Beyond the Feature Tree

For 35 years, CAD has been built on the same foundation: boundary representation (B-rep). You sketch, you extrude, you fillet — and the software stores surfaces, edges, and vertices. This paradigm is mature, powerful, and deeply embedded in industrial workflows. It is also reaching its limits.

NeuroCAD is built on a different mathematical substrate: signed distance fields (SDF). Instead of tracking where the boundary is, NeuroCAD tracks how far every point in space is from the boundary. This is not a rendering trick or a post-processing step — it is the native geometric truth of the entire platform.

The consequences are practical: SDF-native Booleans avoid many failure modes associated with surface intersection and healing in traditional kernels; field operations compose naturally with lattices and implicit features; and design states can be replayed from graph-backed representations. The current platform already has substantial kernel and contract implementation, while the commercial product layer is being hardened for pilots.

This article surveys the NeuroCAD platform — not as a feature list, but as a coherent engineering system.

The Mathematical Foundation: Signed Distance Fields

A signed distance field is a scalar function defined over three-dimensional space. For every point (x, y, z), it returns a single number:

Negative inside the solid
Positive outside the solid
Zero on the surface

Think of a weather map for geometry. Where B-rep stores surface patches and their connectivity, SDF stores a single mathematical function. This means:

Field-level Booleans are structurally robust. Union, intersection, and difference are simple operations on the field, avoiding many brittle surface-surface intersection cases. Exported surface or mesh representations still need validation.

Lattices are native. A gyroid or diamond lattice is just another SDF. Compose it with a solid body via a smooth blend operation, and the result is a single coherent field — no stitching, no watertightness checks.

Gradients become a first-class signal. SDF formulations expose gradients across most of the design domain, with non-smooth junctions treated explicitly. This is the foundation for differentiable CAD and bounded optimization workflows.

Four Graphs, One Truth

NeuroCAD’s architecture is organized around four interconnected graphs, each responsible for a distinct concern:

The SDF Graph — Geometric Truth

Every primitive, every Boolean, every lattice, every blend is a node in a directed acyclic graph whose terminal output is a signed distance field. There is no separate “render mesh” or “analysis mesh” — the SDF is the single source of geometric truth.

The Distributed Constraint Graph — Parametric Intelligence

When you sketch a profile and constrain it — parallel, tangent, coincident, dimension — those constraints form a graph that the solver evaluates incrementally. Change one dimension and only the downstream dependencies recompute. The solver is deterministic: the same constraint graph always produces the same geometry, on any machine, at any time.

The E-Graph — Automatic Optimization

E-graphs (equality saturation) are a technique from program optimization applied to geometry. Given a complex SDF tree with redundant or suboptimal operations, the e-graph discovers equivalent but simpler representations — smaller memory footprint, faster evaluation, same geometry. This runs transparently; the designer never sees it, but benefits from it on every interaction.

The Commit DAG — Deterministic Replay and Collaboration

Every design state is a node in a content-addressed commit DAG. Branches, merges, and history are first-class citizens of the geometry kernel — not bolted-on file versioning. This enables deterministic replay (reconstruct any past state exactly), typed three-way merge for collaboration, and audit trails for regulated industries.

Browser-Native: No Install, No Server Round-Trip

NeuroCAD runs entirely in the browser. There is no installer, no OS lock-in, and no cloud dependency for core modeling.

This is possible because:

Rust compiles to WebAssembly. The entire geometry kernel — SDF evaluation, constraint solving, Boolean operations, lattice generation — runs at near-native speed in the browser. No server round-trip. No latency.
WebGPU provides GPU acceleration. Field evaluation, mesh generation, and real-time rendering run on the GPU through the modern WebGPU API — not the aging WebGL. Compute shaders enable parallel field queries and mesh extraction that would saturate a CPU.
Web Workers keep the UI responsive. The kernel runs in a background thread. The interface never freezes, even during heavy Boolean operations or lattice generation on complex parts.

This architecture also has a security implication: your design data never leaves your browser unless you choose to export it. For European engineering firms subject to GDPR, ITAR, or internal data sovereignty policies, this is a decisive advantage over cloud-rendering CAD platforms.

The Feature Landscape

NeuroCAD is shipping in phases, with each phase building on the mathematical foundation of the previous one.

Core Geometry (Phase 1)

The foundation: sketch and constraint authoring with a deterministic solver, SDF primitives and Boolean operations, field evaluation and preview meshing, STL export, and deterministic replay. Everything a designer needs to create, modify, and export solid geometry.

