Data-driven models for weather forecasting
The field of data-driven weather forecasting has exploded in the last couple of years. Since the publication of this great review by ECMWF in mid 2023 there's been new models coming out almost every month. It is quite hard to keep up to date, and this post will probably be irrelevant in two weeks time :-).
Today it is quite straightforward to download and run models like pangu-weather, graphcast or fourcastnet in your laptop. One time step can last a few minutes in a normal CPU and few seconds if you have a GPU. But you don't even need to download these models directly. You can simply clone the ai-models repository from ecmwf and select the model you want to test. On the "foundational models" front (more on this later) there is also AtmoRep, a freely available general atmospheric model that can do multiple things, from downscaling to seasonal forecasting.
Recently ECMWF recently released a preprint that describes a new neural network model that predicts weather purely from historical observations, hence dropping the requirement of traditional data assimilation methods. The authors claim "promising accuracy for 12-hour forecasts".
New from Google: NeuralGCM, a Python library for building hybrid ML/physics atmospheric models for weather and climate simulation. which was also open sourced, since it was developed in collaboration with ECMWCF. This is Google's third iteration of a data-driven model, after GraphCast and MetNet. One thing all these models have a common is they are global. On the regional model side there is not that many models available (at least to my limited knowledge), but it is worth mentioning neural-lam. This model is based on GraphCast, but only considers a regional domain around Sweden. The Norwegian Meteorological Institute is developing a regional model based on the ai-model infrastructure from ECMWF that does a similar thing, but in this case it takes the global model developed by ECMWF and makes a finer grid around Norway.
Update (2024/09/13): I found this great list of mostly freely available models in github and I thought I'd post it here.