📄 other

Coord-Based Surface Intersection for Semantic CAD Modeling

This paper presents a coordinate-based surface intersection method that represents intersection curves as sequences of residual-audited contact frames containing semantic and geometric metadata, demonstrating superior numerical precision and robust Boolean closure compared to OpenCASCADE Technology across a diverse suite of primitive and complex surface intersection cases.

Original authors: GuoJun Pan

Published 2026-07-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: GuoJun Pan

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are an architect designing a futuristic city, but instead of blueprints, you are building with pure math. In this world, every wall, pipe, and dome is defined by invisible equations. To build a complex structure, you often need to cut one shape with another—like slicing a sphere with a cylinder to make a pipe hole. The line where they meet is called a "surface intersection." For decades, computer programs have been good at drawing this line just to show you where the cut is, kind of like a chalk outline on a sidewalk. But if you want to actually build the pipe, the hole, or the joint, you need more than just a drawing. You need to know exactly how the two shapes touch, the angle of the cut, and whether the math holds up perfectly at every single point. If the math is even a tiny bit off, your digital pipe might leak, or your 3D printer might fail. This is the problem of "surface intersection" in Computer-Aided Design (CAD): how do we find that meeting line not just as a picture, but as a precise, trustworthy guide for building?

This paper introduces a new way to find that meeting line, called the "Coord-Based Surface Intersection" method. Think of the old way as taking a photo of the cut line; it looks right, but if you zoom in, the pixels might be a little fuzzy. The new method is like sending a team of surveyors to walk along that line. At every step, each surveyor stops, plants a flag, and records a full "contact frame." This frame isn't just a dot; it's a tiny, 3D compass that tells you exactly where you are, which way the line is going, and what the two walls are doing right there. The authors call this a "residual-audited contact frame," which is a fancy way of saying, "We checked our math against the original rules, and we are confident we are exactly on the line."

The researchers tested this new surveyor team against the industry standard, a massive software engine called OpenCASCADE (OCCT), which is like the "Google Maps" of CAD geometry. They set up a series of challenges, from simple spheres meeting cylinders to tricky shapes like tilted cylinders hitting donuts (toruses) and cones. In the simple, perfect cases where they knew the exact answer (the "ground truth"), the new surveyor team was incredibly precise, hitting the mark with an error so small it was around $3.7 \times 10^{-14}$ . The industry standard was also very good, but slightly less precise, with errors around $3.6 \times 10^{-8}$ . To put that in perspective, the new method was millions of times closer to the perfect mathematical line in these tests.

However, the paper is careful not to claim that this new method has "won" or replaced the old one. The authors explicitly state that their method is not a universal fix for every possible shape or every tricky, messy situation in the real world. They tested it on specific families of shapes (spheres, cylinders, cones, etc.) and found that while their method produces a "contact frame" that is mathematically tighter and more transparent for building features like Boolean seams (the joints where shapes merge), the old industry standard is still a very strong, reliable baseline for general use. The new method shines because it gives the builder more information at every point, making it easier to create clean, crack-free joints in complex designs.

In a final test, the authors used their new method to build two specific types of 3D objects: a sphere with a cylinder cut out of it, and a sphere with a donut shape cut out. They ran 30 different variations of these cuts, changing the angles and sizes. Every single one passed a strict "topology audit," meaning the resulting 3D shapes were perfectly closed with no gaps, no dangling edges, and no weirdly flipped surfaces. The "seam" where the cut happened had zero gap. This proves that for these specific, exact cuts, the new method can take the math all the way from the intersection line to a finished, buildable 3D object without breaking.

So, what's the big takeaway? The paper doesn't say the old way is broken; it says the new way is a more "honest" and detailed tool for specific jobs. It turns a simple line into a rich, data-packed guide that tells you exactly how the shapes interact. While it hasn't solved every problem in the universe of 3D modeling, it shows that by treating intersection lines as a sequence of detailed, math-checked frames, we can build more reliable and precise digital structures. It's a step forward in making sure that when we say two shapes meet, they really, truly meet, down to the tiniest fraction of a pixel.

Technical Summary: Coord-Based Surface Intersection for Semantic CAD Modeling

1. Problem Statement

Surface-surface intersection is a fundamental operation in CAD, Boolean modeling, trimming, and feature generation. While classical workflows typically return a geometric locus (a curve, edge, or polyline), downstream semantic modeling often requires richer data. A Boolean seam, hole boundary, or pipe branch needs to know not just the point location, but also:

The source surfaces that produced the curve.
The local tangent direction compatible with both surfaces.
The individual source normals at the contact point.
Residual information quantifying how well the point satisfies the original surface constraints.
Topological and semantic labels connecting the point to its construction history.

Standard representations (e.g., detached polylines or mesh intersections) often lose this semantic and geometric context, or their accuracy is bound by mesh resolution rather than the original analytic equations.

2. Methodology

2.1 Contact-Frame Representation

The paper proposes representing an intersection curve not as a set of points $\Gamma = \{p_i\}$ , but as a sequence of contact coordinate frames $\Gamma_C = \{F_i\}$ . Each frame $F_i$ is defined as:
$F_i = (p_i, t_i, n_{A,i}, n_{B,i}, \Gamma_i, \epsilon_i)$
Where:

$p_i$ : The contact point.
$t_i$ : The intersection tangent.
$n_{A,i}, n_{B,i}$ : The normals of the two source surfaces.
$\Gamma_i$ : Semantic constraint labels (e.g., construction provenance).
$\epsilon_i$ : Residual diagnostics and conditioning metrics.

