Reduce the distance dependence of shifted potentials for free energies if possible
The implementation issues of the new potential modifiers + free energies in #1463 were resolved, but there appears to be underlying algorithmic issues with shifted cutoffs, which still have moderate distance dependence (see the data in #1463). Researchers at D. E. Shaw appear to have found similar issues.
In the meantime, I would recommend that potential-switch be used for free energy calculations rather than potential-shift.
#1 Updated by Erik Lindahl over 5 years ago
The problem with switch is that it is more expensive to compute compared to just adding the shift constant, and the force impulse close to the cutoff might do bad things to dynamics in systems where we are sampling-limited. This gets even worse if one selects different switch settings (either rswitch or cutoff) for electrostatics and Van der Waals interactions - then we're literally doubling the cost compared to shift again.
For the long run, I think the difficult question might not necessarily be which form is most accurate - that's mostly relevant for the special case of small molecules in solvent where the runtime is no longer a big concern. For the general case where we are also limited by sampling, the question might rather be whether we benefit more from a better interaction format, or ~30% more sampling.
#2 Updated by Michael Shirts over 5 years ago
Well, yes, that's what I want to reduce the distance dependence of shift longer-term so one doesn't have to use switch to get consistent results with cutoff. The D. E. Shaw results were for protein-ligand binding, where it made even more of a difference (though I don't know exactly what parameters they were using - I'll try to get the information from them).
If one wants quick and dirty calculations to rank molecules, it doesn't matter. If one wants reliable results, then switch will be more reliable. Whether it's worth the computational cost -- that's probably something users should decide. Better yet, we find a way to make it so they don't have to.
I think that a multiscale LJ-PME will resolve some of these differences as well, as well as the more general problems caused by heterogeneous media (like protein-ligand binding) where the analytical correction will not be good enough. Shifting by etol will be irrelevant.
We've looked at the force-impulse from switch a bit (Paliwal and Shirts JCTC 2013), and it's pretty negligible for LJ as long as it's out at 1.0 nm or so, and the switch is 0.5 nm. No real problems were noticed.