Emerging directions in limited-area AI-Driven Weather Forecasting
Leif Denby, Danish Institute of Meteorology
16 May 2025
Machine learning (ML) is rapidly transforming weather forecasting, offering new approaches that can replace traditional numerical weather prediction (NWP) models with faster and more flexible alternatives. With this presentation we wish to give an overview of recent developments and outlook with respect to weather forecasting at the Danish Institute of Meteorology.
This presentation will first describe graph-based limited-area weather forecasting models currently under development that can emulate and fully replace traditional NWP methods. Graph neural networks efficiently represent complex spatial dependencies and temporal evolution in meteorological data, allowing for high-resolution, scalable, and accurate forecasts. We will then discuss ideas on how to improve interpretability, physical consistency, uncertainty quantification and predictive capabilities in this framework using operator learning and physics-informed components, and analysing these models from the perspective of traditional numerical simulation in fluid dynamics applications. Specific efforts involve heterogeneous data integration, ensuring that models effectively utilize diverse datasets, from remote sensing to in-situ observations initially by developing techniques to build common embedding representations of heterogeneous datasets.
The second part of the talk will explore image-based (gridded) deep learning nowcasting models for precipitation and solar radiation. These models aim to provide highly localized, short-term predictions critical for applications such as renewable energy forecasting, hydrology, and severe weather warnings. We are seeking to develop models that will deliver accurate, high-temporal-resolution forecasts beyond the capabilities of traditional methods, by investigating methods for data fusion, physics-informed ML constraints, and uncertainty estimation.
About the speaker
Leif Denby is Senior Research Scientist at the Danish Institute of Meteorology (DMI) and Visiting Research Fellow at University of Leeds. He holds a MSci in Physics, MPhil in Scientific Computing and PhD in Atmopsheric Physics from University of Cambridge. His research focusses on the dynamics and effects of convective clouds on climate, and has studied these through use of satellite observations, numerical weather prediction models, large-eddy simulations and self-supervised machine learning. Since joining DMI in 2023 he has formed the DMI SciML community, co-created the mllam and mlcast communities (focussed on limited-area data-driven forecasting and deep-learning based nowcasting). He lead a DMI team to create the first Graph Neural Network trained on DANRA 2km 30-year reanalysis using the Gefion supercomputer.