram of the input and output of the machine learning force field standardization provided by the present invention;

[0024] Figure 3 The graphical interface for the general machine learning force field framework provided by this invention;

[0025] Figure 4 The structural diagram of the general machine learning force field framework provided by this invention is shown.

[0026] Figure 5 A schematic diagram of the large language model mechanism for machine learning force fields provided by this invention;

[0027] Figure 6 The process for selecting and constructing a force field model based on a large language model is provided for this invention;

[0028] Figure 7 The machine learning force field of Au based on the S7 process and the phonon dispersion curve calculated by DFT are provided for this invention.

[0029] Figure 8 The flowchart for force field model selection and construction based on large language model provided by this invention. Detailed Implementation

[0030] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] This invention discloses a convenient development method for machine learning force field models oriented towards application scenarios, including:

[0032] S1. Construct standardized input and output interfaces for various machine learning force field models, unify the units and data structures of the microstructure data, topological connection information, and output data such as energy, force, and stress required for different model inputs, and ensure that different force field models can be quickly embedded into relevant frameworks.

[0033] S2 is a modular and pluggable model training framework built upon S1, supporting rapid configuration, training, validation, and inference of various mainstream model architectures (including but not limited to Multilayer Perceptron (MLP), Graph Neural Network (GNN), kernel methods, etc.). The training framework includes core components such as data processing, parameter setting, logging, model training, and extension interfaces. Users can flexibly combine and replace these components through configuration files or APIs to adapt to different physical systems and task requirements.

[0034] S3. Construct a force field model database and knowledge base to realize a quantitative evaluation and intelligent matching mechanism for different force field models. The mechanism constructs a multi-dimensional evaluation standard for measuring the performance and applicability of different machine learning force field models, including an application scenario description table, a material system description table, and a machine learning force field model performance evaluation table, supporting quantitative comparison and matching analysis of multiple force field models.

[0035] S4. Based on S3, a force field selection and recommendation mechanism assisted by a large language model is proposed. Based on the quantitative evaluation and intelligent matching mechanism of the force field model, by accessing the large language model, the user's input needs are understood and parsed, and the most suitable force field model is automatically retrieved from the force field model database and knowledge base. Matching suggestions and model explanation information are generated, thereby improving the intelligence and automation level of force field model selection.

[0036] S5. A force field model optimization method based on distillation and fine-tuning is proposed. A general model is used to generate high-quality label data of the target system. The professional model achieves knowledge transfer by fitting the high-quality label data and is fine-tuned by combining a small amount of first-principles (DFT) data. This reduces training costs while improving prediction accuracy and enables rapid deployment of professional models in specific application scenarios.

[0037] S6. Construct an adaptation interface for mainstream molecular simulation platforms (such as LAMMPS, ASE, etc.); and support the efficient operation of models on heterogeneous computing platforms such as GPUs and CPUs. Through a unified inference interface and resource scheduling mechanism, ensure the portability and execution efficiency of the models.

[0038] like Figure 8 As shown, in addition to further protecting the critical paths and operational specifications of this invention in the actual application of the model, a force field model selection and construction process based on a large language model is also provided, which specifically includes the following steps:

[0039] S701: Users submit questions or task objectives described in natural language to the system through a graphical interface or API, based on the material system, reaction type, or simulation task of interest. These questions typically include information such as the combination of target elements, required accuracy, and simulation timescale.

[0040] S702: The system calls the embedded large language model to perform semantic parsing and intent recognition of user needs. It then combines the built-in machine learning force field model database and knowledge base to perform automatic retrieval and matching analysis, generating several recommended solutions and providing model adaptation suggestions. If an existing model that meets the requirements is found, the process proceeds to S703; otherwise, if no suitable model can be used directly, the process jumps to S707 to initiate the custom model creation process.

[0041] S703: The user selects the recommended model and connects to the actual computational process, such as performing molecular dynamics simulations, structure optimization, or reaction pathway search. If the simulation results meet the prediction accuracy and application objectives, the process ends; otherwise, it proceeds to the subsequent model optimization path.

[0042] S704: The user further configures and adjusts the selected model. The system calls a general model to generate high-quality labeled data (such as energy, force, etc.) for specialized models to undergo distillation-style training. If the model's performance is insufficient during this process, the nature of the problem needs to be determined.

