Medium-Frequency Transformer Design Optimization Using FEA-Supervised Machine Learning

This project focuses on the rapid design and prototyping of single-phase transformers for magnets power converters of the FCC (Future Circular Collider). The approach combines analytical modeling, multi-frequency FEA simulations, thermal and mechanical models, and supervised ANN-based surrogate models for an effective multi-objective optimization process (targeting accurate leakage and magnetizing inductances, AC losses, and thermal constraints). Validation was performed on transformers rated between 500 kVA and 3 MVA, operating from 50 Hz to 1200 Hz. The methodology achieves up to a ×100 reduction in computation time compared to classic FEA approaches, while maintaining sufficient accuracy for rapid prototyping and industrial design. Integrating neural-network-based surrogate models into non-linear optimization (NLO) loops resulted in more robust and efficient gradient descent and demonstrated more consistent convergence to optimal solutions.