The goal of this project is to use deep learning to perform statistical downscaling of simulated climate data, from a global grid with a large distance between discretization points to a regional grid with a smaller distance between discretization points.
We have already created a neural network that is based on the following paper:
* Configuration and intercomparison of deep learning neural models for statistical downscaling
This paper describes how to train a network to downscale ERA-5 input data with E-OBS data as ground truth, and we have made improvements upon the network architecture described in it.
Upcoming, we plan to support more datasets, including using the CERRA dataset as ground truth where we today use the E-OBS dataset, as CERRA is both bigger and more highly resolved than E-OBS.
We will also attempt to train one network such that we can switch between different input datasets coming from different global climate models, by drawing inspiration from the following paper:
* Downscaling Multi-Model Climate Projection Ensembles with Deep Learning (DeepESD): Contribution to CORDEX EUR-44
If we can successfully do this, we expect that we can substitute this form of statistical downscaling for some of the dynamic downscaling that we do today; the latter basically requires a lot of extra simulation, and is much more computationally expensive than the former. We therefore expect to be able to reduce the amount of computation that we need to do by a significant amount.