Quench / simulated annealing

A quench run changes the temperature during the simulation rather than holding it fixed. The usual purpose is simulated annealing - start hot so the system can explore freely, then cool slowly so it settles into a low-energy, well-equilibrated state (assembled droplets, folded structures, ordered phases). The same machinery runs in reverse, so you can also heat a configuration to melt or dissolve it.

Because a quench deliberately walks the temperature along a schedule, it is one of the few features here that does change the ensemble being sampled - that is the point. Within any single temperature window the moves are the ordinary detailed-balance moves; the quench simply resets the temperature every so often.

Turning it on

Set QUENCH_RUN : True and provide the full set of quench keywords. When a quench is active the plain TEMPERATURE keyword is ignored - the starting temperature comes from QUENCH_START instead:

QUENCH_RUN      : True
QUENCH_START    : 200      # initial temperature (TEMPERATURE is ignored)
QUENCH_END      : 40       # final temperature
QUENCH_FREQ     : 100      # change the temperature every 100 steps
QUENCH_STEPSIZE : 5        # by this much each change (always a positive number)
QUENCH_AS_EQUILIBRATION : False

Keyword	Type	Meaning
`QUENCH_RUN`	bool	Master switch. When `True`, all of the keywords below must be set.
`QUENCH_START`	float	Temperature the run begins at (replaces `TEMPERATURE`).
`QUENCH_END`	float	Target temperature the ramp finishes at.
`QUENCH_FREQ`	int	Number of steps between successive temperature changes.
`QUENCH_STEPSIZE`	float	Size of each temperature change. Always positive - the direction is inferred from `START` vs `END` (see below).
`QUENCH_AS_EQUILIBRATION`	bool	If `True`, the ramp is the equilibration phase (see below).

Cooling vs heating

You never specify a direction explicitly - PIMMS infers it from the endpoints:

QUENCH_START > QUENCH_END → a cooling run (simulated annealing).
QUENCH_START < QUENCH_END → a heating run (melting/dissolving).

QUENCH_STEPSIZE is given as a positive magnitude in both cases; internally the step is negated for a heating run so the temperature moves the right way. Every QUENCH_FREQ steps the temperature is nudged by one QUENCH_STEPSIZE toward QUENCH_END. When the next step would overshoot the target, the temperature is clamped exactly to QUENCH_END instead, and from that point on the quench flag is switched off and the remainder of N_STEPS runs at a constant QUENCH_END - so a quench always finishes with a stretch of ordinary fixed-temperature production at the final temperature.

Sizing the ramp

The ramp occupies roughly

\[\text{ramp steps} \;\approx\; \frac{|\,\text{QUENCH\_START} - \text{QUENCH\_END}\,|} {\text{QUENCH\_STEPSIZE}} \times \text{QUENCH\_FREQ}\]

steps, after which the simulation continues at QUENCH_END for whatever is left of N_STEPS. For a good anneal you generally want the ramp to be slow relative to how quickly the system relaxes: prefer many small steps (small QUENCH_STEPSIZE, generous QUENCH_FREQ) over a few large jumps, and leave enough steps after the ramp for the system to equilibrate at the final temperature.

Startup constraints (all checked before the run begins):

QUENCH_STEPSIZE must be positive.
The START → END span must be at least one QUENCH_STEPSIZE.
The ramp must fit inside N_STEPS (it cannot request more steps than the run has).

Using the ramp as equilibration

With QUENCH_AS_EQUILIBRATION : True the temperature ramp is the equilibration phase. The equilibration period is sized to cover the ramp, and once the target is reached the temperature is held at QUENCH_END for the production phase. This is the natural choice for “anneal, then measure”: all of your analysis then comes from the constant-temperature production stretch at QUENCH_END rather than from the non-equilibrium ramp.

With QUENCH_AS_EQUILIBRATION : False the ramp runs through the normal production accounting, so output written during the ramp reflects the changing temperature.

Output

Every temperature change is logged to QUENCH.dat, one tab-separated row per change (step, temperature, energy; no header line is written):

100       195.00         -1423.0000
200       190.00         -1502.0000
...

This lets you plot the energy against temperature directly - the classic view for spotting a transition (a sharp drop in energy, or a peak in its fluctuations, as the system assembles on cooling).

Interaction with TSMMC

If you combine a quench with the TSMMC moves, the TSMMC coordinator is rebuilt at the new base temperature each time the quench updates - so the temperature excursions always heat relative to the current simulation temperature. The TSMMC_FIXED_OFFSET keyword is especially convenient here, because it defines the jump temperature as an offset above the current temperature rather than as a fixed absolute value that might fall below the (falling or rising) base temperature during the ramp.

Worked example: anneal to assemble

Cool a mixture from a well-mixed hot state down to an assembly temperature, using the ramp as equilibration and then measuring at the bottom:

DIMENSIONS      : 40 40 40
PARAMETER_FILE  : params.prm
CHAIN           : 200 AABB

QUENCH_RUN      : True
QUENCH_START    : 250
QUENCH_END      : 30
QUENCH_FREQ     : 500
QUENCH_STEPSIZE : 2
QUENCH_AS_EQUILIBRATION : True

N_STEPS         : 4000000
MOVE_CRANKSHAFT : 0.8
MOVE_SLITHER    : 0.2

Here the ramp spends (1 + (250-30)/2) × 500 = 55500 steps cooling from 250 to 30 as the equilibration phase (the 1 + accounts for the initial temperature point), then the remaining steps run as production at T = 30, where the analysis is collected.