Thanks for opening this issue. I'm going to cross-link it with #25896 which is related as it is about supporting dataframes other than pandas. I think it makes sense to try and have an overview of "all the demands" when trying to come up with a solution. At least I hope we can come up with a general abstraction that makes it simpler to add new in-/outputs.

I'll leave it to you to decide if you want to keep this issue open or close it&post a short comment about xarray in the linked issue.

I think this in particular is also very much related to the discussion we had during the sprint in Paris, where we talked about a heterogeneous container.

I don't remember all the details anymore, but I think @thomasjpfan has a good overview.

At SciPy 2023, I spoke with Tom about supporting XArray directly in scikit-learn as another "output" option. Compared to Panda's DataFrames, XArray has a bit more metadata that is "sample aligned" to be passed to the output. I'm planning to open a PR proposing a general abstraction to easily support other output containers.