Brief Analysis of the Advantages of Complex Neural Networks in Metamaterial Micro-Structure Design

In this paper, we propose a complex neural network that takes the real and imaginary parts of electromagnetic response simultaneously and extracts extra signature contributed by both parts. Compared with using two neural networks to train real and imaginary parts separately, the complex neural network prevails in less numbers of parameters and less training time. Once the complex neural network is designed, it usually takes less than a few hundred epochs in training. In the inference stage, either MSE or MAE of the EM response of 1001 sampled frequencies between 1 and 18 GHz is around 0.008. The network provides electromagnetic response with reasonable accuracy.