[0043] S705: If the problem is insufficient generalization ability or accuracy deviation, proceed to this step. Enrich the training set by supplementing it with a small amount of first-principles calculation data (such as DFT) and fine-tune the professional model to improve accuracy.

[0044] S706: If the model's expressive power is insufficient (i.e., the structural design is limited), proceed to this step. Users can adjust the model architecture based on the standardized model building interface, such as introducing new features or modifying the neural network structure.

[0045] S707: If the model database cannot provide a model that meets the requirements, the user creates a new model structure using the standard interface, and then completes the entire process from data preparation, model training to evaluation and validation within the modular training framework, generating a new usable force field model. Then, proceed to step S703 for actual deployment and use.

[0046] Based on the technical solution of the present invention described above, this embodiment provides a general, modular, and scalable method and apparatus for creating, filtering, and optimizing machine learning force field models, the overall architecture of which is as follows: Figure 1 As shown, it includes:

[0047] S1, such as Figure 2 As shown, this paper describes the development based on Python and the PyTorch machine learning framework, constructing a unified input / output interface for various machine learning force field models. The microstructure and topological connectivity information are standardized into a fixed-format tensor input: if the number of atoms to be calculated is n, and the maximum nearest neighbor number is m, then the atom types... The element type of each central atom; the neighboring atom type tensor The type of the m nearest neighbors of each atom; the neighbor index tensor Record the neighboring atom numbers corresponding to each central atom; distance and direction information tensor Store the relative distances between the central atom and each of its neighboring atoms, as well as the Cartesian components (r). ij ,x ij ,y ij ,z ij ).

[0048] At the output end, the output is uniformly composed of five tensors, including: atomic energy atomic force Atomic stress With overall stress The output and output unit specifications are both based on mainstream units, with energy expressed in eV and distance in eV. This standard interface can serve as a common data layer in the data processing workflow, thereby improving interoperability between models and the flexibility of platform deployment.

[0049] S2. Based on S1, this embodiment further develops a highly modular and pluggable machine learning force field training framework, with a GUI as follows: Figure 3 As shown. This framework is built around PyTorch and adopts a component-oriented design approach, such as... Figure 4 As shown, data processing, parameter settings, logging, model building, and extension interfaces are decoupled and encapsulated, allowing users to flexibly combine different functional modules through the input.json configuration file and API interfaces. The training framework supports fast loading and training of various mainstream force field models (including but not limited to Behler–Parrinello atomic symmetry function (BP), Deep Potential (DP), etc.), and internally integrates multiple loss functions (energy, force, stress loss and their weighted combinations), commonly used optimizers (Adam, SGD, Adamw), and multi-metric evaluation systems (RMSE, MAE, R). 2 It also provides breakpoint recovery, early stop detection, and training control functions. The entire training process is managed by a unified process controller and can run efficiently in CPU or GPU environments.

[0050] S3. By constructing a force field model database and knowledge base, a quantitative evaluation and intelligent matching mechanism for the performance of different force field models is achieved. This mechanism includes the following core components: an application scenario description table, used to understand and characterize the simulation task requirements proposed by the user, including simulation scale (atomic level, nanometer level), accuracy requirements, and time scale; and a material system description table, used to record basic information of the force field dataset, such as elemental composition, structural type, and application direction. Based on the above information, combined with user input, a complete task semantic vector is generated. To ensure the consistency and comprehensiveness of the evaluation, this invention defines a force field model performance evaluation table, covering key dimensions such as model accuracy (energy RMSE, force RMSE), computational efficiency (e.g., time per step), applicability (elemental types, structural complexity), generalization ability (performance in new structure testing), and model size. These components constitute a multi-dimensional evaluation mechanism.

[0051] S4, such as Figure 5 The paper presents a force field screening and recommendation mechanism based on the S3 large language model. This mechanism introduces natural language understanding and semantic matching capabilities on the basis of multi-model evaluation, realizing intelligent recommendation throughout the entire process from user needs to model matching. The integrated large language model DeepSeek is used to perform semantic parsing and vectorization encoding of user input, generating corresponding task semantic vectors. Subsequently, the system uses a constructed intelligent matching mechanism to calculate and rank the similarity between the task semantic vectors and the evaluation metrics of various models in the database, automatically retrieving the set of candidate models that best meet the user's needs, and outputting matching scores and merit-based information to assist users in making quick decisions. The system uses the large language model to generate recommendation reasons and model explanations in natural language, including the applicable basis, performance advantages, usage precautions, and possible limitations of the recommended models.

