NeMF: Neural Microphysics Fields

Area of Science:

• Scientific imaging
• Atmospheric science
• Computational physics

Background:

• Inverse problems in scientific imaging aim to characterize heterogeneous materials using physical quantities like microphysics.
• Accurate 3D microphysics of clouds are crucial for understanding cloud dynamics, lifetime, albedo, and their impact on Earth's energy balance and rainfall.
• Existing methods provide limited representations of cloud microphysics.

Purpose of the Study:

• To develop a novel method for 3D volumetric recovery of cloud microphysical parameters.
• To introduce the neural microphysics field (NeMF) for enhanced characterization of cloud properties.
• To improve the accuracy and detail of microphysical retrievals from multi-view polarization images.

Main Methods:

• A deep neural network, NeMF, is employed, taking multi-view polarization images as input.
• NeMF is pre-trained using supervised learning, incorporating polarized radiative transfer and noise modeling for polarization-sensitive sensors.
• The method focuses on recovering microphysical parameters, including droplet effective variance.

Main Results:

• NeMF achieves unprecedented recovery of 3D cloud microphysical parameters.
• The model demonstrates robust performance in rigorous simulations.
• Successful application to real-world polarization-image data validates its effectiveness.

Conclusions:

• NeMF offers a significant advancement in retrieving detailed 3D cloud microphysics.
• This technology has the potential to improve climate modeling and weather forecasting.
• The method provides a powerful tool for analyzing cloud properties from observational data.