AI-Based Equation Discovery and Deep Learning for Predicting Large-Scale Climate Dynamics and Precipitation Extremes

Extreme rainfall is occurring more frequently and with greater intensity due to global warming resulting in an increasing number of extreme flooding events worldwide. Extreme flooding is one of the most serious consequences of climate change having profound environmental and socioeconomic impacts in areas such as public health, infrastructure, financial services, and ecosystems (Hu et al. 2018). It is therefore important to improve the forecasting capability of extreme precipitation and to better understand the physical mechanisms behind their emergence. The recent rapid advancement of artificial intelligence (AI) algorithms has attracted a lot of interest within the climate science community. To this point, it has shown remarkable capabilities for weather prediction when compared to conventional numerical models (Lam et al. 2022) as well as the ability to accurately identify precursors of extreme events including heatwaves and precipitation (Miloshevich et al. 2023). Machine learning models such as deep neural networks are fundamentally based on a statistical fitting procedure, where predictions of a quantity are learnt from a subset of training data and tested on the remaining data. This can lead to problems where the model may lack interpretability (where the model cannot point to underlying physical processes) and lack generalisability (the model does not predict well outside its training subset). Therefore, using these methods to predict the emergence of severe weather should ideally be performed together with physical understanding (McGovern et al. 2017), which is not readily available from such methods.
Separate from these more conventional statistical-based AI tools is AI-based equation discovery which is a way of discovering equations that govern nonlinear dynamical systems from data. This offers the potential to provide clear process understanding of the climate system from reduced order models, and in some cases may even uncover hidden physics which is then able to replicate conventional AI models. When the governing equations of a dynamical system are known explicitly, they allow for more robust forecasting, control, and the opportunity for analysis of system stability and bifurcations through increased interpretability. Furthermore, if a mathematical model accurately describes the processes governing the observed data, it therefore, can generalize to data outside of the training domain.

This project focuses on AI-based equation discovery techniques in conjunction with deep learning methods to build an improved process-based understanding of the climate system for potentially better and more robust predictions of the dynamics of large-scale climate processes which influence global weather events. These models will then be extended to include precipitation allowing predictions of precipitation statistics and probabilities of extreme events.