AI-assisted design optimisation of permanent magnet synchronous machine for e-bike

Abstract

This thesis investigates a data-efficient multi-objective optimization workflow for the electromagnetic design of a permanent magnet synchronous motor (PMSM) intended for a mid-drive electric bicycle (e-bike) powertrain. In this application, the motor must meet competing requirements in compactness, efficiency, and torque quality, since torque ripple directly impacts pedaling smoothness, noise/vibration, and controllability. High-fidelity assessment of these trade-offs generally requires finite-element modeling (FEM), yet FEM-driven design exploration is computationally expensive and often limits the number of candidate geometries that can be evaluated. To address this challenge, a modular simulation–optimization framework is developed to couple MATLAB with ANSYS Maxwell and automate parametric FEM evaluations. The considered motor is a 48 V interior PMSM featuring a 48-slot/8-pole configuration and a Δ-shaped buried-magnet topology. The design is parameterized by key stator-slot dimensions and magnet geometric variables, and the optimization is formulated with three objectives: maximize average torque, maximize efficiency, and minimize torque ripple. The proposed approach combines Bayesian optimization (BO) with a Gaussian-process surrogate model and augments the BO loop with a retrieval-augmented generation (RAG) large language model (LLM) acting as a memory-based design agent. Using an internal database of prior FEM results and trend summaries, the LLM proposes candidate designs and incorporates qualitative natural-language rules to steer exploration without repeated manual retuning of objective weights. Comparative studies against a reference multi-objective metaheuristic demonstrate that the LLM guided BO strategy reaches a competitive Pareto region while reducing the number of expensive FEM evaluations. The results support the use of RAG-LLM assistance to improve sample efficiency and interpretability in FEM-based multi-objective PMSM geometry optimization for light electric mobility.

Date	2 Apr 2026
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Qingsong Wang (Supervisor)