Deep learning techniques applied to an attribution study for heatwaves in the Iberian Peninsula

Extreme event attribution quantifies the influence of climate change on a particular extreme event (EE). Understanding the extent to which climate change is responsible for particular EE is of paramount importance because of the vulnerability of society and ecosystems to these events, especially when it comes to heatwaves that have become more frequent and intense in many parts of the world in recent decades. This led the scientific community to focus its efforts on attribution analysis and the implementation of new techniques for its study. Attribution studies of temperature EE using machine learning (ML) methods are scarce in the specialized literature. Most attribution studies perform statistical comparison between the probability of occurrence of an event today with its probability in the pre-industrial past, making it possible to determine how much more likely that event is due to climate change and how much severe it could be. However, some limitations of these classical methodologies are the difficulty in understanding the links between the physical processes responsible for the occurrence of extreme events and anthropogenic forcing and the impossibility of detecting new trends associated with this forcing. The CLImate INTelligent (CLINT) project aims, among its objectives, to design ML algorithms to improve classical attribution methodologies in some of the aforementioned limitations for three european hot-spots located in Spain, Italy and Netherlands. In this framework, this work presents a preliminary attribution analysis for summer heatwaves focused in Iberian Peninsula and based on deep learning tools such as anomaly detection with autoencoders. The autoencoder is an unsupervised method that comprises two neural networks, one to encode information and the other to decode it. The autoencoder is fed with pre-industrial realizations integrated in the framework of the Coupled Model Intercomparison Project in its sixth version (CMIP6) in such a way that it allows detecting variabilities and trends that are present in the historical run and not in the pre-industrial one. In addition, the influence of climate change for a particular temperature EE could be associated with the AE anomaly for this EE.