Autonomous Analysis of Thermoelectric Materials Big Data by Multi-Agentic AI and Construction of a Knowledge Base and Specialized AI

Yusuke Hashimoto, Takaaki TomaiarXiv 2026 · arXiv

TL;DR

Researchers have developed a Multi-Agentic AI (MAA) system that autonomously analyzes massive experimental datasets to discover new thermoelectric materials. By organizing AI agents into a hierarchical "corporate" structure—complete with a President, Boss, and specialized Workers—the system builds a searchable knowledge base where every scientific insight is traceable back to its raw data source, enabling a new level of auditability in AI-driven research.

What Problem Does This Paper Address?

Materials Informatics (MI) has traditionally excelled at handling computational data (like DFT calculations), but experimental data is a different beast. It is often messy, scattered across thousands of papers, and requires complex, multi-step reasoning to interpret.

While Large Language Models (LLMs) are great at summarizing text, they struggle with the rigorous, iterative nature of scientific data analysis and are prone to "hallucinations." The authors tackle the challenge of creating an autonomous system that can perform high-level data extraction, machine-learning modeling, and knowledge synthesis on experimental "big data" while ensuring that every conclusion can be verified by a human researcher.

Methods

The core of the study is the Multi-Agentic AI (MAA), a system of nine specialized AI agents (implemented using Anthropic’s Claude models) working in a three-layer hierarchy:

  1. Strategic Layer (President): Selects research themes and initiates the analysis cycles.
  2. Management Layer (Boss): Plans the analysis, allocates tasks to workers, and performs quality checks (QC) at every phase.
  3. Worker Layer: Six specialized agents handling Data (extraction/cleaning), Physics (constraint validation), ML (modeling), Implementation (prediction scripts), Review (paper scoring), and Scribe (documentation).
The MAA architecture and its five-phase workflow, including quality checks (QC) at each transition.

The system operates in cycles, where knowledge gained in one round (e.g., feature importance or material relations) is passed to the next. This allows the AI to "deepen" its understanding over time. Finally, the authors built a Specialized AI using Retrieval-Augmented Generation (RAG) that searches the accumulated records to answer specific materials design questions.

Datasets

  • Starrydata2 (SD2): A massive public database of experimental materials data. The authors used a snapshot containing over 4 million measurement points from at least 10,000 papers, specifically focusing on thermoelectric properties.
  • MAA Knowledge Base: A newly created dataset consisting of 6,393 entries (entities, relations, and findings) generated by the MAA during its autonomous runs.

Key Insights and Ideas

  • Traceability is the "Killer Feature": The authors argue that in science, an AI's answer is only as good as its citations. By forcing the MAA to save detailed analysis records, they could perform "retrospective checks." They found that while the AI still made errors (like unit mismatches), these errors were findable because the system documented its entire reasoning path.
  • Emergent Research Strategy: In later cycles of the experiment, the "Boss" agent autonomously decided to change the prediction target from the Seebeck coefficient to the figure of merit ($zT$) to better address the research theme, despite not being explicitly told to do so by the "President."
  • Knowledge Accumulation: Knowledge isn't just about more data; it's about connections. The authors used network analysis to show how the AI's "vocabulary" and understanding of relations grew over 10 cycles. Interestingly, about 10% of the relations in the final knowledge graph were only discovered in the later stages of the experiment.
Knowledge graphs showing how entities and relations accumulate from a single analysis (a) to a complex network of 100 analyses (d).

Key Results

  • Scale of Analysis: The system successfully completed 100 autonomous analyses across 10 different thermoelectric themes (e.g., BiTe-based materials, Oxides, Zintl phases).
  • Model Performance: The AI-developed machine learning models achieved $R^2$ values typically between 0.60 and 0.90. However, the authors noted that $R^2$ alone is a poor metric for "knowledge," as it can be inflated by data leakage—something their "Physics" agent was specifically tasked to watch for.
  • Specialized AI Utility: The RAG-based specialized AI was able to propose specific material design guidelines (e.g., for $Mg_3Sb_2$) by synthesizing results from multiple previous analysis cycles, providing a "one-stop shop" for researchers.
The three-layer design of the specialized AI, which uses an index to find relevant analyses and then reads the full records to generate an answer.

Limitations and Future Work

  • Hallucinations Persist: Even with a multi-agent setup, LLMs can still produce incorrect outputs. The authors emphasize that the system makes errors traceable, not absent.
  • Physical Validity: While the AI can find statistical correlations, it doesn't "understand" physics in the human sense. The design guidelines it proposes still require experimental validation by human scientists.
  • Feature Engineering: The current system relies heavily on composition-based features. Future versions could integrate 3D crystal structure data from first-principles calculations to improve predictive power.
  • Error Propagation: If an error occurs early in a cycle (like a unit conversion mistake), it can spread through the knowledge base. Developing more robust "integrator" agents to check for cross-theme consistency is a key next step.

Summary updated

Abstract

Extracting knowledge from large databases of experimental materials data requires repeated analyses and reusable records of their choices and results. We developed Multi-Agentic AI (MAA), a multi-agent system in which AI agents with distinct roles analyze experimental thermoelectric data under predefined instructions and quality checks, saving analysis records traceable to their sources. In four experiments varying theme and cycle counts, we measured what accumulated in these records. In the largest experiment (100 analyses across 10 themes and 10 cycles), concept co-occurrences were concentrated within themes, vocabulary was shared across themes and some relation descriptions were first detected in later cycles. A specialized AI that searches this experiment's knowledge bases and reads their analysis records answered materials questions with cited sources. Retrospective checks, including tracing an answer generated with a preliminary corpus back to the source data, exposed errors: traceability makes errors findable but not absent.
BibTeX
@misc{hashimoto2026autonomous,
  author = {Yusuke Hashimoto and Takaaki Tomai},
  title = {{Autonomous Analysis of Thermoelectric Materials Big Data by Multi-Agentic AI and Construction of a Knowledge Base and Specialized AI}},
  year = {2026},
  eprint = {2611.00627},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.00627},
}
Figure