2.2 Residual-Driven Numerical Solver

The core solver uses a predictor-corrector approach:

Prediction: A point is predicted along the tangent direction.
Correction: The point is projected onto the simultaneous zero set of the two original surface equations ( $f_A(p)=0, f_B(p)=0$ ) using a minimum-norm projection (Newton-like update).
Auditing: Every sample is audited for its implicit residual and transversality (conditioning).

2.3 Topology-First Validation Protocol

The paper introduces a "hard-feasible set" protocol. An output is only admitted for accuracy comparison if it:

Completes the operation.
Is non-empty.
Satisfies a declared residual threshold.
Passes closure and component-count gates (e.g., correct number of closed loops).
Invalid or uncertified geometry is rejected before accuracy ranking.

3. Experimental Setup and Benchmarks

The study compares the Coord prototype against OpenCASCADE Technology (OCCT) 7.9.3.1 (via cadquery-ocp) and a PyVista/VTK mesh-first baseline.

3.1 Principal Matrix (8 Cases)

Spanning five primitive families (sphere, ellipsoid, torus, cylinder, cone), including:

Regular Analytic: Sphere vs. Tilted Cylinder (provides analytic ground truth).
Non-Closed-Form: Ellipsoid vs. Tilted Cylinder, Torus vs. Tilted Cylinder, Skew Cylinder vs. Cylinder, Cone vs. Tilted Cylinder.
Conditioning Probes: Near-tangent cases and scale sensitivity ( $10^{-3}$ to $10^3$ ).

3.2 Robustness Portfolio (20 Cases)

A second set perturbs dimensions, offsets, and orientations across the same families to test robustness without relying on simple analytic ground truths.

3.3 Implementation-Level Boolean Test

Direct testing of BoEngine 0.2.0 (C++17) on two exact Boolean branches:

Sphere minus Cylinder.
Sphere minus Torus.
Tests cover multiple poses and resolutions (16–256 segments) to verify shared seams, feature loops, and face normals.

4. Key Results

4.1 Accuracy and Residuals

Coord Method: Achieves near-machine-precision residuals across all cases.
- Max normalized distance-proxy residual: $4.2 \times 10^{-15}$ to $4.0 \times 10^{-14}$ .
- On the analytic sphere-cylinder case, max distance to ground truth: $3.7 \times 10^{-14}$ .
- Results are scale-invariant after normalizing the stopping rule.
OCCT: Produces high-quality CAD-kernel results but shows scale sensitivity in the tested configuration.
- Max normalized residual ranges from $6.5 \times 10^{-11}$ (large scale) to $9.1 \times 10^{-6}$ (small scale).
- On the analytic case, max distance to ground truth: $3.6 \times 10^{-8}$ .
VTK/Mesh: Errors are resolution-dependent and significantly higher (e.g., $10^{-2}$ at low resolution), confirming it is a discretization control rather than a CAD-kernel peer.

4.2 Topology and Boolean Closure

Hard Gates: Both Coord and OCCT passed all 8 principal and 20 portfolio hard gates (non-empty, closed loops, residual thresholds).
BoEngine: All 30 positive test runs (2 branches $\times$ $\times$ 3 poses $\times$ $\times$ 5 resolutions) passed.
- Zero open loops, dangling nodes, or shared-seam gaps.
- Worst face-normal alignment was 0.960222 (at coarsest resolution), improving monotonically with refinement.
- Negative controls (mismatched centers, undersized cutters) were correctly rejected.

4.3 Conditioning and Near-Tangency

In near-tangent cases, the Coord method maintained low residuals, while the transversality sine ( $\sigma_t$ ) dropped to $\approx 0.10$ , indicating ill-conditioning. OCCT's analytic error increased in these cases ( $1.93 \times 10^{-6}$ ), demonstrating that residual-only success is insufficient without explicit transversality and topology audits.

5. Claims and Significance

The paper makes modest, bounded claims rather than asserting universal superiority over mature CAD kernels.

Primary Claim: In the declared finite suite of test cases, residual-driven contact coordinates provide a transparent, auditable, and controllable mechanism for pointwise constraint satisfaction. The method explicitly attaches semantic and geometric metadata (normals, tangents, residuals) to the intersection curve, which is often lost in standard CAD edge representations.
Implementation Claim: The BoEngine 0.2.0 implementation successfully demonstrates that intersection constraints can be propagated through shared Boolean seams to produce audited, crack-free shell outputs for specific exact branches (Sphere-Cylinder, Sphere-Torus).
Limitations: The paper explicitly states it does not claim:
- Universal superiority over OpenCASCADE or other kernels.
- Robustness for all singularities, exact tangencies, cusps, or branch bifurcations.
- Complete handling of general trimmed NURBS face topology.
- Global loop classification in all arbitrary cases.

Conclusion: The work establishes that a "coord-based" approach offers a rigorous, residual-audited bridge from surface constraints to downstream construction for specific, well-defined scenarios. It complements, rather than replaces, existing B-rep kernels, offering explicit control over conditioning and semantic labeling where standard kernels may treat these as internal implementation details.