mmgpy 0.16.1

Python bindings for the MMG software

mmgpy

mmgpy brings the power of MMG mesh adaptation to Python. Generate, optimize, and refine 2D, 3D, and surface meshes through a native PyVista accessor.

import pyvista as pv
import mmgpy  # noqa: F401  -- registers the .mmg accessor + Medit reader/writer

mesh = pv.read("input.mesh")
remeshed = mesh.mmg.remesh(hmax=0.1)
remeshed.save("output.vtk")

Try It

No installation needed, run directly with uvx:

# Remesh a mesh file from the command line
uvx mmgpy input.stl -o output.mesh -hmax 0.1

# Launch the interactive UI
uvx --from "mmgpy[ui]" mmgpy-ui

Installation

The recommended way to install mmgpy:

uv pip install mmgpy

This uses pre-built wheels from PyPI that bundle all native libraries (MMG, VTK), no compiler needed.

Other install methods

# pip
pip install mmgpy

# conda-forge
conda install -c conda-forge mmgpy

# With UI support
uv pip install "mmgpy[ui]"

# With elasticity-based displacement propagation
uv pip install "mmgpy[fem]"

PyVista accessor (mesh.mmg.remesh(...))

The .mmg accessor and the Medit .mesh / .meshb reader/writer (see How it works) activate whenever pyvista >= 0.48 is importable in the same environment as mmgpy. Either of these gives you that:

pip install mmgpy pyvista        # mmgpy + latest pyvista
pip install "mmgpy[pyvista]"     # same effect, version-pinned to >=0.48,<1

If pyvista isn't installed (or is older than 0.48, which predates the plugin entry-point system), the accessor simply isn't registered. import mmgpy still works, and the in-memory MmgMesh2D / MmgMesh3D / MmgMeshS API and mmgpy.mmgs.remesh(...) file path stay available.

Without PyVista

For headless / server / CI use, the slim install (pip install mmgpy, no pyvista, no VTK) exposes the C-binding mesh classes directly. Build a mesh from numpy arrays, attach fields via item syntax, and remesh in place:

import numpy as np
from mmgpy import MmgMeshS

mesh = MmgMeshS(vertices, triangles)        # numpy arrays, shape (Nv, 3) and (Nt, 3)
mesh["metric"] = sizing_array               # optional: per-vertex isotropic size
mesh.remesh(hmin=0.01, hmax=0.1, hausd=0.005)

verts_out = mesh.get_vertices()
tris_out = mesh.get_triangles()

MmgMesh2D (planar triangular) and MmgMesh3D (tetrahedral) follow the same pattern. File-based round trips are also available without pyvista via mmgpy.mmg2d.remesh(in_path, out_path, options={...}) and its mmg3d / mmgs siblings.

Using uv for project management

uv add mmgpy                 # add to project dependencies
uv tool install mmgpy        # install CLI tools globally
uv tool install "mmgpy[ui]"  # install CLI tools + UI globally

PyPI vs conda-forge

	PyPI (pip/uv)	conda-forge (Linux/macOS)
Install speed	Fast (pre-built wheels)	Slower (solver + download)
Dependencies	Bundled (self-contained)	Shared across packages
Disk usage	Larger (duplicate VTK/libs)	Smaller in conda environments
Best for	Quick setup, isolated use, CI	Scientific stacks sharing VTK/NumPy

Use PyPI (uv pip install) for the fastest setup. Use conda-forge when you already have a conda environment with VTK, PyVista, or other scientific packages.

How it works

Importing mmgpy registers a PyVista plugin that adds two things to every pv.UnstructuredGrid and pv.PolyData:

• A .mmg accessor that exposes the full MMG API: remesh, remesh_optimize, remesh_uniform, remesh_levelset, move, validate, element_qualities, and more.
• A Medit reader/writer for .mesh and .meshb files (with auto-loading of companion .sol files into point_data / cell_data).

Every accessor call returns a fresh PyVista dataset, so the result composes with the rest of the PyVista API (slicing, plotting, IO).

Features

• Multi-dimensional, 2D triangular, 3D tetrahedral, and surface meshes (auto-detected from cell types via dataset.mmg.kind).
• Local refinement, sphere / box / cylinder / point-based sizing, passed as local_sizing=[...] on remesh.
• Anisotropic adaptation, metric tensors in point_data["metric"], including least-squares Hessian recovery from a scalar field.
• Level-set discretization, extract isosurfaces from implicit functions via mesh.mmg.remesh_levelset(...); multi-material splits via set_multi_materials.
• Lagrangian motion, move boundaries and remesh through mesh.mmg.move(displacement, ...), with a Laplacian propagator or an optional elasticity backend (fedoo).
• Required entities, lock vertices, edges, triangles, or tetrahedra during remeshing via kwargs (required_triangles=...) or mmg_* data tags.
• Companion .sol I/O, scalar / vector / tensor fields via load_sol, save_sol, load_all_sols, save_all_sols.
• Validation & quality, mesh.mmg.validate(detailed=True) returns a ValidationReport; mesh.mmg.element_qualities() returns MMG's in-radius ratios.
• 40+ file formats, native Medit, plus everything PyVista supports (VTK, STL, OBJ, GMSH, MED, Abaqus, etc.; install pyvista[io] for meshio-backed formats).

