delmesher 1.0.1

Robust tetrahedral meshing via surface chamfering (SIGGRAPH 2026)

Surface Chamfering for Robust Tetrahedral Meshing

delmesher turns a triangulated input surface into a high-quality tetrahedral mesh of its interior, and is robust to the inputs that defeat classical meshers — in particular faces that meet at sharp/acute angles and faces with arbitrarily low-quality (even near-degenerate) triangulations.

An input polyhedral surface whose faces meet at acute angles and carry a low-quality triangulation (left); the chamfered surface with the acute angles removed (mid-left); the volume meshed into tetrahedra with all face angles above 14° (mid-right); and the final mesh, made exactly conformal to the input by allowing a few tetrahedra with smaller angles (right).

This is the reference implementation of the paper:

Surface chamfering for robust tetrahedral meshing. Lorenzo Diazzi, Jiacheng Dai, Daniele Panozzo, and Marco Attene. ACM Transactions on Graphics 45, 4 (SIGGRAPH 2026), Article 146. doi:10.1145/3811395 · paper (PDF)

Overview

Robust tetrahedral meshing usually forces a trade-off. Boundary-conforming methods reproduce the input surface exactly but tend to choke on sharp features and bad triangles; boundary-approximating methods are robust but no longer match the input. This work takes a hybrid route that keeps robustness while modifying the input only in a tightly controlled way. The pipeline has three stages:

1. Chamfering — the input surface is locally and minimally modified to remove the acute angles between faces that are the main obstacle to robust, well-shaped meshing.
2. Delaunay refinement — the volume bounded by the chamfered surface is tetrahedralized into well-shaped elements, with a guaranteed lower bound on the face angles (> 14°).
3. Enriched CDT — exact conformance to the original input surface is restored, introducing a few lower-quality tetrahedra only where strictly necessary.

Every geometric decision is taken with exact predicates, so the pipeline is numerically robust regardless of how degenerate the input is.

Building

Requirements

• CMake ≥ 3.10
• A C++20 compiler — tested with GCC, Clang/AppleClang and MSVC 2019+
• Network access on the first configure: dependencies are fetched automatically (SIMDe on ARM, Catch2 for the tests)

Build

git clone https://github.com/MarcoAttene/DelOptim
cd DelOptim
cmake -S . -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build --config Release

This produces the delmesher executable (build/delmesher on Linux/macOS, build/Release/delmesher.exe on Windows).

Build in Release. A build without NDEBUG keeps the internal assertions enabled and is several times slower; use -DCMAKE_BUILD_TYPE=Release for any real run.

CMake options

Option	Default	Description
CMAKE_BUILD_TYPE	(unset)	Set to Release for production runs (see above).
DELMESHER_BUILD_TESTS	ON	Build the Catch2 test suite (fetches Catch2). Set OFF to build offline / skip tests.
LGPL	ON	Build the variant covered solely by the LGPL (default). Set OFF to enable the USE_MAROTS_METHOD code path.
DELMESHER_TEST_MAX_VERTICES	2000	Refinement cap (-m) applied to each test run; empty string for unbounded test runs.

On x86, AVX2 is enabled by default; for older CPUs flip ENABLE_AVX2 / ENABLE_SSE2 at the top of CMakeLists.txt. On ARM (e.g. Apple Silicon) those x86 SIMD intrinsics are transparently emulated on NEON through SIMDe.

Running

Run with no arguments to print the full list of options:

./build/delmesher

Process the bundled example model and write the angle-bounded tetrahedral mesh:

# chamfer -> Delaunay refinement; write the well-shaped output (out_mesh.tet)
./build/delmesher -a -z -v input_models/boeing_part.off

A full run refines until convergence and can take a few minutes; pass -m N to stop after N inserted vertices (the test suite uses -m 2000). The program prints Execution correctly COMPLETED. and returns 0 on success.

Command-line options

Flag	Argument	Effect
(none)		run the full pipeline: chamfer → Delaunay refinement → enriched CDT
-a		skip the final enriched CDT phase; the Delaunay-refined output is angle-bounded (min face angle > 14°) but its constrained faces are not conformal to the input
-c		compute the minimum local feature size (LFS) of the input triangulation
-d		enable sliver removal during Delaunay refinement
-e	int	ignore input constraints whose LFS is below 10⁻ᵉ × the bounding-box diagonal (larger e ⇒ output closer to a plain CDT of the input)
-m	int	stop Delaunay refinement after this many inserted vertices
-h		print angle histograms
-l		logging mode: append a row to delOpt_log.csv
-v		verbose output

Output flags (each writes into the current directory):

Flag	Needs	Output file	Type
-b		<input>_rebuilt.off	surface
-u		chamfered_plc.off	surface
-w		DR_interface.off	surface
-x		DR_mesh.tet	volume
-y	-a	constrainedFaces.off	surface
-z	-a	out_mesh.tet	volume

Python bindings

delmesher ships nanobind Python bindings that drive the exact same pipeline as the CLI, but take and return NumPy arrays instead of OFF / .tet files.

