A Physics-Informed Neural Network as a Digital Twin of Optically Turbid Media

Abstract

Digital Twin

In article number 2400574, Mohammadrahim Kazemzadeh, Massimo De Vittorio, Ferruccio Pisanello, and co-workers introduce a novel methodology for characterizing and creating a digital twin of turbid media using only intensity measurements. Unlike previous deep learning approaches that function as black boxes, this physics-informed framework prioritizes interpretability, providing a clearer understanding of light propagation through scattering media and enabling the estimation of the transmission matrix. A key advantage of this digital twin is its differentiable nature, allowing gradient-based optimization for solving inverse problems, such as retrieving the initial wavefront of light that passed through the medium. The results demonstrate superior accuracy compared to models directly optimized for this task, highlighting the precision and robustness of the proposed approach. This advancement paves the way for future developments in neuromorphic and deep learning computation using photonic and optical systems. The authors sincerely thank Patricia Bondia for the artwork that beautifully visualizes their research.