Inverse design of optical lenses enabled by generative flow-based invertible neural networks

Abstract

Developing an optical geometric lens system in a conventional way involves substantial effort from designers to devise and assess the lens specifications. An expeditious and effortless acquisition of lens parameters satisfying the desired lens performance requirements can ease the workload by avoiding complex lens design process. In this study, we adopted the Glow, a generative flow model, which utilizes latent Gaussian variables to effectively tackle the issues of one-to-many mapping and information loss caused by dimensional disparities between high-dimensional lens structure parameters and low-dimensional performance metrics. We developed two lenses to tailor the vertical field of view and magnify the horizontal coverage range using two Glow-based invertible neural networks (INNs). By directly inputting the specified lens performance metrics into the proposed INNs, optimal inverse-designed lens specifications can be obtained efficiently with superb precision. The implementation of Glow-assisted INN approach is anticipated to significantly streamline the optical lens design workflows.

Robust Reconstruction of the Void Fraction from Noisy Magnetic Flux Density Using Invertible Neural Networks

“…[ 53 ] for inverse problem in morphology, Ref. [ 54 ] for inverse problem in medical imaging, or [ 55 ] for inverse design of optical lenses. However, INNs remain largely unexplored in the field of solving inverse problems in process tomography.…”