Volume scans going to very high densities

[JPDarby]

Pre-review checklist for PR author

PR author must check the checkboxes below when creating the PR.

• I've confirmed the contribution guidelines.

Summary

This is similar to the diatomic test for the two-body but uses 3D crystals to probe higher body terms. For single element structures, 5 prototypes are formed (fcc, sc, bcc, hcp and diamond). A further 5 structures are randomly generated, subject to sensible constraints, using the pyxtal package. Multi-element terms are probed by picking 10 random binary and 10 random ternary compositions, then forming 10 random structures for each composition. The structures are then isotropically scaled across a wide range, covering extremely high and low densities. The energy and pressure are computed for each structure and these are used to form similar metrics to the diatomic test e.g. Pressure sign flips etc.

Linked issue

Resolves #308

Progress

• Calculations
• Analysis
• Application
• Documentation

Testing

Tested all models below using a subset of the elements. The

• mace-mp-0a
• mace-mp-0b3
• mace-mpa-0
• mace-omat-0
• mace-matpes-r2scan
• orb-v3-consv-inf-omat
• pet-mad

New decorators/callbacks

Discussion Points

• I've included the code to generate all structures in calc_compression.py. They are then generated by running this file directly. This could be removed but including the code here allows the benchmark to be tweaked more easily in the future.
• Currently showing one plot at a time (due to multiple curves on each plot) rather than the periodic table style used to show all dimer curves. This could be adjusted.
• Some materials should have more than one minimum in the (isotropic scaling) energy vs volume curve. See this preprint on Ce. Given this, some of the metrics, e.g. number of pressure flips, are "usually" useful but break down in some rare cases. The "Holes" metric does not suffer from this issue and identifies many-body holes at close approach in multiple models.

[ElliottKasoar]

Thanks for this, @JPDarby!

I need to go through everything much more carefully, but generally it's looking great!

I've left some (largely trivial) suggestions for now

[ElliottKasoar]

If this is very slow even on a GPU, you could consider some sort of parameterisation, similar to #388, to make it easier to parallelise.

[JPDarby]

So this allows different models to be run in parallel? The structure generation (now moved to run_compression) uses pyxtal and I can't guarantee that will be deterministic and produce the same random structures each time. Given this, I think it's best to avoid this for now. If the test really is too slow we could move back to generating the structures once and for all and then parametrize.

[ElliottKasoar]

Also note there are a couple of conflicting files to resolve, both hopefully pretty straightforward since it doesn't involve the code.