PhD Studentship: Machine Learning the Building Blocks of the Weather Forecast

Weather forecasts are essential for informing decisions made by the public, industry and the government. Recent technological advances have led to a proliferation of machine learning (ML) weather models that are outperforming traditional models by some measures, although challenges remain. As ML models are not designed based on physical principles they are not explainable and do not perform reliably outside of regimes seen in their training data. This project aims to address these challenges by breaking the weather forecast into its fundamental building blocks and applying different ML architectures to each. Key questions are: In matching the structure of the ML architecture to the underlying mathematical structure of the model component, can we improve the representation of physical processes? To what extent does this approach result in a model that is more efficient, stable and robust outside of its training data, for example, under a warming climate? This approach aims to improve understanding of how ML models represent different building blocks of the weather forecast, progressing towards a model that is human-understandable rather than a black box. This is crucial to increase trust in ML methods as they move towards operational use. The Researcher will receive training through taught and self-study modules, directed reading, supervisory discussions, hands-on coding, and regular hackathons. They will acquire specialist skills in geophysical fluid dynamics and ML techniques, as well as sought-after transferrable skills in software development, reproducibility, data management and scientific communication. As the project progresses, the Researcher will be encouraged and supported to take the lead in steering the direction of the project. Through the CASE partnership and a decade-long supervisor-Met-Office collaboration, they will gain experience in working closely with Met Office scientists, relating their research directly to weather forecasts with the potential to shape future Met Office forecast system development.