CR2025_27 Using AI-based weather forecast models to improve representation of tropical cyclones in climate simulations

Lead Supervisor: Kieran Hunt, National Centre for Atmospheric Science and Department of Meteorology, University of Reading

Email: k.m.r.hunt@reading.ac.uk

Co-supervisors: Pier Luigi Vidale, National Centre for Atmospheric Science and Department of Meteorology, University of Reading; Kevin Hodges, National Centre for Atmospheric Science and Department of Meteorology, University of Reading; James Warner, Met Office

Artificial intelligence (AI)-based weather prediction models (AIWPs) have recently emerged as a promising complement to traditional numerical weather prediction models (NWPs). NWPs, currently used by operational forecasting centres around the world, simulate weather by integrating time-dependent, prescribed physical equations describing the dynamics and thermodynamics of the atmosphere as well as its interactions with land, ocean and cryosphere. NWPs are accurate, but computationally expensive, especially at the high resolutions needed to correctly simulate the mechanisms important for tropical cyclone (TC) development and maintenance, such as convection.

In contrast, AIWPs use a range of state-of-the-art machine learning techniques – transformers, graph neural networks, and diffusion models – to predict the weather. These models are trained on vast amounts of NWP-processed historical weather data, known as reanalyses, learning relationships without explicitly solving the underlying physical equations. As a result, the current generation of AIWPs can produce weather forecasts with comparable skill to NWPs in large-scale fields but at a minuscule fraction of the cost. Some forecasting centres, such as ECMWF, have thus started using them as a complementary addition to their existing NWP forecasts.

The current generation of climate models use very similar methods to NWPs. Historically, due to their long simulation times, these were often run at much coarser spatial resolutions, making the simulation of TCs in current and future climates challenging as their development and maintenance depend on small-scale processes such as convection and turbulent heat exchange. However, global climate models can now be run at very high resolutions – so-called Global Storm Resolving Models (GSRMs) – that can partially resolve these processes, leading to considerable improvements in the representation of TCs and their statistics (e.g., Vidale et al, 2021).

One recently discovered and exciting application of AIWPs is in downscaling climate model output (Koldunov et al, 2024). AIWPs can be used to post-process climate model output by treating them as initial conditions for short-term AI forecasts. This method has already been shown to be a very effective way of cheaply and accurately increasing the resolution of near-surface temperature, wind, and humidity (Figure 1). Previous work has also shown that AIWPs are able to simulate some, but not all, small-scale features in extratropical cyclones (Charlton-Perez et al, 2024). In this project, you will investigate how effective this approach is at improving the representation of TCs and their impacts in climate models, with results that can be compared directly to GSRM output from pre-existing projects. You will approach this through following research questions:

• Q1) What forecast lead time and which AIWP are most appropriate for downscaling TCs? This must balance the need for the AIWP to ‘spin up’ from coarse initial conditions against the tendency for AIWPs to drift towards smoother fields at longer lead times.
• Q2) Are TC characteristics (e.g., genesis location, intensification from weaker storms, tracks, structure, peak intensity, including the location of peak intensity) improved by AIWP downscaling, and if so by how much? How important is AIWP resolution – for example, Aurora can be run at either 25 km or 10 km – in these improvements? How sensitive are improvements to any underlying climate model biases, particularly circulation biases, and how do the results compare with those from GSRMs used in DYAMOND, NextGEMS and the EU’s Destination Earth Digital Twins, which are run over a wide range of resolutions, from 20km to 2km?
• Q3) How do results differ between AIWP downscaling a large ensemble of low-resolution climate models versus downscaling a smaller ensemble of GSRMs? This will help understanding of uncertainty and variability in the response of TCs to future climate scenarios.
• Q4) Can this method be used to find unprecedented TC events in coarse climate model outputs, such as unusual tracks or extremely high wind speeds?
• Q5) How might the improved representation of TCs from AIWP downscaling impact our understanding of TC behaviour under future climate scenarios? What are the projected changes in TC frequency, intensity, and tracks?

Training opportunities: 

Training opportunities include opportunities to undertake extended visits at the Met Office, where the student can work with leading scientists in both climate modelling and AI NWP development. This will provide hands-on experience with state-of-the-art models and exposure to real-world applications of AI in weather and climate science.

The student will have the opportunity to undertake further training offered by NCAS in their Introduction to Atmospheric Science course and Climate Modelling Summer School. The student will also have opportunities to present their work at international conferences and workshops, and to engage with the vibrant research community at the University of Reading and the National Centre for Atmospheric Science. 

Student profile: 

This project would be suitable for students with a degree in atmospheric science, meteorology, climate science, physics, mathematics, computer science, or a closely related environmental or physical science. Strong quantitative skills are essential, and experience with programming (e.g., Python, MATLAB) and machine learning techniques would be highly beneficial. An interest in AI applications and a background in data analysis would also be advantageous. 

Co-Sponsorship details: 

The project will be applying for a CASE award with the Met Office.

References:

• Charlton-Perez, A. J., Dacre, H. F., Driscoll, S., Gray, S. L., Harvey, B., Harvey, N. J., Hunt, K. M. R, … & Volonté, A. (2024). Do AI models produce better weather forecasts than physics-based models? A quantitative evaluation case study of Storm Ciarán. npj Climate and Atmospheric Science, 7(1), 93.
• Koldunov, N., Rackow, T., Lessig, C., Danilov, S., Cheedela, S. K., Sidorenko, D., … & Jung, T. (2024). Emerging AI-based weather prediction models as downscaling tools. arXiv preprint arXiv:2406.17977.
• Vidale, P. L., Hodges, K., Vannière, B., Davini, P., Roberts, M. J., Strommen, K., … & Corti, S. (2021). Impact of stochastic physics and model resolution on the simulation of tropical cyclones in climate GCMs. Journal of Climate, 34(11), 4315-4341.