Toward Automated\nSimulation Research Workflow through\nLLM Prompt Engineering Design

The advent of Large Language Models (LLMs) has created\nnew opportunities\nfor the automation of scientific research spanning both experimental\nprocesses and computational simulations. This study explores the feasibility\nof constructing an autonomous simulation agent (ASA) powered by LLMs\nthrough prompt engineering and automated program design to automate\nthe entire simulation research process according to a human-provided\nresearch plan. This process includes experimental design, remote upload\nand simulation execution, data analysis, and report compilation. Using\na well-studied simulation problem of polymer chain conformations as\na test case, we assessed the long-task completion and reliability\nof ASAs powered by different LLMs, including GPT-4o, Claude-3.5, etc.\nOur findings revealed that ASA-GPT-4o achieved near-flawless execution\non designated research missions, underscoring the potential of methods\nlike ASA to achieve automation in simulation research processes to\nenhance research efficiency. The outlined automation can be iteratively\nperformed for up to 20 cycles without human intervention, illustrating\nthe potential of ASA for long-task workflow automation. Additionally,\nwe discussed the intrinsic traits of ASA in managing extensive tasks,\nfocusing on self-validation mechanisms, and the balance between local\nattention and global oversight.