pySMESH

Python bindings to SALOME SMESH and Open CASCADE (OCCT), packaged as one self-contained Windows wheel. No SALOME platform. No CORBA. No GUI.

pySMESH gives a CFD or CAD preprocessing pipeline direct access to a production meshing and geometry kernel, through plain NumPy arrays and BREP bytes. It does not replace SALOME. It exposes the parts of SMESH and OCCT a pipeline needs, as a normal pip-installable module.

Its core is SMESH's structured meshing: mapped quadrangle faces, block-structured hexahedra, swept and extruded prisms, radial O-grids and projected mesh patterns, with exact control of the node spacing on every edge and boundary layers grown from named walls. A model is meshed per sub-shape, so structured blocks and free or body-fitted regions share one conformal mesh. See Meshing capabilities.

Meta:

• License: LGPL-2.1-only (see LICENSE, NOTICE.md)
• Platform: Windows x64, CPython 3.11 to 3.14
• Runtime dependencies: NumPy. Nothing else. SMESH, NETGEN, OCCT, Boost and VTK all ship inside the wheel.

What it covers

pySMESH covers three areas:

• Session: a stateful CAD modelling session. Build and edit a shape (primitives, booleans, fillets, chamfers, sweeps, healing). Every entity keeps a persistent id across every edit. See CAD modelling.
• Mesher: SMESH's full meshing pipeline. Assign algorithms and hypotheses per sub-shape: the structured, body-fitted and free algorithms, the 1-D spacing laws and the boundary layers listed in Meshing capabilities. Compute a mesh, then edit, query, and check its quality. It also accepts a mesh it did not build: a discrete body with no B-rep goes in as plain arrays. See Mesh generation and editing and Discrete meshes.
• Standalone OCCT geometry operations: STEP and IGES import/export, tessellation, offsets, distance and leak checks, point-in-solid classification, and geometry queries. See Geometry operations.

Two entry points serve one-shot work outside a session. compute_viscous_layers grows boundary-layer prisms on a surface mesh that you supply. unify_same_domain wraps ShapeUpgrade_UnifySameDomain for B-rep face merging that removes STEP import seams.

Meshing capabilities

pySMESH binds SALOME SMESH V9_16_0: its algorithms, its hypotheses and its assignment model. Each algorithm is a typed class in pysmesh.mesher, and each hypothesis is a frozen dataclass whose fields are the SMESH parameters. An algorithm and its hypotheses go on the whole model or on one sub-shape (a solid, a face, an edge, a vertex). So one model can mix the families below.

Structured meshing

These algorithms build structured meshes: each face or block is mapped onto a logical grid, and the node spacing follows the edge discretisation exactly.

The 1-D discretisation drives every structured mesh. Regular1D takes one spacing law per edge:

• NumberOfSegments: a fixed count, spaced uniformly, by a scale factor, by a density table, by a density expression, or by the beta law that clusters nodes towards a wall.
• Arithmetic1D, Geometric1D, StartEndLength: graded segments between stated lengths. FixedPoints1D: nodes at named positions.
• reversed_edges on each graded law makes a chain of edges grade in one direction, whatever the orientation of each edge.
• Propagation carries an edge's law to every opposite edge of a structured region. PropagOfDistribution carries its relative spacing instead.

from pysmesh import Session, load_brep
from pysmesh.mesher import (BlockRenumber, Distribution, Hexa3D, Mesher, NumberOfSegments,
                            Propagation, Quadrangle2D, Regular1D, SubShape, SubShapeKind)

s = Session()
s.add_box(1.0, 2.0, 3.0)
mesher = Mesher(load_brep(s.export_handoff().brep))
mesher.assign(Regular1D())
mesher.assign(NumberOfSegments(count=10))
# Cluster 40 nodes towards one end of edge 1, and carry that spacing to its opposite edges.
wall = SubShape(SubShapeKind.EDGE, 1)
mesher.assign(NumberOfSegments(count=40, distribution=Distribution.BETA_LAW, beta=1.02), on=wall)
mesher.assign(Propagation(), on=wall)
mesher.assign(Quadrangle2D())
mesher.assign(Hexa3D())
mesher.assign(BlockRenumber())
report = mesher.compute()    # 40 x 10 x 10 = 4 000 hexahedra

