Machine learning for complex spatio-temporal processes

Spatio-temporal processes are ubiquitous in science and technology, with applications in fluid dynamics, epidemiology, and weather and climate.

Recently, sophisticated machine learning models have been developed, capable of accurately predicting the evolution of such complex spatio-temporal processes.

This development is inspired by the ongoing AI revolution in fields such as computer vision and natural language processing.

In these fields, key technological advances such as self-supervised representation learning, generative modeling, and multimodal learning, today act as pillars on which modern machine learning is built.

However, there are still many challenges and open questions related to how these core technologies can be transferred and adapted to the spatio-temporal setting.

The purpose of this project is therefore to develop novel machine learning methodology for consolidating the pillars of contemporary machine learning in the spatio-temporal domain.

The focus is on generic method development, but particular attention will be given to applications related to weather and climate.

The project will result in enabling technology for the next generation of spatio-temporal foundation models, more accurate and reliable probabilistic forecasts (with implications for extreme event prediction), and novel multimodal data assimilation methods for incorporating the vast amount of observational data that is available today, such as satellite measurements of the atmosphere.