Scattering-physics-constrained neural network framework for retrieving dust microphysical properties from scattering matrix measurements

Abstract

Inverting light-scattering data to retrieve the microphysical properties of irregular particles like dust is a classic, ill-posed challenge. This study presents a scattering-physics-constrained deep learning framework that overcomes this by directly and robustly inferring dust properties, including refractive index, size distribution, and shape related parameters, from multi-angle scattering matrix measurements. Our approach synergizes a particle-size-comprehensive forward simulator for super-spheroids with a custom loss function that embeds physical constraints. By unifying the invariant imbedding T-matrix, improved geometric optics, and deep neural network methods, the simulator seamlessly bridges optical properties across all particle sizes. Validated on experimental data, the developed framework provides an efficient and accurate approach for retrieving dust aerosol microphysics.