Toward Self-Improving MEMS Research Agents through Experience-Informed Procedure Refinement

Abstract Microelectromechanical systems (MEMS) research requires translating device concepts into geometries with physically meaningful behavior under specified structural constraints. Although research agents increasingly automate scientific workflows, successful numerical execution alone does not establish operating-mode identity and dynamic behavior, identify loss mechanisms, or verify geometric compliance. Here we introduce the MEMS Research Agent Framework, which brings agent-based research and experience-informed procedure refinement into an expert-guided MEMS workflow. Agents execute investigations using reusable instructions and scripts, while experts define objectives, assess physical interpretations, and review methodological changes. In a researcher-guided study of a silicon double-ended tuning-fork resonator, the proposed framework carries a literature-derived thermoelastic-dissipation hypothesis through geometry construction, coupled-field simulation, and operating-mode identification. The selected segmented-slot design achieves a modeled thermoelastic quality factor approximately 5.2 times the baseline at a resonant frequency 20% lower. The slotting mechanism is consistent with a published experimental study reporting a 5.9-fold increase in measured quality factor in a different geometry. Development histories trace a procedural route toward self-improvement: recorded failures inform expert-reviewed changes to research instructions and diagnostic checks reused in subsequent runs. A registered comparison across 25 runs evaluates research outcomes and evidence completeness under five frozen configurations. The final author assessment gives 1/5 successful runs without the proposed framework and 5/5 with it; the initial model-judge assessment gives 1/5 and 2/5, respectively, reflecting differences in evidence acceptance. These studies demonstrate the feasibility of delegating parts of MEMS inverse-design iterations to agents, laying a foundation for self-improving research automation in device design and discovery.