Overcoming Combinatorial Explosion in Alloy Design via Hierarchical Multi-Agent Systems

Keywords: Hierarchical Multi-Agent System (MAS), Autonomous Materials Discovery, Alloy Design Optimization, Surrogate Modeling

Abstract: Traditional AI-driven materials discovery pipelines employ a monolithic architecture where a single surrogate model is trained, scalarized, and deployed statically, creating a brittle interface with physical experimentation. We present a hierarchical multi-agent system (MAS) that fundamentally redesigns this architecture through three innovative mechanisms: (1) furnace-to-agent feedback loops enabling continuous online learning, (2) a curiosity-annealing scheduler for adaptive exploration-exploitation balance, and (3) memory-injected composition generators that leverage historical success. This architectural approach reduces required physical lab iterations by seven-fold compared to both single-agent and static multi-agent baselines. The system identified 21 novel Pareto-optimal alloys that outperform canonical benchmarks (Ti-6Al-4V, Inconel-718, Cantor HEA) while maintaining 97% metallurgical feasibility. These gains are attributable not to larger models or increased compute, but to specific architectural innovations that enable distributed, adaptive, and physics-informed optimization.

Submission Number: 238