Boundary layers

Viscous layers grow from named walls, with the first layer, the growth factor and the total thickness in a closed form (first_layer_thickness). There are three ways to build them:

• Inside a Mesher: ViscousLayers on a solid, with Hexa3D, CompositeHexa3D, PolyhedronPerSolid3D, Cartesian3D, Netgen3D or Netgen1D2D3D. ViscousLayers2D on a face, with Quadrangle2D, QuadFromMedialAxis1D2D, Mefisto2D, PolygonPerFace2D, Netgen2D or Netgen1D2D. With NETGEN, prisms grow on the chosen walls and tetrahedra fill the rest. An algorithm that cannot build layers refuses them by name.
• In two steps, with ViscousLayerBuilder: Mesher.shrink_geometry returns the shape shrunk by the layer thickness. Mesh it with any algorithm, then Mesher.add_layers adds the layers onto the original shape.
• On your own surface mesh: compute_viscous_layers grows prisms on a classified surface mesh. See examples/box_bl.py.

Body-fitted and free meshing

Free 1-D sizing: LocalLength, MaxLength, AutomaticLength, and the curvature-driven Deflection1D and Adaptive1D. MaxElementVolume bounds 3-D cells, and QuadraticMesh makes second-order elements. NETGEN (netgen 6.2.2101, SALOME's NETGENPlugin V9_16_0) is built into _core; see the NETGEN guide.

Reporting

compute() returns a ComputeReport with the element counts per sub-shape and the warnings SMESH raised. A failed compute raises PysmeshError that names each failed sub-shape and the reason: a missing algorithm, a missing hypothesis, or the algorithm's own error. Hypotheses that conflict where their sub-shapes meet are refused at assign, naming the shared sub-shape.

Install

pip install pysmesh

That is the whole procedure. The wheel is self-contained: SMESH, OCCT, Boost and VTK all ship inside it. NumPy is the only thing pip pulls in.

Platform: Windows x64, CPython 3.11 to 3.14. There is one wheel per interpreter and none for other platforms. Pip picks the right one, and refuses to install on anything unsupported rather than land something that cannot import.

macOS and Linux are not built. The blocker is packaging, not the code: PyPI requires Linux wheels to be manylinux-tagged, and this build resolves Boost and VTK from conda-forge, which is a different ABI baseline. OCCT is already built from source (ci/build_occt.py), but on Windows only. Supporting Linux properly means building all three from source inside a manylinux container. That is planned separately, not skipped by oversight.

Wheels are also attached to every GitHub Release, for pinning a build by exact file.

Upgrading from 5.0: 5.1.0 adds NETGEN and completes the boundary layers. No public name is removed. A few calls that returned a wrong mesh in silence now raise PysmeshError, for example a compute that leaves a sub-shape unmeshed. See the changelog.

Upgrading from 4.x: 5.0.0 moves to SMESH 9.16 and OCCT 8.0.1 and fixes the defects of 4.2.2. No public name is removed. Some results change: 1-D distributions, Adaptive1D, bounding boxes and default measures. Some calls that returned a bad shape or mesh in silence now raise PysmeshError. Read the changelog before you upgrade.

Upgrading from 3.x: 4.0.0 removes the shared VTK requirement. Earlier versions linked the host environment's VTK and refused to import unless it was exactly 9.6.2. That constraint is gone. pySMESH now carries its own private VTK, so it no longer cares which VTK you have, or whether you have one at all. If you install with --no-deps, or gate on _build_info.VTK_VERSION, that check is now obsolete and always passes.

