An AI/ML-Augmented Toolkit for Materials Intelligence through Knowledge Fusion and Reasoning

Author(s):

Xiongjun Wu and B. Chad Hornbuckle

Abstract:

High-throughput materials discovery requires the assimilation of structured and unstructured data distributed across public literature, proprietary records, external databases, and internal experimental observations. Although large-language models (LLMs) offer powerful capabilities for processing technical information, their unconstrained use risks hallucinations, weak provenance, and insufficient domain rigor. To address these challenges, this technical report describes the concept, architecture, and an example implementation of an AI/ML Augmented Toolkit for materials intelligence that enables structured knowledge fusion and multistep reasoning. This toolkit combines retrieval-augmented generation, an explicit materials ontology, a typed, provenance-aware knowledge graph (KG), and federated materials database search across external repositories (e.g., OPTIMADE, JARVIS) within an orchestrated tool-calling workflow. The ontology guides the extraction and normalization of materials entities and relationships, while the KG supports relational traversal across distributed evidence. Within the tool-calling loop, the LLM proposes structured tool requests, the Orchestrator executes the selected tools, and the resulting literature passages, graph relationships, database records, or conversation context are returned to the LLM for synthesis. This architecture constrains foundation models using explicit tools, domain boundaries, and semantic extraction to ensure auditable, cross-source validation. The toolkit is demonstrated using refractory high-entropy alloys, whose complex design space requires integrated reasoning over composition, processing, phase stability, characterization, and properties. Accessed via an interactive widget, the workflow ingests literature, normalizes extracted entities, links relationships to source evidence, and synthesizes findings for evidence-based candidate prioritization and gap identification. By bridging otherwise isolated evidence sources through a unified, orchestrated interface, the toolkit provides a trustworthy, reproducible, and scalable materials-intelligence layer that supports hypothesis generation, candidate screening, and experiment planning for accelerated Army materials-discovery missions.