Field-Driven Shape Evolution in CAD

Traditional CAD modeling is a construction process: you sketch, extrude, fillet, chamfer, and Boolean your way to a target shape. Every operation is an explicit geometric command. The designer decides exactly what happens at every step.

Field-driven shape evolution works differently. You define a physics-based process, the geometry evolves under that process, and you control the result by tuning the parameters of the evolution. The shape emerges from the physics rather than from explicit construction.

This is not a replacement for parametric modeling. It is a complement that handles design problems where the optimal shape is not known in advance and must be discovered through simulation-like processes.

The Foundation: Implicit Geometry as a PDE Domain

In signed distance field representation, geometry is a scalar field f(x, y, z) over three-dimensional space. The surface is the zero-level set f = 0. This representation is naturally compatible with partial differential equation (PDE) operators because the field is a continuous function defined everywhere, not a collection of discrete surface patches.

PDE-based shape evolution modifies the field over time according to a differential equation:

df/dt = F(f, grad(f), laplacian(f), ...)

where F is a velocity function that depends on the field, its spatial derivatives, and potentially external forces. The field evolves, and the zero-level set (the surface) moves accordingly.

This is the level set method, introduced by Osher and Sethian in 1988 and widely used in computational physics. Applying it to CAD geometry is a natural extension that implicit geometry enables.

Curvature Flow: Smoothing by Physics

Mean curvature flow evolves a surface so that every point moves in the normal direction at a speed proportional to the local mean curvature:

df/dt = kappa * |grad(f)|

where kappa is the mean curvature. Points on convex features (high positive curvature) move inward. Points on concave features (high negative curvature) move outward. Flat regions do not move.

What Curvature Flow Does

Smoothing. Small surface irregularities (noise, tool marks, tessellation artifacts) are smoothed out because they have high curvature relative to their size. Large-scale features with low curvature are preserved.

Sharp edge rounding. Sharp edges have infinite curvature and evolve fastest, naturally producing fillets. The fillet radius is controlled by the evolution time. Unlike a CAD fillet command that requires selecting specific edges, curvature flow rounds all sharp features simultaneously.

Convexification. Extended curvature flow eventually shrinks any convex shape to a sphere and then to a point (the Gage-Hamilton-Grayson theorem). For CAD applications, the evolution is stopped long before this, using either a fixed time limit or a curvature threshold.

Practical Applications

Post-optimization smoothing. Topology optimization outputs have rough surfaces from the discrete density field. Curvature flow smooths these surfaces while preserving the structural topology.

Automatic filleting. For parts with hundreds of edges that need consistent fillets, curvature flow applies a uniform smoothing more reliably than selecting and filleting each edge individually. This is especially useful for organic shapes where edge identification is ambiguous.

Mesh repair. Imported STL meshes from scanning or other CAD systems often have noise, spikes, and discontinuities. Converting to SDF and applying mild curvature flow cleans the surface without manual repair.

Reaction-Diffusion: Pattern Formation

Reaction-diffusion systems couple a diffusion process (spreading) with a reaction process (local interaction between species). The classic system is the Turing equations:

du/dt = D_u * laplacian(u) + f(u, v)
dv/dt = D_v * laplacian(v) + g(u, v)

where u and v are concentrations of two chemical species, D_u and D_v are diffusion coefficients, and f and g are reaction terms.

Depending on the ratio of diffusion coefficients and the reaction kinetics, these systems produce remarkable spatial patterns: spots, stripes, labyrinths, and hexagonal arrays. These are the same patterns that appear in animal markings, shell structures, and many biological forms.

CAD Applications

Surface texturing. Reaction-diffusion patterns on a surface create organic textures that are mathematically defined and parametrically controllable. Varying the diffusion ratio produces different pattern wavelengths. Varying reaction parameters switches between spots, stripes, and labyrinthine patterns.

Ventilation pattern design. Reaction-diffusion patterns can generate perforated surfaces where the hole spacing and size are naturally optimized for uniform distribution and structural continuity.

Biomimetic design. For products where organic appearance is desired (consumer products, architectural elements, biomedical devices), reaction-diffusion generates patterns that are inherently natural-looking because they follow the same mathematics that nature uses.

