Running Bris-inference over the UK domain

[emanuelegent]

Hi,

Thanks to the excellent support here, I successfully ran Bris over the original Scandinavian domain. I am now attempting to extend Bris to a UK regional domain, which I think could also be useful for other users as more Anemoi inference models become available and we aim to avoid retraining models on GPUs for similar domains.

I have a few technical questions:

1. Regional grid definition and domain awareness
   Does Bris (or regional Anemoi inference models in general) require the high-resolution grid to be explicitly defined on the target regional domain (e.g. cut out or remapped to the UK grid) as part of the input?
   If not, how does the model “recognise” or adapt to the regional domain during inference?

2. Interpolation of IFS analysis to 2.5 km
   Is there a recommended method or toolchain to interpolate the IFS 9 km analysis (or other analyses) to 2.5 km for Bris input?
   For example: conservative vs bilinear vs nearest-neighbour, or any best practice used in Bris workflows.

3. MARS disk quota issue
   When retrieving data from MARS, I encounter a “disk quota exceeded” error.
   Is there a recommended strategy to handle this (e.g. chunking requests, streaming, temporary storage, or alternative workflows)?

4. Checkpoint–input compatibility
   What is the best way to verify that my input dataset is compatible with the Bris checkpoint I am using (e.g. variables, levels, grid, normalization, metadata)?
   Is there a checklist or diagnostic tool to validate compatibility before running inference?

For reference, the checkpoint I am using is:
https://huggingface.co/met-no/bris_cloudy-skies/tree/main/2025-02

Thanks in advance for your help, I’m happy to share any findings if useful for documentation or other users.

Best regards,
Emanuele