Machine Learning in Climate Science

We've seen a boom in DL-based weather prediction models recently, each with their own unique DL recipe. But what makes these models tick? We went through 100's of papers and wrote a review so you don't have to: arxiv.org/abs/2304.04664 #dlwp #weather #forecasting #deeplearning

Here's what's surprising: the Niño 4 box in the Central Pacific (b) is very weakly teleconnected during both Eastern Pacific and Central Pacific events. Teleconnections of both El NIño types are primarily seen only in the Eastern Pacific. Holds true also for CMIP6 model output

Now to El Niño types: Climate network curvature of daily SAT data nicely separates the +ve regional links from the -ve teleconnections. Moreover, Eastern Pacific El Niños (h) show mainly tropical teleconnections while Central Pacific EL Niños show tropical-intertropical links

1st a toy example: We use artificial data in 4 regions ensuring strong regional & weaker long-range correlations, as seen in the estimated climate network (a,b,c). Network curvature identifies dense regional links as +ve curved and weaker interregional links as -ve curved (d,e)