Deep learning for light scattering computation: Reconstructing light scattering fields from 1-D randomly rough surfaces as an example

Abstract

As targets can often be treated as randomly rough surfaces in optical bands, light scattered from them carries rich information, therefore, light scattering analysis has been used in various applications such as remote sensing and target identification. Compared to experimental light scattering detection, numerical light scattering computation is still an essential tool to understand the functions of randomly rough surface features on scattered light in the theoretical level. However, classical numerical methods rely on statistical process consisting of massive randomly rough surface generation, scattered light calculation and ensemble average computation to obtain stable results of light scattering fields, therefore they suffer from heavy computation load and low processing efficiency. To solve such problem, deep neural networks are designed here to directly obtain the stable light scattering fields but completely avoid the statistical process; and an example of reconstructing light scattering fields from 1-D randomly rough surfaces is reported. Compared to the classical Kirchhoff approximation method in high accuracy but low efficiency, the proposed deep learning method not only provides accurate light scattering results, but also remarkably accelerates the processing speed, thus opens a new way for numerical light scattering computation in broader scopes.