AlloyGraph: A Multi-Agent AI Platform for Nickel-Based Superalloy Design
Designing Nickel-based superalloys for aerospace and energy applications
Nickel-based superalloys are critical to aerospace and energy applications, yet their development remains constrained by fragmented data and costly experimental iteration. We introduce AlloyGraph, an open-source platform that integrates a knowledge graph of 77 experimentally characterized superalloys, physics-informed machine learning models, and a multi-agent large language model orchestration layer for property prediction, inverse alloy design, and natural language querying. The platform employs an OWL 2 DL ontology with HermiT reasoning for automatic alloy classification and data validation, a dual-database backend coupling an RDF triplestore with a vector database for hybrid retrieval, and a sequential Analyst--Reviewer agent pipeline that triangulates ML predictions, physics-based estimates, and knowledge graph experimental data through confidence-weighted fusion. Evaluated on 88 independent alloys across three alloy classes (solid solution, precipitation hardened, and single crystal/directionally solidified), the full system achieves yield strength MAE of 80.6 MPa and UTS MAE of 95.2 MPa, outperforming both the ML-only baseline and a fine-tuned GPT-4.1-mini model on strength properties. A retrieval-augmented research assistant achieves 91% accuracy on 250 factual questions compared to approximately 50% for ungrounded LLMs, and the inverse design pipeline produces metallurgically plausible compositions meeting 72% of property targets with no critical topologically close-packed (TCP) phase stability violations. Built entirely from open-access data and open-source software, AlloyGraph demonstrates that structured domain knowledge and multi-agent reasoning can deliver competitive superalloy property prediction while lowering barriers to reproducible, AI-guided materials design.
This result is a collaboration between researchers from AIRi@UTCN (Alexandru Lecu, Adrian Groza), Digital Science & Research Solutions Ltd, London, UK (Lezan Hawizy) and University of Portsmouth, Portsmouth, UK (Soran Birosca).
