Procurement engineers have lived for two decades with should-cost tools that read STEP files, walk the B-rep face graph, and emit a quote. NeuroCAD has no face graph. So we had to ask a question nobody in the should-cost literature has had to ask before: how do you recognise a hole, a pocket, or a fillet directly from a signed distance field?

The procurement engineer’s problem in one sentence

A buyer at an aerospace tier-one supplier opens a CAD file from an OEM. They need a number — a defensible, auditable estimate of what the part should cost to make — within minutes, before the negotiation call.

The way that number is produced today, across DFMA Concurrent, aPriori, ShouldCost.com, DigiFabster, and every other commercial should-cost platform, is the same:

Read the STEP / Parasolid / SolidWorks file.
Walk the B-rep boundary representation. Identify faces and edges.
Pattern-match face/edge configurations against a library of machining features — through holes, blind holes, pockets, slots, fillets, chamfers, bosses.
Map each feature to a process cost (drill cycle, mill cycle, tool change, setup) using a Boothroyd-Dewhurst-style cost model [1, 2].
Sum, add overhead, return the quote.

Step 3 is the keystone. Get feature recognition wrong and every cost downstream is wrong. NIST published the canonical reference on this — Feature Recognition for Manufacturability Analysis, NIST PUB ID 821984 [3] — and an entire academic subfield has grown around it, most recently with deep-learning approaches [4].

Every published algorithm assumes B-rep input.

NeuroCAD has no B-rep.

What “no B-rep” means in the workshop

NeuroCAD is built on signed distance fields (SDF). A solid is a function f(x, y, z): negative inside, positive outside, zero on the surface. There is no list of faces. There is no list of edges. There is no Edge#42 to attach a fillet to. There is just a function, evaluated at any point in space.

This buys NeuroCAD some properties that B-rep CAD has wanted for thirty years and never quite got — variable-topology editing without naming faults, native smooth blending, lattice infill, differentiable simulation. We have written about those elsewhere.

But it costs us, on day one, the ability to plug into the existing should-cost ecosystem. A B-rep feature-recogniser is asking the question “is this face cylindrical, of finite axial extent, with a planar cap on one end?” — that is, a hole. We cannot answer that question directly, because we do not have faces. We have a single number f(p) and its gradient ∇f(p), evaluable everywhere.

So the question we had to answer is: what is a hole, in SDF terms?

The reformulation: features as topology of the gradient field

A through hole, in a B-rep, is a cylindrical face plus two circular edges plus two planar cap faces.

A through hole, in an SDF, is a region of space where:

the field’s negative region (interior of the part) has been locally subtracted by something whose own field is a cylinder,
the gradient of the part’s field, sampled along the cylinder’s axis from one side of the part to the other, never reverses sign,
the field’s medial axis (the locus of points equidistant from two surfaces) contains a straight segment of finite length whose two endpoints lie outside the part’s bounding box.

That is, a hole has a topological signature in the SDF and its derivatives that does not require a face list to detect. The signature is built from:

Sign-change topology of the field along candidate axes. A through hole produces a signed-distance dip-then-rise pattern when you sample the field along a line that pierces the part.
Gradient-direction clustering. The interior wall of a cylindrical hole has a gradient field that all points radially toward (or away from) a common axis. A pocket has gradients that point toward a planar floor. A fillet has gradients whose directions interpolate smoothly between two parent feature gradients. Each feature class has a different signature in the histogram of gradient directions.
Medial-axis topology. The medial axis of an SDF is computable directly from the field and its Laplacian. Holes contribute straight segments. Pockets contribute planar patches. Fillets contribute curved skeletons of consistent local radius. Bosses contribute closed loops.

None of these require a face graph. All of them are local, evaluable on a regular spatial sampling grid, and parallelisable on the same GPU shader path that already ray-marches the SDF for display.

This is what NeuroCAD commits to as a unique capability: SDF-native feature recognition. It is also, as far as we have been able to find, unprecedented in the published feature-recognition literature.

Why this matters for procurement, not just for CAD theorists

The classical objection to non-B-rep CAD has always been: “fine, your representation is more flexible, but you have cut yourself off from the entire mechanical-engineering ecosystem. Manufacturing partners read STEP. Should-cost tools read STEP. ERP systems consume B-rep features.”

That objection is right in the short run. It is solvable, but only in two steps.

Step one is the bridge: NeuroCAD ships a converter that round-trips between SDF and B-rep so a NeuroCAD file can be quoted by an existing should-cost tool. This is the ecosystem-compatibility track. It works the way every dual-representation CAD has worked for years.

Step two is the leapfrog: NeuroCAD’s own feature recogniser, working directly on the SDF, can answer the should-cost question on geometry that has no valid B-rep representation. That includes:

Lattice-infill parts for additive manufacturing. The lattice cells contribute hundreds of thousands of micro-features that are individually meaningless but jointly determine print time and material cost. A B-rep extractor cannot tractably enumerate them. An SDF gradient-topology analyser can integrate over them.
Field-blended freeform parts. A turbine cooling channel that smoothly transitions from a circular cross-section to an elliptical one over a curved path has, in B-rep, a single tortured spline surface with no recognisable hole feature. In SDF, the channel’s medial axis is a clean curve and its local cross-section can be costed as a 5-axis pocket.
Topology-optimised brackets. Output from a topology-optimisation solver (e.g. nTopology, Altair OptiStruct lattice export) is natively closer to an SDF than to a B-rep. Costing it through a B-rep recogniser usually requires a smoothing/decimation step that loses 5–15% of the geometric detail. Costing it directly on the SDF preserves the detail.

