Two Philosophies of Shape

The geometry representation you choose determines everything downstream: what operations are fast, what operations are fragile, what you can manufacture, and what you can optimize. In modern CAD, two representations dominate the conversation.

Boundary Representation (B-rep) describes a solid by its bounding surfaces. A cube in B-rep is six planar faces, twelve edges, and eight vertices, stitched together by topological relationships. The representation is explicit — you know exactly where every surface begins and ends.

Signed Distance Fields (SDF) describe a solid by a scalar function over space. Every point in 3D space has a value: its signed distance to the nearest surface. The surface itself is the set of points where the function equals zero. The representation is implicit — the geometry is defined by a mathematical condition, not by enumerated surfaces.

Both are valid. Neither is universally better. But they have profoundly different strengths, and understanding those differences matters for anyone choosing or building a CAD system.

Data Structure Comparison

B-rep Internals

A B-rep model consists of:

Topology: vertices, edges, loops, faces, shells, and their connectivity (typically stored as a half-edge data structure)
Geometry: the underlying mathematical surfaces (planes, cylinders, NURBS surfaces) and curves (lines, arcs, NURBS curves) associated with faces and edges
Attributes: per-face colors, per-edge tolerances, feature tags

The topology and geometry are deeply intertwined. Moving a single vertex can invalidate face adjacency, break tangency conditions, or create self-intersections. B-rep kernels spend enormous engineering effort maintaining topological consistency.

SDF Internals

An SDF model consists of:

A function tree (or expression graph): each node is either a primitive (sphere, box, cylinder) or an operation (union, intersection, difference, blend, transform)
Parameters: dimensions, positions, blend radii, field modifiers
Evaluation: given a point (x, y, z), the tree is evaluated bottom-up to produce a scalar distance value

There is no topology in the traditional sense. The surface exists only as the zero-level set of the function. Connectivity, face adjacency, and edge identity are not explicitly stored.

Boolean Operations

This is where the difference is most dramatic.

B-rep Booleans

Computing the union, intersection, or difference of two B-rep solids requires:

Surface-surface intersection to find trim curves
Splitting faces along those curves
Classifying which portions of each solid’s faces lie inside, outside, or on the boundary of the other
Rebuilding topology: sewing surviving faces into a new shell

Each step is algorithmically complex and numerically fragile. Tolerance issues at near-tangent intersections, sliver faces, and degenerate edges are the daily reality of B-rep kernel development. Commercial kernels like Parasolid and ACIS have decades of accumulated fixes for these corner cases, and they still encounter failures.

SDF Booleans

Computing the union of two SDFs A and B:

f_union(p) = min(f_A(p), f_B(p))

That is the entire algorithm. No surface-surface intersection. No topology rebuild. No tolerance issues. The result is a new function that can be evaluated at any point.

Smooth blending is equally simple. A smooth union with blending radius k:

f_smooth_union(p) = smooth_min(f_A(p), f_B(p), k)

This produces a fillet-like blend between the two shapes without any explicit fillet geometry.

The trade-off: the resulting SDF does not give you an explicit surface representation. If you need a triangle mesh or NURBS surface for manufacturing, you must extract it — typically via Marching Cubes, Dual Contouring, or variational surface reconstruction.

Modeling Capabilities

Where B-rep Excels

Exact analytics: a B-rep cylinder is exactly a cylinder. Its surface area, volume, and moments of inertia can be computed analytically.
Feature identity: “the top face of the boss” is a persistent, addressable entity in a B-rep model. This matters for machining (select a face, apply a surface finish), assembly (mate two faces), and drawing (dimension from face to face).
Precise edges: fillets, chamfers, and blends are explicit geometry with exact tangent continuity. You can measure a fillet radius directly.
Industry compatibility: STEP, IGES, Parasolid XT, ACIS SAT — the entire manufacturing supply chain speaks B-rep.

Where SDF Excels

Organic shapes: smooth blends, variable-radius fillets, and free-form deformations are natural operations on scalar fields. There is no topology to break.
Lattice structures: triply periodic minimal surfaces (TPMS), gyroid lattices, and stochastic foams are compact implicit functions. As B-rep, they would require millions of faces.
Multi-material fields: an SDF can be paired with material property fields (density, stiffness, thermal conductivity) that vary continuously through the volume. B-rep has no concept of volumetric material variation.
Topology changes: when a design parameter changes the number of holes in a part (think of an optimization that opens vents in a heat exchanger), B-rep topology must be rebuilt. SDF topology changes are continuous — holes open and close as the field evolves.
Gradient computation: the gradient of an SDF with respect to design parameters can be computed analytically or via automatic differentiation. This enables gradient-based shape optimization, where you can ask “how should I change this radius to minimize mass while maintaining stiffness?”

