Autonomous Materials Simulation with an LLM Agent: Modeling Exothermic Fronts in 2D Ferroelectrics

Recent advances in large language models (LLMs) enable autonomous agents that integrate reasoning, planning, and scientific tool use. Building on the ReAct framework [1] and Retrieval-Augmented Generation, RAG [2], we present an LLM-driven agent designed to orchestrate various materials computation tools. This agent autonomously generates aligned input files, analyzes calculation outputs, and executes complex research workflows. We demonstrate its capability by modeling exothermic front dynamics in 2D ferroelectric materials (e.g. CuInP2S6) under different high electric fields E. Guided by humans, the agent plans and manages the generation of density functional theory data for the ferroelectrics, orchestrates the training of a machine learning force field and deploys this potential for large-scale molecular dynamics simulations to model the front propagation under various E (exothermicity ~eaE, a is lattice parameter). Of interest is the transition from the atomistic domain wall dynamics to the heat-dominated phenomenological ZFK front [3]. This work demonstrates the LLM agent's capability to accelerate scientific discovery by automating the complex workflow.

[1] S. Yao et al. ReAct: Synergizing Reasoning and Acting in Language Models. ICLR 2023.

[2] P. Lewis et al. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. NeurIPS 2020.

[3] Y. Zeldovich & D. Frank-Kamenetskii. The theory of thermal propagation of flames. Zh. Fiz. Khim, 12, 100-105, 1938.