Lang2MLIP: A Multi-Agent Framework for Natural Language-Driven Development of Machine Learning Interatomic Potentials

Lang2MLIP is a multi-agent framework that enables end-to-end development of machine learning interatomic potentials (MLIPs) via natural language input. It models MLIP development as a sequential decision-making problem, where large language models (LLMs) automatically select optimization actions without predefined workflows. Its core goal is to lower the barrier for non-experts to develop MLIPs.

Machine learning interatomic potentials (MLIPs) are a rapidly developing interdisciplinary technology in materials science, enabling large-scale molecular dynamics simulations while maintaining quantum mechanical accuracy. But developing high-quality MLIPs faces three key challenges:

1. Professional knowledge barriers: Requires simultaneous expertise in three areas—atomic simulation (molecular dynamics, density functional theory, etc.), machine learning (model architecture selection, hyperparameter tuning, etc.), and workflow design (orchestrating data generation, training, validation, etc.);
2. Complexity of iterative active learning: MLIP development usually adopts an iterative active learning paradigm, requiring manual decisions on when to stop, expand datasets, or adjust model architectures, which is highly dependent on experience;
3. Limitations of existing automation solutions: Most existing automated pipelines assume fixed phase sequences or rely on expert intervention, leading to low efficiency or even failure when dealing with heterogeneous material systems.

Lang2MLIP adopts a multi-agent architecture, with the core concept of 'describing requirements in natural language and letting the intelligent system automatically complete the entire MLIP development process.'

• Decision agent observation space: Includes current dataset status (distribution, coverage, annotation quality, etc.), model status (architecture, parameters, training progress), evaluation results (validation set performance metrics, uncertainty estimation), and execution logs (records of previous steps, success and failure information);
• Action space: The agent can choose actions such as generating new training data, adjusting model architecture/hyperparameters, expanding/reducing datasets, performing validation tests, and backtracking for corrections;
• Self-correction capability: When failures or performance bottlenecks are detected, the system can automatically backtrack to previous subsystems to re-execute or adjust steps, adapting to the needs of different material systems.

The technical implementation of Lang2MLIP has two key aspects:

1. Natural language interface: Converts users' natural language requirements into structured task specifications. For example, if a user inputs 'Develop an MLIP to simulate the SEI layer of lithium-ion batteries, accurately describing the interface interactions between organic and inorganic components,' the system will parse key information such as material type and accuracy requirements;
2. No predefined pipeline: Completely abandons the traditional fixed phase sequence, treating development stages as optional actions dynamically selected by the agent. This brings three advantages: adaptability (different material systems follow optimal paths), robustness (flexible strategy adjustment when failures occur), and efficiency (avoiding unnecessary resource waste).

The research team validated the effectiveness of Lang2MLIP on the solid electrolyte interface (SEI) system of lithium-ion batteries.

• Experimental setup: The target system is an SEI layer containing multiple organic and inorganic components. Evaluation metrics include energy prediction accuracy, force prediction accuracy, and molecular dynamics stability. The baseline for comparison is the traditional fixed pipeline method;
• Key findings: The agent successfully identified the data gaps that most affect model performance, adjusted the priority of data generation and training based on intermediate results, and performed excellently in modeling organic-inorganic interface interactions.

Significance and Impact

• Materials science field: Promotes the democratization of computational tools, lowers the barrier to MLIP development, and allows more researchers to explore new materials;
• Agent systems: Demonstrates the potential of LLM-based multi-agent systems in automating complex scientific workflows, providing a reference paradigm for other professional fields.

Limitations

• High computational cost (large overhead for LLM inference and iterative training);
• Decision-making process interpretability needs improvement;
• Generalization ability needs to be validated in more material systems.

Future Directions

• Introduce reinforcement learning to optimize decision-making strategies;
• Develop visualization and interpretation tools for the decision-making process;
• Expand to more types of material simulation tasks.