Process-based machine learning for observational constraints on temperature in past and future climates

Wilkinson, Sophie (2023) Process-based machine learning for observational constraints on temperature in past and future climates. Doctoral thesis, University of East Anglia.

Preview

PDF Download (23MB) | Preview

Abstract

This thesis develops a novel method using machine learning (ML) to combine existing climate model output with observations in order to impose new constraints on regional near-surface temperature anomalies from the Coupled Model Inter-comparison Project Phase 6 (CMIP6) in Northern Hemisphere summertime. The ML model is trained with reanalysis data to predict near-surface temperature anomalies from a set of process-based dynamic and thermodynamic predictor variables. Test predictions over the historical period capture day-to-day variation in temperature anomalies well, including the magnitude of more extreme events (for example the Europe 2003 heatwave) in most land regions. The ML method is applied for bias correction of historical climate model output and to constrain uncertainty in future warming projections. This future warming constraint points to a potential over-sensitivity of several CMIP6 models with the constraint tending to result in a small downward correction of the projected future temperature range. The ML bias correction technique performs competitively with a traditional variance scaling approach and, using interpretable AI methods, can be decomposed into contributions from each predictor variable to reveal the presence of potential compensating biases in climate models. The ML method is also applied in a climate model evaluation context by additionally training CMIP6 ML emulators using historical climate model data. Learnt coefficients are then compared between climate models and reanalysis to identify process simulation differences. Finally, the reanalysis-based model is applied as a tool for investigating historical heatwave drivers which are compared with existing literature. The ML method can also be used for climate model evaluation in this context, evaluating the representation of extreme event processes in climate models and testing the ability of climate models to reproduce the magnitude of historical heatwave events when provided with predictor variables from reanalysis.

Item Type:	Thesis (Doctoral)
Faculty \ School:	Faculty of Science > School of Environmental Sciences
Depositing User:	Chris White
Date Deposited:	12 Jun 2024 13:58
Last Modified:	12 Jun 2024 13:58
URI:	https://ueaeprints.uea.ac.uk/id/eprint/95596
DOI:

Downloads

Actions (login required)

View Item