CB - Are these redundant internal coordinates so that they have linear dependence? Or are they non-redundant like a Z matrix? If they have linear interdependence, it’s harder to write down gradients.
TG - I generate the gradients in terms of linearly independent eigenvectors so that you get fewer eigenvectors than internal coordinates.
CB - What about coupling between bonded and nonbonded coordinates?
TG - The eigenvectors are derived from delocalized internal coordinates that can include coupling between bonded and nonbonded coordinates.
LPW - Like a Z matrix as a transformation of Cartesian coordinates, you can included e.g. a distance between nonbonded atoms as an internal coordinate. When you diagonalize the matrix of redundant internal coordinates, you get some zeros in the diagonals that take care of the linear dependencies. You end up with 3N or 3N-6 coordinates depending on whether you include global translation and rotation.
Slide 7
CC - What’s on the x axis here?
TG - Snapshot from doing NEB between pairs of conformers with 12 snapshots per NEB trajectory
Slide 9
LPW - I’m not surprised that MM gradients are higher than QM gradients. But I’m not sure what you’re plotting on the y axis.
TG - Norm over all DLC coordinates, i.e. sqrt of the sum of square magnitudes.
Slide 11
PB - Can you apply weights to gradients by DLC?
TG - Yes, e.g. you can weihgt them by their eigenvalue in the diagonalization
PB - What about by category, e.g. Bonds vs Angles vs Torsions? TG - Yes, but I haven’t tried that.
CB - Are you optimizing both equilibrium values and force constants?
TG - Yes, and I needed to change force constants to get a good fit.
LPW - This is a complex case. You may have better luck debugging this protocol by working with simpler molecules like hydrogen peroxide, then ethane, then butane.
CB - Since there are mismatches between valence configurations in QM and MM, you might be able to resolve the difference in gradient magnitude by doing an MM minimization with Cartesian restraints to let the valence coordinates relax. Maybe 1 kcal mol^-1 angstrom^-2 for non-hydrogen atoms.
Slide 12
CB - 12-6 LJ potential could explain this behavior. Perhaps the DEXP potential will be softer in the repulsive regime, and this will let you fit torsions more precisely.
CB - Can you look for correlations between redundant internal coordinates to find instances of steric interactions driving high torsion barriers?
LPW - I like this idea. It’s hard to do energy decompositions in QM, but we already have the decomposition in the MM energies.
LPW - The linear combination of DLCs is not unique. If this idea is not working, it could point to a suboptimal choice of DLCs.
Slide 17
TG - How do I fit NB parameters in gas phase?
PB - Dimer interaction energies
DM - Chris Raleigh and Adam Hogan have worked on this.
PB - David Sherrill has done this with SAPT.
CB - You need to change the functional form to soften the repulsive wall.
DM - I’m worried that this will balloon into a large project that is not conducive to Trevor graduating soon. Maybe we should sync with Danny Cole during the Newcastle check in.
LPW - I don’t think Trevor’s results indicate that he needs to switch functional forms and start from scratch to be successful.
LPW - Even though your MM gradients are larger than QM gradients here, the magnitude of the MM gradient is reasonable for a force expected during a simulation. The QM gradient is small because the conformer is the QM minimum.