Predicting Scattering From Complex Nano-Structures via Deep Learning

Abstract

Existing numerical electromagnetic (EM) solvers are usually computationally expensive, time consuming, and memory demanding. Recent advances in deep learning (DL) techniques have demonstrated superior efficiency and provide an alternative pathway for speeding up simulations by serving as effective computational tools. In this paper, we propose a DL framework for real-time predictions of the scattering from an isolated nano-structure in the near-field regime. We find that, to achieve precise approximation of the optical response obtained from numerical simulations, the proposed DL framework only requires a small training data set. The fully trained framework can be three orders of magnitude faster than a conventional EM solver based on the finite difference frequency domain method (FDFD). Furthermore, the proposed DL framework has demonstrated robustness to changes in design variables which govern the nano-structure geometry and material selection as well as properties of the incident wave, shedding light on universal scattering predictions at the nano scale via deep learning techniques. This framework increases the viability of the design and analysis of complex nanostructures, offering great potential for applications pertaining to complex light-matter interaction between electromagnetic fields and nanomaterials.

Deep learning in nano-photonics: inverse design and beyond

“…The prediction of physical effects is not limited to extinction, transmission or other far-field effects. It has been shown that also nearfield effects can be approximated accurately, for instance around nanowires of complex shape [56].…”

“…At least not if no time constraint is set for the iterative optimization. Well trained and optimized data-driven ANNs usually produce errors in the order of a few percent [28,56]. Furthermore, it is virtually impossible to completely suppress outliers in the network predictions [65].…”

Intelligent metaphotonics empowered by machine learning

“…At the same time, accuracy of prediction reaches the values about 95%. Together with developing ML-supplemented methods for solving both forward and inverse scattering problems [37][38][39][40], such studies suggest that AI technologies may constitute conventional numerical methods resulting in the development of novel computational tools. For instance, the recent results [41] suggests real-time web-based tools for designing far-fields of arbitrary-shaped structures.…”

Deep Learning Based Thermal Stress and Deformation Analysis of Satellites

“…In addition, Cheng et al [12] predicted the 2D velocity and pressure fields around arbitrary shapes in laminar flows by deep learning based surrogate model. In optics, Li et al [13] proposed a deep learning framework for real-time predictions of the scattering from an isolated nano-structure in the neared regime. The literature above has demonstrated the powerful regression capabilities of surrogate models endowed with deep learning techniques.…”