We are trying to find an alternative method to DFT for molecular crystals structures reproduction. We start with experimentaly verifed structure and check, whatever the geometry optimization does not corrupt the input. The first tested method was pGFN-ff. The results were not acceptable. So we tested gaff2 + abcg2 charges. The results were often nonsence due to wrong charges. So we tested gaff2 + RESP charges (charges based on QM potential fitted on VdW surface - the best what can be done). The results are again useless - converting correct input to totaly different arangement. Does it mean current MM ff are simply not good enough to reproduce reality for week bonded molecular crystals ?
I think you’re very optimistic to expect any universal/general force field to accurately describe weakly bonded molecular crystals across the board. The description of “useless” is also not a scientific approach since it’s better to report statistical measures (MAE) across a definited test set, as there is no universal quantitative definition of “useless”. It should also be remembered that DFT is not likely to work brilliantly for weakly bound systems since the treatment of dispersion is often empirical and so not unlikely a force field for a key part of the interaction. If DFT + dispersion corrections works well then in some cases the dispersion correction part could be applied as a force field term with a fitted repulsion. Then there’s the whole can of worms associated with thermal and zero point effects for weakly bonded systems & so if these are not handled correctly then you have another source of discrepancy. Anyway, there is a very large literature on crystal structure prediction out there which will give you quantitative errors on standard data sets for a range of methods & so it’s possible to benefit from all this prior knowledge about the relative mean quality of different approaches.
My conclusions are based on RMSD values comparison from a test set we had created ourself (random sample from CSD, verified by DFT). I did not find any large and enough complex standard set of crystal structures for benchmarking the methods - the X23 does not reflect the behaviour of common structures. A DFT method with dispersion gives typicaly 10 lower RMSCD than MM and primary it never gives totaly different conformation with RMSD out of range. I have even succesfuly participated in the 7-th blind test of structure prediction, so I know the alternative methods. I am just surprised MM work so bad with no progres in last 20 years :-(.
Another issue is luck of periodic boundary conditions support in the energy minimization software. As far as I know only GULP, BIOVIA Forcite and LAMMPS support correctly periodic crystal structures. So the most “up to date” forcefield with e.g. polarizability can not be validated against experimental crystal structure at all. If I will be not forced to implement them myself (as I had partialy done for gaff2 and GULP in the form of atom recognition engine, latest parametr .lib generation, molecule specific .lib generation and Orca/RESP charges egnie).