MSc thesis · Siemens AG, Munich
Instance recognition of machining features in B-rep CAD models
My MSc thesis, from five months as an applied AI research intern at Siemens. It's about a step CAM engineers still do by hand: looking at a CAD part and deciding which of its faces make up one machining feature. I built three ways of automating it. Two were measured on real industrial parts, and the third only on generated ones.
The problem
A machined part gets designed twice. First it's a shape in a CAD system, then a process in a CAM system, which works out how a machine tool will cut that shape from a block of metal. In between, an engineer reads the solid in manufacturing terms. This bore and the cone under it are one countersunk hole, these five faces are one pocket, that thin band is a chamfer on its rim. Tool choice, sequencing and set-up planning all work from that reading, and it's redone by hand for every new part.
Recognising a feature means naming it and grouping its faces. I worked on the grouping: given the faces of a part, decide which ones belong to the same feature. The score is strict. A feature only counts if the predicted group has exactly its faces, because a group missing a face machines an incomplete pocket and a group with an extra one cuts into material that has to stay. Two things make it hard. Labelled real parts are scarce, because companies don't publish them. And the people who label parts don't fully agree on what counts as one feature.
Three approaches
A million generated parts. There's no large labelled corpus of realistic parts, so I made one. Zero-to-CAD is about a million CAD programs written by a language model. I ran each one under an instrumented CAD kernel that records which operation created every face, so the parts label themselves. That gave Z2C-Seg, 961,243 labelled solids. I then continued the self-supervised pretraining of a Point Transformer V3 encoder on it, for 5.6 days on two GPUs. When labels were scarce, pretraining was worth ten times as many: with it, 300 labelled parts scored 0.408 instance F1, against 0.337 for 3,000 parts without it. The gain peaked after two to four epochs of ten and then faded.
A few hundred real parts. This extends FeatureFox, the recogniser my supervisor and three colleagues built. Gradient-boosted trees score pairs of faces, a threshold keeps the likely links, and each connected group is a feature. I ran it as a ledger where every experiment changed exactly one thing and was scored against its parent on the same five folds. Exact recognition on the curated real parts went from 51% to 76%. Better labels gave 13.6 points of that, more parts 11.2, new features 1.0 and a different decision rule −0.5. A lot of my time went into those labels. I labelled or corrected several hundred parts by hand in Siemens NX. The version I handed over for deployment is trained on the full 283-part pool, where it scores 78.9%.
No training at all. An LLM agent works inside Siemens NX through exact geometric tools. It can ask what lies between two faces, what's coaxial with a bore, or which faces are copies of each other, and it reads the labelling conventions as text. Four geometric rules propose a group of faces, the model decides the spots where those rules are known to go wrong, and the faces take their feature's colour in NX as it goes, until every face is placed. On 119 held-out parts it got 76.6% of features exact, against 60.6% for the trained classifier on 117 of them. Most of that is the rules, which score 74.0% on their own. A part takes about a minute and costs under a cent.
What limited all three
All three ran into the same limit, which is how consistently the parts are labelled. In the team's corpus, about a tenth of the coaxial pairs of adjacent faces (10.9%) are labelled the opposite way from a geometrically identical pair in another part. Those pairs are unwinnable for any model that only sees geometry. My supervisor designed an A/B test that changed one rule, whether a chamfer belongs to the feature it cuts into or stands on its own, on the same 20 parts labelled twice with identical folds. Exact recognition went from 60.8% to 70.2%, more than any modelling change I made. The thesis argues that the labelling convention is a design decision, and that it should be written down and tested on fixed folds like a model change.
Caveats
The real-part corpora are a few hundred parts each, so a handful of parts can move the mean, and one gain I measured reversed on the next data step. The data steps also grew the test folds, so the 11.2-point gain from more parts is generous. The pretrained encoder was never tested on real parts, so the tenfold figure only holds on generated ones. Every held-out agent number comes from a single run, and parts over 80 faces were left out for cost. The chamfer rule that won the A/B also makes more and simpler groups, which an exact-match metric rewards, so its nine points are an upper bound.