[0052] S5. This embodiment proposes a force field model optimization method based on distillation and fine-tuning. First, based on the problem to be solved and the target system, the most suitable model is selected from a general force field model library to generate high-quality labeled data including energy, force, and stress. Then, a lightweight specialized model is selected or constructed, and distillation training is performed by fitting the prediction results of the general model, thereby achieving effective transfer of general knowledge to a specific system. Based on this, according to the accuracy performance of the preliminary model and the complexity of the research problem, an appropriate amount of high-precision data obtained through first-principles calculation (DFT) is introduced to fine-tune the specialized model, further improving its accuracy and generalization ability under the target system. During the fine-tuning process, a joint loss function is used, comprehensively considering the fitting effect of distillation supervision and real labels, thereby ensuring efficient and stable model convergence.

[0053] S6. To achieve efficient integration between the machine learning force field model and mainstream molecular simulation platforms (such as LAMMPS and ASE), this invention uses PyTorch's JitScript technology to export the model as a static graph after training. This exported model has excellent cross-language compatibility and can be directly called in environments such as Python and C++, making it suitable for subsequent high-performance computing or large-scale simulation tasks. Meanwhile, the ASE interface integrates the model through Python extension modules, while the LAMMPS interface loads the model inference library via a C++ plugin. Based on the native features of the PyTorch framework, the model can automatically adapt to run on CPU or GPU platforms during deployment, supporting heterogeneous computing resource scheduling and switching.

[0054] like Figure 6 As shown, the process of force field model selection and construction based on large language model assistance will be illustrated by constructing machine learning force fields for Au and Na respectively. It is worth noting that the protection of this invention should not be limited to the system in the embodiments, but to various machine learning force fields created based on this process.

[0055] Taking element Au as an example, the specific implementation process of step S7 is shown below:

[0056] S701. The user submits a task request to the system in natural language via the API interface, for example, inputting: "Can you provide a machine learning force field suitable for face-centered cubic (Fcc) structures of Au, requiring moderate computation speed and reliable accuracy?" This request is recorded by the system and submitted to the integrated large language model DeepSeek for processing.

[0057] S702: The system calls the DeepSeek large language model to perform semantic parsing and intent recognition of user needs. It then combines this with the pre-built machine learning force field model database and knowledge base to perform automatic retrieval and matching analysis, generating several recommended solutions and providing model adaptation suggestions. If an existing model that meets the requirements is found, the process proceeds to S703.

[0058] S703. The user selects the recommended model and connects it to the actual computational optimization path, such as performing a phonon spectrum task. The result is as follows: Figure 7 As shown;

[0059] S704. The user further configures and adjusts the selected model. The system calls the general model to generate high-quality labeled data (energy and force of the structure) for the specialized model to undergo distillation training. If the model performance is insufficient during this process, the nature of the problem needs to be determined:

[0060] If the problem is insufficient generalization ability or accuracy deviation, proceed to S705; if the problem is insufficient model expressive ability (i.e., limited structural design), proceed to S706.

[0061] S705: Perform DFT calculations using VASP to enrich the training set and fine-tune the professional model to improve accuracy.

[0062] S706. Users can build interfaces to adjust the model architecture based on standardized model processes, such as introducing new features and modifying the neural network structure.

[0063] Taking element Na as an example, the specific implementation process of step S7 is shown below:

[0064] S701. The user submits a task request in natural language through the API interface provided by the system. For example, the user may input: "Can you provide a machine learning force field suitable for face-centered cubic (FCC) structures of Na, requiring high computational speed and slightly lower prediction accuracy?" The natural language request is received by the system and passed to the DeepSeek large language model for further processing.

[0065] S702 invokes the DeepSeek large language model to perform semantic parsing and intent recognition on the above task request. Through natural language understanding, the system accurately extracts the core elements of interest to the user, including element type (Na), crystal structure (FCC), accuracy tolerance, and performance preference (biased towards high speed). Based on the built-in force field model database and knowledge base, the large language model intelligently generates suggestions by considering the performance of currently trained models. For example, if the system does not find an existing model that simultaneously meets the requirements of high speed and low accuracy, it returns the following prompt: "No force field model that fully meets the requirements was found in the current database. It is recommended to use classical potential functions (such as Lennard-Jones potential, LJ potential) for rapid fitting and application. The LJ potential function performs well in terms of speed and is suitable for preliminary simulations with moderate accuracy requirements." Since no readily available model was found, the process proceeds to step S707 to initiate the custom model creation process.

