Uncertainty Quantification for Parametrisation, including machine learning approaches
A key theme is the development of 'stochastic parametrisations' - these are representations of processes such as clouds and thunderstorms which explicitly account for uncertainties. This means our weather and climate predictions become Monte-Carlo type simulations, where we see how uncertainties in the small-scale processes influence the rest of the climate system.
Peter is working on using machine learning tools such as recurrent neural networks to predict the ideal stochastic perturbations within a machine-learnt parameterisation of sub-grid physics.
Zhixiao is working on developing a stochastic, data-driven parameterisation of organised convective storms to be used in the UK Met Office's operational weather forecasting system.
Greta is working on developing a machine learning stochastic parameterization for the convection trigger in climate models, directly trained on observations.
Laura is breaking down uncertainties associated with machine learning parameterisations by source: epistemic uncertainty (from uncertainty in the machine learning model) and aleatoric uncertainty (from uncertainty in the training dataset).
Edward works on the Model Uncertainty- Model Intercomparison Project. It compares small-scale process representations across several international modelling centres to better characterise uncertainties in these processes
Model validation
A complementary approach to improving model approximations is increasing the resolution of our simulators, so that they explicitly capture more processes. But just how good are these high-resolution computer models? To answer this question, we develop novel approaches to compare models with observations.
Lilli is evaluating how clouds and convection are represented in global km-scale climate models, using satellite observations and a variety of methods including fractal analysis and machine learning.
Simon is working to understand the benefits of high-resolution modelling with machine learning-based data analysis. A key focus is on the relative roles of atmosphere and ocean resolution
Kristian tries to understand how and why increased model resolution affects the representation and predictability of midlatitude dynamics, such as blocking and weather regimes.
Predictability
The theory of predictability and prediction, encompassing ideas such as chaos and turbulence theory, underpins much of the wider research in the group. What can we predict using today's computer models? And what could we theoretically predict using technologies of the future?
Salah is working on applying coarse-graining analysis to atmospheric data and simulations to gain a deeper understanding of the dynamics and predictability of the weather.
Emma is using machine learning tools to understand drivers of extreme flooding, in collaboration with with scientists in the School of Geography and the Environment.
Edward works on predictability of deep convective thunderstorms using ultra-high-resolution simulations.
Machine Learning for Earth-system prediction
Recent developments in Machine Learning have opened up the possibility of bypassing traditional computer simulators entirely when making predictions of future weather and climate. This is a very exciting and fast-moving field! Our work explores the potential (and limitations) of these new ML models, with a particular interest in what they can teach us about the real world.
Bradley is working on a world model approach to modelling tropical Pacific climate variability. A key goal is gaining mechanistic understanding of processes involved using the machine learning approach.
Bobby is investigating how machine learning models can be used to improve the quality of Earth-system model simulations, specifically focused on making the initialisation of the ocean more efficient by using machine learnt emulators.
Andrew is working on data-driven dynamical modelling of large-scale climate variability and precipitation in the midlatitudes with the goal of improving process understanding.
Maren is stress-testing machine learning-based weather prediction models, by investigating whether they generalise spatially.