Accessibility navigation

Machine learning, emulation and Bayesian dimension reduction for climate change projection

Mansfield, L. (2021) Machine learning, emulation and Bayesian dimension reduction for climate change projection. PhD thesis, University of Reading

Text - Thesis
· Please see our End User Agreement before downloading.

[img] Text - Thesis Deposit Form
· Restricted to Repository staff only


It is advisable to refer to the publisher's version if you intend to cite from this work. See Guidance on citing.

To link to this item DOI: 10.48683/1926.00107021


Climate change projection under different greenhouse gas and aerosol emission scenarios is crucial for informing societal adaptation and mitigation measures. This traditionally relies on computationally expensive global climate models (GCMs) run on decadal to centennial timescales. One of the goals of this thesis is in exploring machine learning models and emulators trained on the output of global climate models, that can assist in this endeavour by providing rapid estimations of the climate response. Two statistical models are developed, one of which emulates the global short-term climate response to an emissions perturbation and one which learns the mapping from the short-term climate response to the long-term climate response. Different perspectives are taken so that the short-term response is predicted with a probabilistic emulator which interpolates between known and unknown data points, while the global patterns of long-term response are predicted with machine learning methods. Both models are shown to accelerate climate change projections and also provide new insights into the main drivers of climate change through sensitivity analysis to different emission perturbations and by uncovering consistent early indicators of long-term climate response. Discovering structures in climate data that can explain patterns and behaviour is another focus of this thesis, addressed through a dimension reduction technique to simplify large datasets. This is approached from a Bayesian perspective which could allow a complete quantification of uncertainty when making predictions through an emulator trained on a reduced dataset. Reversible jump Markov chain Monte Carlo and Sequential Monte Carlo algorithms are developed for a latent factor model to infer the probability distribution on both the number of underlying dimensions and the structure of these. Sequential Monte Carlo is found to be significantly more effective at determining these and is demonstrated on weather observations to reveal underlying factors governing the weather behaviour.

Item Type:Thesis (PhD)
Thesis Supervisor:Voulgarakis, A., Everitt, R., Nowack, P. and Dance, S.
Thesis/Report Department:School of Mathematics and Statistics
Identification Number/DOI:
Divisions:Science > School of Mathematical, Physical and Computational Sciences > Department of Mathematics and Statistics
ID Code:107021


Downloads per month over past year

University Staff: Request a correction | Centaur Editors: Update this record

Page navigation