Performance Characteristics

Evaluation Speed

B-rep point-in-solid classification is O(log n) with spatial indexing, where n is the number of faces. For simple models, this is fast. For models with millions of faces (lattice structures, scanned geometry), it becomes a bottleneck.

SDF evaluation depends on the depth of the function tree. A model composed of 100 primitives with 50 Boolean operations evaluates in constant time proportional to the tree depth, regardless of geometric complexity. A gyroid lattice that would require 10 million B-rep faces evaluates as a single trigonometric expression.

Memory Footprint

B-rep memory scales with geometric detail. A finely tessellated freeform surface consumes more memory than a plane. Lattice structures can consume gigabytes of B-rep data.

SDF memory scales with the number of operations in the expression tree, not with geometric detail. A lattice structure that repeats a unit cell a million times is stored as a single periodic function — a few kilobytes regardless of repetition count.

Robustness

B-rep operations fail. Surface-surface intersection is an unsolved problem in the general case, and practical implementations use tolerances that occasionally produce invalid geometry. Every B-rep kernel maintains a “healing” subsystem to fix broken models.

SDF operations do not fail in the traditional sense. The min/max operations that implement Booleans are unconditionally valid. However, SDFs can become non-Euclidean (the gradient magnitude deviates from 1.0) after operations, which can cause artifacts in downstream algorithms like sphere tracing. Maintaining field quality requires re-normalization or careful operator design.

Manufacturing Workflows

Subtractive Manufacturing (CNC Machining)

CNC toolpath generation requires explicit surface geometry. The cutter follows edges, faces, and curves that must be precisely defined. B-rep is the natural representation here. SDF-based CAD must extract B-rep or mesh geometry before handoff to CAM software.

Additive Manufacturing (3D Printing)

Slicing a model into layers is the fundamental operation in additive manufacturing. With B-rep, slicing requires computing plane-face intersections for every layer. With SDF, slicing evaluates the field at the layer height — a parallel operation that scales trivially to GPU hardware.

For lattice-filled parts, SDF slicing is orders of magnitude faster than B-rep slicing because the lattice field is evaluated procedurally rather than intersected geometrically.

Inspection and Metrology

CMM (Coordinate Measuring Machine) workflows compare measured points against nominal geometry. B-rep provides exact nominal surfaces. SDF provides exact distance values at any point, which is arguably more natural for deviation analysis — the SDF value at a measured point is literally the deviation from nominal.

When to Use Which

Choose B-rep When

Your manufacturing process is subtractive (CNC milling, turning, EDM)
You need to exchange data with existing supply chains via STEP/IGES
Your workflow depends on persistent face/edge references
You are producing traditional engineering drawings with GD&T

Choose SDF When

You are designing for additive manufacturing
You need lattice structures or graded material properties
You are running topology or shape optimization
You need smooth blends and organic geometry
You want differentiable geometry for design automation

Choose Both When

You are building a modern CAD platform that serves diverse manufacturing workflows. Design in the implicit domain for maximum flexibility, convert to B-rep for downstream interoperability.

How NeuroCAD Approaches This

NeuroCAD’s kernel is SDF-native. All geometry is represented as signed distance fields evaluated over a bounded domain. Boolean operations, smooth blends, lattice generation, and field-driven material assignment are first-class operations, not post-processing steps.

For downstream interoperability, NeuroCAD includes surface realization capabilities that extract explicit boundary representations from implicit fields. This means engineers can design in the implicit domain — where organic shapes, lattices, and optimization are natural — and export to STEP or mesh formats when manufacturing toolchains require it.

The kernel also maintains a distributed constraint graph (DCG) that tracks dependencies between operations, enabling incremental re-evaluation when parameters change. This gives NeuroCAD the parametric editing experience that engineers expect from traditional CAD, built on an implicit geometry foundation.

The Bottom Line

SDF and B-rep are not competitors — they are complementary representations optimized for different stages of the engineering workflow. The question is not which one wins, but which one your specific problem needs, and whether your CAD system can bridge both worlds when required.

SDF vs B-rep: Which Geometry Wins?