Project: Machine Learning Approaches in Bayesian and Ensemble Data Assimilation
Supervisors: Eviatar Bach (UoR), Jochen Broecker (UoR) & Andrew Duncan (Alan Turing Institute)
Project Description:
Probabilistic data assimilation (DA) is the process of combining models with observations to obtain the filtering distribution—the conditional probability over states given past and present observations. Due to computational limitations, typically only rough approximations of the true filter are tractable. This project proposes to use machine learning (ML) to learn new DA algorithms that better approximate the true filter, holding the potential to improve forecasts and quantify their uncertainty. This will be done using strictly proper scoring rules, skill metrics with appealing theoretical properties for this purpose.
This project will focus on learning ensemble DA algorithms for use in high-dimensional chaotic systems such as the atmosphere. Initial application will be to idealized problems, but scaling up these methods to operational weather prediction will also be explored. Theoretical issues about learnability and comparisons to other methods will also be considered.
The combination of ML with DA is an active and quickly expanding area of research. However, learning DA algorithms is an underexplored field and has the potential to significantly improve on current DA methods used for weather and climate forecasting. The student would thus be at the frontier of high-impact DA research, at world-leading institutions on research in DA (Reading) and ML (Turing Institute).
