Materealize: a multi-agent deliberation system for end-to-end material design and synthesis

Technical Summary: Materealize

Problem Statement
Despite significant advancements in computational chemistry and machine learning (ML) for materials discovery, a critical gap remains between computational design and practical experimental realization. Current computational tools and ML models are fragmented, task-specific, and lack a unified framework that seamlessly connects structure generation, property evaluation, synthesizability assessment, and synthesis planning. This fragmentation necessitates substantial domain expertise to navigate, limiting accessibility for experimental chemists and materials scientists who are not proficient in computational software. Furthermore, existing agentic AI systems often rely on a passive "tool-calling" paradigm where the final output is strictly bounded by the accuracy of predefined tools, lacking the capacity for higher-level scientific reasoning, cross-checking, or mechanistic insight.

Methodology
The authors propose Materealize, a unified multi-agent system designed for end-to-end inorganic materials design and synthesis planning. The system integrates a suite of specialized domain tools within a natural-language interface and operates in two complementary modes:

1. Instant Mode: Designed for rapid task execution (1–2 minutes), this mode orchestrates a pipeline of connected tools to solve diverse inorganic tasks. It utilizes:

   • Generative Models: ChargeDIFF and Chemeleon for unconditional and property-conditioned (e.g., bandgap, magnetic density) crystal structure generation.
   • Synthesizability Predictors: PU-CGCNN (crystal-graph neural network) and SynCry (LLM-based on CIF-strings) to filter candidates.
   • Property Predictors: ALIGNN for property estimation.
   • Synthesis Planners: ElemwiseRetro for predicting precursors and heating temperatures.
   • Diagnostic/Redesign Tools: StructGPT for explaining synthesis difficulties and SynCry for redesigning unsynthesizable structures.
   • Data Utilities: Tools for handling CIF files, formula parsing, and descriptive text generation.

2. Thinking Mode: Designed for deeper, reasoning-driven analysis (~20 minutes), this mode employs a Multi-Agent Debate (MAD) framework to refine synthesis routes and generate mechanistic hypotheses. It features five specialized agents:

   • Precursor Agent: Focuses on precursor selection and accessibility, grounded in tools like ElemwiseRetro.
   • Thermodynamics Agent: Evaluates bulk thermodynamics, phase stability, and formation energy using ML interatomic potentials (MACE bulk, MatterSim bulk).
   • Surface & Kinetics Agent: Assesses surface energetics, diffusion, and reaction barriers using tools like MACE slab and NEB calculations.
   • Literature Agent: Grounds the debate in existing knowledge via web search and literature retrieval for fact-checking and mechanistic comparison.
   • Judge Agent: Aggregates arguments from the four domain agents over multiple rounds (up to five) to produce a final, structured synthesis report including precursors, stepwise procedures, and mechanistic narratives.

Key Contributions

• Unified Framework: Materealize is the first system to integrate structure generation, synthesizability assessment, and synthesis planning (including mechanistic insight) into a single agentic framework for inorganic materials.
• Dual-Mode Operation: It offers a trade-off between speed (Instant mode) and depth of reasoning (Thinking mode), catering to both rapid screening and complex, high-stakes experimental planning.
• Mechanistic Insight: Unlike previous agents that merely output recipes, Materealize's Thinking mode generates and validates mechanistic hypotheses by comparing them against literature and physics-grounded simulations.
• Extensibility: The system is designed to be modular, allowing for the seamless integration of new tools and expansion of functionality.

Results

• Task Capabilities: The system successfully demonstrated four key tasks: (1) conditional end-to-end design of synthesizable candidates with recipes; (2) diagnosis and repair of unsynthesizable structures; (3) generation of detailed synthesis procedures and mechanisms; and (4) unconditional data augmentation of synthesizable materials.
• Performance Benchmarks:
  • Structure Generation: Materealize produced a significantly higher fraction of valid structures (chemically valid formulas, correct oxidation states, thermodynamic stability) compared to tool-free LLM baselines (gpt-4o, gemini-2.5-flash), which frequently violated physical constraints.
  • Synthesis Prediction: In precursor prediction, Materealize (Thinking mode) achieved top-3, top-4, and top-5 accuracies of 86.2%, 90.2%, and 91.4%, respectively, outperforming both tool-free agents and single-agent tool-connected baselines.
  • Synthesis Reports: Evaluated against 20 experimentally elucidated case studies, Materealize achieved the highest overall score (0.484) among all tested systems, outperforming tool-free reasoning and single-agent tool-connected reasoning. Ablation studies confirmed that removing any specific agent role or reducing debate rounds degraded performance, validating the necessity of multi-perspective debate.
• Physics-Grounded Validation: For a novel target (Mg2TeSe), the synthesis conditions suggested by Materealize were validated via MACE-based molecular dynamics (MLP–MD) simulations. The simulated annealing process reproduced the dominant Bragg features of the target phase, confirming the chemical plausibility of the proposed precursor-to-product transformation.
• Design Space Expansion: Materealize expanded the "synthesis-realizable design space" by 8.3% compared to manually constructed fixed pipelines, identifying valid synthesis routes for materials (e.g., BaCaBO3F, MgV2O4) where tool-only execution failed.

Significance
The paper claims that Materealize represents a significant step toward autonomous, synthesis-aware materials design. By combining tool-level accuracy with reasoning-level integration, it bridges the gap between computational discovery and practical experimental realization. The system lowers the barrier for non-experts to access complex computational workflows while providing expert-level, chemically grounded guidance for experimentalists. The authors emphasize that while the system improves reliability through cross-checking and debate, it remains dependent on the coverage and accuracy of its underlying tools, noting that future progress requires unbiased data collection including negative outcomes, potentially via self-driving laboratories. The authors have released a web-based demo and a dataset of 100 synthesizable candidates with detailed guidelines to facilitate practical adoption.