Improving Climate and Weather Models Outputs using Statistics and Machine Learning

Speaker: Bastien François, KNMI

In order to better anticipate the short and long-term future, weather and climate models have been developed to produce weather forecasts and climate projections, based on knowledge of the important physical processes involved. However, in spite of scientific progress, weather and climate model outputs can present errors ーor statistical biasesー compared to observations, i.e., they can fail to predict short-term atmospheric conditions, or to provide an adequate representation of the longer-term climate system. These errors may concern the model output values themselves (e.g. mean or variance), but also more complex multivariate statistics such as the correlation between variables. Consequently, weather and climate models require some form of post-processing to improve the quality of their outputs and enable them to be used for subsequent applications. Can we adjust these systematic biases using statistical methods? And can we do better with recent Machine Learning tools? In this colloquium, I'll answer these questions by giving you an overview of my work on statistical biases in climate simulations, as well as the work I do at KNMI on statistical post-processing of forecast extremes. I will be giving practical advice on how to correct climate simulations and presenting the results obtained by applying machine learning methods to improve weather and climate outputs.

CV:

After studying applied statistics in Toulouse (France), Bastien François did his PhD at the “Laboratoire des Sciences du Climat et de l’Environnement” (Paris, France). During his PhD, his main research interests focused on machine learning and statistical methods for analyzing and understanding past, present and future climate variability, including extreme events and multivariate dependencies (compound events). After completing his PhD in September 2022, he joined KNMI to work as a Data Scientist to improve extreme weather forecasts using machine learning tools.