Statistical space-time models for our climate

Spatio-temporal statistics is concerned with modelling, predicting and forecasting data collected in space and time, taking advantage of information contained in the correlations and dependencies between datapoints in this (up to) four-dimensional space. Classical theory and models exist spanning several decades of research, but many outstanding challenges remain in terms of implementing them on real-world datasets:
Modelling large data volumes. Large datasets are challenging to model, where the potential for capturing additional features increases with data size, but become increasingly masked by other features, non-stationarity, and observation noise. Classical methods often miss the predictive potential gained from increasing volumes of data. This project will build new multivariate and non-stationary spatio-temporal models relevant to climate and geological sciences.
Inference, Forecasting and Prediction. Then, even after finding a great model, there is usually an enormous computational bottleneck specific to spatio-temporal data, where classical inference and prediction methods are often based on matrix inversion which becomes increasingly prohibitive as the dimensions of the data grow. This project will build novel machine learning techniques to solve these issues.
Irregular and Missing data. Many existing spatio-temporal methodologies assume data is perfectly observed on a regular grid in space and time, this is not realistic in many applications. This project will build novel models, inference and prediction tools that are robust to irregular and missing data, including techniques for uncertainty quantification that reflect the uncertainty enduced by the irregular nature of the data.
The above-mentioned challenges are ubiquitous in the climate-related datasets and scientific challenges that BGS has. In this project you will work closely with BGS on these datasets to build novel predictions, insights and conclusions.