Design Optimization of Magnetic Material Distribution by Using Encoder-Decoder with Additive Mixing for Design Conditions

Recently, the deep learning technology attracts much attention in various industrial fields. In our previous research, we developed an Encoder-Decoder precisely reproducing the optimization process of conventional optimization method, that is, the level-set method which is one of the gradient methods, by means of Convolutional Neural Network (CNN) and Long Short-term Memory (LSTM). The developed method enables us to implement high speed search of solutions, which means the possibility of better and effective optimization starting with various initial shapes. This method can deal with only the initial shape as design parameter for optimization. Thus, it is necessary to re-train the Encoder-Decoder when the design conditions change, e.g., the displacement of permanent magnet. To overcome this drawback, we have developed a novel network structure to incorporate the design conditions into the training data. Finally, to confirm the validity of the proposed method, we evaluate its calculation time and computational accuracy by using a magnetic circuit design model.