Is there any chance to add to GULP support for MACE ML force field ? MACE is alredy integrated in ASE, Lamps, OpenMM. My test in ACE shows MACE outperforms any MM force field in quality of results. And atom typing is automatic - build in, no atom type nor bond type recognition engine required. But it have issues with electrostatic energy …
We’re working with MACE as part of a collaborative project and I agree it’s an excellent baseline for further machine learning. I’m certainly thinking about whether to link this to GULP, but there are some technical issues to overcome to link the codes…
I was benchmarking MACE on crystal structure prediction blind test data (molecular crystals). I used the MACE OFF23 Large pre-trained set. The results were identical to DFT with about 1000x less computational cost. GAFF2 and other clasical MM fields were almost useless. The MCE results were bad only for salts. MACE have in principle issue with long range electrostatic … ASE is not well designed to work with CIF with no support for electrostatic … So combine GULP and MACE can give sense …