This repository contains the trained checkpoints for Cerrora, an AI-based regional weather model for Europe based on Microsoft's Aurora. The model is trained on the Copernicus European Regional Reanalysis (CERRA) dataset provided by ECMWF. You can find the training and inference code, as well as information about how to use the checkpoints in our GitHub repository. For more detailed information about the model, the training procedure, and the evaluation results, read our technical report (TODO: MISSING LINK).

In addition, this repository also contains a small preprocessed excerpt from the CERRA dataset and the IFS boundary conditions to facilitate easy testing of the models. Both CERRA and IFS-HRES are provided by ECMWF under the the Creative Commons Attribution 4.0 International license. The full text of this license can be found here. Compared to the original versions, our data is altered in some regards, for example we replace some variables with derived variables for compatibility with the Aurora model. This concerns the 10 meter wind, which is supplied as wind strength and direction in CERRA, and converted to u and v components. We also calculate the specific humidity from the relative humidity supplied in CERRA Additionally, we add a soil type static variable by interpolating the ERA5 soil type variable. The scripts to replicate this preprocessing can be found in the GitHub repository.

Aurora is a 1.3B parameter model consisting of an encoder that projects the input into a fixed-size latent, a backbone to process the latent, and a decoder to recreate the original data shape. The encoder and decoder use a Perceiver architecture, while the backbone is a Swin Transformer. It is pretrained on a variety of weather and climate datasets, with the goal of providing a foundation model that can be finetuned for diverse downstream tasks. For more information, read the Aurora paper. Cerrora is based on the 0.25° pretrained model with 6h lead time. We mostly leave the model architecture unchanged, with the exception of the patch size, which we increase from 4 to 8.

We adopt a two-stage training procedure consisting of a 6h pretraining stage, and a rollout finetuning stage. In the first stage, the model is trained on CERRA data to forecast the weather state in 6 hours. This done to adapt the model to the change in data domain and input resolution. The 6h model is published as cerrora-base.ckpt.

In the second stage, we finetune the model to autoregressively roll out predictions for lead times up to 30 hours. As we train a regional model, we face the issue that the CERRA input data lacks the global context necessary for performing longer forecasts. We address this by using the IFS-HRES forecasts provided in the WeatherBench2 GCP bucket as lateral boundary conditions. The rollout trained model is published as cerrora-rollout.ckpt.