- Representing model uncertainty: lessons from hybrid modelling by Martin Brolly (University of Edinburgh)
7 November 2025
13:00 to 14:00 GU01, Brian Hoskins Building
Abstract: Uncertainty arises when numerical models fail to represent certain scales and processes. Stochastic parameterisations provide a means to represent this uncertainty explicitly. However, constructing accurate and physically-consistent stochastic parameterisations is highly non-trivial. While approaches based on stochastically perturbing deterministic parameterisations have been used widely, the accuracy that can be obtained with such schemes is limited by (i) statistical assumptions made in their construction and (ii) challenges in their calibration.
Using the test problem of turbulence closure in a model of quasi-geostrophic turbulence, I will demonstrate how common statistical assumptions about model error degrade forecast skill. I will also highlight the importance of online learning for calibrating parameterisations. Finally, I will discuss an approach to improving the representation of memory in Earth system models, which offers a promising route to extending the range of predictability in forecasting.
- Designing Machine Learning Tools to Characterize Multi-stationarity of Fully Open Reaction Networks by Dr. AmirHosein Sadeghi Manesh (Coventry University)
26 November 2025
14:00 to 15:00 Maths 314