Add initial emulator base class for The Well

[sgreenbury]

Closes #828.

Update

I think this now ready to merge with the following additions following discussion:

• Following discussion relating to Different AR methods for Autoemulate #830, the trainer is now sublcassed to enable custom batch formatting (e.g. a temporal encoder) and loss functions (e.g. with rollout at training time) efb0599
• Notebooks updated 4c5c33e...f18110f

Next steps

These include:

• Refining the models/hyperparams in order to get better performance on advection-diffusion and BOUT++
• Exploring whether the API requirements specified in Different AR methods for Autoemulate #830 can be accommodated with the Trainer and API here. Adding trainable temporal weights in the loss with a look ahead for e.g. m steps can be an option to test the flexibility of the implementation
• Explore potential limitations from adding the-well as a dep Explore and discuss potential limitations adding the_well as a dep #852

If the above look possible, we later might consider moving the training into the SpatioTemporalEmulator base class so this API for training and model structure is available by default in subclasses of our spatiotemporal emulator.

Notes

The deps need to be updated:

 uv sync --extra spatiotemporal --extra dev 

For the BOUT++, the torch tensor is split beforehand into the following expected files to be read for train, valid, test:

• base_path/train/data.pt
• base_path/valid/data.pt
• base_path/test/data.pt

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB