Hello @asmeyer2012,

Thank you for providing the PR and describing the changes. I have not tested the behavior for this different kind of structural arrays, but from what I can tell, the adjustments make sense to me.

Independent of the structural parts, the meta data adjustments should be implemented for both the fft_file_complex_structural_array and the original method. I would slightly adjust it though to prevent a trailing & char in the scenario where meta is not None but max_momentum is :

meta = dset.attrs.get("meta", "")
if max_momentum is not None:
    meta = str(meta) + "&" if meta else ""

From what I can tell, the proposed change for structural arrays leaves all the core functionality of fft_file in place but changes the underlying assumptions of the dset shape and the variables arr, axes and axis.

I believe now (different compared to the comment in the issue I have made), that it is cleaner to factor this out (including assumptions) in a new infer_fft_array function along the lines of

def infer_fft_array(dset, **kwargs):
   """Computes axes, axis and shape parameters for fft array"""
  # scenario one (old case)
  if dset.shape == ...
    array = dset[()]
    array = (array.T[0] + array.T[1] * 1j).T
    axes = (-1, -2, -3)
    axis = {0: "x", 1: "y", 2: "z"}
  # structural array case
  elif ...:
     dtype_shape = tuple()
     while len(dt.shape) > 0:
         ...
   else:
       raise ValueError("Data set shape and structure did not match expectations. (e.g., ...)"
  return array, axes, axis

This allows to just keep one FFT method simplifying future maintenance

The FFT method would then look like

def fft_file(...):
    ...
     if chunk_size is None:
        arr, axes, axis = infer_fft_array(dset)
    ...

It might be possible to re-use the infer_fft_axes method on other parts.

I am not entirely sure yet if we want to automate decision making by just parsing shapes (what would the elif in infer_fft_array look like?) or if we should make it explicit (i.e., based on a kwarg which is also specified at the fft_file level like dset_kind="structural"?

Also it may make sense to unify the output of infer_fft_array to always return non-structural arrays (so that eventual data is of the same type).

What do you think?

If this makes sense to you, I'd create an update based on your PR.