mace_mp vs MACECalculator, and dispersion

[mbatgh]

Hi,

I've performed energy optimizations of crystal slabs using ASE and MACE. I first used the ACE command mace_mp to load a local copy of a model (MH-1 with the omol head), and set some parameters (see below for details). Then, I tried using the command MACECalculator instead. Both commands were used with the same model file, and the same parameters, and the same input structures. Turns out the energies I get with the two commands and for the same input geometries, are very different. The absolute values differ by two orders of magnitude. Relative values (eventually I calculate surface energies) still differ by about a factor 2 (with error bars that are much smaller than that). I have a couple of questions:

1. Am I supposed to use mace_mp or MACECalculator to load/initialize the model, what is the difference between these two commands?
2. I understand the MH-1 model with the omol head was pre-trained on OMAT-24 (no dispersion correction) and then fine-tuned using data obtained at the ωB97M-VV10 (with dispersion correction) level of theory. Given that, how am I supposed to deal with dispersion correction, do I want to set that True or False?
3. Is MH-1 + omol the optimial choice for organic solids that are SALTs (organic cation + Cl-), if not what is a better alternative?

Details below, thanks for any help/advice!

michael

the system: an organic molecular crystal, a HCl salt of an nitrogen heterocycle,
the two versions of the command used for loading the model:

1. mace_calc = mace_mp(
   model='/usr/local/models/mace-mh-1.model',
   device="cuda:0"
   head="omol",
   dispersion = False,
   default_dtype="float64"
   )

2. mace_calc = MACECalculator(
   model_paths="/usr/local/models/mace-mh-1.model",
   device="cuda:0"
   default_dtype="float64",
   dispersion = False,
   head="OMOL"
   )

[ilyes319]

Hey @mbatgh,
0. the head is case sensitive, so you need to use "omol" and not "OMOL" hence the difference.

1. mace_mp enables you to load download models automatically from just pre-defined name while MACECalculator requires you download the model yourself. if you want to use mh-1, i encourage you use it just like explained here: https://huggingface.co/mace-foundations/mace-mh-1
2. If you use the omol head, put dispersion to false as it is already trained on dispersion (VV10). if you use the omat head, you need to activate the dispersion.
3. I would say mh1 omol is one of the best for that yes. other alternatives are the mace-omol model (which is trained on the full omol dataset) or the r2scan head of mh-1 or mh-0. I would benchmark these and pick the best. Note that mace-omol will be much slower than the other models.

[mbatgh]

thank you for the swift reply!
one more question: since the head keyword is case sensitive, as you suggest, does that mean that if I write head="OMOL" the head is undefined, and then the default (omat_pbe) is used?
regards
Michael

[ilyes319]

yep it defaults to omat_pbe, there should be a warning saying that.

[mbatgh]

well that's the point, I get no warning, which made me assume that OMOL works.

written to stdout is:
"Using head OMOL out of ['matpes_r2scan', 'mp_pbe_refit_add', 'spice_wB97M', 'oc20_usemppbe', 'omol', 'omat_pbe']"

and in stderr there are only two, presumably harmless, warnings:
/share/micb/mace/lib64/python3.12/site-packages/e3nn/o3/_wigner.py:10: UserWarning: Environment variable TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD detected, since theweights_only argument was not explicitly passed to torch.load, forcing weights_only=False.
_Jd, _W3j_flat, _W3j_indices = torch.load(os.path.join(os.path.dirname(file), 'constants.pt'))
/share/micb/mace/lib64/python3.12/site-packages/mace/calculators/mace.py:197: UserWarning: Environment variable TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD detected, since theweights_only argument was not explicitly passed to torch.load, forcing weights_only=False.
torch.load(f=model_path, map_location=device)

cheers,
michael