BW: tried not freezing the atoms, restrained the dihedral by 1000 kJ, positions look the same
LW: might be good to pull out the function and check the code is running as expected
MS (in Slack):
So, another question as I try to get better instincts for this - are these parts of the potential energy surface the right ones to be matching, since the system spends less time in these higher energy configurations? I realize, of course, that the relaxed portions of the MM energy surface will not be relaxed ones of the QM system . . . I realize, also, that you all have been thinking about this plenty already, in order to make sure the QM and MM are actually being compared at correctly corresponding coordinates, rather than non-corresponding coordinates.
benchmarking infrastructure repo?
bzipping, and dumping QM coordinates, cut database size into a quarter (~50-60 MB)
MT: looking into new formats, e.g. JSON
BW: can now upload files to Zenodo
BW: could use clarification on status and how far along this will be, i.e. when it will be published. Happy to maintain
MT/BW meeting: AWS could run this for $10-15
Statistics at the end take ~1 hour
smee
BW: keep running out of memory when loading dataset
BW: looks like issue is in QCPortal and caching. Otherwise will write data to disk and reload it
LW: is batching connected to this? DCole and JHorton found this helped.
BW: that’s probably during training, this is during data formatting
LW: working with scale factors too
datasets
BW: need to get back to FPickard, MShirts about updates
