Guidelines on setting ewaldErrorTolerance

[lewis martin]

Hi all,

Is there a set of guidelines for changing ewaldErrorTolerance and changing the nonbonded cutoff to increase simulation speed? I'm considering a simulation where I just push everything to its limit or beyond to points considered not very sensible. One of the easy knobs to twist is to "just increase ewaldErrorTolerance" but I'm not sure if increasing it say, ten times (to 0.005 instead of 0.0005), is catastrophic, not recommended, or probably OK for some situations. Any guidance or a reference to follow up would be really appreciated. Or perhaps some context behind the choice of default - 0.0005
Thanks
Lew

[Peter Eastman]

If you increase the tolerance, that will make the forces less accurate. How far you can push it depends on what you're trying to compute. If you just want a lot of conformational exploration and you don't care if the sampling diverges a bit from a Boltzmann distribution, you could probably push it pretty far. If you care about accurate sampling or energies, that wouldn't be a good idea.

The default choice of 5e-4 was taken from the original PME paper (https://aip.scitation.org/doi/10.1063/1.470117). That was the relative force error they aimed for in their tests, and my own experiments found that it worked pretty well as a default: accurate enough for many (but not all) applications while still being pretty fast.