Quick example

import pysmesh
from pysmesh import Session
from pysmesh.session import EntityKind

s = Session()
s.add_box(3.0, 7.0, 11.0)
s.fillet(edge_ids=s.entities(EntityKind.EDGE), radius=0.5)

handoff = s.export_handoff()  # brep bytes + per-entity id arrays, ready for Mesher

mesher = pysmesh.Mesher(pysmesh.load_brep(handoff.brep))
mesher.assign(pysmesh.mesher.Regular1D())
mesher.assign(pysmesh.mesher.NumberOfSegments(count=8))
mesher.assign(pysmesh.mesher.Quadrangle2D())
mesher.assign(pysmesh.mesher.Hexa3D(), on=pysmesh.mesher.SubShape(
    pysmesh.mesher.SubShapeKind.SOLID, 1
))
mesher.compute()
mesh = mesher.mesh()

CAD modelling

Session owns one live shape and gives every entity a persistent id, so edits, undo, and mesh handoff all stay correct as the shape changes.

from pysmesh import Session
from pysmesh.session import EntityKind

s = Session()
s.add_box(3.0, 7.0, 11.0)
s.fillet(edge_ids=s.entities(EntityKind.EDGE), radius=0.5)
mark = s.snapshot()          # O(1)
s.restore(mark)              # O(1)

handoff = s.export_handoff()  # brep bytes + per-entity id arrays, ready for Mesher

Session covers primitives, curve and surface construction, sweeps, booleans with history, fillet and chamfer, transforms, healing, defeaturing, imprinting, tessellation, and geometric queries. See src/pysmesh/session/__init__.py for the full API.

Two queries answer the questions a feature filter asks. surface_parameters reads a face's analytic parameters off its surface: a radius, a cone's taper, a torus's two radii. face_wires splits a face's boundary into its loops, so an inner loop (a hole) is distinguishable from the outer one. A parameter the surface type does not define reads NaN, never a stand-in value.

import numpy as np

faces = s.entities(EntityKind.FACE)
params = s.surface_parameters(faces)

# every cylindrical face under 1 mm across: fillets and small bores
small = params.ids[(np.array(params.types) == "Cylinder") & (params.radius1 < 1.0)]

wires = s.face_wires(faces)
for row in np.flatnonzero(~wires.is_outer):        # one row per hole
    lo, hi = wires.edge_range[row]
    hole_edges = wires.edge_id[lo:hi]

Three queries answer about edges and about the space between entities. curve_geometry is the curve-side surface_parameters: a line's direction, a circle's or an ellipse's centre, plane normal and radii, with NaN on any free-form curve. curve_at gives the point and unit tangent at each parameter of one edge — along increasing parameter, never flipped for a reversed edge, because an edge shared by two faces is FORWARD in one and REVERSED in the other. distance is the exact minimum distance between any two entities, with the witness point on each.

edges = s.entities(EntityKind.EDGE)
curves = s.curve_geometry(edges)

# every circular edge under 1 mm across: the rims of small bores
is_circle = np.array(curves.type) == "Circle"
rims = curves.edge_id[is_circle & (curves.radius[:, 0] < 1.0)]

first, last = s.edge_parameter_bounds([edges[0]])[0]
sample = s.curve_at(edges[0], np.linspace(first, last, 32))
direction = sample.tangents[sample.defined]

# the clearance between two entities, with the witness point on each
gap = s.distance(faces[0], faces[2])   # gap.distance, gap.point_a, gap.point_b

Mesh generation and editing

Mesher builds a volume or surface mesh from a shape, using SMESH's own algorithm and hypothesis model. Assign an algorithm and its hypotheses to a sub-shape. Different sub-shapes can use different algorithms. Compute the mesh, then read it back as NumPy arrays.

from pysmesh import load_brep
from pysmesh.mesher import Hexa3D, Mesher, NumberOfSegments, Quadrangle2D, Regular1D
from pysmesh.mesher import SubShape, SubShapeKind

mesher = Mesher(load_brep(handoff.brep))
mesher.assign(Regular1D())
mesher.assign(NumberOfSegments(count=8))
mesher.assign(Quadrangle2D())
mesher.assign(Hexa3D(), on=SubShape(SubShapeKind.SOLID, 1))
report = mesher.compute()
mesh = mesher.mesh()

Once a mesh exists:

• Quality controls measure and classify cells: aspect ratio, skew, warping, the scaled Jacobian, orientation, and more.
• Groups name sets of elements. A group survives edits, so a wall named on a coarse mesh is still the wall after conversion to second order.
• The editor smooths, merges coincident nodes, reorients cells, splits and fuses faces, splits volumes, converts between linear and quadratic, builds the missing boundary faces or edges, sews free borders, offsets a surface, and deletes elements and nodes.
• Search locates elements at a point, casts rays through the faces or the volume cells of the mesh, finds sharp edges, and classifies a point as inside or outside a closed surface.
• The medial axis of a face reports its centreline and local wall thickness. A face can also be decomposed into blocks or have a pattern mapped onto it.
• Boundary layers come in three forms; see Boundary layers.

See src/pysmesh/mesher/__init__.py for the full model and src/pysmesh/_core.pyi for the typed API.

Discrete meshes (no CAD)

A Mesher does not need a shape. Mesher() starts empty and is filled from arrays, which is how a body that never had a B-rep gets in: an imported STL, OBJ or PLY, a shrink-wrap result, the boundary another mesher produced, or a mesh read back from a file.

from pysmesh import Mesher

mesher = Mesher.from_arrays(points, triangles)   # (N, 3) float64, (M, 3) row indices

# Divide the surface into patches. Without CAD faces, this is what a viewport picks on.
edges = mesher.sharp_edges(angle=40.0)
patches = mesher.separate_faces_by_edges(edges, name_prefix="patch_")

# Delete one patch. The report names every id that went, including the freed nodes.
gone = mesher.remove_elements(patches.at(2), free_nodes=True)
print(gone.elements, gone.nodes)

Four things are worth knowing:

• Ids are the handle. Nodes and elements keep their ids for as long as they exist, and nothing is ever renumbered. add_nodes and add_elements return the ids they created. Mesher.from_mesh(mesh_data) rebuilds a live mesh from a harvest and keeps every one of them, which is the way back from read_gmf.
• A patch index is not stable. A patch group is. Each call to separate_faces_by_edges re-derives the partition, so indices can shift once faces have been deleted. Passing name_prefix stores each patch as a group, and SMESH maintains that membership itself: a deleted element leaves the group, survivors keep their place.
• The one algorithm it takes. NetgenRemesher2D, on the whole mesh, meshes the triangles again with NETGEN, chart by chart. The new mesh replaces the old one.
• What such a mesher cannot do. Anything that resolves a sub-shape ordinal: compute with any other algorithm, assign/unassign, add_group_on_shape, the pattern mapping, smooth(in_uv_space=True), and the ElementsOnShape and Deflection2D controls. Each refuses by name. Everything else, the editor, search, quality controls, groups by id or by filter, behaves identically. Check with mesher.has_shape.

Geometry operations

Standalone Open CASCADE operations. All are headless (no VTK, no SMESH), take and return BREP bytes and NumPy arrays, and key every result to the same 1-based ordinals Shape.faces() / .edges() / .solids() use.

• read_step_xde / write_step_xde: STEP import/export through OCCT's XDE stack, preserving product names, per-face names and colours, and the file's length unit. Session.write_step names faces by session id.

• read_iges / write_iges: IGES import/export, on the same contract.

  Both readers return the geometry in the file's native unit, plus length_unit (metres per unit) and unit_name. Both writers take that unit as a required argument and declare it in the header without rescaling, so a file cannot arrive silently normalised or leave silently mislabelled. The ten accepted names are the keys of IGES_UNITS.

  Neither reader touches OCCT's global Interface_Static unit, and neither does the IGES writer. The STEP writer is the one exception: OCCT exposes no per-writer control over the STEP header's unit, so write_step_xde sets that global for the duration of one export and restores the previous value on every exit path, including an exception. Nothing leaks out of the call.

• tessellate: render-ready triangulation with per-vertex normals.

