Relaxations and quenches

Command line

fastmace relax NaCl.cif --model mh1 --head matpes_r2scan --optimizer gpu-LBFGS --fmax 0.01
fastmace relax NaCl.cif --model mh1 --head matpes_r2scan --optimizer gpu-FIRE --cell-filter frechet --pressure 0
fastmace quench run/ --into gpu --optimizer gpu-LBFGS

Optimizer name	Runs on
gpu-LBFGS (default)	the graph engine
gpu-FIRE	the graph engine
ase-LBFGS, ase-BFGS	ASE's optimisers on fm.calc
LBFGS, BFGS	ASE's (so existing protocol files keep their meaning)

gpu-LBFGS reproduces the annealing-benchmark minima of ASE's LBFGS.

fastmace relax writes the relaxed structure (--output, default <stem>.relaxed.extxyz) with a .json summary and a .log.

fastmace quench RUN_DIR quenches frames of a finished run, chosen as a protocol's quench picks them. Each frame keeps its own cell. --into NAME writes RUN_DIR/quench-NAME/, so one run can be quenched with several optimisers side by side. A GPU quench relaxes all its frames on one engine.

Python: Simulation.relax

import fastmace as fm

model = fm.load("mh1", head="matpes_r2scan")
opt = fm.optimize.LBFGS()                                  # ASE 3.27's defaults
opt = fm.optimize.FIRE(cell_filter=fm.optimize.FrechetCell(scalar_pressure_GPa=0.0))
sim = fm.Simulation(atoms, model, optimizer=opt, seed=0)
relaxed, summary = sim.relax(fmax=0.01)

A relaxation Simulation takes optimizer= instead of integrator= and writes no output files (pass no output=).

Python: ASE-style optimisers

from ase.filters import FrechetCellFilter
from fastmace.ase import LBFGS

opt = LBFGS(FrechetCellFilter(atoms), model="mh1", head="matpes_r2scan")
opt.run(fmax=0.01)

fastmace.ase.LBFGS and FIRE take ASE 3.27's arguments. use_line_search=True, position_reset_callback and the deprecated maxmove are refused. They take Atoms or FrechetCellFilter; other cell filters are refused with the reason.

Anneal and quench protocols

schedule=fm.protocols.PROFILES["anneal"].schedule (or the profiles slow, fast, ultrafast; in TOML just schedule = "anneal") runs a temperature schedule; fm.protocols.ramp and fm.protocols.hold build your own:

anneal = fm.protocols.ramp(0, 1000, ps=20) + fm.protocols.ramp(1000, 250, rate_K_per_ps=-3)
sim.run(schedule=anneal)

A protocol file's [relax] and [quench] tables take the same optimizer names.

See also: Integrators and optimisers, Command line and TOML.