Difficulties running inference with AIFS v1.1
Hello everyone,
I’m trying to run inference with AIFS v1.1, following the approach in run_AIFS_v1.1.ipynb. I’ve tested two different setups:
- Running a forecast based on IFS data with minimal modifications to the notebook.
- Adapting the code to read ERA5 data (converted into the required AIFS input format) and performing hindcast inference on historical data.
In both cases, the inference diverges quickly from the initial conditions and evolves toward an unphysical atmospheric state. Below, I’ve included examples showing this behavior for temperature and the U component of wind in the forecast inference for a lead time of 3 days. For the hindcast experiments based on ERA5, I also compared the results against ERA5 itself, finding very large discrepancies.
Has anyone experienced a similar issue, or could suggest possible causes for this divergence?
Any insights or advice would be greatly appreciated.
Thanks in advance!
Alberto
yeah, I have ever tested, I also found the same issues, anyone else?
Hello both
Thanks for raising this. Please could you confirm the anemoi versions that you are using when running the notebook, and if they correspond to the package versions suggested at the top of the notebook?
Also, you said "minimal modifications", it would be helpful to specify these.
Thanks
Hello both
Thanks for raising this. Please could you confirm the anemoi versions that you are using when running the notebook, and if they correspond to the package versions suggested at the top of the notebook?
Also, you said "minimal modifications", it would be helpful to specify these.
Thanks
Hello,
Thanks for your help.
Yes, I was using the exact same versions of anemoi, earthkit and torch_geoemetric than in the notebook. Regarding to the modifications:
- In one set up I did not modify anything other than the data ingest. As the GPU cluster I use does not have internet access, I download the initial conditions in a separate script in a different machine and then I read them from the notebook. Other than that, no further differences with the notebook exist. The model version I use is the one given downloaded from this checkpoint under the name "aifs-single-mse-1.1.ckpt".
- In a different set up, we modified the data processing so it uses ERA5 data instead, then run using the same .ckpt.