dt/reset variants

Hello all,

This is just a spit-ball idea, I thought I would throw out there. I assume it may be of use to others as well. The application that I am thinking of is to attractive granular systems, where the range over which attraction is present is much smaller than the repulsive ‘hard’ core(typically also in constant shear-rate).

At present, as I understand it, dt/reset is used to adjust the time-step to limit the maximum displacement of an individual particle between time integration steps. The main problem utility-wise for this fix, as I see it, is that it is applied on regular coordinates(laboratory frame). This fix does nothing to address the root of the problem, the fix is non-GI whereas most interparticle potentials are GI. One can easily demonstrate this by asking how one could expedite a (highly)sheared system using this fix. When there is a separation of length-scales such as the systems I treat, the need for this type of GI time-step adjustment becomes more apparent.

I expect that the computational expense of adding this may be quite a bit, as it involves evaluating an n(n-1)/2 matrix(without neighborlist data taken into account). But I think, it is of quite a bit of interest for these ‘not-so-smooth’ potentials(esp. when systems are athermal).

Anyway, thought I would throw this out there, for non-equilibrium, scale-separated systems this could be quite useful.

What is GI? I normally think of only using fix dt/reset when you
have high-velocity collisions (like a knock-on cascase with KeV particles),
or when un-overlapping a highly overlapped system. Why do you

need a variable timestep, and if you do dt/reset may not be the

method of choice.


By GI I just meant Galilean invariant. The main point was that in sheared systems far from equilibrium the mean motions(at the top of the box) can be much faster than the local relative velocities, which I am more concerned with. Rendering dt/reset, in its present form, quite useless.

The point may be moot… I was just trying to get around flying sand cubes in cases where my attractive granular system condenses into a network, without resorting to such small time-scales for entire simulations(they are already pretty small). This may be pathological, thought I’d throw it out.

i think there is some credit to it and there could be quite a "market"
for this in several multi-scale modeling approaches where one would
time integrate motions not in absolute terms but as a delta relative
to some other motion on a different time scale. e.g. one could model
an overall particle flow through a non-local model and then model the
specific detail within the reference frame of that vector field. that
would be conceptionally similar to, say, how you compress movies.
where you pick reference frames, then try to figure out how you can
"interpolate" between them and then only store the delta to this
interpolation. with fourier transforms and throwing away higher
coefficients (i.e. detail) for good measure.