Search

This study proposes a novel methodology for topology optimization, employing an explainable deep neural network (XDNN). This innovative approach utilizes Shapley additive explanations (SHAPs), a tool designed to elucidate the predictive reasoning of convolutional neural networks (CNNs). The contributing regions to the characteristics are revealed by SHAP and topology optimization is performed in restricted regions to reduce the computational cost during the search while ensuring an efficient search. As a practical demonstration of the efficacy of the proposed method, it is applied to an interior permanent magnet synchronous motor. The results comprehensively demonstrate the effectiveness and potential implications of the novel methodology.

This study proposes a novel approach using convolutional neural networks (CNNs) to predict the torque characteristics of an interior permanent magnet (IPM) motor. The proposed method allows CNNs to predict the average torque and torque amplitude under various magnet configurations using the magnetic flux density distribution of the stator core. The proposed method, compared with methods trained only on the magnetic flux density distributions of the rotor core, improves the generalization performance in predicting torque characteristics. This significantly reduces the training data. In addition, the topology optimization (TO) time using the proposed method is reduced by 45% compared with the conventional method that only utilizes the finite element method (FEM).