This github repo provides the analysis and code for a scientific machine learning analysis of depositional ice growth models. We provide a PyTorch implementation of the code for the paper "Discovering How Ice Crystals Grow Using Neural Ordinary Differential Equations and Symbolic Regression".
Here we explore how neural ordinary differential equations and equation discovery can be applied to observations of times series of the mass of single ice crystals grown in a levitation diffusion chamber.
Data sets for the levitation diffusion chamber experiments are described in Pokrifka et al. 2020 and Pokrifka et al. 2023.
The preprocessing.py script creates a pytorch data loader containing all experiments used in this analysis.
Synthetic data for the levitation diffusion chamber experiments assumes a functional form the depositional ice growth model, and the script synthetic_data.py
takes the initial conditions and detrended noise from the real experiments to initiate the synthetic data sets.
Data sets from the IsoCloud experiments in the AIDA Cloud Chamber are described in Lamb et al. 2017, Clouser et al. 2020, and Lamb et al. 2023. Pre-processing scripts for the AIDA data sets can be found here.
The main.py script is used for training the NODE model against the experimental observations or synthetic data sets and for finding a symbolic equation. It includes weak, medium, and strong assumptions for the physical constraints
used in the NODE model. The torchdiffeq library is used to implement the NODE models and the PySR library is used for symbolic regression.
The preprint for this paper can be found at:
@article{Lamb2025,
title={Discovering How Ice Crystals Grow Using Neural Ordinary Differential Equations and Symbolic Regression},
author={Lamb, K.D. and J. Harrington},
journal={},
doi = {DOI:},
year={2025}
}