Simulating Arctic Snow using Process-based and Machine Learning Approaches for Future Carbon Flux Projections

Abstract

Arctic permafrost soils store ~1700 Gt of carbon and are increasingly vulnerable to climate change, with high latitude environments warming up to four times faster than the global average. Arctic warming is particularly pronounced in winter, and soil carbon emissions during winter form a crucial part of the annual carbon budget. Seasonal snow acts as a thermal insulator that strongly influences soil temperature which regulates winter soil emissions of carbon dioxide (CO2) and methane (CH4). However, key snowpack properties, such as bulk density and effective thermal conductivity (Keff), are often misrepresented in earth system models (ESMs) which limits the reliability of winter carbon flux projections.
This thesis combines process-based modelling with machine learning (ML) to improve simulations of Arctic snow properties and assess their downstream impacts on soil carbon dynamics. Firstly, an Arctic-specific Keff parameterisation, derived from field observations, was implemented into the Community Land Model (CLM5.0) to simulate soil and carbon fluxes from present day to 2100. CLM5.0 simulations were driven by an ensemble of 33 meteorological projections from the North American Coordinated Regional Downscaling Experiment (NA-CORDEX). Secondly, a Random Forest (RF) model was trained on meteorological data and snowpack simulations to predict bulk snow density, underpinned by field measurements from Trail Valley Creek, NWT, Canada. Finally, RF-predicted snow densities were assimilated into CLM5.0 subroutines to test their influence on snow insulation, soil temperature, and carbon fluxes under future climate scenarios (RCP 4.5 and 8.5).
Replacing the default Keff parameterisation with an Arctic-relevant scheme increased simulated winter CO2 emissions by up to 130% and CH4 emissions by 50% by 2100 under RCP 8.5, with soil temperatures 4 – 7 °C warmer. Notably, Keff representation exerted a control on projected Arctic carbon emissions as large as the uncertainty introduced by the NA-CORDEX projections. Further, the RF model accurately reproduced tundra snow density evolution and was in close agreement with in situ observations. The assimilation of RF-derived densities into CLM5.0 reduced cold soil biases, and produced future soil carbon flux responses comparable in magnitude to those resulting from Arctic-relevant changes in Keff parameterisation. This thesis demonstrated that better representation of snow processes in ESMs is essential for reliable prediction of permafrost–carbon feedbacks in a warming Arctic and their contribution to the global carbon budget.

Date of Award	19 Feb 2026
Original language	English
Awarding Institution	Northumbria University
Supervisor	Nick Rutter (Supervisor) & Leanne Wake (Supervisor)