ecmwf

/

aifs-single-1.1

like 11

Follow

ECMWF 159

Hi AIFS/Anemoi team,

I am new to Anemoi and I am still learning the recommended workflow for fine-tuning AIFS. I am opening this as a question, not as a bug report.

I am running a small research/learning experiment to better understand how AIFS fine-tuning works and how sensitive it is to dataset and training configuration choices.

In this experiment, I created a 2024 ERA5-based Anemoi dataset and replaced the 2t field with SAMeT observed 2 m temperature regridded to the N320 grid. The goal was to test whether using observed near-surface temperature could improve forecasts, or at least help me understand the correct fine-tuning workflow.

I started from an AIFS v1 checkpoint, trained again using this modified dataset, and then compared forecasts from:

1. the original AIFS v1 checkpoint;
2. the checkpoint produced by my fine-tuning experiment.

In my evaluation, all metrics became worse with the fine-tuned checkpoint.

My main question is whether this degradation is expected in this kind of setup, or whether it suggests that I likely made a mistake in the training configuration or dataset preparation.

The training configuration I used was based on generic Anemoi defaults, with:

data: zarr
dataloader: native_grid
graph: multi_scale
model: gnn
num_channels: 128
training data: 2024-01-01 to 2024-11-30
validation data: 2024-12-01 to 2024-12-31
max_epochs: 5
limit_batches.training: 100
rollout.max: 1
learning rate: 5e-6
warm_start/input_checkpoint: aifs-single-mse-1.0.ckpt

I am unsure whether this setup is actually fine-tuning the AIFS v1 architecture correctly. In particular, I wonder whether I should instead start from the AIFS-specific configuration in anemoi-configs, rather than using the generic model: gnn setup with num_channels: 128.

I also wonder whether replacing only 2t with SAMeT observations, while keeping all other variables from ERA5, may create physically or statistically inconsistent atmospheric states and therefore degrade the model.

Could you advise what is the most likely issue here?

• an incorrect fine-tuning configuration for AIFS v1;
• an incompatibility between my model settings and the AIFS v1 checkpoint;
• a problem in the way I constructed the modified ERA5/SAMeT dataset;
• or an expected limitation of this type of small fine-tuning experiment?

I can provide the full training YAML, package versions, and dataset inspection outputs if useful.

Any guidance on the minimal recommended setup for a small AIFS v1 fine-tuning experiment would be very helpful.

Thank you.