multi_replica prd
This command selects parallel replica dynamics (PRD) [Voter1998]. The discrete-time clock update follows [Aristoff2014]; see prd_events.out.
Syntax
multi_replica prd replicas <number> [<per_gpu>] event <interval> dephase <steps> correlate <steps> distance <threshold> [quench <force_tolerance> <max_steps>] [max_dephase_retries <number>] [resume <prefix>]
replicas: positive replica count; see multi_replica for GPU allocation.event: positive interval in MD steps between event checks.dephase: nonnegative number of consecutive MD steps required in the basin. Velocities are randomized once at the start of each attempt, not between event checks. Escaping replicas return to their dephasing starting configuration and retry with new velocities. Zero disables dephasing.correlate: nonnegative number of consecutive steps without an escape required before dephasing and parallel dynamics.distance: positive maximum atomic displacement threshold in Angstrom between quenched configurations, using periodic minimum images. Must be smaller than half the shortest periodic box thickness when periodic directions are present.quench: positive FIRE force tolerance in eV/Angstrom and positive step limit; defaults are1e-4 1000. Convergence requires the largest absolute force component among the \(3N\) force components to be smaller than the tolerance. Unconverged quenches stop the run.max_dephase_retries: positive limit on dephasing trial rounds, including the first attempt; default 1000.resume: restart prefix.
Example
potential nep.txt
time_step 1
ensemble nvt_bdp 300 300 100
multi_replica prd replicas 8 event 10 dephase 5000 correlate 1000 distance 0.5
dump_xyz -1 0 1 prd_events.xyz
run 100000
Caveats
Choose the basin threshold, dephasing, and correlation times for the system of interest. The displacement test does not align translations, rotations, or atom permutations.
A detected escape triggers step-by-step block replay; simultaneous escapes select the lowest replica index. An escape followed by a return before a block-end check can be missed. Use
event 1to check every step.runcounts correlation and parallel MD steps, not accelerated physical time; see run.