We developed a deep-learning model that emulates surface winds from a complex atmospheric model. It hence permits to increase the spatial resolution of modeled wind fields in mountains (from 1300m to 30m) with low computing requirements and competitive results!

Here is a small example: DEVINE (our model) is on the left, ERA5 (large scale model used before) is on the right. The study domain is in Himalaya.

DEVINE is based on a Unet like architecture that uses convolutional neural network.

For more information, please see: https://journals.ametsoc.org/view/journals/aies/2/1/AIES-D-22-0034.1.xml