[0066] S707, an example of creating a new model using the API interface is as follows: `class LJ(FieldModel):` The potential function is created through this class. Then, the user sets training parameters (such as learning rate, number of training epochs, loss function type, etc.) by uploading or editing a standardized `input.json` configuration file. The platform's modular force field training framework automatically completes the entire process, including data loading, model initialization, training scheduling, and result evaluation. To reduce the user's burden, the platform calls a general large model to generate structure and label data under the target system (such as Na-FCC) to quickly build a high-quality training set. After model training is complete, the process transitions to S703 for actual deployment and use.

[0067] S703. The user connects the trained LJ potential model to an actual molecular simulation task, such as molecular dynamics simulation and structure optimization. If the model performs well in the target task and meets the accuracy and performance requirements, the task process ends. If performance bottlenecks or large error deviations still exist, the system will suggest that the user proceed to subsequent model optimization paths (such as model fine-tuning or structural improvement) to continuously improve model performance.

[0068] Although the present invention has been described in detail above with general descriptions and specific embodiments, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the present invention fall within the scope of protection claimed by the present invention.

Claims

1. A convenient development method for machine learning force field models for application scenarios, characterized by: Includes the following steps: Step 1: Standardize the input and output interfaces for building unified machine learning force field models; the input interface includes standard formats for microstructure data, topological link information, and physical quantity label input data; the output interface includes standard structures and units for energy, force, and stress data; construct a standard model training framework and modularize it. This training framework supports the rapid configuration, training, validation, and inference of machine learning force field models for application scenarios. The module functions include a model definition module, a loss function module, an optimizer module, and an evaluation metric module. Step 2: Based on the created application scenario description table, material system description table, and machine learning force field model performance evaluation table, a force field evaluation and matching mechanism is formed to build a force field model database and knowledge base; By leveraging large language models to vectorize user needs, we can match machine learning force field models with user-submitted application scenarios and material systems. Step 3: Use a general model to quickly obtain high-quality data of the target system, and then fine-tune it by combining a small amount of first-principles DFT data. This will improve the prediction accuracy while maintaining low computational cost, so as to realize the acquisition and application of professional models in specific scenarios. By calling the standardized input / output interface and modular training framework described in step 1, the professional force field model is modified and optimized to support the user's specific application scenarios and material systems. Step 4: Based on the standardized interface and modular training framework described in Steps 1 and 3, construct a code organization structure that supports unified management of multiple types of force field models, so as to improve the R&D efficiency and adaptability of machine learning force field models in specific systems and application scenarios.

2. The convenient development method for machine learning force field models oriented to application scenarios according to claim 1, characterized in that: The application-scenario-oriented machine learning force field model described in step 1 includes multilayer perceptron (MLP), graph neural network (GNN), and kernel methods.

3. The convenient development method for machine learning force field models oriented towards application scenarios according to claim 1, characterized in that: Step 4 specifically includes: (1) Decouple the functions of data processing, model definition, training process and evaluation index to form a pluggable and scalable modular system so as to realize the sharing and reuse of various force field models; (2) Supports the development of multiple models coexisting and switching, and unifies the management of configuration files, training and application processes of various models, making it convenient for model comparison and deployment; (3) Provide a collaborative development interface between models to support the collaborative calling and management of multiple models in model distillation, integration training and structure transfer tasks, so as to improve the R&D efficiency and adaptability of machine learning force field models in specific systems and application scenarios.

4. The convenient development method for application-scenario-oriented machine learning force field models according to claim 1, characterized in that: Based on a unified code organization structure for multiple models, the force field model is integrated and interacts with mainstream molecular dynamics simulation software, including ASE and LAMMPS. Through interface encapsulation and data adaptation, the resulting model can be used for molecular dynamics simulation tasks, and supports automated structure transfer and calculation calls. Based on the code organization structure that supports unified management of multiple types of force field models, the interface between force field models and heterogeneous computing hardware is realized. Through a unified backend calling mechanism, the distributed scheduling and efficient management of model training and inference tasks on different hardware resources are supported to ensure computing performance and scalability.

5. A convenient development device for machine learning force field models for application scenarios, characterized in that: The device performs the method according to any one of claims 1-4.

Images