Abstract
Achieving high-performance interior permanent magnet machines with reduced rare-earth content typically requires complex topology or shape optimization paired with costly finite element analysis simulations. Bitmap-based surrogate models accelerate the process, however, suffer from substantial data demands for training. To overcome this problem, we explored low-dimensional B'ezier curve-based magnet shaping along with surrogate modeling for a 22-MW generator. Several models including support vector regression, fully connected neural networks, and a novel two-stage surrogate that integrates a least-volume autoencoder with a feedforward neural network were evaluated. The two-stage surrogate framework is shown to achieve a normalized RMSE = 1.0 in prediction errors in several key performance metrics, including demagnetization, efficiency, power, and load torque. A Globol Sobol sensitivity analysis further identified dominant shape- and magnetic features in the design influencing machine performance.
| Original language | American English |
|---|---|
| Number of pages | 11 |
| State | Published - 2026 |
| Event | International Conference on Electric Machines - Madeira, Portugal Duration: 6 Sep 2026 → 9 Sep 2026 |
Conference
| Conference | International Conference on Electric Machines |
|---|---|
| City | Madeira, Portugal |
| Period | 6/09/26 → 9/09/26 |
NLR Publication Number
- NLR/CP-5000-99005
Keywords
- autoencoder
- Bezier curves
- parametric shaping
- Sobol sensitivity
Fingerprint
Dive into the research topics of 'Surrogate-Based Sensitivity Analysis for Low-Dimensional Freeform Generator Designs: Preprint'. Together they form a unique fingerprint.Cite this
- APA
- Author
- BIBTEX
- Harvard
- Standard
- RIS
- Vancouver