Geometry foundation model

A geometry foundation model would be a large model pretrained on broad 3D or CAD geometry data, intended to be adapted to many downstream geometric tasks, by analogy with how a language foundation model is pretrained broadly and then adapted to specific language tasks. As of this writing, this remains a nascent research direction without an established, peer-reviewed foundational reference specific to CAD geometry.

Why this term carries more uncertainty than others on this page

This is a Subduxion research interpretation, describing a direction the field is moving toward rather than a settled, citable body of work; treat any specific claim about a geometry foundation model's capability with more caution than an established technical term.

Scope of this page

Not applicable; this describes a potential future class of machine-learning model, not something specified on a drawing.

Common mistakes

Treating "geometry foundation model" as an established, well-defined technical term with a settled body of peer-reviewed research behind it, the way "foundation model" is established for language. The geometry-specific version is still an emerging research direction.

Does a geometry foundation model already exist?

Not as an established, widely validated technology; related research (such as B-rep learning) points in this direction, but a broadly capable geometry foundation model remains a research goal, not a shipped capability.

Is this the same as B-rep learning?

B-rep learning is a specific technique for learning from CAD topology; a geometry foundation model would be a much larger, more broadly pretrained model, potentially built using such techniques.

Does Subduxion claim to have built a geometry foundation model?

No; this page describes a research direction the wider field is exploring, not a specific claim about a Subduxion product capability.