Add structure_property attribute to `Structure`

[utf]

I'd like to add a structure_property attribute to the Structure class. Sometimes it is useful to have structure-wide properties that can be stored alongside the structure. One example is the Poscar.predictor_corrector_preamble which is necessary to write structures to continue MD simulations. Currently, only site-specific properties can be stored (e.g., the velocities and predictor-corrector coordinates).

With the present Structure class, it is not possible to store structures that can be used to continue MD simulations without storing this data separately which is quite unwieldy. I'm sure there are other instances where I've wanted to do this.

@shyuep would you support me adding this feature? The API would be the same as for adding site properties. E.g.,

structure.add_structure_property("property_name", property_value)

This is related to materialsproject/atomate2#185

[shyuep]

I don't see a big issue with it, but what is the difference with structure.hello = "world"? Unlike sites and Elements which are designed to be immutable for the most part, there are no restrictions with structure.

[utf]

The only issue with setting properties like that is they won't be serialised through structure.as_dict(). I'd like the information to be serialised so that it could be stored in a database along side the structure and then used in future MD jobs.

[janosh]

I was also considering proposing this feature. The ability to store metadata on structures like DFT params used is one thing I like about AseAtoms that pymatgen is currently lacking. It would also help with provenance.

[Andrew-S-Rosen]

@utf/@janosh, I would also really love this if it could be added eventually. I'm expecting something like Structure.structure_properties with a type dict. It should also be added to Molecule.

[janosh]

Big open question is serialization. I think it would be straightforward to read and write structure_properties from/to JSON. Don't know enough about other formats like CIF, POSCAR, etc. whether they support this.

[Andrew-S-Rosen]

I don't think it really has any relevance to the CIF or POSCAR. It should only be read/written out to JSON (or perhaps, pandas Dataframe as well but nobody really uses that I imagine).

Note that if the user stores an Atoms object in structure_properties, there is a way to serialize/deserialize it even though it's not MSONable. You would do:

from ase.io.jsonio import decode, encode

encode(atoms) # serialize
decode(atoms) # deserialize

[shyuep]

I would say this:

1. There are already structure properties (e.g., charge).
2. In effect, the Structure.properties dict would essentially be for "second class" properties that are deemed not important enough to be a first class Structure.charge like property.
3. I also think that using inspection, there would be a way to serialize a structure.hello = "world" property. Might be hard to do in a super robust way, but might solve all these problems once and for all.

In any case, I have no real objection to adding properties to Structure. (But pls do not call it Structure.structure_properties).

[janosh]

But pls do not call it Structure.structure_properties).

Agreed. 😄