PKU_MDAgent

About

MDAgent came out of Peking University to automate molecular-dynamics simulation setup, where materials scientists otherwise hand-write and iteratively debug LAMMPS scripts to extract thermodynamic parameters like heat capacity and thermal conductivity. Built on Microsoft AutoGen, a Planner decomposes the task, a Worker agent generates LAMMPS code executed in Docker containers, and an Evaluator scores each script on a 0–10 deduction rubric, looping until scripts score at least 8, with humans able to intervene mid-loop. Domain knowledge enters through QLoRA fine-tuning on a published 167-script LAMMPS dataset (with an expert-scored benchmark, LEQS, alongside) and optional RagFlow-based retrieval. The paper reports a 42.22% reduction in task time, and the code, datasets, and a Panel web UI shipped alongside the Scientific Reports publication. Its users are computational materials-science researchers, and the code is a research artifact rather than a maintained product.