Compressing the Validation Bottleneck: An Agentic Self-Driving Lab for Metal Additive Manufacturing

Kyunghoon Hur, Chihun lee

Track: Track 3: AI Scientist Proposal Competition

Keywords: Self-Driving Labs, Agentic AI, AI for Science, Bayesian Optimization, Active Learning, Multi-fidelity Transfer, Additive Manufacturing

TL;DR: A single agent compresses the validation bottleneck of self-driving labs along two cost axes: trial count via prior-aware DoE, and per-trial cost via cross-domain transfer from cheap SLA polymer printing to expensive metal additive manufacturing.

Abstract: Agentic AI-for-Science systems have walked the upstream research pipeline end-to-end — ideation, planning, experimental design, and feedback loops are no longer the long pole. The bottleneck has slid downstream onto the bench: each candidate must still be made and measured, and self-driving labs (SDLs) amortise human labour but not the underlying material, machine-time, and characterisation cost of one trial. We compress this bottleneck along two orthogonal cost axes with a single agent. (i) A prior-aware agent design-of-experiments (DOE) loop ingests physical context and trial history to cut trials-to-target below grid, random, and vanilla Bayesian optimisation; both prior and outcome are routed through a verifier and surrogate so feedback measurably affects selection. (ii) Cross-domain augmentation supplements scarce target-domain runs with a structurally aligned cheap surrogate; domain choice is itself an action of the same agent. We instantiate the framework for metal additive manufacturing with SLA resin printing as the surrogate, at orders-of-magnitude lower cost per trial.

Submission Number: 332