Long short-term memory neural network for directly inverse design of nanofin metasurface.

Abstract

In this Letter, the neural network long short-term memory (LSTM) is used to quickly and accurately predict the polarization sensitivity of a nanofin metasurface. In the forward prediction, we construct a deep neural network (DNN) with the same structure for comparison with LSTM. The test results show that LSTM has a higher accuracy and better robustness than DNN in similar cases. In the inverse design, we directly build an LSTM to reverse the design similar to the forward prediction network. By inputting the extinction ratio value in 8-12 µm, the inverse network can directly provide the unit cell geometry of the nanofin metasurface. Compared with other methods used to inverse design photonic structures using deep learning, our method is more direct because no other networks are introduced.