rmc

Reverse Monte Carlo structural refinement: move atoms at random, and keep a move when it brings the structure's computed G(r), S(Q) or bond-angle distribution closer to experiment. The engine is C++; rmc drives it from Python, with coordinates as NumPy arrays and the refinement running with the GIL released.

pip install reverse-monte-carlo

The distribution is reverse-monte-carlo and the import is rmc. Wheels are for Linux x86-64 (manylinux_2_28, CPython 3.11–3.14) and need an AVX2 CPU; Boost is linked in statically and oneTBB rides inside the wheel. Elsewhere — Windows included — build from source (below); Windows is supported and tested, it simply has no published wheel.

Refine from a configuration

RMCConfig holds the same knobs as the RMC_run command line, validated; refine builds the engine, runs it and writes the result:

import rmc

config = rmc.RMCConfig(
    pdb_path="input.pdb",               # or lammps_path / vasp_path
    pdf_path="experimental_gr.dat",     # and/or sq_path, adf_path
    rho0=0.033,
    box_override=(20.0, 20.0, 20.0),    # or "inf"
    steps=100_000,
    out_path="refined.vasp",            # the extension picks the format
)
engine = rmc.refine(config, chi2_csv="chi2.csv")
print(engine.stats, engine.constraints.error_breakdown())

refine also takes a sampler, a selector and a step_callback. rmc.refine_ensemble(config, 4) runs four replicas in parallel and keeps the best.

Drive the engine

start = rmc.make_random_amorphous(["Cu", "Zr"], [32, 32], seed=7)
engine = rmc.Engine(start.structure, start.periodic_bc())
engine.build_atomic_groups(0.0, 0.2)            # one translation group per atom

pdf = rmc.PairDistributionConstraint()
pdf.set_experimental_data(rmc.read_xy_data("experimental_gr.dat"))
pdf.set_number_density(0.06)
pdf.set_elements(engine.structure)              # bind to engine.structure
engine.add_constraint(pdf)

engine.set_sampler(rmc.MetropolisSampler(0.01))
engine.set_track_best()
engine.run(50_000)                              # releases the GIL

best = engine.best_structure                    # a copy
curve = engine.constraints[0].concrete().computed

engine.structure.coordinates is an (N, 3) float64 view into the engine, not a copy; writing into it moves atoms.

Every class and function has a docstring, and the package ships type stubs.

Constraints and moves written in Python

Any object with compute_error(coords, moved) -> float is a constraint, and any object with generate(coords, indices) is a move:

import numpy as np

class StayNearOrigin:
    cost = 0.5                                  # optional: cheapest run first
    def compute_error(self, coords: np.ndarray, moved: np.ndarray) -> float:
        return float(np.sum(coords[:, :2] ** 2))

engine.add_constraint(rmc.Constraint(StayNearOrigin()))

rmc.ConstraintProtocol and rmc.MoveGeneratorProtocol list the optional hooks (rigid, name, modifies_species, rejection_override, ...). Each call takes the GIL, so a Python constraint costs about a microsecond more per step than a C++ one.

Ensembles

def make(i: int) -> rmc.Engine:
    engine = rmc.Engine(start.structure, start.periodic_bc())
    engine.build_atomic_groups(0.0, 0.2, seed=100 + i)
    ...                                         # constraints
    return engine

best = rmc.run_ensemble(make, 4, 20_000)        # replicas on threads, best one back

run_ensemble_cooperative shares the best structure between replicas every sync_every steps.

Analysis

g = rmc.compute_gr(start.structure, start.periodic_bc(), params=rmc.GrParams(r_max=8.0))
adf = rmc.compute_adf(start.structure, start.periodic_bc())
columns = g.as_dict()                           # {"r", "total", "Cu-Cu", ...}: pandas.DataFrame(columns)

Special quasirandom structures

import numpy as np
from rmc import mcsqs

sqs = mcsqs.enumerate(mcsqs.parse_lattice("rndstr.in"), np.diag([2, 2, 2]), {2: 3.0})
engine = mcsqs.sqs_engine(sqs, sampler=rmc.AnnealingSampler(), seed=7)
engine.run(100_000)
rmc.write_pdb(engine.best_structure, "bestsqs.pdb")

Keeping borrowed data valid

• The engine copies the structure it is given. Bind constraints to engine.structure, not to the structure you passed in.
• Constraints borrow a structure's per-atom arrays. A structure keeps its atom count once built; setters that would change it raise ValueError.
• Gradient moves point into engine.constraints. Build them after every constraint is added: engine.build_langevin_groups(...), or run_ensemble(..., prepare=...).

Every library failure raises a subclass of rmc.RmcError (IoError, AnalysisError, ConfigError, RandomStructureError, McsqsError).

Building from source

GCC 15 (or MSVC 19.43+) and CMake 3.28+. Boost, oneTBB, Catch2 and spdlog are all fetched when they are not installed:

git clone https://github.com/reach2sayan/ReverseMonteCarlo && cd ReverseMonteCarlo
CXX=g++-15 pip install .

On Windows, from an x64 Native Tools prompt:

set CMAKE_GENERATOR=Ninja
pip install .

For development, the python CMake preset builds the extension in place and runs the tests; see the repository README.