Bounded Evolution: Controlled Shape Change

Unbounded field evolution is a simulation. Bounded field evolution is a design tool. The distinction is critical.

Time Bounding

The simplest control is limiting the evolution time. Curvature flow for 10 timesteps produces mild smoothing. For 100 timesteps, it produces aggressive rounding. The designer controls the degree of evolution through a single parameter.

Curvature Bounding

Instead of evolving for a fixed time, the evolution can be terminated when the maximum curvature anywhere on the surface drops below a threshold. This guarantees that all features have at least a minimum fillet radius, regardless of their initial sharpness.

Spatial Bounding

The evolution velocity can be multiplied by a spatial mask field. Inside the mask, evolution proceeds normally. Outside the mask, the velocity is zero and the geometry is frozen. This enables localized evolution: smooth only the upper surface of a part, round only the edges in a specific region, apply texture only to a designated area.

Field Magnitude Bounding

The maximum displacement of any surface point can be bounded by clamping the accumulated evolution. If a feature has moved more than 2 mm from its original position, further evolution at that point is suppressed. This prevents small features from being entirely consumed by aggressive smoothing.

Exact-Zone Protection

Real parts have regions that must not be modified: interface surfaces, datum features, precision-machined areas, and sealing surfaces. Field-driven evolution must respect these zones absolutely.

Frozen Zones

A frozen zone is defined as a region of space (a bounding box, a distance field from a surface, or an arbitrary field mask) where the evolution velocity is exactly zero. No matter how long the evolution runs, frozen zones remain unchanged.

Transition Zones

At the boundary between a frozen zone and an evolving zone, a smooth transition is needed to avoid creating artificial discontinuities. A ramp function that blends the evolution velocity from zero to full over a specified distance (typically 2-5 mm) creates a smooth geometric transition.

Implementation

In NeuroCAD’s field-driven architecture, zone protection is implemented as a field multiplication:

v_effective(x) = v_evolution(x) * mask(x) * clamp(displacement(x))

where mask(x) is the frozen-zone protection field (0 in frozen zones, 1 in active zones, smooth transition between them), displacement(x) tracks accumulated surface movement for magnitude bounding, and v_evolution(x) is the raw evolution velocity from the PDE operator.

This is evaluated at every field point during evolution, providing pixel-accurate zone protection without any special-case geometric logic.

Anisotropic Evolution

Isotropic curvature flow rounds features equally in all directions. Anisotropic evolution applies different velocities in different directions, enabling:

Directional smoothing. Smooth along the length of a part (removing longitudinal ripples) while preserving sharp features in the transverse direction. This is useful for aerodynamic surfaces where streamwise smoothness matters but cross-sectional sharpness is desired.

Crystallographic-like features. Anisotropic evolution that preserves or enhances certain surface normals can create faceted, crystalline geometries from smooth inputs.

Flow-aligned features. Coupling the evolution anisotropy to an external vector field (e.g., airflow direction from CFD) creates surface features that align with the flow, naturally producing drag-reducing textures.

Combining Evolution with Constraints

Field evolution and parametric constraints can coexist:

1. Pre-evolution constraints. Dimensional and geometric constraints define the base geometry. Feature positions, interface dimensions, and assembly relationships are locked.

2. Evolution domain. The designer marks regions where field evolution is permitted. All other regions are frozen.

3. Evolution parameters. PDE type, time bounds, curvature bounds, and spatial masks are set.

4. Post-evolution constraints. After evolution, geometric queries measure the result (minimum radius, surface area, volume) and the designer verifies they meet requirements.

This workflow preserves the parametric design intent while allowing physics-based shape refinement where explicit construction would be tedious or impossible.

Why This Matters

Many engineering shapes cannot be efficiently described by explicit construction. The optimal heat fin shape, the smoothest aerodynamic fairing, the most uniform stress distribution surface: these are all solutions to PDE-based optimization problems. Encoding the governing equation as a field evolution operator and letting the geometry converge to the answer is more direct than trying to construct the answer explicitly.

Field-driven shape evolution turns the CAD tool from a construction kit into a physics engine for geometry. The designer specifies the rules, the boundary conditions, and the protection zones. The shape emerges from the mathematics.

This is what “computed, not drawn” means in practice.