A physics-defined recurrent neural network to compute coherent light wave scattering on the millimetre scale

TL;DR

This paper introduces a physics-based recurrent neural network that computes coherent light scattering in biological tissues without training, enabling fast, deterministic, and large-scale wavefield calculations for microscopy applications.

Contribution

A novel recurrent neural network model that directly encodes Maxwell's equations, eliminating training and enabling large-scale, deterministic light scattering simulations.

Findings

Computes light fields in large scattering volumes efficiently.

Eliminates training phase, reducing computation time.

Provides accessible, open-source tool for researchers.

Abstract

Heterogeneous materials such as biological tissue scatter light in random, yet deterministic, ways. Wavefront shaping can reverse the effects of scattering to enable deep-tissue microscopy. Such methods require either invasive access to the internal field or the ability to numerically compute it. However, calculating the coherent field on a scale relevant to microscopy remains excessively demanding for consumer hardware. Here we show how a recurrent neural network can mirror Maxwell's equations without training. By harnessing public machine learning infrastructure, such \emph{Scattering Network} can compute the $633\;\textrm{nm}$-wavelength light field throughout a $25\;\textrm{mm}^2$ or $176^3\;\mu\textrm{m}^3$ scattering volume. The elimination of the training phase cuts the calculation time to a minimum and, importantly, it ensures a fully deterministic solution, free of any training bias. The integration with an open-source electromagnetic solver enables any researcher with an internet connection to calculate complex light-scattering in volumes that are larger by two orders of magnitude