These are not exotic edge cases any longer. The fastest-growing categories of CNC and additive jobs in the 2020s — orthopaedic implants, aerospace topology-optimised brackets, additive heat exchangers — are the ones where the B-rep representation breaks down first. Every should-cost vendor we have surveyed handles them by falling back to volumetric / weight-based estimation, which is dramatically less accurate than feature-based costing.

NeuroCAD is positioning to do feature-based costing on exactly the parts where the incumbents cannot. That is the procurement-side value proposition.

What is shipped, what is proposed, and the honest gap

We are not going to overclaim. The Costing/DFM/DFA workbench in NeuroCAD is a placeholder today. The SDF-native feature recogniser described in this post is a proposed test obligation in our REV11.5 spec. It is publishable — explicitly because it has no precedent — but it is not yet shipped.

What is shipped:

The SDF kernel itself with composition operators, gradient sampling, and Lipschitz-bounded field evaluation.
The dual-kernel bridge that lets us hand a B-rep export to a third-party should-cost tool today.
The DCG (Distributed Constraint Graph) machinery that will track cost-evolution audits when a parameter changes.

What is proposed:

The feature recogniser itself (gradient-direction clustering + medial-axis extractor + classifier).
The cost-model integration (Boothroyd-Dewhurst handling/insertion times, machining process selection, additive-overhang scoring).
The benchmark corpus (NIST DFM benchmark parts, plus a NeuroCAD-original lattice/topology-optimisation extension).
The accuracy target: recall ≥ 0.9 against ground-truth feature labels on the NIST DFM benchmark, ±15% should-cost reproduction against published reference quotes.

This blog is the public-facing argument that the work should exist.

The bigger picture

For thirty years, “should-cost” has meant “should-cost for B-rep parts”. That assumption was invisible because there was no other option. Single-kernel CAD had won; B-rep was the only representation in production use; should-cost had to start from B-rep faces.

The CAD landscape is moving. Implicit-modelling tools (nTopology, Carbon Design Engine, Spatial’s 3D ACIS Modeler with implicit extensions, NeuroCAD) are entering production for the geometries B-rep cannot represent cleanly. As they do, the should-cost question gets reopened. Which features can you recognise? With what accuracy? Against what cost model? Audited by whom?

NeuroCAD’s claim is narrow and concrete: gradient-topology analysis of an SDF can recover the same machining-feature classes that a B-rep recogniser recovers, on the parts where B-rep extraction is reliable, and can recover them on the parts where B-rep extraction fails. The same algorithm covers both regimes. The cost model on top is the standard Boothroyd-Dewhurst pipeline, unchanged.

If the claim survives evaluation, it is the missing piece for SDF-native CAD to enter procurement-grade workflows. If it does not survive evaluation, we will have learned exactly which feature classes resist gradient-topology classification, and that is publishable too.

Either way, the work needs to be done in public, with a benchmark corpus, with the cost model exposed for audit, and with a regression suite that procurement engineers can run on their own parts. That is what this work commits to.

References

[1] Boothroyd, G. and Dewhurst, P. Product Design for Manufacture and Assembly. 3rd edition. CRC Press, 2010. [2] Boothroyd Dewhurst Inc. DFMA Software documentation. https://www.dfma.com. [3] National Institute of Standards and Technology. Feature Recognition for Manufacturability Analysis. NIST PUB ID 821984. [4] Zhang, Z., Jaiswal, P., Rai, R. FeatureNet: Machining feature recognition based on 3D convolution neural network. Computer-Aided Design, vol. 101, 2018.

Status

Artifact	Status
Costing/DFM/DFA workbench	Placeholder — no production code
SDF-native feature recogniser (gradient-direction clustering + medial-axis extractor)	Proposed
Cost-model integration (Boothroyd-Dewhurst + ~30 machining models)	Proposed
Benchmark corpus (NIST DFM benchmark + NeuroCAD lattice extension)	Proposed
Recall ≥ 0.9 on NIST DFM, ±15% should-cost reproduction	Target metric, not yet evaluated
4-mode B-rep ↔ SDF bridge (Analytic, Hybrid, Numerical, Lossless)	Shipped
SDF gradient sampling + Lipschitz-bounded field evaluation	Shipped
DCG cost-evolution audit	Proposed

This work is published as a research direction, not as a finished product. The publishable contribution is the formulation itself — feature recognition directly from SDF gradient topology, with no B-rep extraction step — because the published feature-recognition literature has no SDF-native precedent.

Should-Cost Software Needs Feature Recognition. Every Published Algorithm Assumes B-Rep. SDF Needs a Different Approach.