Topological Data Analysis for Image-based Machine Learning: Application to Electric Motors

Many finite-element simulations are required to fully evaluate the performance of a motor design candidate at different operating points. In this work, we investigate deep learning based surrogate modeling technique for motor design optimization to reduce simulations required. In particular, we introduce topological data analysis to electric machine design, which extracts topological features from motor design images for the training of machine learning models. We introduce the process of computing persistence homology and Betti sequences, which serve as vectorized input data for machine learning models. We propose two-channel deep learning models, with one convolutional network branch built for motor image data, and another multi-layer perceptron branch for Betti sequences. We show with numerical tests that two-channel models perform better in prediction accuracy and generalization capability compared with models without topological feature input. The results show that the proposed strategy is effective for image- based deep learning problems.