The road to 500 years of multi-member, seasonal climate hindcasts

Understanding potential drivers of seasonal prediction skill as well the non-stationarity behaviour of prediction skill itself over time is key to the development of a trustworthy, operational climate forecast system. That said, most prediction systems, either statistical or physical, are tuned on the climate of the last 30-40 years. Going into a new climate state, it is important to evaluate the underlying predictability assumptions over multiple climate states.

We present initial output of a data set version 1.0, which covers the years 1421-2008 C.E., has 100 members for each forecast step, covers the variables sea level pressure, 2m temperature and 500 hPa geopotential height and will be produced for the months January, February, June, July, August and December. This data set is produced using rather simple convolutional neural networks as architecture (same as in the initial WeatherBench approach) and is trained on reanalysis-infused atmosphere-ocean general circulation model data.

Exchanging parts of the model chain, such as model architecture, training data and initial conditions will allow the community to develop better and better versions of this data set eventually.

This data set and its future versions should be understood as an open-science, community-driven project. The code and output data behind this data set will be published openly. An exchange platform for interested community members will be highlighted during the presentation.