On Thu, Sep 26, 2024, Pengfei Li via AMBER wrote:
>I am trying to perform minimization for one photosystem embedding with POPG
>lipid bilayer membrane. First 900 cycles of CPU minimization using pmemd
>and the next 60000 cycles of GPU minimization using pmemd.cuda can work
>well. But when switched to the last 50000 cycles of CPU parallel
>minimization using pmemd.MPI, this task failed.
Oh my! I know this may be a big system, but why are you trying to do
100000(!) steps of minimization? That's close to two orders of
magnitude more than I would think would be necessary. Is the energy
actually changing much after the first few thousand steps?
Two ideas here:
1. Run sander with ntmin=3 and maybe 1000 to 3000 steps. That should get
you a lot of relaxation away from initial bad contacts.
The default minimizer in Amber was written 50 years ago, does not understand
how to handle SHAKE restraints, and is quite inefficient. In spite of these
defects, it "works" OK for many systems, in the sense of generating a
structure that can subsequently be used for MD. But just giving it more and
more cycles is rarely a good option.
2. If the system is still not low enough in energy, run MD with something
like ntt=3, temp0=0, gamma_ln=1.0, dt=0.001. Again, only use a few thousand
steps. Play around with the exact parameters, but such a strategy can also
fix some problematic initial structures.
There are cases, such as chains passing through rings, that can only be
fixed by finding such problems and manually fixing them. Or, if the force
field is suspect (especially having hydrogens with zero vdW radius that can
still get close to other atoms with a negative partial charge), then the
force field description will need fixing.
...hope this helps...dac
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Wed Sep 25 2024 - 22:00:05 PDT
