S1-MatAgent: Materials Discovery & Logistics

• S1-MatAgent System is a multi-agent framework that integrates a Planner–Executor architecture with ML-driven, gradient-based optimization to enhance material discovery, achieving a 27.7% performance improvement.
• It deploys an adaptive, decentralized MAS design in industrial logistics, utilizing real-time communication and task decomposition to boost throughput by up to 53%.
• The system's modular architecture enables easy extension to new scientific domains, supporting rapid workflow reconfiguration and autonomous task orchestration across varied applications.

The S1-MatAgent System is a class of multi-agent systems (MAS) designed to address two domains: automated material discovery and reconfigurable industrial logistics. In both implementations, S1-MatAgent embodies a modular, scalable architecture that leverages agent hierarchies, dynamic task decomposition, and domain-specific optimization strategies to achieve autonomous workflow generation, efficient resource allocation, and adaptive operational control (Wang et al., 18 Sep 2025, Fischer et al., 2022).

1. Planner–Executor Architecture for Materials Discovery

In the context of materials informatics, S1-MatAgent is instantiated as a Planner-driven MAS for inverse design and experimental validation of candidate materials. The core architectural innovation is a two-tiered Planner–Executor pattern. The Planner is implemented as an LLM-based agent capable of hierarchical task network (HTN) decomposition. It accepts high-level research goals (e.g., "Design highly active HEA catalysts for alkaline HER") and recursively decomposes them into primitive subtasks, constructing a directed acyclic graph with dependency edges representing execution order.

Primitive subtasks generate Executor agents, each dynamically configured with a custom prompt, a minimal required toolset (e.g., LiteratureExtractor, HEARecommender, MLIPevaluator, ExperimentLogger), and task-specific policies, such as composition validity checks. Executors operate within four canonical workflow tiers: literature mining, composition recommendation, performance optimization using MLIP-based gradients, and ingestion of experimental results. Executors run either in parallel or sequence per the HTN, and results are aggregated by the Planner to fulfill the initial research goal (Wang et al., 18 Sep 2025).

The process is codified by the following pseudocode for hierarchical decomposition:

$\mathrm{Proj}_\Delta$4

This approach decouples workflow structure from hard-coded scripts, enabling rapid adjustment to new inverse-design challenges.

2. Gradient-Based Optimization Algorithm

S1-MatAgent introduces a novel differentiable optimization strategy to refine material compositions. Given a candidate represented by the atomic fraction vector $\mathbf{c} = (c_1, ..., c_n)$, with $\sum_i c_i = 1$ and $c_i \geq 0$, the system defines the performance objective $J(\mathbf{c})$—such as the negative HER overpotential—using a fine-tuned machine learning interatomic potential (MLIP, specifically a MACE model).

Gradients $\nabla_c J(\mathbf{c})$ are computed with respect to the composition. Projected gradient ascent is then performed:

$$\mathbf{c}^{(t+1)} = \mathrm{Proj}_\Delta \left( \mathbf{c}^{(t)} + \eta\,\nabla_c J(\mathbf{c}^{(t)}) \right)$$

where $\mathrm{Proj}_\Delta$ projects the vector to the simplex of valid compositions. Two update heuristics are combined: (1) ratio adjustment (reallocating elemental fractions according to derivative magnitude), and (2) element-type swap (replacing the identity if gain is predicted). Empirically, this method achieves $27.7\%$ average performance improvement over heuristic or genetic search baselines, converging more rapidly and reliably across $400$ test cases (Wang et al., 18 Sep 2025).

3. Automated Material Discovery Workflow: HEA Catalysts Case Study

The S1-MatAgent workflow as applied to high-entropy alloy (HEA) catalyst discovery for alkaline HER proceeds through closed-loop execution as follows:

• Literature mining: 1,231 articles are mined to extract $\sim$2,800 formulas, prioritized to identify the top ten metals (Ni, Pt, Co, Fe, Mo, Ru, Cu, Ir, Pd, Rh).
• Candidate enumeration: All quinary combinations from 25 metals yield $\sum_i c_i = 1$020 million formula candidates.
• LLM-driven recommendation: The ScienceOne model proposes statistically justified initial formulas; code-based validators enforce chemical viability.
• Gradient optimization: Each candidate undergoes iterative refinement as described.
• Experimental validation: 13 top-ranking HEAs are downselected; five are synthesized and characterized.

