feat: 402 time interpolation mass conserving interpolation for accumulated fields

[Rilwan-Adewoyin]

closes #402

Description

Context and Discussion on PR -->: #402

What problem does this change solve?

• Non Breaking change only affects Time Interpolator

• Implements the mass conservation for temporal interpolation of accumulated variables

• Implements a processor to ensure input accumulated data has the correct format

• Updates the interpolator config

Summary of Changes

• Mass conservation logic
• pydantic checks to ensure that mass_conservation is setup correctly, with helpful error statements
• ZeroOverwriter to set LHS boundary to 0 for specific variables
• Adds Tests for ZeroOverwiter postprocessor, Mass Conservation logic
• Adds Docs: Getting Stated with interpolator
• Adds Docs: ZeroOverwriter Postprocessor
• Ratify config.data.timestep purpose

What issue or task does this change relate to?

#402

Additional notes

As a contributor to the Anemoi framework, please ensure that your changes include unit tests, updates to any affected dependencies and documentation, and have been tested in a parallel setting (i.e., with multiple GPUs). As a reviewer, you are also responsible for verifying these aspects and requesting changes if they are not adequately addressed. For guidelines about those please refer to https://anemoi.readthedocs.io/en/latest/

By opening this pull request, I affirm that all authors agree to the Contributor License Agreement.

@Magnus-SI @radiradev @frazane @sahahner

---

📚 Documentation preview 📚: https://anemoi-training--544.org.readthedocs.build/en/544/

---

📚 Documentation preview 📚: https://anemoi-graphs--544.org.readthedocs.build/en/544/

---

📚 Documentation preview 📚: https://anemoi-models--544.org.readthedocs.build/en/544/

[anaprietonem]

Thanks Rilwan for addressing the comments and the nice work. Before merging it would be good if you can update the branch with main and check tests still pass https://github.com/ecmwf/anemoi-core/actions/runs/18645927148 and https://github.com/ecmwf/anemoi-core/actions/runs/18645957033.

@Rilwan-Adewoyin currently there seems to be some errors with flash_attention - the integration tests are trying to picked that up and failing. Please check that before merging

[anaprietonem]

the integration tests errors are likely coming from this. We don't test yet the transformer as part of our integration tests, as flash attention can't be installed and we had issues when compiling flex attention. I'd change this back to graph-transformer for now

[Rilwan-Adewoyin]

noted, updated in b16e868 - I'll check if tests are now passing soon

[cosunae]

Thanks for this contribution @Rilwan-Adewoyin this is very valuable.
Are there scientific results for the evaluation that you could attach to the PR? Perhaps a comparison of an interpolator training with and w/o this mass conserving functionality?

[frazane]

This now doesn't match the signature of the abstract method on the parent class.

[Rilwan-Adewoyin]

Thanks for highlighting this - It turns out the abstract method had the wrong signature - it seems that when step is called even the base class expects three outputs in training_step, validation_step as shown below:

train_loss, _, _ = self._step(batch)

I updated the base class _step signature in 99923bc

[Rilwan-Adewoyin]

Thanks for this contribution @Rilwan-Adewoyin this is very valuable. Are there scientific results for the evaluation that you could attach to the PR? Perhaps a comparison of an interpolator training with and w/o this mass conserving functionality?

Hi @cosunae - hope to have experiments detailing this specific thing you've mentioned done by next week. Will update when ready