Attention-Based Encoder-Decoder Network for Prediction of Electromagnetic Scattering Fields

Abstract

To reduce the computation time cost by the numerical methods for electromagnetic scattering field calculation, this paper proposes an attention-based encoder-decoder neural network (AEDNNet) to predict the electromagnetic fields scattered by complex scatterers. The structure of AEDNNet comprises attention mechanism and residual learning strategy, in which the attention mechanism is utilized to improve the accuracy of the network, and the residual strategy makes the network converge quickly and avoid network degradation. The magnitudes of the scattering fields under the illumination of plane waves with various incident angles are used as the training set. Numerical results on the test set show that the mean relative error of the method is less than 1%.