Diffusion Monte Carlo (`mmml dmc`)¶

Estimate vibrational ground-state energies with diffusion Monte Carlo on a PhysNetJax potential. Walker energies are evaluated in parallel via jax.vmap (chunked by --max-batch).

CLI reference: mmml dmc.

Quick example (acetone dimer)¶

The repo ships a 20-atom acetone dimer geometry at mmml/generate/dmc/examples/acetone_dmc.extxyz.

# Resolve a PhysNetJax checkpoint path (sets hints for $MMML_CKPT)
mmml env

mkdir -p runs/dmc_acetone_smoke

mmml dmc \
  --natm 20 \
  --nwalker 64 \
  --stepsize 5e-4 \
  --nstep 200 \
  --eqstep 50 \
  --alpha 1200.0 \
  --max-batch 64 \
  --seed 0 \
  --checkpoint "$MMML_CKPT" \
  --input mmml/generate/dmc/examples/acetone_dmc.extxyz \
  --output-dir runs/dmc_acetone_smoke

For a longer production-style run:

mmml dmc \
  --natm 20 \
  --nwalker 512 \
  --stepsize 5e-4 \
  --nstep 5000 \
  --eqstep 1000 \
  --alpha 1200.0 \
  --max-batch 512 \
  --seed 0 \
  --checkpoint "$MMML_CKPT" \
  --input mmml/generate/dmc/examples/acetone_dmc.extxyz \
  --output-dir runs/dmc_acetone

What the command does¶

Minimises the first input frame with the PhysNetJax ASE calculator (BFGS).
Perturbs that geometry with Gaussian noise (--random-sigma, default 0.02 Å).
Diffuses / branches walkers for --nstep imaginary-time steps.
Averages the reference energy after --eqstep equilibration steps.

Outputs¶

Written to --output-dir (default: current working directory), using the input stem (e.g. acetone_dmc):

File	Contents
`*.pot`	Step, alive walkers, (V_\mathrm{ref}) (hartree and cm⁻¹)
`*.log`	Run metadata and average energy
`configs_*.traj`	ASE trajectory of the last 10 steps
`defective_*.xyz`	Walkers that dipped below the reference minimum

Pass / fail smoke check¶

Process exits 0 and prints nalive each 10 steps.
*.pot has nstep + 1 lines (including step 0).
Average energy in *.log is finite (cm⁻¹ column populated).

Inputs¶

Any ASE-readable geometry (xyz, extxyz, …).
--natm must match the atom count of the first frame.
Currently supported element types: H, C, O.
--fbohr 1 if coordinates are already in Bohr (default assumes Å).

Tunables¶

Flag	Typical role
`--stepsize`	Imaginary-time step (a.u.); smaller is stabler / slower
`--alpha`	Population feedback; often ≈ `1/stepsize` scale
`--nwalker` / `--max-batch`	Walkers vs GPU/CPU memory; lower `--max-batch` if OOM
`--seed`	Reproducible noise + branching draws

Diffusion Monte Carlo (mmml dmc)¶