CR2025_05 Realising the benefits of next generation sub-km ensemble weather forecasts

Lead Supervisor: Suzanne Gray, Department of Meteorology, University of Reading

Email: s.l.gray@reading.ac.uk

Co-supervisors: Humphrey Lean, Met Office; Todd Jones, Department of Computer Science, University of Reading; Lewis Blunn, Met Office.

The cutting-edge in operational regional environmental prediction involves forecasting systems that use numerical weather prediction models to solve equations on grids with lengths of only a few kilometres. These models, called km-scale or convective-scale models, can represent deep convective processes explicitly rather than approximating their effects; this resolution has led to a step change in the realism of forecasts of localised high impact weather events. The chaotic nature of the atmosphere leads to the divergence of forecasts with near identical starting conditions, and error growth is most rapid at small spatial scales. To represent this forecast uncertainty, operational centres perform not just one but tens of forecasts, a so-called ensemble, with small variations in the starting conditions or the model physical equations in each ensemble member. The next forecasting frontier is even finer, sub-km, models. The Met Office has a project to develop these models and is expecting to start running experimental 300-m grid length ensembles for London and other cities in the next few years. However, despite advances in supercomputing capability, sub-km modelling trade-offs will still have to be made, e.g., in selecting the grid length, number of ensemble members, and domain size, as each halving of grid length leads to an approximately 10-fold increase in computational cost. Investigation of how to optimise ensemble modelling capability whilst balancing computational cost is therefore essential.

The aim of this project is to reveal effective ways of using km-scale ensembles to gain the benefits of sub-km scale ensembles.

An ensemble approach is essential for sub-km scale forecasts due to the scales of interest being small compared to the unpredictable (convective and larger) scales being forecast. Probabilistic forecasts are required to make effective hazard warnings, e.g., for flooding events. Within the (UK) Met Office, ensemble forecasts with 300-m grid length have been trialled in summer 2022 (London), summer 2023 (southern England, see Figure 1) and summer 2024 (Paris, for the Olympics) ahead of planned operational implementation as a “trailblazer” configuration in 2025/26. These sub-km ensembles potentially provide benefits in situations where surface heterogeneity and convective-scale dynamics are important, but their extraordinary computational costs raise questions about whether they are the most effective use of resources.

To address the question of the most effective use of km-scale ensembles while gaining the benefits of sub-km scale models, the project will take three strands:

(1) the characteristics of km-scale versus sub-km scale ensembles will be evaluated,

(2) methods of reducing the computational cost of conventional sub-km ensembles will be explored (e.g., through clustering the parent model ensemble members and only running a subset at subkm scale), and

(3) machine learning post-processing techniques will be used to statistically downscale km-scale ensembles to obtain sub-km scale ensembles.

You will investigate the weather conditions (such as weather regimes) and forecast variables (e.g., rainfall, wind) for which sub-km ensembles provide additional value compared to km-scale ensembles and examine how the ensemble skill compares to its spread, i.e., the divergence of the forecasts with time. You will consider the mechanistic sources of forecast improvements, e.g., through upscale development of convection and/or the improved representation of orographic and land surface processes. Error sources in the km-scale parent forecasts will be identified. You will explore the potential to exploit machine learning approaches. Can the parent km-scale ensemble forecasts be clustered into representative groups, limiting the number of necessary sub-km forecasts? Are sub-km forecasts able to simulate different forecast evolutions to the km-scale “parent” forecast, or do they just add realistic “noise” that could be emulated by machine learning downscaling methods?

This research will support the Met Office strategy of developing the next generation of very high resolution global and regional environmental prediction systems to improve hazard warnings while developing and exploiting ensemble-based systems. There is also strong synergy with Met Office plans to develop machine learning modelling and post-processing approaches that will make the forecasting workflow more efficient in terms of person power and computing resource.

Training opportunities:

This project will provide opportunities to develop mathematical modelling and data analysis skills, particularly in numerical weather model prediction. You will also develop scientific insight in forecasting systems and predictability as well as weather systems and their associated hazards. The project offers opportunities to attend postgraduate MSc modules and summer schools. You will have CASE support including a placement at the Met Office headquarters in Exeter. This placement will provide experience of models used operationally for atmospheric forecasting, machine learning applications, and an industrial research environment.

Student profile:

This project would be best suited to a student with a strong physical sciences or mathematical background. The student will not need to have prior coding or data analysis expertise as full training and support will be given, but they should be keen to perform careful analysis of the outputs of complex weather forecast models.

Co-Sponsorship details:

The project will receive a CASE award from the Met Office.

References:

• Lean, H.W., Theeuwes, N.E., Baldauf, M., Barkmeijer, J., Bessardon, G., Blunn, L., et al. (2024) The hectometric modelling challenge: Gaps in the current state of the art and ways forward towards the implementation of 100-m scale weather and climate models. Quarterly Journal of the Royal Meteorological Society, 1–38. Available from https://doi.org/10.1002/qj.4858
• Hanley, K.E. and Lean, H.W. (2024) The performance of a variable-resolution 300-m ensemble for forecasting convection over London. Quarterly Journal of the Royal Meteorological Society, 150, 3737–3756. Available from https://doi.org/10.1002/qj.4794