Watertight Processing Method

Updated 31 December 2025

Watertight processing is a computational technique ensuring closed, topologically correct geometric representations free from gaps and non-manifold edges.
It employs methods such as CSG, octree voxelization, learning-based mesh generation, and CAD-CAM surface repair to achieve robust and precise outputs.
Applications span 3D CAD, video synthesis, microfabrication, and simulation, enabling reliable point-in-membership testing, rendering, and FEM/CFD analysis.

A watertight processing method is any computational technique that guarantees the closedness and topological integrity (absence of gaps, holes, or non-manifold edges) of geometric representations in modeling, simulation, or physical fabrication workflows. In 3D CAD, graphics, video synthesis, microfabrication, and simulation domains, watertightness is essential for reliable point-in-membership testing, rendering, mesh-based physical simulation (FEM, CFD), and robust downstream processing. The following sections present key approaches, algorithmic principles, data structures, and recent advances in watertight processing methods, referencing major developments in embedded domain simulation, surface extraction, mesh repair, video geometry, CAD-CAM integration, and microfluidic fabrication.

1. Constructive Solid Geometry (CSG) and Embedded Domain Simulation

Watertightness is guaranteed in Constructive Solid Geometry (CSG) through hierarchical Boolean operations on a set of closed primitives (cubes, spheres, cylinders, sweeps, etc.). In contrast, Boundary Representation (B-Rep) models often suffer from dirty geometries (gaps, slivers, misaligned faces), which break numerical methods for simulation and mesh extraction.

The design-through-analysis workflow using CSG and the Finite Cell Method (FCM) proceeds as follows:

CSG Tree Input: The domain $\Omega_\text{phy}$ is defined via a tree of Boolean operations on closed solid primitives.
Domain Embedding: A larger, easy-to-discretize domain $\Omega_\cup \supset \Omega_\text{phy}$ is constructed (e.g., a bounding box).
High-Order Grid Discretization: $\Omega_{\cup}$ is discretized into cells (Legendre/B-spline basis).
Point-In-Membership Testing: Gaussian integration points $x_i$ in each cell are recursively checked for membership in $\Omega_\text{phy}$ using the CSG tree.
Fictitious Domain Penalization: The characteristic function $\chi(x)=1$ in $\Omega_\text{phy}$ , $10^{-q}$ in $\Omega_\text{fict}$ ( $q \in [5,10]$ ) is used for variational assembly.
Boundary Condition Enforcement: Dirichlet and Neumann conditions applied via penalty/Nitsche or surface integration.

The core algorithm for point-in-membership recursively traverses the CSG tree:

def Inside(node, P):
    if node is leaf:
        transform P to local
        return testPrimitive(node.prim, P_local)
    else:
        A = Inside(node.child1, P)
        B = Inside(node.child2, P)
        if node.op == UNION: return A or B
        if node.op == INTERSECT: return A and B
        if node.op == DIFFERENCE: return A and (not B)

This method yields intrinsic watertightness, avoids downstream mesh repair, supports parametric updates, and achieves high-order FE accuracy. Interface quadrature and stress-singularity treatment may require adaptive octree refinement or enrichment (Wassermann et al., 2018).

2. Mesh Extraction and Topological Guarantees

Several processing methods rely on volumetric voxelization, adaptive octrees, and isosurface extraction (marching cubes style) to produce watertight manifold meshes from arbitrary input geometry.

Robust Octree-Based Pipeline (Huang et al., 2018):

Voxelization: The mesh is normalized and embedded in an enlarged bounding box. An adaptive octree is built, with occupied and empty cells.
Sign Classification: Boundary empty cells are labeled as exterior ( $\phi >0$ ), interior cells as $\phi <0$ via BFS flood.
Manifold Surface Extraction: Faces are added between occupied and exterior cells, with explicit resolution of T-junctions and ambiguous edges.
Projection and Smoothing: Final mesh vertices are iteratively projected to the original triangles and smoothed (Laplacian), yielding sub-voxel accuracy.
Topological Checks: Each edge must be incident to exactly two faces; the mesh is a closed 2-manifold.

This produces watertight meshes suitable for simulation and analysis, handling thin parts and holes, and ensuring no face-flip or non-manifold edges in routine geometric datasets.

3. Learning-Based Mesh Construction and Topological Safety

Voronoi-based boundary extraction enables learnable, discrete mesh representations that are guaranteed watertight by design (Maruani et al., 2023). The VoroMesh algorithm:

Initializes $N$ Voronoi generators $G = \{g_1, \dots, g_N\}$ , each assigned occupancy $\sigma(g_i) \in \{\text{inside}, \text{outside}\}$ .
The boundary surface $\partial\Omega$ is the union of Voronoi faces $F_{ij}$ where $\sigma(g_i) \neq \sigma(g_j)$ .
Optimization proceeds via VoroLoss, minimizing the sum of squared distances between ground-truth samples and the closest Voronoi planes.
Mesh extraction (CGAL) and occupancy assignment create closed polytopes whose boundaries automatically meet manifold, watertight criteria.
Learning pipelines predict generator positions and occupancy by sparse 3D CNNs and MLPs.

VoroMesh matches or exceeds classical methods (Marching Cubes, Dual Contouring) in Chamfer Distance, F1-score, and Normal Consistency under strict watertightness constraints.

