This talk presents a multi-agent framework that combines domain-specialized agents with independent physics-based verification to accelerate materials discovery while maintaining scientific rigor.

Our manager-orchestrator architecture coordinates specialist agents—each augmented with domain knowledge vaults and validated skills—to answer complex research questions. When experimental data is unavailable, the system escalates to first-principles calculations; when answers are generated, an independent Correctness Agent traces citations to source data and applies drift-detection to predictions.

Demonstrated through nuclear fuel materials research coordinating TRISO fabrication, fuel performance, and DFT calculation agents, the verification framework reduces grading error from 22.5% to 4.3% and detects critical physics errors including 50× deviations in diffusion models. Across five foundation models, the system improves answer quality by 1.2×-1.5× relative to cost. This work provides a generalizable blueprint for trustworthy multi-agent scientific systems.

Link to my session: https://conferences.iaea.org/event/454/contributions/43609/

Link to the slides: AIforFusion.pdf