Jeff Whitaker

@shoyer we've got an initial version of a cycling EnKF for neuralgcm running and assimilating surface pressure observations. The results suggest that the orography that we are using for the...

Thanks @shoyer, using the learned orography really moved the needle for the data assimilation.

@shoyer for the deterministic 0.7 degree model it looks like the learned orography is in `stochastic_modular_step_model/~/dimensional_learned_weatherbench_to_primitive_with_memory_encoder/~/learned_weatherbench_to_primitive_encoder_1/~/learned_orography` and `correction_scale` should be 2.e-6. Is this correct?

Thanks @kochkov92, I was able to find `'stochastic_modular_step_model/~/stochastic_physics_parameterization_step/~/custom_coords_corrector/~/dycore_with_physics_corrector/~/learned_orography'`. However, when I use it with the `get_orography` function @shoyer provided ```python @hk.transform def get_orography(): base_orography = functools.partial( orographies.FilteredCustomOrography, orography_data_path=None, renaming_dict=dict(longitude='lon', latitude='lat'),...

please disregard my previous question - it works if I use `longitude_wavenumbers=254`

don't know why that error messages flags a non-existant numpy version. Looks like the fundamental problem is that the HDF5 library is not found. The [docs](https://unidata.github.io/netcdf4-python/) describe how to tell...

I am not familiar with pyinstaller, but I imagine it's tricky to get it to work correctly when there are many dependencies with C extensions that link external C libraries.

how about a conda environment bundled in a container (like [apptainer](https://apptainer.org/))?

not sure you should expect the quantization to leave nans alone, unless the _FillValue or missing_value is set to nan

did you report this [upstream](https://github.com/Unidata/netcdf-c)?