Background

Probabilistic machine learning allows us to learn from data while quantifying uncertainty, and is increasingly important in climate science. However, these methods rely on statistical models that are inevitably imperfect. For example, satellite measurements of the polar regions may contain anomalous observations due to calibration problems, retrieval errors or other sources of contamination. Standard methods can be surprisingly sensitive to such errors, leading to distorted predictions and misleading uncertainty estimates. Robust probabilistic machine learning seeks to develop methods that continue to perform reliably when data or modelling assumptions are imperfect. Recent advances suggest that robustness can be achieved without sacrificing computational efficiency. The challenge is to turn these ideas into practical tools for the scale and complexity of modern climate datasets.

PhD project

This PhD will develop new statistical and machine-learning methodology for making climate predictions more robust to imperfect observations and modelling assumptions. The student will focus on methods such as Gaussian processes and Kalman filtering, investigating how ideas from generalised Bayesian inference and robust statistics can be used to improve their reliability. The work will combine mathematical theory with the development of scalable algorithms, and will study questions such as how robustness affects prediction, uncertainty quantification and computational cost. The methodology will be developed and evaluated in collaboration with climate scientists, with a particular focus on satellite observations of the cryosphere and their use in understanding changes in the polar climate system. These applications provide challenging test cases in which data are sparse, irregularly distributed across space and time, and often affected by measurement errors.

The student will be primarily based in UCL’s Fundamentals of Statistical Machine Learning research group, providing a strong environment for research at the interface of statistics and machine learning. The project will be co-supervised by Prof. Michel Tsamados (UCL Earth Sciences), an expert in polar observation, remote sensing and modelling. This offers the opportunity to develop new statistical methodology while working closely with climate scientists on problems directly relevant to monitoring and understanding climate change in the polar regions.

Applicant Profile

Students should have a strong mathematical background and strong programming skills. This might, for example, include applicants with undergraduate and postgraduate degrees in mathematics, statistics, computer science, or a related discipline. The ideal candidate will also have some previous research experience in statistics or machine learning, for example through a research thesis or summer research placement. Applicants with additional relevant experience from industry are also strongly encouraged to apply. A strong interest in climate and environmental applications is essential, but an existing background in these areas is not.

Other Information

Personal webpage of primary supervisor: https://fxbriol.github.io
Webpage describing recent methodological advances: https://fxbriol.github.io/research/bayesian-robustness/
Group webpage: https://fsml-ucl.github.io