Electrochemical metrics for these catalysts are summarized below.

Composition	$\sum_i c_i = 1$1 (mV)	Tafel (mV/dec)	$\sum_i c_i = 1$2 (Ω)	$\sum_i c_i = 1$3 (mF/cm$\sum_i c_i = 1$4)	Stability
Ni$\sum_i c_i = 1$5Co$\sum_i c_i = 1$6Cu$\sum_i c_i = 1$7Mo$\sum_i c_i = 1$8Ru$\sum_i c_i = 1$9	18.6	65.6	2.51	121.9	97.5% @ 500 h
Ni$c_i \geq 0$0Mo$c_i \geq 0$1Ru$c_i \geq 0$2Pd$c_i \geq 0$3Pt$c_i \geq 0$4	20.4	64.1	3.12	93.5	95.0% @ 300 h
Ni$c_i \geq 0$5Fe$c_i \geq 0$6Mo$c_i \geq 0$7Ru$c_i \geq 0$8Pd$c_i \geq 0$9	22.7	69.6	3.85	54.2	96.1% @ 200 h
Ni$J(\mathbf{c})$0Mo$J(\mathbf{c})$1Ru$J(\mathbf{c})$2Rh$J(\mathbf{c})$3Pt$J(\mathbf{c})$4	24.9	72.3	4.21	47.6	95.8% @ 250 h
Ni$J(\mathbf{c})$5Fe$J(\mathbf{c})$6Mo$J(\mathbf{c})$7Pd$J(\mathbf{c})$8Pt$J(\mathbf{c})$9	26.1	78.0	4.68	45.2	94.3% @ 150 h

The leading composition, Ni$\nabla_c J(\mathbf{c})$0Co$\nabla_c J(\mathbf{c})$1Cu$\nabla_c J(\mathbf{c})$2Mo$\nabla_c J(\mathbf{c})$3Ru$\nabla_c J(\mathbf{c})$4, demonstrated an overpotential of $\nabla_c J(\mathbf{c})$5 mV at $\nabla_c J(\mathbf{c})$6 mA cm$\nabla_c J(\mathbf{c})$7 and $\nabla_c J(\mathbf{c})$8 activity retention over $\nabla_c J(\mathbf{c})$9 hours at $$\mathbf{c}^{(t+1)} = \mathrm{Proj}_\Delta \left( \mathbf{c}^{(t)} + \eta\,\nabla_c J(\mathbf{c}^{(t)}) \right)$$0 mA cm$$\mathbf{c}^{(t+1)} = \mathrm{Proj}_\Delta \left( \mathbf{c}^{(t)} + \eta\,\nabla_c J(\mathbf{c}^{(t)}) \right)$$1 (Wang et al., 18 Sep 2025).

4. Multi-Agent System Framework in Industrial Logistics

In manufacturing and automated material flow systems (aMFS), S1-MatAgent embodies a metamodel-driven MAS with explicit decomposition into Automated Material-Flow Modules (aMFMs), each controlled by a dedicated module agent running on a PLC. The system is structured as follows:

• aMFM layer: Each module combines physical logistics hardware with a software agent responsible for configuration, real-time control, and inter-module negotiation.
• Coordinator layer: Provides FIPA-style MAS services, system-wide registration, and strategic route optimization (Agent Management System, Directory Facilitator).
• Dynamic topology management: Upon hot-plug or removal, module agents negotiate system graph updates and route recalculations, eliminating the need for full system recompilation or retesting (Fischer et al., 2022).

Decision logic within each agent encompasses four levels: configuration, material-flow scheduling, internal functional planning, and low-level hardware execution.

