Through the AI of the Storm

: Leveraging Generative AI for High-Resolution Tropical Cyclone Rainfall

Abstract

Tropical cyclone rainfall considerably impacts human life, causing severe
damage to properties and displacing large populations. For instance, Cyclone
Freddy devastated East Africa in early 2023, with Mozambique receiving over
half a meter of rain, leading to widespread flooding, $1.53 billion in damages
and the displacement of 1.4 million people across six countries. At time of
submission, Hurricane Beryl, the earliest category 5 hurricane, made landfall
in Jamaica causing widespread devastation. The scientific consensus suggests
that tropical cyclone precipitation rates rise by 14% per 2°C of warming. Current
climate change mitigation ambition corresponds to a 2.5-2.9°C global
warming above pre-industrial levels by 2100, thus tropical cyclone rain rates
will likely become more intense as the century progresses. Climate models
should be run at “convection permitting" resolutions to adequately resolve the
inner core processes of tropical cyclones, but this requires large computational
resources which restricts simulation output and limits the possible risk-related
conclusions.

The overarching purpose of this thesis is to address the gap in available
high-resolution tropical cyclone rainfall datasets by using machine learning
to downscale storm rainfall to scales of kilometres. Each chapter builds on
the complexity of the model to create a generalised AI approach that can
achieve this purpose. We first compare a set of machine learning methods
on their ability to enhance a coarsened (∼100 km) observational rainfall to
approximately 10 km resolution. In this idealised experiment, the Generative
Adversarial Network with Wasserstein loss (WGAN) outperformed the
other methods with mean biases within 5% of observations. We then generalise
the WGAN to generate rainfall on a large set of synthetic storm tracks
covering 360 simulation years across three climate scenarios. By combining
our rainfall dataset to global population estimates, we find that between 288-
325 million people are likely to be newly exposed to extreme tropical cyclone
rainfall under likely climate change conditions.

Date of Award	1 Oct 2024
Original language	English
Awarding Institution	University of Bristol
Supervisor	Dann M Mitchell (Supervisor), Peter A G Watson (Supervisor), Laurence Aitchison (Supervisor) & Raul Santos-Rodriguez (Supervisor)