A collaborative agent with two lightweight synergistic models for autonomous crystal materials research

Abstract: Current LLMs require hundreds of billions of parameters yet struggle with domain-specific reasoning and tool coordination in materials science. Here, we present MatBrain, a lightweight collaborative agent system with two synergistic models specialization for crystal materials research. MatBrain employs a dual-model architecture: Mat-R1 (30B parameters) as the analytical model providing expert-level domain reasoning, and Mat-T1 (14B parameters) as the executive model orchestrating tool-based actions. Entropy analysis confirms that this architecture resolves the conflict between tool planning and analytical reasoning by decoupling their distinct entropy dynamics. Enabled by this dual-model architecture and structural efficiency, MatBrain significantly outperforms larger general-purpose models while reducing the hardware deployment barrier by over 95%. MatBrain exhibits versatility across structure generation, property prediction, and synthesis planning tasks. Applied to catalyst design, MatBrain generated 30,000 candidate structures and identified 38 promising materials within 48 hours, achieving approximately 100-fold acceleration over traditional approaches. These results demonstrate the potential of lightweight collaborative intelligence for advancing materials research capabilities.

• The paper introduces MatBrain, a dual-model agent combining a 30B analytical model and a 14B executive model to autonomously discover crystal materials.
• The paper details a robust architecture that decouples deterministic analytical inference from adaptive tool planning, overcoming entropy collapse in traditional LLMs.
• The paper demonstrates state-of-the-art performance with superior metrics and a 100-fold acceleration in catalyst design, validated by experimental synthesis and testing.

Lightweight Collaborative Agent Design for Autonomous Crystal Materials Research

Introduction

The paper presents "MatBrain," a dual-model lightweight collaborative agent tailored for the autonomous discovery of crystal materials. The MatBrain system is motivated by the inefficacy of large, general-purpose LLMs in capturing domain-specific, context-sensitive reasoning and orchestration requirements in materials science. The design leverages an analytical model (Mat-R1, 30B parameters) and an executive model (Mat-T1, 14B parameters), each specialized for scientific reasoning and tool-driven task execution, respectively. The architecture decouples two orthogonal cognitive regimes—deterministic, low-entropy analytical inference and high-entropy, adaptive tool planning—addressing the entropy collapse and inefficiency present in monolithic LLMs.

Architecture and Methodological Innovations

The core architectural framework consists of two synergistic LLMs:

Mat-R1 (Analytical Model):

A 30B parameter model based on Qwen3-30B-A3B, fine-tuned with a 252K curated corpus (Mat-252K-SFT) synthesizing structured database entries and literature-derived knowledge. This model focuses on domain reasoning, mapping natural language to crystallographic knowledge representations for hypothesis generation, plausibility assessment, and result interpretation.

Mat-T1 (Executive Model):

A 14B parameter model initialized from Qwen3-14B. It is trained using RL (DAPO algorithm) on the Model Context Protocol (Mat-MCP)-standardized tool suite. It manages workflow planning, tool selection, and multi-step tool chaining—optimized with a composite reward signal targeting process fidelity, syntax, and interactional depth.

Mat-MCP Tool Ecosystem:

Tool fragmentation is resolved via containerized integration of generative design (CrystaLLM, MatterGen), simulation (MEGNet, CHGNet), validation (pymatgen, PhaseDiagram), and retrieval platforms within a unified protocol, supporting robust and scalable tool invocation.

Collaborative Workflow:

Mat-T1 translates queries into tool-driven action sequences, iteratively executing workflows. Mat-R1 scrutinizes results, provides expert analysis, and determines the need for further investigation, establishing a dynamic "generation-verification-feedback loop." The process leverages shared global state, concurrency, and limit-enforced graph routing (max_iterations=6).

Data Curation and Training Pipeline

A meticulous data curation pipeline aligns structured CIF/JSON database entries with literature-derived contextual information, constructing high-fidelity instruction tuning pairs. The generate-distill-validation paradigm ensures data quality, with multi-model ensemble chunking, prompt-based question generation, DeepSeek-R1 reasoning distillation, and dual-layer human-machine factual validation. Training adopts large context windows (up to 16k tokens), hybrid and expert parallelism, and precision optimization pipelines. Mat-T1's RL training utilizes a diverse, high-complexity Mat-20K-RL dataset, engineered to preclude reward hacking and shortcut exploitation.

