Atmospheric blocking plays a role in defining the general circulation pattern and contributes to many extreme weather events, such as atmospheric rives or heat waves. Accurately detecting atmospheric blocks from climate model data is important, as it will help the study and prediction of blocking in changing climate.

One of the roadblocks in assessing atmospheric blocking across different climate models is the amount of time and computer resources needed to detect blocking across the different models and scenarios, such as the CMIP project, with the traditional algorithms. While machine learning (ML) can be computationally expensive to train, it is fast to apply once the training is done. This is why we developed a global Deep Learning blocking detection algorithm that can be applied across the CMIP6 model outputs and experiments. The algorithm utilizes segmentation framework and transfer learning to detect atmospheric blocking based on 5-day 500hPa geopotential height field. The ML model was trained on binary labels generated with the TempestExtremes tracking algorithm for three separate CMIP6 models. The model was trained with daily 500 hPa geopotential height fields from 1950-2100 to guarantee the generalization capabilities of the model in the changing climate. The ML detection algorithm’s performance will be shown for individual block detection, seasonal climatologies, and standard ML performance metrics. Discussion will also briefly touch on the benefits and possible future uses for ML in detection of extreme events.

We will present results of current and future blocking climatologies across all CMIP6 data available at PCMDI for the historical and future scenario simulations (1979-2010, 2014-2100). These climatologies will provide more robust understanding of our abilities in modeling atmospheric blocks and give us more in-depth understanding of future changes given the vast amount of data used to produce these climatologies. Results will be split seasonally, and emphasis will be given to future changes and their robustness across the model ensembles and different scenarios.