Could deep learning methods replace numerical weather models?

Over the last year, there has been a rise in deep learning methods being used to produce highly accurate weather predictions. These methods include deep learning techniques such as Transformers and Graph Neural Networks amongst others, and have been applied by some of the world's leading tech companies. Some works have made claims that these deep learning methods are more accurate than the existing state-of-the-art numerical weather models. But is this claim justified? And if it's not, will there be deep learning methods in the future that can provide a better forecast than numerical models? In this talk, I will discuss the methods used and their advantage and limitations over existing numerical weather models. 

Bio:

Mariana Clare is a researcher at the European Centre for Medium Range Weather Forecasts (ECMWF), where she works on building a machine learning model for weather forecasting. She is particularly interested on how to capture the model uncertainty in these data-driven approaches. She recently received a PhD from Imperial College London, focussing on developing advanced numerical and statistical techniques to quantify uncertainty in coastal ocean models. By training she is a mathematician, having done her undergraduate degree in Mathematics at the University of Oxford.