Running FiniteT Mode

Running FiniteT Mode

finiteT mode is designed for sequential finite-temperature Monte Carlo sampling.

Concept

Unlike search mode, finiteT mode is a strict Markov chain:

current state -> one trial -> slow relaxation -> accept/reject -> next trial

This mode generally uses one slow worker and no fast workers.

Basic FiniteT Run

paipai \
  --input struc.in \
  --mode finiteT \
  --device cuda \
  --ngpu 1 \
  --fast 0 \
  --slow 1 \
  --steps 200000 \
  --temp 700

Resume from Search

A common workflow is:

1. run search mode;
2. choose or use the latest accepted state;
3. start finiteT from that state.

paipai \
  --mode finiteT \
  --resume-state search/mcprocess \
  --root finiteT_700 \
  --device cuda \
  --ngpu 1 \
  --slow 1 \
  --steps 200000 \
  --temp 700

If --resume-state points at an mcprocess directory, PAIPAI selects the latest numbered child.

Local Interstitial Hops

Local hop moves are useful when global swaps move interstitials too far.

--p-hop-inter 80 \
--p-swap-inter 20 \
--intsite-hop-cutoff 3.0

FiniteT includes a Hastings correction for local-hop proposals because the number of valid forward and reverse hop edges can differ.

Relaxed Interstitial Migration

In finiteT mode, if relaxation moves an interstitial atom to a different reference site, the trial is rejected. This keeps the sampled ensemble well-defined in reference-site occupation space.