How do we predict the weather next week or the climate next century?
The answer is that we build computer simulators of the Earth system – the atmosphere, the oceans, the land surface and more. These simulators encode our physical understanding of the key processes at work. To prepare for hazardous weather or changing risks due to climate change, we need predictions which accurately capture the chance of extreme events. To improve our predictions, we must improve our computer simulators. To improve our computer simulators, we must first improve our knowledge.
Our group works across this whole pipeline. We use theory and data to develop an improved understanding of the atmosphere. We then use our new understanding to improve our computer models to get better predictions from days to decades. Finally, we collaborate with users of weather and climate data, such as the National Energy System Operator, to ensure best-use of the resultant predictions.
As our group name suggests, our focus is on understanding atmospheric processes and their role in the climate system. We are most interested in smaller scale phenomena – such as clouds, thunderstorms, or turbulence. These phenomena are difficult to capture in our computer simulators, as our simulators typically don’t have the detail to explicitly capture them. Instead, they must be approximated.
The assumptions and simplifications made in this approximation process are a large source of error and uncertainty in weather and climate predictions. We are interested in characterising this uncertainty, so we can make well-calibrated probabilistic forecasts. This ensures we have reliable predictions of the chance of different weather events.
Finally, a growing area of interest in our group is the application of Machine Learning (ML) tools to environmental problems – around half of us use these tools in one way or another. We use a wide range of tools, ranging from simple yet effective random forests and neural networks through to graph neural nets, autoencoders, and equation discovery. Central to our exploration with ML is the goal of using these tools to learn new physics.