Autonomous AI generator for Machine Learning Interatomic Potentials

Machine learning interatomic potentials (MLIPs), which predict the system potential energy given an atomic configuration, have profoundly improved the accuracy and efficiency of molecular simulations. Learning from high-fidelity quantum chemistry-based data while interfacing with classical simulation frameworks, MLIPs inherit the quantum-level accuracy and the efficiency of classical molecular dynamics, achieving the best of both worlds. In recent years, MLIPs have witnessed great research efforts and various types of MLIPs have been proposed, tested, and documented. However, the development of MLIPs has largely been a manual, <i>ad hoc</i> process. Additionally, effective MLIPs are either scattered in online repositories such as Git-Hub or provided as supplementary materials of publications, making them difficult to locate and retrieve when needed. Here, we aim to create an autonomous AI generator for MLIPs, which consists of a Searcher, a Trainer, and an Evaluator. Given a request, for example, “generate an MLIP for Indium phosphide system”, the Searcher is first up to the task, looking for available MLIPs online via web scraping or APIs. If such MLIPs do not exist, the Trainer then performs the following tasks in sequence autonomously: select an MLIP template (such as a neural network or a SNAP MLIP, upon request or by certain criteria), generate DFT-MD training trajectories, and train the MLIP. For an MLIP returned either by the Searcher or the Trainer, the Evaluator calculates various accuracy metrics such as the mean absolute errors (MAEs) and R-squared values. It is also viable to request training multiple types of MLIPs and compare their performances. Lastly, we propose a large language model (LLM)-based agent named ChatMLIP which accepts textual user requests and delivers appropriate responses. By way of AI, the present work may accelerate the development of MLIPs and potentially benefit boarder research fields like carbon capture, protein engineering, drug discovery, among others.