Skip to content

LAMMPS integration

Experimental

LAMMPS support is experimental and far less tested than the ASE path. For most work, use RootstockCalculator from Python. If you try the LAMMPS styles and hit issues, reach out.

Rootstock ships two LAMMPS styles, sharing one worker protocol. Both spawn a worker subprocess that runs the MLIP in its pre-built environment and exchange positions/forces over a Unix socket each timestep:

  • pair_style rootstock — the MLIP as a genuine pair style. It contributes to thermo pe, compute pair, and the pressure natively, and composes with pair_style hybrid/scaled. Tools that assume the potential is a pair style (e.g., calphy) require this form. Use this when the MLIP is the only potential.
  • fix rootstock — adds MLIP forces on top of whatever pair style is defined. Use this to combine an MLIP with another potential; energy is exposed as the fix scalar f_<id>.

Both require units metal and a single MPI rank (the worker sees all atoms and computes its own neighborhoods; runs on more than one rank are rejected at startup).

Building the styles

The styles ship as six files (rootstock_ipi.{h,cpp}, fix_rootstock.{h,cpp}, pair_rootstock.{h,cpp}) with no dependencies beyond the C++ standard library and POSIX sockets. Copy them into your LAMMPS src/ directory and rebuild:

./lammps/install.sh /path/to/lammps/src
cd /path/to/lammps/build
cmake ../cmake [your usual flags]
make -j 4

Rootstock must be installed and on PATH so the styles can call rootstock resolve and rootstock serve:

pip install rootstock

pair_style rootstock

pair_style rootstock cluster <name> checkpoint <ckpt> \
    [device <dev>] [timeout <sec>] [cutoff <r>]
pair_coeff * * <e1> <e2> ...
Keyword Required Default Description
cluster yes — Cluster name (e.g., della)
checkpoint yes — Canonical checkpoint id (e.g., mace-mp-0-medium)
device no cuda cuda or cpu
timeout no 120 Seconds to wait for worker startup
cutoff no 1.0 Neighbor/comm bookkeeping only; the worker builds its own neighborhoods

The element symbols on pair_coeff map atom types in order (type 1 = first element, and so on).

A minimal input fragment:

units metal

pair_style rootstock cluster della checkpoint mace-mp-0-medium device cuda
pair_coeff * * Cu

thermo_style custom step temp pe press

The MLIP energy appears in pe directly — no placeholder pair style, no f_<id> bookkeeping. Per-atom energy and stress (compute pe/atom, compute stress/atom) are not provided; the worker reports global quantities only.

fix rootstock

fix <id> <group> rootstock cluster <name> checkpoint <ckpt> \
    device <dev> [timeout <sec>] elements <e1> <e2> ...
Keyword Required Default Description
cluster yes — Cluster name (e.g., delta)
checkpoint yes — Canonical checkpoint id (e.g., mace-mp-0-medium)
device no cuda cuda or cpu
timeout no 120 Seconds to wait for worker startup
elements yes — Element symbols mapping atom types (must be last)

A minimal input fragment (standalone use, with a placeholder pair style):

units metal
pair_style zero 6.0
pair_coeff * *

fix mlip all rootstock cluster delta checkpoint mace-mp-0-medium device cuda elements Cu

thermo_style custom step temp pe f_mlip press
  • Forces are added (+=) to existing forces — that is the point of the fix form; don't combine with a real pair style unless you intend to sum potentials.
  • Energy is exposed as f_<id> (variable e equal f_mlip to capture it). fix_modify <id> energy yes folds it into pe.
  • Ordering matters with force-rescaling fixes (e.g. fix ti/spring): define fix rootstock first, so its forces are present in the array when the rescaling fix runs.

Notes

  • checkpoint takes the same canonical id as RootstockCalculator and the CLI. Run rootstock list --root <root> to see what is registered on a cluster.
  • Barostats (npt, nph) work with both styles: the virial is passed through from the worker.
  • Single-node only. MPI-parallel runs are rejected; launch with mpirun -np 1 on MPI builds.