Preliminary Study on RL-Based IPMSM Rotor Design for Torque Enhancement

Takahiro Maezono, Mitsuru Endo, Yukio Tsutsui

The Sixth International Conference of IFToMM ITALY IFIT2026

pp. 99-107

• 論文本文 / Full paper (Springer Nature Link)

---

背景 – Background

モータ設計では，寸法や動作仕様などの要求が変わるたびに設計の見直しが必要となる．本研究では，設計経験を学習し，要求の変更に適応するモータ設計を目指して，強化学習による回転子形状設計を検討する．その第一段階として，固定した設計条件の下で，埋込磁石同期モータ（IPMSM）のトルクを改善する形状変更を学習できるかを調べる．

Motor design often requires redesign when dimensions or operating specifications change. This study investigates reinforcement learning for rotor-shape design as a step toward motor design that adapts to changing requirements through learned design experience. As an initial step, it examines whether an agent can learn torque-improving shape modifications for an interior permanent magnet synchronous motor (IPMSM) under a fixed requirement.

Fig. 1 Overview of the IEEJ D-model used as the base motor model

---

手法 – Method

IEEJ Dモデルを基に，永久磁石・固定子・巻線の形状を固定し，回転子鉄心を設計対象とする．初期形状には空気領域を設けず，エージェントが円形の空気領域の位置と大きさを選び，d軸に対して対称に順次追加する．各形状の平均トルクを2次元有限要素解析で評価し，初期形状からのトルク増加を報酬とする．連続的な設計操作を扱うため，学習にはSoft Actor-Critic（SAC）を用いる．

Based on the IEEJ D-model, the rotor core is designed while the permanent magnets, stator, and coils remain fixed. Starting from a rotor core without air regions, the agent sequentially inserts circular air regions by selecting their positions and sizes, with each modification applied symmetrically about the d-axis. Two-dimensional finite element analysis evaluates the average torque, and the increase relative to the initial geometry defines the reward. Soft Actor-Critic (SAC) is used to learn these continuous design actions.

Fig. 2 Rotor-core design area for sequential insertion of air regions

---

結果 – Result

強化学習により，初期形状から平均トルクを改善する回転子形状が得られ，逐次的な形状変更を学習する設計手法の実現可能性が示された．一方で，学習中に得られた高トルクの形状を，最終的な方策で安定して再現する点には課題が残る．

The learned design process produced a rotor geometry with improved average torque relative to the initial geometry, demonstrating the feasibility of the sequential RL-based design formulation. Reliably reproducing the high-torque designs found during training with the final policy remains a challenge.

---

結論 – Conclusion

固定条件下でのIPMSM回転子設計を，強化学習による逐次的な形状変更問題として定式化した．今後は，形状表現・報酬設計・学習の安定性を改善し，設計要求の変更に適応するメタ強化学習への展開を目指す．

This study formulated IPMSM rotor design under a fixed requirement as a sequential shape-modification problem using reinforcement learning. Future work will improve the shape representation, reward design, and learning stability, and investigate meta-RL extensions for adaptation to changing design requirements.

---

The content on this page is not freely available for reproduction or redistribution. Unauthorized use may lead to legal consequences.