Agentic AI for accelerated fuel qualification framework: Demonstrating workflows with the AERIS enthalpy reasoning model

Abstract

Nuclear fuel qualification remains a fundamental challenge for new fuel deployment. The Accelerated Fuel Qualification (AFQ) framework proposes to address this by shifting from serial to parallel testing, leveraging advanced separate-effects experiments and multiscale modeling, yet a persistent obstacle remains: the orchestration problem — connecting atomistic, mesoscale, and engineering-scale in an efficient way. We propose an Agentic AFQ architecture in which domain-specific models and nuclear-materials datasets are exposed as reusable reasoning modules through Model Context Protocol (MCP) servers, so that any validated dataset or surrogate model becomes an interoperable node in a broader AFQ ecosystem. To demonstrate this architecture on concrete fuel-materials problems, we deploy AERIS Enthalpy — a ML model trained on DFT formation enthalpies and exposed as an MCP server — where ΔHf prediction directly enables nuclear fuel materials tasks as screening of phase stability, phase precipitation, and fuel–cladding chemical interaction. Two independent agents reproduce published DFT studies: one recovers the U–Si convex hull including U3Si2; the second identifies Mo and W as the most stable UN claddings and V, Ta as the least, consistent with nitride-layer formation tendencies reported experimentally. These results establish that agentic workflows provide a viable orchestration layer for AFQ, and that the same architecture offers a structured and reproducible pathway toward support future qualification roadmap.