[論文レビュー] From Natural Language to Materials Discovery:The Materials Knowledge Navigation Agent

Genmao Zhuang, Amir Barati Farimani|arXiv (Cornell University)|Feb 11, 2026

Machine Learning in Materials Science被引用数 0

ひとこと要約

MKNAは自然言語の科学的意図を実行可能な手順へ変換する言語主導の自律システムで、データ取得・性質予測・構造生成・安定性検証を通じて高Debye温度セラミックの発見を行う。

ABSTRACT

Accelerating the discovery of high-performance materials remains a central challenge across energy, electronics, and aerospace technologies, where traditional workflows depend heavily on expert intuition and computationally expensive simulations. Here we introduce the Materials Knowledge Navigation Agent (MKNA), a language-driven system that translates natural-language scientific intent into executable actions for database retrieval, property prediction, structure generation, and stability evaluation. Beyond automating tool invocation, MKNA autonomously extracts quantitative thresholds and chemically meaningful design motifs from literature and database evidence, enabling data-grounded hypothesis formation. Applied to the search for high-Debye-temperature ceramics, the agent identifies a literature-supported screening criterion (Theta_D > 800 K), rediscovers canonical ultra-stiff materials such as diamond, SiC, SiN, and BeO, and proposes thermodynamically stable, previously unreported Be-C-rich compounds that populate the sparsely explored 1500-1700 K regime. These results demonstrate that MKNA not only finds stable candidates but also reconstructs interpretable design heuristics, establishing a generalizable platform for autonomous, language-guided materials exploration.

研究の動機と目的

• Translate open-ended scientific queries into a multi-stage discovery workflow.
• Ground qualitative intent with literature-derived quantitative criteria.
• autonomously construct datasets and perform structure modification for exploration.
• Predict properties with ML surrogates and validate candidates with physics-based models.
• Identify thermodynamically stable, high–Debye-temperature material candidates.

提案手法

• Stage I grounds queries via literature mining and a Map–Reduce extraction to derive quantitative screening thresholds.
• Stage II autonomously retrieves properties with LLM-generated code and trains surrogates (e.g., CGCNN) for target properties.
• Stage III proposes modified structures and validates stability with M3GNet, using energy-above-hull as the stability criterion.
• Structure modification uses group-wise, valence-preserving substitutions with Gaussian perturbations to generate thousands of candidates.
• Debye temperature is estimated from elasticity data when not directly available, enabling data-driven screening and training.
• “Theta_D>800 K” is established as the screening threshold and used to prioritize high–stiffness materials.

Figure 1 : Overview of the proposed agentic workflow for materials discovery.

実験結果

リサーチクエスチョン

• RQ1 Can an autonomous language-driven agent infer quantitative design thresholds from literature for materials discovery?
• RQ2 Can MKNA integrate retrieval, prediction, and validation to identify thermodynamically stable, high–Debye-temperature candidates?
• RQ3 How well can LLM-generated code retrieve missing properties and support end-to-end screening?
• RQ4 What design motifs (e.g., Be–C–rich frameworks) emerge when exploring high–Theta_D materials?

主な発見

• MKNA reconstructs interpretable design heuristics and identifies a screened criterion of Theta_D>800 K for high Debye temperature.
• Canonical ultra-stiff materials (diamond, SiC, SiN, BeO) are rediscov­ered through literature grounding and screening.
• Modified Be–C–rich frameworks emerge as thermodynamically stable, high–Theta_D candidates in the 1500–1700 K range.
• CGCNN predictor achieves RMSE ~247 K and R^2 ~0.68, enabling efficient pre-screening before M3GNet validation.
• Stability filtering (E_hull<0.05 eV/atom) concentrates candidates in the high–Theta_D regime compared with database-only distributions.
• MKNA demonstrates the emergence of Be–C motifs and potential new materials while providing transparent, literature-grounded reasoning.

Figure 2 : Workflow for LLM-generated property retrieval. If a property is not available in a database, GPT-5-mini synthesizes and repairs a custom retrieval routine until valid outputs are obtained.