Question about hyperdynamics/bond boost theory

[mh11]

Hi,

I'm reading the configuration options on http://theory.cm.utexas.edu/eon/hyperdynamics.html and just want to make sure I understand them correctly. I'm very new to hyperdynamics, though I've done a lot of molecular dynamics simulations previously.

I've had a read of the two referenced papers, and my understanding is the two most important parameters are:

bb_dvmax: the magnitude of the bond-boost bias potential

and

bb_stretch_threshold: defining the bond-boost dividing surface

If I understand correctly, bb_dvmax should be of the order of the lowest activation barrier for any infrequent events you're interested in: is this a strict upper bound, or would you still get useful results as long as the number is of the correct order? The Voter paper seemed to suggest the latter, but I'm not sure how the simulation 'breaks' if your bias potential is too large. How does one go about optimising these parameters given the output EON makes available? Do you just keep making them larger until you observe unrealistic behaviour? Or do you need to manually enumerate each possible 'event', find its transition state, activation barrier, bond lengths, etc., and use those? (For my system this might be quite difficult to do every combination of events, though I know roughly what a 'typical' event will be like.)

Also, is there any danger in using a too-large bb_rcut? Presumably you want to set bb_rcut to just include nearest neighbours?

Finally, how does one use EON to find the connection between statistical MD time and 'real' time?

Many thanks for your help.

With best regards,

Matthew

[xph]

For the bb_dvmax, the ideal value is close to the lowest barrier height so that hyper potential energy surface (original plus bias) is as flat as possible. If bb_dvmax is set too high, it will prohibit efficiently sampling of boost factor in the reactant state. That means you may easily see a transition, but cannot easily get an accurate average <exp(dV/kT)>. In the end the variance of escaping time will be large.

For bb_rcut, I think including the second nearest neighbors is probably a good choice. If only the nearest neighbors are included, processes involving bond breaking with the second nearest neighbor, like two atom exchange, are not boosted. If bb_rcut is too large, some irrelevant soft motion may turn off the bias and the average boost you gain becomes small. In either case, it is purely an efficiency issue.

I didn't quite get the third question. If the "statistical MD time" means the time spent on the biased surface, EON will calculate the average boost factor and report the corresponding time on the original potential.

[mh11]

Thanks very much for this response. I'll try out my system once I have EON working with LAMMPS potential support before I ask any follow-up questions, just to get a better sense for how it works.