S1-MatAgent: A planner driven multi-agent system for material discovery
A Planner driven Multi Agent System for Material Discovery
🔗 Homepage | 📑 Paper
🌟Introduction
S1-MatAgent is a planner-driven multi-agent system (MAS) that overcomes the limitations of predefined, inflexible MAS for materials discovery.
Its automated Planner-Executor architecture dynamically decomposes complex design tasks and configures specialized tools, drastically reducing manual effort.
We demonstrated its power by autonomously completing the full-cycle inverse design of high-entropy alloy catalysts for the hydrogen evolution reaction.
To bridge the design-to-target gap, a novel gradient-based optimization algorithm was employed, achieving a 27.7% performance improvement.
From 20 million candidates, S1-MatAgent designed 13 high-performance catalysts, with Ni₄Co₄Cu₁Mo₃Ru₄ exhibiting an exceptional overpotential of 18.6 mV and outstanding durability.
This work provides a universal and scalable framework for autonomous material discovery.
🔧Prepare
📦 Build an environment for MACE server
conda create -n server python=3.10
conda activate server
pip install mace-torch==0.3.13
# you may need to adjust the torch version according to your device
pip install ase
📦 Build an environment for S1-MatAgent
conda create -n matagent python=3.11
conda activate matagent
pip install -r requirements.txt
⚙️ Configure API
Open the .env file in your preferred text editor and insert your API keys in the corresponding fields.
🚀 Quick Start
1. Start MACE server
conda activate server
python model/mace_server.py
The MACE server needs to be kept running when using S1-MatAgent. It use a fine-tuned MACE model on HEAs mace_finetuned.model and a general one mace_agnesi_medium.model (you need to download it from mace-foundations, and place it under the model directory).
2.1 Run in browser
conda activate matagent
chainlit run app.py -w --host [your host] --port [your port] #e.g. chainlit run app.py -w --host 0.0.0.0 --port 8501
# Then open the browser
Three sample tasks of S1-MatAgent are predefined below the input bar: "Material Q&A", "Material Calculation", "Material Design". Click to execute.
When performing "Material Design" task, you need to upload a "literatures.csv" file. The file contains relevant literature information on high entropy alloy catalysts for hydrogen evolution reaction under alkaline conditions, and should contain at least the columns title, date, authors, abstract and publications
S1-MatAgent has good extensibility and can define other tools to enable it to perform more tasks.
2.2 Run in Terminal
conda activate matagent
python run.py "your task"
# python run.py "Summarize the latest application of alloy materials in the catalytic field"
# python run.py "Calculate the energy of Ni(HO)2, and demonstrate its structure"
# python run.py "Design highly active high entropy alloy (HEA) catalysts for hydrogen evolution reaction (HER) in an alkaline environment based on the literature file (path='your file path')"
🤝 Acknowledgements
🔍 License
This project is licensed under the Apache-2.0 License.
🚩Citation
If you find our repository useful in your research, please kindly consider cite:
@article{
title={S1-MatAgent: A planner driven multi-agent system for material discovery},
author={Xinrui Wang, Chengbo Li, Boxuan Zhang, Jiahui Shi, Nian Ran, Linjing Li, Jianjun Liu, Dajun Zeng},
year={2025},
eprint={2509.14542},
archivePrefix={arXiv},
url={https://arxiv.org/abs/2509.14542},
}