Entropy Dynamics and Information-Theoretic Analysis

A central theoretical contribution is the analysis of token-level Shannon entropy, demonstrating that post-training, Mat-R1 exhibits strongly convergent, low-entropy output (mean ≈ 0.483), while Mat-T1's entropy increases (mean ≈ 0.974), indicating effective decoupling. Kernel density estimation indicates a high-entropy peak for Mat-T1 at ~3.3 bits—corresponding to the expanded action space of multi-tool invocation—distinct from the monotonic decay of Mat-R1. The coupling of high-entropy exploratory search with deterministic reasoning enables information gain to be efficiently transformed into actionable scientific confidence.

Benchmarking and Empirical Performance

Comparative Performance:

MatBrain exhibits superior performance on the Mat-252K test set compared to state-of-the-art general-purpose models (DeepSeek-R1 671B, GPT-5, Gemini-2.5-Pro), particularly in tasks requiring geometric syntax, structure manipulation, and rigorous property regression. Monolithic LLMs, even when agentified, fail at high-precision syntactic and geometric tasks, and display catastrophic regression failure.

Numerical Results:

• Regression (energy/property): $R^2$ values of 0.97 (formation energy), 0.99 (energy above hull), 0.84 (Fermi level) with low MAEs—all significantly superior to general LLMs.
• Deployment/Hardware Efficiency: Over 95% reduction in hardware requirements (H100 cluster to dual RTX 4090 workstation), democratizing state-of-the-art AI for laboratories with modest resources.
• Throughput: In catalyst design, MatBrain autonomously generated and reasoned over 30,000 candidate structures, identifying 38 promising materials within 48 hours—a ~100-fold acceleration over traditional, months-long workflows.

End-to-End Case Studies and Practical Validations

Materials Research Lifecycle:

MatBrain autonomously executes structure generation (including fractional occupancy, Vegard's Law extrapolation), property prediction, synthesis design, and theoretical characterization. The system demonstrates adaptability for non-stoichiometric materials and robust expert-level reasoning in high-dimensional, multi-objective scenarios.

Autonomous Catalyst Discovery:

A seven-stage pipeline, from bioinspired hypothesis generation to synthesis protocol computation, reduces 30,000 raw candidates to 38 previously unreported, stable, and functionally plausible vanadium-based sulfides. Finalists are subject to workflow-governed, tool-executed stability, redundancy, and electronic property checks. Experimental synthesis and characterization of CoV$_4$S$_8$ validate the computational pipeline, with the system's predicted synthesis protocol, structural properties (phase, morphology, composition), and N$_2$ reduction activity confirmed in laboratory tests.

• NH$_3$ yield peaks at 34.6 $\mu$g h$^{-1}$ mg$^{-1}$ at -0.55 V; Faradaic efficiency peaks at 4.3% at -0.35 V, with high selectivity and long-term operational stability.
• No hydrazine byproduct and negligible background contamination validate functional selectivity and experimental robustness.

Theoretical and Practical Implications

The research demonstrates that small, modular architectures, when carefully specialized and decoupled by cognitive mode, can match or exceed outsized general-purpose models in domain and tool-centric scientific workflows. The entropy-based design principle is empirically validated, suggesting a pathway for next-generation agents where submodels are optimized for orthogonal objectives and information fusion is achieved via structured interaction protocols.

On a practical level, the public release of the tool integration (Mat-MCP), datasets, and source code ensures extensibility and reproducibility, facilitating direct translation and adoption in the materials research community. The architecture is forward-compatible with multi-modal (e.g., microscopy, spectroscopy) data integration and closed-loop laboratory robotics, constituting a scalable platform for autonomous science.

Conclusion

MatBrain establishes a new methodological paradigm in scientific LLM deployment, wherein domain knowledge reasoning and workflow execution are functionally decoupled across lightweight, RL-optimized agents. This dual-model architecture overcomes the entropy collapse and scaling inefficiencies of monolithic LLMs, achieving state-of-the-art accuracy, efficiency, and accessibility for autonomous crystal materials research. The information-theoretic analysis, empirical benchmarking, and closed-loop experimental validation provide compelling evidence for the extensibility and practical utility of collaborative lightweight AGI architectures in scientific domains.