4. Video Synthesis and Depth Watertight Meshes

In monocular video synthesis under extreme viewpoint, the Depth Watertight Mesh (DW-Mesh) representation is used to maintain geometric consistency (Hu et al., 5 Jun 2025).

Videos are first converted to per-frame dense depth maps. Each frame is reconstructed into a mesh $M_t = \{V_t, F_t, T_t, O_t\}$ via depth unprojection with enforced boundary padding ( $D_\text{max}=100$ ) to guarantee closure.
Triangle connectivity is built on regular pixel blocks and supplemented by large boundary-closing triangles.
Per-face occlusion is tagged via angle and depth continuity; degenerate triangles are labeled as occluded.
The DW-Mesh is rendered to generate color and occlusion priors for simulation masking.
A lightweight LoRA-based video diffusion adapter is trained using simulated occlusion masks for physical consistency and temporal coherence.

Boundary padding and morphological mask dilation are crucial for maintaining watertight geometry under extreme camera rotations.

5. CAD-CAM Surface Watertightization at Trim Boundaries

Surface-surface intersection in CAD (particularly with trimmed NURBS or Bézier patches) presents persistent watertightness problems due to parametric domain irregularity and poorly conditioned intersection curves (Li et al., 2023).

Trimmed patches with non-rectangular domains are reparameterized into standard $[0,1]^2$ Bézier domains via univariate polynomial mapping $T$ that straightens curved boundaries.
The intersection curve $C(w)$ between two adjacent patches is represented by a low-order spline. Its control net is used to overwrite boundary control points of both patches.
This guarantees exact $C^0$ continuity (watertightness) across patches, removing trimming artifacts and ensuring compatibility for downstream CAE (mesh stitching, simulation).
Computational complexity scales with degree elevation and patch subdivision.

This direct replacement of boundary control nets via the true intersection curve closes gaps between CAD and CAE representations.

6. Microfabrication: Watertight Physical Sealing

Microscale fabrication of leak-tight assemblies in microfluidics is achieved using SU-8 photoresist walls patterned, bonded, and characterized for water pressure, ionic leak, and vacuum compatibility (Pashayev et al., 2023).

Spin-coating, UV exposure, and high-temperature bake of SU-8 yield vertical walls (thickness $35$– $135~\mu$ m).
Mechanical compression and curing bond glass/SU-8 to SiO $_2$ substrates with no use of plasma or aggressive chemicals.
Burst pressures ($1.5$–$5.5$~bar), ionic conductivity ( $\leq 1~\mu$ S/cm), and vacuum leak rate ( $\leq 6.3\times10^{-4}~\text{mbar}\cdot\text{L}/\text{s}$ ) demonstrate robust watertight and long-term stability.

Cross-linked SU-8 offers stable, nano-porous-free walls for microchip isolation in liquid or vacuum environments.

7. Unsupervised Mesh Generation from Noisy Point Clouds

Watertight mesh generation from unorganized, incomplete sensor data employs unsupervised learning (Growing Neural Gas) with mesh-fixing post-processing (Fromm et al., 2016):

Surface generation: Modified GNG network grows a mesh on the point cloud, with node insertion driven by local error accumulation and edge age criteria.
Post-processing: Removal of close-by edges, filling holes (MeshFix), duplicate and isolated vertex elimination (VCG), and optional simplification/tetrahedralization.
Parameters (learning rates, thresholds) are automatically optimized via PSO against mesh consistency and edge-length ratio.
Final outputs are watertight, colored surface (and optionally volume) meshes suitable for physics simulation even from data with large missing regions.

Comparative Overview of Key Methods

Method	Input Type	Watertight Principle	Exemplary Domain
CSG + FCM (Wassermann et al., 2018)	Procedural CAD model	Boolean-closed sets, recursive PIM test	Mechanical simulation
Octree Isosurface (Huang et al., 2018)	Mesh	Occupied/exterior voxel BFS, marching cubes	ShapeNet, graphics
VoroMesh (Maruani et al., 2023)	Points/SDF/Grid	Voronoi faces between opposite occupancy	Learned 3D mesh
DW-Mesh (Hu et al., 5 Jun 2025)	Monocular video	Depth unprojection, padded boundary	Extreme viewpoint video
Bézier Patch (Li et al., 2023)	Trimmed NURBS/Bezier	Degree-elevated boundary curve overwrite	CAD-CAE integration
SU-8 Sealing (Pashayev et al., 2023)	Photolithography	Crosslinked photoresist wall assembly	Microfluidic chips
GNG + MeshFix (Fromm et al., 2016)	Point Clouds	Growing neural graph + hole filling	Physics simulation

References

A design-through-analysis approach using the Finite Cell Method (Wassermann et al., 2018)
Robust Watertight Manifold Surface Generation Method for ShapeNet Models (Huang et al., 2018)
VoroMesh: Learning Watertight Surface Meshes with Voronoi Diagrams (Maruani et al., 2023)
EX-4D: EXtreme Viewpoint 4D Video Synthesis via Depth Watertight Mesh (Hu et al., 5 Jun 2025)
Watertightization of Trimmed Surfaces at Intersection Boundary (Li et al., 2023)
Quantifying the performances of SU-8 microfluidic devices (Pashayev et al., 2023)
Unsupervised Watertight Mesh Generation for Physics Simulation Applications (Fromm et al., 2016)