• offset_shape / make_thick_solid: B-rep offset and hollowed thick-solid operations.

• shape_distance / free_boundary_edges: exact minimum distance between two shapes, and the naked edges that localise a hole in an open shell.

• point_in_solid: exact inside test against a solid.

• unify_same_domain: real B-rep face merging that deletes the shared seam from the topology, instead of leaving a mesher hint that still forces nodes along it.

• Shape: per-entity metadata. Surface type, face adjacency, solids, and centroid-based face matching.

import pysmesh

imp = pysmesh.read_step_xde("blade.step")
shape = pysmesh.load_brep(imp.brep)
shape.faces()[0].surface_type   # "Plane" / "Cylinder" / "Cone" / "Sphere" / "Torus" / ...

mask = pysmesh.point_in_solid(imp.brep, points, tol=1e-7)

# Both formats carry the same unit contract. Coordinates stay native, the unit comes back
# with them, and a re-export declares the unit it was handed.
imp.length_unit                              # 0.001 for an MM file, 1.0 for a metre file
pysmesh.write_step_xde(imp.brep, unit=imp.unit_name)   # round trip, unit-exact

igs = pysmesh.read_iges("housing.igs")
igs.length_unit                              # 0.001 for an MM file, 0.0254 for an INCH file
pysmesh.write_iges(igs.brep, unit=igs.unit_name)

read_iges takes the content as bytes or a path, like read_step_xde. OCCT ships no IGES stream reader, so bytes go through a temporary file. No call writes to stdout or stderr: pySMESH removes the console printer from the default messenger of its private copy of OCCT, so the transfer banners and the IGES entity count are not printed.

See src/pysmesh/_core.pyi for the full typed API. mypy --strict type-checks against it.

Packaging model

SMESH and OCCT are mature, production-grade CAD and meshing libraries, but neither ships a standalone Python wrapper. Both are normally reached only through the full SALOME platform, wrapped through CORBA/SWIG. That pulls in the entire SALOME GUI and KERNEL stack. pySMESH strips both down to the static library set a CFD preprocessing pipeline needs, and exposes them as a plain, pip-installable module.

Doing that also solves a packaging problem. SMESH pulls in OCCT, Boost and VTK as dependencies. Installing those directly into a host application's environment can trigger a dependency solver cascade that downgrades unrelated packages. pySMESH's build makes that impossible by construction.

• SMESH, KERNEL, netgen and NETGENPlugin are statically linked into a single _core.pyd.
• OCCT, Boost and VTK are private to that binary. Their DLLs are bundled into the wheel and name-mangled, so they never appear in the host environment and cannot collide with the host's own copies.

Net effect on the host environment: installing the wheel adds one pip entry (pysmesh), plus NumPy. It constrains nothing else. Your application is free to use any VTK it likes, including a different version, because pySMESH never touches it.

That privacy rests on one property, which is worth stating plainly because it is the thing the design protects: no VTK object crosses the Python boundary. Every result leaves as a NumPy array or BREP bytes. A vtkUnstructuredGrid built inside _core would be an instance of a class from the private, name-mangled copy, and therefore a different C++ type from the host's, even at an identical version string. tests/test_vtk_privacy.py fails the build if a binding ever exports one.

Binary size: the wheel is 39.6 MB, holding 52 bundled DLLs. OCCT is the largest share at 18.8 MB, and private VTK costs 15.7 MB. That is the deliberate trade for zero native footprint in the host environment. _core links only three VTK components (CommonCore, CommonDataModel, FiltersVerdict), so the bundle carries 17 VTK DLLs and no rendering, IO or Python-wrapper module. netgen and NETGENPlugin are linked into _core.pyd and add no DLL; they add 1.5 MB to the wheel. CI reports the breakdown on every build and fails if the wheel would exceed PyPI's 100 MB limit.

