System Overview
Design → Generate → Evaluate, closed by a GA loop over rotor images
Multi-Agent based Non-parametric Motor Design Optimization
Three LLM agents close an evolutionary loop over rotor images instead of geometry parameters.
3.24×hypervolume vs baseline
−25.3 %iron loss (torque within 10 %)
300FEA solves · 5 generations
PIPELINEdesign → generate → evaluate ↻ next generation
AGENT 01
Design
Problem-card agent
Chats with you and turns a reference image, the rotor/stator envelope and the objectives into a Design Problem Card.
image + envelope + objectives → problem card
Open
AGENT 02
Generation
Explore · exploit · filter
Breeds a generation from Pareto parents: zero-shot topology moves and fine-tuned text edits, then a feasibility gate and embedding-based filtering.
parents → 200 images → gate → N11 + FPS → 60
Open
AGENT 03
Evaluation
CAD → FEA → ε-dominance
Converts images to STEP, solves Maxwell 2D, judges ε-dominance vs the parent, writes improvement directions and updates the operator scoreboard.
RGB → STEP → FEA → ε vs parent → scoreboard
Open
GA loop · one generationclick to play Design Optimization
ParentsPareto front of the FEA archive · 4–8
→
Explorationzero-shot T2I · topology move
+
Exploitationfine-tuned editor · library + transfer
→
Gateinfeasible magnets dropped
→
FilterN11 embedding · FPS · 60 kept
→
FEAimage → STEP → Maxwell 2D
→
ε-dominancevs parent · 5 % iron / 1 % torque
→
ScoreboardBeta posterior · Thompson shares
↻improved designs → directions → next generation's parents and shares








