I would like to use GULP for calculating electrostatic lattice energy as an external correction term to energy calculated by MACE in ASE environment. The work is targeted on molecular crystals - salts of organic acids.
The idea is:
Points already implemented on my side
export individual molecular ions to Orca
calculate RESP charges from Orca wavefunctions
export the structure with REPS charges to GULP
In this point I was normally running geometry optimization by GAFF2 …
Can I instead just simply run single point and get only electrostatic lattice energy ?
I need to add this energy to MACE calculated energy - structure prediction ranking correction.
Personally I don’t think this will work without re-training the MACE model. Unless you can modify the electrostatics appropriately, this scheme is assuming that electrostatics is an add on to MACE, but MACE was trained against QM data that includes electrostatics (even if it can’t directly describe the long-range part). You would need to retrain the MACE model against the original data with inclusion of your correction so that this is removed from machine-learning fit to avoid duplication. The best thing to look at would be models that have been trained with Wannier functions to explicitly include electrostatic interactions on top of the short-range structural descriptors.
The MACE OFF23/24 parametrization was selectively trained on molecular crystals wit intentional exclusion of salts and ani ions from the training set. As a consequence MACE had failed to predict correctly energy of all salts from the CSD structure prediction tests. The question is whatever MACE wit 6 A cut of for the connected atoms can in principle include electrostatic …
salts were rejected
But I see now other point - if I do QM anyway to get RESP charges, why not directly calculate the energy from QM. And to use MACE only for geometry.