Build from source

Requires MSVC v143, GNU patch, git, and a conda-forge build environment. VTK and Boost come from that environment. OCCT 8.0.1 does not: ci/build_occt.py builds it from the upstream tag, with our fixes from patches/occt801/ and only the toolkits _core needs. All three are build-time only, and end up inside the wheel. The build environment must not contain an occt package. CMake accepts only OCCT 8.0.1 from the prefix you name, and stops on any other.

The SALOME sources are vendored unmodified in extern/, at tag V9_16_0: SMESH, KERNEL, salome_bootstrap and GEOM's GEOMUtils, plus the MEFISTO triangulator carried forward from SMESH V9_9_0. prepare.py copies the compiled parts into staged/ and applies patches/; every patch must apply exactly. PROVENANCE.md lists every source, commit and patch.

conda env create -f ci/environment.yml
conda activate <the env name in ci/environment.yml>

# From an MSVC x64 developer shell. <deps> is any directory outside the checkout.
# A second run with the same inputs reuses the install.
python ci/build_occt.py build --root <deps>/occt-8.0.1

python prepare.py                                # stage extern/ -> staged/ and apply patches
pip wheel . --no-build-isolation --no-deps -w dist \
    -C cmake.define.PYSMESH_OCCT_ROOT=<deps>/occt-8.0.1/install
# CI additionally repairs the wheel with delvewheel, bundling the whole native closure
# (OCCT + Boost + VTK) and name-mangling every DLL:
delvewheel repair --add-path <deps>/occt-8.0.1/install/bin \
    --add-path <env>/Library/bin -w dist/repaired dist/<wheel>.whl

For local development (run tests against a freshly built extension without a wheel):

cmake -G Ninja -S . -B build -DCMAKE_BUILD_TYPE=Release \
      -DCMAKE_PREFIX_PATH=<env>/Library -DPython_EXECUTABLE=<env>/python.exe \
      -DPYSMESH_OCCT_ROOT=<deps>/occt-8.0.1/install
cmake --build build --target _core               # copies _core + _build_info into src/pysmesh
PYSMESH_OCCT_BIN=<deps>/occt-8.0.1/install/bin PYTHONPATH=src pytest tests/ -q

PYSMESH_OCCT_BIN is for this dev layout only. The in-tree _core.pyd loads the OCCT DLLs from the OCCT build, and Python does not search PATH for an extension's DLLs. tests/conftest.py and tests/golden/capture.py therefore add that directory with os.add_dll_directory. The package itself holds no such path: a wheel bundles OCCT and needs no variable. The examples import an installed pysmesh, so run them against the repaired wheel, as CI does.

Capability probe

tests/probe is a build-verification target: it constructs and runs every OCCT class and SMESH capability pySMESH depends on, against the exact link set _core uses. Run it after any change to the OCCT toolkit list, the patch series, or the StdMeshers source set. It turns a missing toolkit, a dead-stripped SMESH object, or an un-built translation unit into a named failure instead of a surprise mid-binding.

cmake -G Ninja -S . -B build -DCMAKE_BUILD_TYPE=Release \
      -DCMAKE_PREFIX_PATH=<env>/Library -DPython_EXECUTABLE=<env>/python.exe \
      -DPYSMESH_OCCT_ROOT=<deps>/occt-8.0.1/install -DPYSMESH_BUILD_V2_PROBE=ON
cmake --build build --target v2_probe
# An executable does search PATH for its DLLs: put the OCCT build's bin first.
PATH="<deps>/occt-8.0.1/install/bin:$PATH" ./build/v2_probe.exe   # exit 0 == every probed capability is usable

Design principles

• Narrow API. Every exported function exists to serve a concrete CAD or meshing pipeline need. No SWIG, no smeshBuilder emulation, no MED/CGNS I/O. Data crosses the boundary as NumPy arrays and BREP bytes.
• Fail loud. Every failure is a typed pysmesh.PysmeshError carrying the underlying SMESH/OCCT message and, where applicable, the offending ids. Never a silent best-effort fallback.

Provenance and licensing

Every vendored source and patch is traced in PROVENANCE.md. The third-party component table is in NOTICE.md.