5. Communication, Ontology, and Decentralized Routing

S1-MatAgent employs a FIPA-ACL based message ontology and two main communication primitives: real-time/reactive (system-level handover, emergency) via EtherCAT/TCP/ADS; and planning/proactive (capacity queries, route reservations) over TCP/ADS or OPC-UA.

Route optimization for material flow is realized with semi-static routes (ssRs), leveraging a two-stage constraint-based algorithm:

1. Capacity pruning: For module $$\mathbf{c}^{(t+1)} = \mathrm{Proj}_\Delta \left( \mathbf{c}^{(t)} + \eta\,\nabla_c J(\mathbf{c}^{(t)}) \right)$$2 and flow $$\mathbf{c}^{(t+1)} = \mathrm{Proj}_\Delta \left( \mathbf{c}^{(t)} + \eta\,\nabla_c J(\mathbf{c}^{(t)}) \right)$$3, ensure

$$\mathbf{c}^{(t+1)} = \mathrm{Proj}_\Delta \left( \mathbf{c}^{(t)} + \eta\,\nabla_c J(\mathbf{c}^{(t)}) \right)$$4

where $$\mathbf{c}^{(t+1)} = \mathrm{Proj}_\Delta \left( \mathbf{c}^{(t)} + \eta\,\nabla_c J(\mathbf{c}^{(t)}) \right)$$5 is the transfer unit (TU) rate for flow $$\mathbf{c}^{(t+1)} = \mathrm{Proj}_\Delta \left( \mathbf{c}^{(t)} + \eta\,\nabla_c J(\mathbf{c}^{(t)}) \right)$$6, $$\mathbf{c}^{(t+1)} = \mathrm{Proj}_\Delta \left( \mathbf{c}^{(t)} + \eta\,\nabla_c J(\mathbf{c}^{(t)}) \right)$$7 encodes path inclusion, $$\mathbf{c}^{(t+1)} = \mathrm{Proj}_\Delta \left( \mathbf{c}^{(t)} + \eta\,\nabla_c J(\mathbf{c}^{(t)}) \right)$$8 is module capacity.

2. Dijkstra-based shortest path: Minimize

$$\mathbf{c}^{(t+1)} = \mathrm{Proj}_\Delta \left( \mathbf{c}^{(t)} + \eta\,\nabla_c J(\mathbf{c}^{(t)}) \right)$$9

where $\mathrm{Proj}_\Delta$0 is average process time per TU for module $\mathrm{Proj}_\Delta$1.

Routes are renegotiated if measured flow exceeds reserved capacity or variability tolerance (Fischer et al., 2022).

6. Evaluation and Scalability

S1-MatAgent has demonstrated high performance in both domains. For materials discovery, it yields state-of-the-art HEA catalysts and enables end-to-end autonomous inverse design, with improved activity gains and experimental success. In material flow logistics, S1-MatAgent's MAS architecture yields throughput increases of $\mathrm{Proj}_\Delta$2 in bidirectional simulations and $\mathrm{Proj}_\Delta$3 in complex plant layouts, with rapid reconfiguration and minimal programming overhead upon module (de)commission (Wang et al., 18 Sep 2025, Fischer et al., 2022).

The Planner–Executor pattern and metamodel-based modularization allow the system to generalize: new scientific domains can be incorporated by extending toolsets, adapting communication schemas, or swapping surrogate models within the gradient optimizer. This modularity and automation fundamentally distinguish S1-MatAgent from legacy static-MAS approaches which lack real-time adaptive planning and task orchestration.

7. Significance and Extensions

S1-MatAgent represents a paradigmatic shift in MAS design for scientific automation, coupling abstract task planning via LLMs with domain-specific differentiable optimization and closed-loop experimental validation. In logistics, its fully decentralized, metamodel-oriented control structure enables safe, auditable, and flexible material flow in heterogeneous industrial environments.

A plausible implication is that further integration with broader AI-driven surrogates (e.g., DFT, microstructure predictors) and cross-domain interoperability will enhance S1-MatAgent's capacity to autonomously address increasingly complex design and manufacturing challenges with minimal manual intervention (Wang et al., 18 Sep 2025, Fischer et al., 2022).