S1-MatAgent: Autonomous Materials Discovery

• S1-MatAgent is a planner-driven multi-agent system that decomposes complex material design tasks and employs MLIP-based gradient optimization for predictive catalyst design.
• It automates literature mining, candidate composition generation, and experimental validation, achieving up to a 27.7% improvement in HER activity.
• The modular Planner–Executor architecture ensures scalability and adaptability, enabling extensive screening and iterative optimization in HEA catalyst discovery.

S1-MatAgent is a planner-driven multi-agent system (MAS) designed for automated and scalable material discovery, notably applied to the inverse design of high-entropy alloy (HEA) catalysts for hydrogen evolution reaction (HER) in alkaline conditions. Distinguishing itself from prior MAS architectures that rely on static workflow configurations and rigid toolsets, S1-MatAgent employs a dynamic Planner–Executor model to autonomously decompose complex tasks, assign specialized agents, and iteratively optimize target properties via machine learning interatomic potentials, resulting in significant improvements in material design efficiency and experimental validation cycles (Wang et al., 18 Sep 2025).

1. System Architecture and Workflow Decomposition

S1-MatAgent adopts a Planner–Executor architecture for flexible and hierarchical workflow construction:

• Planner Module The Planner receives high-level design requests and decomposes them into a hierarchical task network (HTN). This decomposition explicitly captures the temporal and logical dependencies among subtasks, which commonly encompass literature analysis, candidate composition generation, property calculation, and performance optimization. The Planner records this structure in a tree format, with each node referencing a discrete subtask.
• Executor Agents Executors are created per subtask, each provisioned with a dedicated toolset and instructions for their specialized roles (e.g., data extraction, statistical analysis, computation, optimization). Executors operate independently and asynchronously, returning intermediate results to the Planner, which in turn conditionally triggers subsequent subtasks. This dynamic allocation obviates manual configuration, allowing rapid adaptation to a broad range of materials tasks.

The overall closed-loop design cycle proceeds from initial data analysis to composition recommendation, computational/external validation, and property optimization. The agent-based architecture enables task parallelism and workflow scalability by modularizing expertise and tool integration with little manual intervention.

2. Automated Catalyst Discovery Pipeline

S1-MatAgent’s application to HEA catalyst design demonstrates its capability to conduct full-cycle inverse design:

• Literature Mining The system parsed 1,231 scientific articles, extracting chemical formulae and computing element occurrence frequencies. This resulted in prioritized selection of candidate metals (Ni, Pt, Co, Fe, Mo, Ru, Cu, etc.) for subsequent composition generation.
• Composition Recommendation and Justification Executor agents synthesized literature insights to propose initial HEA compositions. For example, agents identified “NiCoMoRuPt” and “NiFeCoMoPt” based on modeled performance metrics for HER. The recommendations are substantiated by network-based frequency analysis and documented design rationales.
• Closed-Loop Validation and Optimization Candidate catalysts progress through experimental synthesis and testing, supported by computational evaluation of HER-relevant properties such as overpotential and activity retention. The system validates not only the predicted performance but also long-term operational stability (e.g., Ni₄Co₄Cu₁Mo₃Ru₄ preserved 97.5% HER activity over 500 hours at 500 mA·cm⁻²).

The reaction mechanism is modeled via Volmer–Tafel steps, with activity descriptors quantifying water adsorption energies and transition state barriers ($\Delta G_{\text{ads}}$, $E_{\text{diss}}$, $\Delta G_{\text{H\,ads}}$). These descriptors subsequently guide computational optimization.

3. Gradient-Based Composition Optimization via MLIP

A distinctive feature of S1-MatAgent is the implementation of a gradient-driven optimization algorithm utilizing a machine learning interatomic potential (MLIP), specifically the MACE model:

• MLIP Integration The MACE MLIP is fine-tuned to predict interatomic forces and energetics for HEA compositions, achieving an RMSE of 3.3 meV/atom in tested scenarios.
• Activity Gradient Computation The algorithm computes the gradient of the HER activity descriptor with respect to atomic composition, $$\nabla A = \partial A(\mathbf{x}) / \partial \mathbf{x}$$, where $\mathbf{x}$ encodes element types and fractional ratios. This informs whether increasing or decreasing a specific element, or substituting one metal for another, will improve catalytic activity.
• Iterative Modification Strategies Two strategies are used: adjusting element abundances within constraints (maximum nickel content, ratios typically bounded between 5–35%) and element type replacement. Compositions are iteratively modified and re-evaluated until no further improvement is detected, upon which the optimized composition advances to experimental validation.

Activity descriptor and optimization can be formally represented as: $$A(\mathbf{x}) = f(\Delta G_{\text{ads}}, E_{\text{diss}}, \Delta G_{\text{H\,ads}}) \qquad \nabla A = \left( \frac{\partial A}{\partial x_1}, \frac{\partial A}{\partial x_2}, \ldots \right)$$

4. Performance Metrics and Experimental Validation

Quantitative metrics substantiate S1-MatAgent’s efficacy:

• Improvement Rate The closed-loop optimization yielded an average 27.7% improvement in HER activity over a set of 400 catalysts. When matched for iteration count, the gradient method outperformed traditional genetic algorithms by a factor of 2.4 in max HER activity and 2.8 in first-round gains.
• Catalyst Properties
  • Overpotential: 18.6 mV at 10 mA·cm⁻²
  • Activity retention: 97.5% after 500 hours at 500 mA·cm⁻²
  • Additional features: low Tafel slope, minimal charge transfer resistance (2.512 Ω)

These results confirm rapid convergence and robust experimental realization of optimized designs.

5. Scalability, Modularity, and Adaptability

S1-MatAgent’s universal framework offers multiple advantages for extensibility:

• Dynamic Task Planning The Planner–Executor model allows rapid augmentation of workflows and integration of new domain-specific modules without manual reprogramming, facilitating extension to other material systems and reaction conditions.
• Modular Agent Design Executor agents are self-contained units with specialized toolkits, enabling scalable deployment across varied design challenges.
• Closed-Loop Feedback The architecture inherently supports iterative learning, where experimental results can feedback into the Planner for further optimization cycles.
• Massive Candidate Screening The system routinely screens upwards of 20 million candidates with high throughput, supporting universal application across large search spaces.

6. Significance and Research Implications

S1-MatAgent demonstrates substantial advancement over conventional MAS approaches by automating task decomposition, modularizing agent expertise, and incorporating machine-learning guided optimization for material discovery. Its successful application to HEA catalyst development for HER attests to the approach’s scalability, efficiency, and adaptability in real-world contexts. The universal architecture, coupled with closed-loop verification and gradient-based optimization, constitutes a generalizable methodology for accelerated and autonomous material design cycles with broad relevance to catalysis, functional alloys, and related domains.