Agentic AI for Optimization of Magnetic Robots

Abstract

Magnetic milli-spinner design spans a large coupled geometry space, yet manual CAD construction, mesh preparation and repeated fluid simulations make exploration labour-intensive. Here we develop an Agentic AI workflow connecting literature-derived constraints, parametric geometry, computational fluid dynamics and persistent research memory. Three Missions progressively reconstruct a published benchmark, optimize within its parameter family and explore new spinner--ring geometries. Matched screening calculations identify a within-family design with improved equilibrium speed at both 100 and 160~Hz. Expanded exploration discovers asymmetric and threefold-symmetric speed leaders. The speed-leading ring-nose design reaches 20.04 and 36.63~cm~s$^{-1}$, exceeding the reconstructed published reference by 10.3\% and 12.1\% while also increasing its internal pressure contrast. Spatial flow diagnostics identify a distinct interaction-potential-leading design. These findings demonstrate physics-guided agentic structural discovery and identify improved computational designs with distinct propulsion and internal-flow advantages.

Keywords