Learning Constraints on Depositional Ice Growth Models from Cloud Chamber Experiments with Physics Informed Neural Networks

Depositional ice growth impacts both ice crystal habit and ice growth efficiency, but the physics of ice growth in atmospheric conditions, particularly at low temperatures, is still poorly constrained. Several different models for the deposition coefficient function have been developed in the context of laboratory experiments of single ice crystals growing from vapor. One major challenge in constraining depositional ice growth models against observations is that experimental measurements are often challenging to interpret, and require assumptions about the functional dependence of the deposition coefficient function or on simplified parameterizations of the shape of the ice crystal.

Here we extend our recent work on evaluating deposition coefficient parameterizations against cirrus cloud simulation experiments in the Aerosol Interactions and Dynamics in the Atmosphere (AIDA) cloud chamber with a parcel model with different assumptions for the deposition coefficient function [Lamb et al. 2023]. Physics Informed Neural Networks (PINN’s) are a recently developed machine learning method that can be used to learn models for physical systems, while respecting any known laws of physics as described by partial differential equations by employing physics-constrained loss functions and automatic differentiation. Here we explore how PINN’s can be used to evaluate model structural uncertainty and unknown parameters in depositional ice growth models against cirrus cloud experiments performed in AIDA over a range of temperatures below the homogeneous freezing limit of water (T = 180 – 235 K).