MKNA: Language-Driven Agent for Autonomous Materials Discovery

Paper: 2602.11123 Authors: Genmao Zhuang, Amir Barati Farimani Categories: cs.LG, cond-mat.mtrl-sci

Abstract

The Materials Knowledge Navigation Agent (MKNA) represents a paradigm shift in computational materials discovery by bridging natural language understanding with materials science workflows. Unlike traditional approaches that rely on expert intuition and expensive simulations, MKNA translates scientific intent expressed in plain language into executable actions spanning database queries, property predictions, structure generation, and thermodynamic stability assessments. The system autonomously extracts quantitative screening criteria and chemical design principles from scientific literature and databases, enabling hypothesis-driven exploration grounded in existing knowledge. In a case study targeting high-Debye-temperature ceramics, MKNA identifies $\Theta_D > 800$ K as a literature-supported threshold, successfully rediscovers established ultra-stiff materials (diamond, SiC, SiN, BeO), and proposes novel thermodynamically stable Be-C-rich compounds occupying the underexplored 1500-1700 K regime.

Key Contributions

• Language-to-Action Translation: Converts natural language scientific queries into structured computational workflows without manual programming
• Autonomous Knowledge Extraction: Mines quantitative thresholds and design motifs directly from literature and materials databases
• Multi-Tool Integration: Orchestrates database retrieval, ML-based property prediction, crystal structure generation, and DFT-based stability evaluation
• Interpretable Discovery: Not only identifies candidate materials but reconstructs the reasoning path and design heuristics behind selections
• Validated Performance: Rediscovers canonical high-stiffness materials and proposes novel stable compounds in sparsely explored chemical space

System Architecture and Methodology

MKNA operates through a modular agent framework that decomposes materials discovery into interpretable subtasks. The system accepts natural language prompts describing target properties or application requirements, then autonomously plans and executes a multi-step workflow.

The core architecture integrates four primary capabilities:

Database Navigation: The agent queries crystallographic databases (Materials Project, OQMD) using chemically informed search strategies. Rather than exhaustive enumeration, MKNA extracts compositional constraints and structural motifs from literature context to focus searches on promising chemical spaces.

Property Prediction: For rapid screening, the system employs pre-trained graph neural networks and composition-based ML models to estimate properties like Debye temperature ($\Theta_D$), bulk modulus, and formation energy. This enables filtering of thousands of candidates before expensive DFT calculations.

Structure Generation: When exploring novel compositions, MKNA interfaces with crystal structure prediction tools to generate candidate atomic arrangements. The agent evaluates multiple polymorphs and selects structures based on symmetry considerations and chemical plausibility.

Stability Verification: Final candidates undergo DFT-based thermodynamic analysis to compute formation energies and assess stability against decomposition. The agent constructs convex hulls to determine phase stability and identifies metastable phases within specified energy thresholds.

Critically, MKNA maintains an internal knowledge graph that accumulates evidence from literature mining and database queries. When tasked with finding high-$\Theta_D$ ceramics, the agent autonomously identifies that Debye temperature correlates with elastic stiffness, extracts the $\Theta_D > 800$ K criterion from materials science literature, and uses this to guide subsequent searches.

Results and Discovery Pipeline

The paper demonstrates MKNA’s capabilities through a concrete case study: discovering ceramics with exceptionally high Debye temperatures, which correlate with thermal conductivity and mechanical stiffness—critical for aerospace and electronics applications.

Validation Phase: MKNA first validates its approach by querying existing databases for known high-$\Theta_D$ materials. The agent successfully identifies diamond ($\Theta_D \approx 2230$ K), silicon carbide ($\Theta_D \approx 1200$ K), silicon nitride, and beryllium oxide—all established ultra-stiff ceramics. This demonstrates the system’s ability to navigate chemical space using physically meaningful criteria.

Novel Discovery Phase: Moving beyond known materials, MKNA explores underrepresented compositions in the Be-C-N-O chemical space. The agent proposes several previously unreported compounds with predicted Debye temperatures in the 1500-1700 K range—significantly higher than most engineering ceramics but below the extreme values of diamond.

Stability Analysis: For each candidate, MKNA performs DFT calculations to verify thermodynamic stability. The proposed Be-C-rich phases show formation energies within 50 meV/atom of the convex hull, indicating they are either stable or metastable phases potentially accessible through synthesis. The agent automatically flags compositions that decompose into competing phases, filtering out thermodynamically unfavorable candidates.

Design Heuristic Reconstruction: Beyond listing materials, MKNA articulates the underlying design logic: high Debye temperature requires light constituent atoms (low mass) and strong bonding (high elastic moduli). The agent identifies that Be and C satisfy both criteria, explaining why Be-C compounds populate the high-$\Theta_D$ regime. This interpretability distinguishes MKNA from black-box optimization approaches.

Implications for Materials Science

MKNA addresses a fundamental bottleneck in computational materials discovery: the gap between domain expertise encoded in natural language (papers, textbooks, expert knowledge) and the formal computational workflows required for materials screening. Traditional approaches require researchers to manually translate scientific intuition into database queries, ML model inputs, and simulation parameters—a process that is time-consuming, error-prone, and difficult to reproduce.

By enabling language-driven exploration, MKNA democratizes access to computational materials tools. Researchers can express high-level scientific goals (“find thermally conductive ceramics for hypersonic applications”) and let the agent handle the technical implementation details. This is particularly valuable for interdisciplinary teams where domain experts may lack computational expertise.

The system’s ability to extract and apply quantitative criteria from literature represents a form of automated knowledge synthesis. Rather than treating each discovery task as independent, MKNA accumulates design principles that transfer across problems. A threshold learned for Debye temperature screening could inform future searches for related properties like thermal conductivity or elastic moduli.

The interpretability of MKNA’s reasoning is crucial for scientific acceptance. Materials scientists need to understand why a particular compound was selected, not just receive a ranked list of candidates. By reconstructing design heuristics and citing literature evidence, the agent produces discoveries that integrate naturally into human scientific workflows.

Limitations and Future Directions

While promising, MKNA faces several challenges. The system’s performance depends heavily on the quality and coverage of underlying databases and ML models. For truly novel chemical spaces with limited training data, property predictions may be unreliable. The agent currently lacks mechanisms to quantify prediction uncertainty or actively request additional data when confidence is low.

The language understanding component, while effective for materials science queries, may struggle with ambiguous or underspecified prompts. Developing more robust natural language interfaces that can engage in clarifying dialogue with users would enhance usability.

Computational cost remains a consideration. While MKNA reduces human effort, the DFT calculations required for stability verification still demand significant compute resources. Integrating more efficient surrogate models or active learning strategies to minimize expensive simulations would improve scalability.

Future extensions could incorporate experimental feedback loops, where synthesis attempts and characterization data refine the agent’s search strategies. Connecting MKNA to autonomous laboratories would enable closed-loop materials discovery from prediction through validation.

Takeaways

1. MKNA demonstrates that natural language can serve as an effective interface for complex computational materials workflows, eliminating the need for manual programming of search strategies.

2. Autonomous extraction of quantitative design criteria from literature enables data-grounded hypothesis formation, moving beyond purely data-driven or purely physics-based approaches.

3. The system successfully balances exploration (discovering novel Be-C compounds) with validation (rediscovering canonical materials), demonstrating both novelty and reliability.

4. Interpretable reasoning paths and reconstructed design heuristics make agent-driven discoveries scientifically credible and actionable for experimental validation.

5. The modular architecture generalizes beyond Debye temperature searches, establishing a platform for language-guided exploration across diverse materials properties and applications.