Advancing Large-Scale Snow Modeling with Data-Driven Parameterizations

Author: Charbonneau, Andrew J.

Year: 2026

Degree: Dissertation (Ph.D.)

Advisor: Schneider, Tapio

Committee Members: Thompson, Andrew F.; Schneider, Tapio; Faraon, Andrei; Deck, Katherine M.

Option: Applied Physics

Abstract

Snow is a key component of the terrestrial cryosphere, governing surface energy exchange, water storage, and hydrological variability across seasonal to multi-decadal timescales. Despite its importance, snow dynamics remain among the most challenging earth system components to model. This difficulty arises from the strong coupling of microphysical, radiative, and mechanical processes across scales spanning orders of magnitude, from grain-scale metamorphism to global atmospheric circulation. This dissertation develops physics-informed parameterizations for snow depth, albedo, and surface temperature for usage within large-scale snow models. For snow depth and albedo, this work develops and evaluates a hybrid physics-machine learning framework, designed to reconcile the expressive flexibility of data-driven methods with the structural constraints required for physically consistent simulation. The proposed approach embeds physical guardrails directly into a neural dynamical system, enabling small, computationally efficient models to learn nonlinear updates while remaining stable, bounded, and compatible with coupled global climate model architectures. Across both site-level and global-scale evaluations, the developed schemes match or exceed established empirical and process-based parameterizations. This performance extends from the modeled variables to downstream effects like snow season timing, and generalizes across a range of climate types. The hybrid parameterization framework substantially improves generalization compared to unconstrained approaches. Further assessment suggests observational efforts and available training ranges remain critical for further improvement, rather than deficiencies in the modeling framework itself. Overall, this work demonstrates that a hybrid framework --- and the choice of simple, yet flexible structures, instead of detailed but strict formulations --- offers a viable and scalable alternative to traditional snow parameterizations, achieving improved accuracy at reduced cost. Outcomes reinforce the idea that embedding even minimal physical structure into data-driven or basic schemes can yield robust improvements in complicated physical systems.