Automatic differentiation meets nano-photonics: Inverse problems and physics informed learning

Abstract:

Deep learning methods have today reached most domains of scientific research. They promise significant computational acceleration and unprecedented capabilities, for instance in the solution of difficult inverse problems. Aside from machine learning, the open source availability of highly optimized automatic differentiation (AD) tool-kits such as tensorflow, pytorch or jax also allows implementing photonics simulations methods with implicit automatic differentiation capabilities.

I will explain how deep learning and automatic differentiation can be used to solve ill posed inverse problems such as the design of nano-photonic devices or metasurfaces. Furthermore, I will present how automatic differentiation and concepts of physics informed learning can be useful to solve nano-photonics scattering problems.

[1] S. Ponomareva et al. SciPost Physics Codebases 60, v0.56 (2025), https://homepages.laas.fr/pwiecha/torchgdm_doc/ 
[2] D. Soun et al. Opt. Express 33, 25945–25958 (2025)
[3] A. Khaireh-Walieh et al. Nanophotonics, 12, 4387–4414, (2023)