Install

pip install .            # builds the extension via scikit-build-core

The build is self-contained (CMake + a C++20 compiler; nanobind and, on ARM, SIMDe are fetched automatically). The package follows the standard scikit-build-core layout, so pip wheel . / python -m build produce wheels ready to upload to PyPI.

Quick start

import delmesher

# vertices: (n, 3) float64;  triangles: (m, 3) integer indices
vertices, triangles = delmesher.read_off("input_models/boeing_part.off")

result = delmesher.tetrahedralize(vertices, triangles, max_vertices=2000)

result.vertices    # (V, 3) float64  -- final output mesh
result.tetrahedra  # (T, 4) uint32   -- indices into result.vertices
delmesher.write_tet("out_mesh.tet", result.vertices, result.tetrahedra)

The notebook examples/delmesher_demo.ipynb walks through the whole pipeline and visualizes every output with PyVista (pip install ".[notebook]").

Options

Every keyword argument of tetrahedralize maps to a CLI flag:

Keyword	Default	CLI flag	Effect
enriched_cdt	True	-a ⇒ False	run the final enriched-CDT phase (exact conformance); False stops after Delaunay refinement
sliver_removal	False	-d	enable sliver removal during refinement
lfs_exponent	None	-e	ignore constraints with LFS below 10⁻ᵉ × bbox diagonal
max_vertices	None	-m	stop refinement after this many inserted vertices
compute_lfs	False	-c	compute the input's minimum LFS (see result.min_lfs)
verbose	False	-v	print progress to stdout

The returned Result exposes every mesh the CLI can write, as NumPy arrays: vertices/tetrahedra (the primary output — enriched CDT, or the Delaunay-refined mesh when enriched_cdt=False), dr_vertices/dr_tetrahedra (-x), cdt_vertices/cdt_tetrahedra (-z), chamfered_vertices/ chamfered_faces (-u) and dr_interface_vertices/dr_interface_faces (-w), plus the scalars input_is_manifold, input_has_interior and min_lfs.

Each call runs the pipeline in an isolated child process, so results are reproducible regardless of call history and invalid input raises a Python exception instead of terminating the interpreter.

Python tests

tests/python/ mirrors the C++ suite (it sweeps every option on every model in input_models/) and additionally checks that the bindings reproduce the CLI's .tet output bit-for-bit. The helper script builds everything into a virtual environment under build/venv and runs the suite:

scripts/build_python_venv.sh

Testing

The Catch2 suite in tests/ runs delmesher on every model in input_models/, switching on each accepted command-line flag one at a time, and checks that the binary exits with code 0:

cmake -S . -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build --config Release
ctest --test-dir build -C Release --output-on-failure

Refinement is capped (DELMESHER_TEST_MAX_VERTICES, default 2000) so the suite stays fast while still exercising every phase; configure with -DDELMESHER_TEST_MAX_VERTICES="" for unbounded runs, or -DDELMESHER_BUILD_TESTS=OFF to skip the tests. Continuous integration builds and runs the suite on Linux (x86-64), macOS (Apple Silicon) and Windows (x86-64), in both Debug and Release, on every push.

Citation

If you use this code in your research, please cite:

@article{diazzi2026chamfering,
  author    = {Diazzi, Lorenzo and Dai, Jiacheng and Panozzo, Daniele and Attene, Marco},
  title     = {Surface Chamfering for Robust Tetrahedral Meshing},
  journal   = {ACM Transactions on Graphics},
  year      = {2026},
  volume    = {45},
  number    = {4},
  articleno = {146},
  month     = jul,
  publisher = {Association for Computing Machinery},
  doi       = {10.1145/3811395},
  url       = {https://doi.org/10.1145/3811395}
}

License

This program is free software, released under the GNU Lesser General Public License (LGPL), version 3 or (at your option) any later version. It is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See lgpl.txt and https://www.gnu.org/licenses/lgpl.txt for the full terms.

Copyright © 2019–2026 IMATI-GE / CNR and the authors. By default the build is covered solely by the LGPL; configure with -DLGPL=OFF to enable the USE_MAROTS_METHOD code path.