Machine Learning for Bias Correction in Climate Models, with Application to Forecasting Heatwaves

Climate models are an imperfect representation of reality. They simplify physical processes and sacrifice spatial and temporal resolutions to stay within the limits of super-computing, introducing biases that diminish their accuracy. This thesis presents a novel machine-learning framework and model to correct these biases and provide more accurate climate statistics, focusing on heatwaves. Current correction models find it challenging to generate accurate climate statistics on heatwaves due to their difficulty in correcting temporal statistics, which require capturing dependencies across multiple consecutive time points. Our method has been tested on two case studies, Abuja (Nigeria) and Tokyo (Japan), showing improved accuracy in estimating the number of heatwaves while main- taining on par results on standard metrics. In technical terms, this thesis concentrates on statistical corrections of climate models for a single physical variable, commonly referred to as homogeneous model output statistic (MOS). Our specific interest will be in estimating temporal statistics for daily maximum temperatures. Additionally, our MOS setup covers corrections where the observed value covers a smaller geographical region than the corresponding climate model output, i.e., a spatial resolution mismatch.