Can AI-based weather and climate models enhance humanitarian response and energy planning? University of Leeds in United Kingdom

Background

As climate change continues to unfold, enhanced early warning systems for natural hazards are a key part of disaster risk reduction. The emergence of AI-based weather and climate models that can outperform numerical weather prediction models (NWP) may be central to these efforts. Once trained, AI-based models require much smaller computational resources, allowing large ensembles to be run quickly for better representation of uncertainty. While exciting, more research is required to understand the extent to which these new models can aid decision making. At present, most of the models do not provide the full suite of atmospheric variables required by end users. For example, only a few models predict total precipitation (essential for early warning), and most models do not provide solar radiation at the earth’s surface – vital for applications in renewable energy. Accordingly, this PhD will explore post-processing methods to get from the output of AI models to user-relevant quantities.

PhD opportunity

This project will combine output from climate models and ensemble runs from AI-based models to explore the contribution of uncertainty in initial conditions to extreme events in both present and future climates.

The first part of the PhD will involve evaluation of the best performing AI weather forecasting models with respect to the atmospheric variables of importance to the humanitarian and renewable energy sectors. Comparisons of skill between the leading AI models and NWP will improve sector specific understanding of the extent to which the AI models can enhance decision making. The student will have the opportunity to meet forecast users in the humanitarian and renewable energy sectors.

The second part of the PhD involves developing post-processing methods to map the quantities predicted by AI-based models to user-relevant quantities in sectors of interest. Examples could include total precipitation or solar photovoltaic energy generation. For categorical events of interest (e.g. extreme rainfall), the student will identify atmospheric precursor fields and develop statistical models that use precursors to derive user-relevant quantities. The precursor fields generated by numerical and AI-based models will then be fed to the statistical models to obtain medium-range to sub-seasonal forecasts of user-relevant quantities.

These probabilistic forecasts will be analysed during periods when they are significantly better or worse than the skill of NWP, thereby shedding light on aspects of epistemic and aleatoric uncertainty.

This PhD will be co-supervised by the University of Leeds and Karlsruhe Institute of Technology and the student will have the opportunity to spend time at atmospheric science departments in both universities. The supervisory team provides practical experience of using machine learning in a humanitarian context (Dr Deva), expertise in AI for weather forecasting (Dr Quinting) and leadership in sub-seasonal to seasonal forecasting (Dr Wainwright).