S1-MatAgent: A planner driven multi-agent system for material discovery

1State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation,
Chinese Academy of Sciences
2School of Artificial Intelligence, University of Chinese Academy of Science 3State Key Laboratory of High Performance Ceramics, Shanghai Institute of Ceramics,
Chinese Academy of Sciences
4Center of Materials Science and Optoelectronics Engineering,
University of Chinese Academy of Sciences
5School of Chemistry and Materials Science, Hangzhou Institute for Advanced Study,
University of Chinese Academy of Sciences

Indicates Equal Contribution

*Corresponding to jiahui.shi@ia.ac.cn, rannian@mail.sic.ac.cn

S1-MatAgent empowers material Q&A, meterial calculation, and material design.

Abstract

The discovery of high-performance materials is crucial for technological advancement. Inverse design using multi-agent systems (MAS) shows great potential for new material discovery. However, current MAS for materials research rely on predefined configurations and tools, limiting their adaptability and scalability. To address these limitations, we developed a planner driven multi-agent system (S1-MatAgent) which adopts a Planner-Executor architecture. Planner automatically decomposes complex materials design tasks, dynamically configures various tools to generate dedicated Executor agents for each subtask, significantly reducing reliance on manual workflow construction and specialized configuration. Applied to high-entropy alloy catalysts for hydrogen evolution reactions in alkaline conditions, S1-MatAgent completed full-cycle closed-loop design from literature analysis and composition recommendation to performance optimization and experimental validation. To tackle the deviations between designed materials and target, as well as high experimental verification costs, S1-MatAgent employs a novel composition optimization algorithm based on gradients of machine learning interatomic potential, achieving 27.7 % improvement in material performance. S1-MatAgent designed 13 high-performance catalysts from 20 million candidates, with Ni4Co4Cu1Mo3Ru4 exhibiting an overpotential of 18.6 mV at 10 mA cm−2 and maintaining 97.5 % activity after 500 hours at 500 mA cm−2. The universal MAS framework offers a universal and scalable solution for material discovery, significantly improving design efficiency and adaptability.

S1-MatAgent Architecture

S1-MatAgent Framework

Fig.1 S1-MatAgent: A planner driven multi-agent dynamic
collaborative system for material reverse design.

S1-MatAgent adopts a Planner-Executor architecture which incorporates dynamic workflow generation and adaptive agent configuration. Acting as a central scheduler, the Planner interacts directly with the user to receive a root task, automatically builds task workflows, and automatically generates several Executors to solve each subtask. Executor here refers to a task-specific customized agent, which incorporates unique system messages and a dedicated toolset.

HEA Design Task Execution

S1-MatAgent Framework

Fig.2 Execution process of all subtasks in the HEA design task.

Executor 1.1 and 1.2 extract catalyst chemical formulas and the ten most frequent metal elements for HEAs based on given literature information, which significantly narrows down the search space (Fig. 2a-d). Executor 2.1 analyzes given literatures by the ScienceOne model and generates potential compositions for further optimization (Fig. 2e and f). After that, Executor 2.2 iteratively optimizes previously recommended components based on mechanism modeling and MLIP gradients, forming the final candidate HEA for selection and synthesis by experimenters (Fig. 2g and h).

HEA Catalyst Evaluation and Optimization

Evaluation and Optimization of HER Activity in HEA

Fig.3 Evaluation and Optimization of HER Activity in HEA.

We provided a HEA structure generation tool and established a HER activity evaluation model based on the Volmer-Tafel mechanism in combination with MLIP (Fig. 3a). Specifically, we fine-tuned a MACE model to achieve precise simulation of the HER process of HEA, and the RMSE for energy prediction of this model on the test set is only 3.3 meV per atom (Fig. 3b). S1-MatAgent is equipped with a novel gradient-based catalyst optimization algorithm, and the one-step process of it is shown in figure 3c. After 10 iterations, our algorithm increased the highest HER activity of a population with an average final activity 2.4 times higher than that of the genetic algorithm and increase in the first round being 2.8 times greater, while the average individual activity improved from -21.12 to -15.28, representing a significant increase of 27.7 % (Fig. 3d).

Experimental Validation

Characterizations and alkaline HER performance

Fig.4 Execution process of all subtasks in the HEA design task.

We selected five materials suitable for laboratory synthesis for experimental verification, and figure 4 shows their electrochemical performances and morphology characterizations. We found Ni4Co4Cu1Mo3Ru4 exhibiting an overpotential of 18.6 mV at 10 mA cm−2 and maintaining 97.5 % activity after 500 hours at 500 mA cm−2.

BibTeX

  @article{s1-matagent2025,
  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},
  journal={arxiv},
  year={2025},
  url={https://arxiv.org/abs/2509.14542}
}