SCGAN-enhanced surrogate modeling for FEM analysis of induction motors: A practical use case

Surrogate modeling is a powerful approach for industrial applications that provides data-driven approximations of complex physical or numerical simulators in exchange for minimal computational cost. In this context, Multi-fidelity frameworks enable the fusion of distinct data sources and varying levels of resolution. Despite their advantages, the application of these techniques in industry is still underexplored, particularly for induction motors. The present work aims to bridge this exact gap, considering the modeling of a radial flux, inner rotor, squirrel cage induction motor. By leveraging the data extracted using an analytical formulation and a FEM model, a Stacking-based Conditional GAN (SCGAN) has been implemented. This architecture has proven capable of dynamically generating back electro-motive force (BEMF), line voltage (on load) (VLL) and torque waveforms within a single conditional model. Compared to standalone surrogates, SCGAN has reduced the Normalized Root Mean Square Error by 25% for BEMF, 29% for VLL, and 31% for torque. The study, supported by a systematic comparison of different low-fidelity information extraction steps and a detailed feature relevance analysis, represents a step forward in the field of surrogate modeling for induction motor characterization. It also clearly demonstrates the potential of multi-fidelity learning to overcome the computational-cost barriers of complex engineering systems; notably, SCGAN's prediction step executes orders of magnitude faster than traditional FEM, paving the way for real-time deployment. • First application of multi-fidelity surrogate modeling to induction motors. • Usefulness of fusing different fidelity level information to reconstruct Back-EMF, line-to-line voltage and torque waveforms. • SHAP-driven feature relevance analysis to identify the most informative low-fidelity descriptors. • Significant reduction in computational cost by comparing FEM model and SCGAN approach.