Nowcasting weather with coupled fluid dynamics and machine learning

"Nowcasting" provides weather forecasting central to issuing alerts and warnings of high-impact weather on timescales typically shorter than 6 hours. On these timescales numerical weather prediction (NWP) models cannot be utilised due to their slow runtime, and therefore forecasters generally rely on directly interpreting observations to make predictions about the evolution of weather conditions. By utilising recent advances in machine learning this challenge presents an opportunity to learn directly from data the near-time evolution of weather and through this gain new insights into atmospheric processes of storm formation. Accurate nowcasting tools are extremely useful in providing warning of imminent flash flood events, potentially providing a few hours' notice in which to take action to prevent risk to life or property.

The aim of the project is to provide precipitation nowcasting using real-time remote sensing observations (ground-based radar and satellite) to predict precipitation for the next hours. Satellite observations have the advantage of near global coverage, particularly valuable in data sparse parts of the world such as Africa, while ground based radar offer higher spatial and temporal resolution observations. The student will develop a novel approach to nowcasting by coupling a deep learning model and a simple 2D fluid dynamics model, placing the research within the realms of emerging field of physics-informed neural networks. A precipitation nowcasting system must predict both the movement and evolution of a storm. Current machine learning approaches to nowcasting tend to use neural network as a black box and learn both the processes together, but this can often lead to solutions which are unrealistically smeared out. In practice we already have considerable physical insight into the motion of storms and physics-informed neural networks provide an exciting new approach to integrating this prior knowledge to improve predictions. By utilising a 2D fluid dynamics model to take care of the horizontal advection of storms, the neural network is tasked with only predicting the formation and evolution of storms.

Based on existing research it is hypothesised that utilising a separate advection model will alleviate the numerical diffusion (smearing) seen in contemporary neural network-only approaches to nowcasting. The decoupling of transport and evolution will also allow for the introduction of physics-based constraints on the neural network solution, for example constraining the horizontal wind field to be approximately divergence free.