add facilities for crystal structures and polymers

[sgbaird]

Feature request

Requires valid distance metrics for crystal structures and polymers that encode chemo-structural novelty and polymeric novelty, respectively as well as structure-based regression models. After that, just some basic plumbing.

[sgbaird]

Probably makes sense to default to MEGNet for ease of use. @sp8rks mentioned that the Liverpool group has crystal similarity measures that we can use based on an attention network. Ideally that crystal similarity measure would be packaged on PyPI (i.e. pip-installable) and have a function that takes two pymatgen Structure objects and returns a scalar (or takes two lists of pymatgen Structure objects and returns a pairwise distance matrix)

[sgbaird]

Some places that need to change:

• anything that hardcodes CrabNet; for example, the following which would need to be changed to regressor_kwargs instead:
  

  mat_discover/mat_discover/mat_discover_.py

  Lines 358 to 365 in 7c41890

  	self.crabnet_kwargs = dict(
  	mat_prop=self.mat_prop_name,
  	losscurve=False,
  	learningcurve=False,
  	verbose=self.verbose,
  	force_cpu=self.force_cpu,
  	epochs=self.epochs,
  	)

• anything that hardcodes ElMD; for example:
  

  mat_discover/mat_discover/mat_discover_.py

  Line 704 in 7c41890

  X_train.append(ElMD(comp, metric=self.novelty_prop).feature_vector)

• probably some other places

Probably best to start by modifying and testing the bare bones example:
https://mat-discover.readthedocs.io/en/latest/examples.html#bare-bones

This is something that a collaborator can modify without knowledge of the mat_discover code architecture.

EDIT: For evaluation metrics, could keep the element metrics and instead of new chemical formula, check if there's a new space group represented. Could also be new space group + new number of sites.

[sgbaird]

Based on email discussion:

Taylor brought up some great points, and I think this is an exciting project. There's been a push/encouragement both internal and external to incorporate structure into the search for high-performing, novel materials, and I think this will be a timely extension of DiSCoVeR.

Weighting

how do we weight these? Should it be tunable? Or should it be a fixed ratio of different in composition as well as structure?

For the weighting, perhaps we could use chimera as the scalarizing function. Alternatively, I think it would be interesting/best practice to use these two as separate objectives in a multi-objective optimization via e.g. expected hypervolume improvement. In other words, a mathematically robust way of collapsing multiple objectives in the context of observed data into a single number. Expected hypervolume improvement is taken care of implicitly with most sophisticated multi-objective optimization platforms.

Another option would be using an expected improvement acquisition function, except where the novelty proxy takes the place of uncertainty predictions.

How do we validate performance?

Interesting idea about recognizing new motifs. The structural prototypes from AFLOW seem relevant, since they're going for a set of canonical prototypes IIUC.

Some other issues related to validating performance

• Compare with Bayesian optimization #66
• Additional performance/novelty metric #65
• add unique elements and unique templates from extraordinary as measure of both performance + novelty #64

Comments on the plumbing to modify for structure:

• add facilities for crystal structures and polymers #33

The easiest place to start testing things out is via the mat_discover bare bones script. Today, I adapted this to use a matbench elasticity dataset with pymatgen Structures, M3GNet instead of CrabNet, and a Euclidean fingerprint-based structural distance instead of ElMD. Everything else is the same.

See the notebook below

https://colab.research.google.com/github/sparks-baird/mat_discover/blob/main/examples/structurally-aware-mat-discover-bare-bones.ipynb

When your structural distance metric of choice is ready, then that can be swapped out with the fingerprint-based structural distance. After that comes the most difficult part - validation (hence Taylor's comments). Validation can proceed in a similar fashion to the original one and/or include some extensions/modifications to how validation is performed.