Abstract
<title>Abstract</title> <p>Additive manufacturing (AM) makes structural-alloy processing increasingly programmable, but exploiting the resulting composition-processing design space requires adaptive reasoning over properties, printability, thermodynamics, incomplete knowledge and evolving objectives. Current workflows still rely on experts to formulate criteria, select physics-based tools, debug codes and revise decisions. Here we introduce AlloyGen, a physics-grounded, self-adaptive multi-agent framework for AM alloy design. AlloyGen orchestrates retrieval-augmented knowledge, executable CALPHAD thermodynamic simulations, agent critique, and dynamic team organization. We evaluate three levels of adaptivity using prototype tasks. An expert CALPHAD agent demonstrates domain-specific tool learning by solving single-equilibrium and Scheil tasks through execution feedback and self-correction. A division-of-labor team demonstrates workflow formation by converting an underspecified Al-Si printability prompt into criteria, simulations and traceable decisions. A team-builder agent demonstrates organizational self-adaptation by assembling task-specific expertise and tools for refractory-alloy research synthesis. These results establish AlloyGen as a computational strategy for adaptive, physics-informed orchestration of alloy-design workflows.</p>