Auxiliary Physics-Informed Neural Networks for Forward, Inverse, and Coupled Radiative Transfer Problems

Abstract

In this paper, we develop and employ auxiliary physics-informed neural networks (APINNs) to solve forward, inverse, and coupled integro-differential problems of radiative transfer theory (RTE). Specifically, by focusing on the relevant slab geometry and scattering media described by different types of phase functions, we show how the proposed APINN framework enables the efficient solution of Boltzmann-type transport equations through multi-output neural networks with multiple auxiliary variables associated to the Legendre expansion terms of the considered phase functions. Furthermore, we demonstrate the successful application of APINN to the coupled radiation-conduction problem of a participating medium and find distinctive temperature profiles beyond the Fourier thermal conduction limit. Finally, we solve the inverse problem for the Schwarzschild-Milne integral equation and retrieve the single scattering albedo based solely on the knowledge of boundary data, similar to what is often available in experimental settings. The present work significantly expands the current capabilities of physics-informed neural networks for radiative transfer problems that are relevant to the design and understanding of complex scattering media and photonic structures with applications to metamaterials, biomedical imaging, thermal transport, and semiconductor device modeling.

References in corpus (8)

• Outlook for inverse design in nanophotonics
• A high-bias, low-variance introduction to Machine Learning for physicists
• Multiple scattering of classical waves: from microscopy to mesoscopy and diffusion
• Physics-informed neural networks for inverse problems in nano-optics and metamaterials
• Deep neural networks for the evaluation and design of photonic devices
• Physics Informed Neural Networks for Simulating Radiative Transfer
• Physics-informed neural networks for imaging and parameter retrieval of photonic nanostructures from near-field data
• Predicting nonlinear optical scattering with physics-driven neural networks

Cited by in corpus (1)

• Multiscale Physics-Informed Neural Networks for the Inverse Design of Hyperuniform Optical Materials