As part of my FORMAS project "NEEDS" (dnr 2020-00982) I trained an LSTM to predict extreme sea level using observational data around the wider North Sea + Baltic region. I now want to use it with climate model projections as input, but these need downscaling to an adequately high resolution first. More and more centres are using CNNs for this purpose, with the downside that one can only downscale one variable at a time. Luckily our work on the LSTM showed that of all the variables we could think of, only three were needed to re-create most of the signal (the 2-D fields of u- and v- components of the wind and sea level pressure). Time- and computing capacity-dependent, I would therefore either downscale more variables or re-train the network to work with these 3 only. Time will also determine how many climate change projections to consider. For my other project, funded by Rymdstyrelsen (dnr 2022-00149), our objective is to detect specific patterns in the Arctic sea ice. My PhD student has detected them the traditional way in 40-years of daily data. A logical next step is to see how well supervised and unsupervised classification methods would perform on this dataset. I am planning of using a CNN there too, but depending on the performance, I may have to move to regional analyses with different methods in each region (e.g. Gaussian Mixture Models north of Svalbard, where the distinction between the pattern that we are looking for and the marginal ice zone is unclear). I have a similar dataset for the Antarctic sea ice from a past project, so if all goes well and fast, I would then try to apply the same networks to the Southern Ocean, and see what we can learn from where it works / fails. The FORMAS project is the most time pressing, therefore the clearest here.