We are a research team at Microsoft Research. We are pleased to share that our project, Aurora, has very recently been made fully open source, including model weights. We've attempted to package the model and weights in a way that makes it as easy as possible for anyone to get started.
On medium-term weather prediction, it sets a state-of-the-art amongst operational AI models outperforming GraphCast significantly. However, this is not what's most exciting.
Aurora has been pretrained on a large variety of data sets, unlike other AI weather models, which makes Aurora ideal to be fine-tuned for downstream atmospheric prediction tasks, e.g. by adding in new variables. In the paper (link below), we fine-tune the model for very-high-resolution weather prediction and for air pollution prediction, two areas yet largely untouched by large AI weather models.
Our hope is that the community can use Aurora by either directly leveraging its predictions or by fine-tuning it for specific applications. Feedback would be much appreciated!
We are a research team at Microsoft Research. We are pleased to share that our project, Aurora, has very recently been made fully open source, including model weights. We've attempted to package the model and weights in a way that makes it as easy as possible for anyone to get started.
On medium-term weather prediction, it sets a state-of-the-art amongst operational AI models outperforming GraphCast significantly. However, this is not what's most exciting.
Aurora has been pretrained on a large variety of data sets, unlike other AI weather models, which makes Aurora ideal to be fine-tuned for downstream atmospheric prediction tasks, e.g. by adding in new variables. In the paper (link below), we fine-tune the model for very-high-resolution weather prediction and for air pollution prediction, two areas yet largely untouched by large AI weather models.
Our hope is that the community can use Aurora by either directly leveraging its predictions or by fine-tuning it for specific applications. Feedback would be much appreciated!
Link to the paper: https://arxiv.org/abs/2405.13063