Usage

Basic remeshing

import pyvista as pv
import mmgpy  # noqa: F401

mesh = pv.read("input.mesh")
remeshed = mesh.mmg.remesh(hmax=0.1)

q_before = mesh.mmg.element_qualities()
q_after = remeshed.mmg.element_qualities()
print(f"Quality: {q_before.mean():.2f} -> {q_after.mean():.2f}")

remeshed.save("output.vtk")

Local sizing

Refine inside specific regions without touching the rest of the mesh:

remeshed = mesh.mmg.remesh(
    hmax=0.1,
    local_sizing=[
        {"shape": "sphere", "center": [0.5, 0.5, 0.5], "radius": 0.2, "size": 0.01},
        {"shape": "box", "bounds": [[0, 0, 0], [0.3, 0.3, 0.3]], "size": 0.02},
        {"shape": "cylinder", "point1": [0, 0, 0], "point2": [0, 0, 1],
         "radius": 0.1, "size": 0.01},
        {"shape": "from_point", "point": [0.5, 0.5, 0.5],
         "near_size": 0.01, "far_size": 0.1, "influence_radius": 0.3},
    ],
)

Typed options

from mmgpy import Mmg3DOptions

opts = Mmg3DOptions(hmin=0.01, hmax=0.1, hausd=0.001)
remeshed = mesh.mmg.remesh(opts)

# Or use presets
remeshed = mesh.mmg.remesh(Mmg3DOptions.fine(hmax=0.05))

Anisotropic metrics

Drop a per-vertex metric on point_data["metric"] and remesh() picks it up:

import numpy as np
import mmgpy.metrics as metrics

sizes = np.full(mesh.n_points, 0.05)
mesh.point_data["metric"] = metrics.create_isotropic_metric(sizes)

remeshed = mesh.mmg.remesh()

For solution-adaptive remeshing, recover a Hessian and convert it to a metric:

from mmgpy.metrics import compute_hessian, create_metric_from_hessian

hessian = compute_hessian(vertices, triangles, field)
mesh.point_data["metric"] = create_metric_from_hessian(
    hessian, target_error=5e-3, hmin=3e-3, hmax=8e-2,
)
remeshed = mesh.mmg.remesh(hgrad=2.0)

Level-set discretization

import numpy as np

levelset = (
    np.linalg.norm(mesh.points - [0.5, 0.5, 0.5], axis=1) - 0.3
).reshape(-1, 1)

discretized = mesh.mmg.remesh_levelset(levelset)

Lagrangian motion

Apply a per-vertex displacement and remesh to maintain element quality:

import numpy as np

displacement = np.zeros((mesh.n_points, 3))
displacement[:, 0] = 0.1

moved = mesh.mmg.move(displacement, hmax=0.1)

Pass only boundary values plus propagate=True to fill the interior. The default is a Laplacian smoother; pass propagation_method="elasticity" to use the fedoo-backed linear-elasticity solver (uv pip install "mmgpy[fem]").

Locking entities

Keep specific vertices, edges, triangles, or tetrahedra fixed during remeshing:

remeshed = mesh.mmg.remesh(
    hmax=0.1,
    required_triangles=np.array([3, 7, 11], dtype=np.int32),
)

Or attach the constraint to the dataset (it travels through save / copy):

mask = np.zeros(mesh.n_cells, dtype=bool)
mask[[3, 7, 11]] = True
mesh.cell_data["mmg_required_triangles"] = mask
remeshed = mesh.mmg.remesh(hmax=0.1)

Visualization

remeshed.plot(show_edges=True)

The accessor returns a regular PyVista dataset, so anything PyVista does (slicing, integration, custom plotters) works directly on the result.

Command Line

MMG executables are bundled with the wheel:

# Auto-detect mesh type
mmg input.mesh -o output.mesh -hmax 0.1

# Or use specific commands
mmg3d input.mesh -o output.mesh -hmax 0.1
mmgs surface.stl -o refined.mesh -hausd 0.001
mmg2d domain.mesh -o refined.mesh -hmax 0.05

# Check versions
mmg --version

The _O3 suffix variants (mmg3d_O3, etc.) are also available for compatibility.

Gallery

Documentation

kmarchais.github.io/mmgpy

• Quick Start
• Tutorials
• API Reference
• Examples

Contributing

Contributions are welcome. See CONTRIBUTING.md for development setup, coding standards, and the pull request process.

License

MIT