Beware!

When using Aurora, there are a few things to watch out for.

Did you experience an issue that should be listed here, but currently is not? Please let us know by opening an issue!

Sensitivity to Data

Our hope is that Aurora generally produces sensible predictions. However, there is no guarantee that it will.

If you require optimal performance, then the data that you use needs to be exactly right. This means that you should provide exactly the right variables at exactly the right pressure levels from exactly the right source.

This also means that the performance of the model will be sensitive to how the data is regridded. For optimal performance, you should ensure that the data is regridded exactly like the data seen during pretraining and fine-tuning.

HRES IFS T0 Versus HRES IFS Analysis

HRES IFS T0 is not the same as HRES IFS analysis. Crucially, the analysis product includes an additional surface assimilation step.

Specific versions of Aurora require specific versions of HRES IFS: Aurora 0.25° Fine-Tuned requires IFS HRES T0, and Aurora 0.1° Fine-Tuned requires IFS HRES analysis.

Deterministic and Reproducible Output

If you require deterministic and reproducible output, you should do two things:

1. Set torch.use_deterministic_algorithms(True) to make PyTorch operations deterministic.

2. Set model.eval() to disable drop-out.

Loading a Checkpoint Onto an Extended Model

If you changed the model and added or removed parameters, you need to set strict=False when loading a checkpoint Aurora.load_checkpoint(..., strict=False). Importantly, enabling or disabling LoRA for a model that was trained respectively without or with LoRA changes the parameters!