Autoencoder Based Optimization for Electromagnetics Problems

In this work a novel approach is presented for topology optimization of electromagnetic devices. In particular a surrogate model based on Deep Neural Networks with encoder-decoder architecture is introduced. A first autoencoder learns to represent the input images that describe the topology, i.e., geometry and materials. The novel idea is to use the low dimensional latent space (i.e., the output space of the encoder) as the search space of the optimization algorithm, instead of using the higher dimensional space represented by the input images. A second neural network learns the relationship between the encoder outputs and the objective function (i.e., an electromagnetic quantity that is crucial for the design of the device) which is calculated by means of a numerical analysis. The calculation time for the optimization is greatly improved by reducing the dimensionality of the search space, and by introducing the surrogate model, whereas the quality of the result is slightly affected.