Features and Surfaces (Phase 2)

This is not only a roadmap layer. Sketch-driven extrude and revolve paths are implemented as field operations. The surface layer includes NURBS admission/evaluation, loft/sweep execution contracts, surface workflow state, and quality/healing diagnostics. SDF Booleans and smooth variants are implemented, and feature-preserving Boolean evidence/reporting exists in the core enabler layer. The remaining work is product hardening: UI workflows, pilot evidence, broader validation, and robust surface-field conversion.

Organic and Optimization (Phase 3)

Where NeuroCAD diverges from traditional CAD: field-native organic modeling with protected exact-zone gates, gated diffusion/neural design operations, and differentiable CAD signals for gradients, curvature, and SDF deltas. Material fields are represented with deterministic geometry coupling. Multi-scale composition coordinates macro structure, micro lattice, morph/warp fields, and material grading as a single design state.

Industrialization (Phase 4)

Industrialization is the validation layer. Today the system has strong lattice engineering evidence across TPMS, strut, Voronoi, plate, conforming, blended, and certified lattice workflows. It also has measurement, tolerancing, GD&T, QIF/DMIS, inspection evidence loops, manufacturing-cost heuristics, additive/subtractive workflow checks, BOM/PDM, export gates, and industrial validation gates. Topology optimization is present as a deterministic workflow scaffold under hardening, not yet a production SIMP/level-set optimizer.

Pure Rust: Memory Safety, Determinism, Zero Runtime Cost

The NeuroCAD kernel is written in pure Rust — a deliberate engineering choice with tangible consequences:

Memory safety without garbage collection. Rust’s ownership system eliminates entire categories of bugs (use-after-free, buffer overflows, data races) at compile time, without the runtime overhead of a garbage collector. For a CAD kernel that must run predictably in a browser, this is essential.

Deterministic execution. Rust’s strict aliasing rules and lack of undefined behavior align with NeuroCAD’s deterministic replay guarantee. The same code produces the same result, on any platform, every time.

WASM as a first-class target. The wasm-pack toolchain compiles Rust directly to WebAssembly. There is no intermediate layer, no emulation, no performance penalty.

Modular architecture. The kernel is organized into 13 crates with clear boundaries: SDF evaluation, constraint solving, e-graph optimization, BVH acceleration, mesh generation, and more. Each crate can be tested, benchmarked, and evolved independently.

Manufacturing Governance: Governed, Not Bolted On

Traditional DFM (design for manufacturing) works as a post-design check: you model the part, then run an analysis tool, then fix the violations, then re-analyze. This loop is slow and error-prone.

NeuroCAD’s manufacturing governance inverts the workflow. Process constraints — minimum wall thickness for metal AM, draft angles for injection molding, overhang limits for FDM — are represented as rules inside the design state, not as a separate late-stage checklist. The current system has deterministic validation gates and evidence loops; the pilot-hardening work is turning those gates into broader interactive workflows across more manufacturing processes.

This is the core governance idea: manufacturing knowledge should shape the design path, not arrive only after export.

The Integration Surface

NeuroCAD is not an island. The platform provides:

Multi-format export: STL for 3D printing, STEP with semantic overlays for traditional CAD interchange, and field-native formats for downstream simulation.
MOOSE bridge: Direct coupling to the MOOSE finite element framework for authoritative structural, thermal, and multiphysics validation.
AI orchestration boundary: A defined interface where AI-generated geometry proposals enter the system. Proposals are validated against the SDF truth, manufacturing constraints, and design history — never accepted as source of truth.
Scripting and automation: A headless mode with full SDK access for batch processing, design space exploration, and integration into automated workflows.

What This Means for Engineers

NeuroCAD is not “another CAD tool.” It is a different category of engineering platform:

Geometry with fewer brittle edit paths. SDF Booleans reduce reliance on fragile surface-surface intersection workflows, while imported surfaces and exact/surface workflows still require validation and, where needed, healing.
Designs that never diverge. Deterministic replay means every design state is reproducible — across machines, across time, across teams.
Optimization signals across the design graph. Differentiable CAD exposes gradients, curvature, and SDF deltas as first-class signals, creating a foundation for lawful optimization workflows.
Freedom from installation. Open a browser tab. Start designing. No IT ticket, no license server, no OS dependency.
European data sovereignty. Your design data stays in your browser. No cloud upload, no foreign jurisdiction, no compliance headache.

NeuroCAD is built from mathematics, not legacy. That difference matters.

NeuroCAD: The Next-Generation CAD Platform Built on Mathematics, Not Legacy