More mesh-to-CAD and a full stack of agentic EDA news

Key Takeaways

• CADFit is an open‑source mesh‑to‑CAD engine that reconstructs editable feature trees from STL/OBJ data.
• It runs ≈3× faster than a fully manual workflow thanks to a lightweight ML sketch detector.
• The tool currently handles extrusions, revolutions, fillets, chamfers, unions, cuts, and all basic sketch primitives.
• Development is funded by Altair (now Siemens); a joint Siemens‑MIT patent is pending.
• Compared with commercial rivals Backflip AI and Autodesk AutoTimeline, CADFit offers free non‑commercial use, broader feature‑level reconstruction, and a transparent research‑grade code base.

Introduction

Engineering Paper’s latest roundup spotlights a new wave of mesh‑to‑CAD converters. While commercial offerings such as Backflip AI and Autodesk’s AutoTimeline have attracted attention for their “AI‑first” promises, the MIT‑Siemens collaboration led by Ph.D. candidate Ghadi Nehme introduces CADFit, an open‑source alternative that blends optimization, geometry validation, and a modest machine‑learning front‑end.

How CADFit Works

From Mesh to Sketches

1. Mesh ingestion – The algorithm accepts standard polygonal formats (STL, OBJ, PLY).
2. Sketch extraction – A shallow neural network classifies edge loops into likely sketch entities (lines, arcs, splines). Nehme reports this step cuts processing time by ~3× versus a purely optimization‑driven pipeline.
3. Parametric fitting – Detected sketches are fed into a hybrid optimizer that iteratively proposes CAD operations (extrude, revolve, etc.) and validates them against the original mesh using geometric distance metrics.

Feature Assembly

• Individual solids are merged through union or cut Boolean steps.
• The system re‑evaluates the assembled model, identifies residual discrepancies, and repeats the fitting loop until the deviation falls below a user‑defined tolerance (typically ≤0.01 mm for high‑resolution scans).

Supported Operations

CADFit vs. Commercial Counterparts

Real‑World Demonstration

A 2‑minute YouTube walkthrough (ICML 2026 presentation) walks viewers through a full reconstruction: starting from a noisy scan of a turbine blade, CADFit isolates the airfoil profile, extrudes it, adds fillets at the leading edge, and finally validates the result against the original mesh. The video underscores the iterative “fit‑validate‑refine” loop that distinguishes CADFit from black‑box AI generators.

Limitations & Future Roadmap

• Operation scope – Currently limited to basic solid modeling; surface‑only features (e.g., lofts, sweeps) are not yet supported.
• Commercial licensing – Free use is restricted to non‑profit or academic projects; a paid tier is under discussion for industrial partners.
• Scalability – While effective on meshes up to 5 M triangles, larger point clouds demand pre‑processing (decimation) to stay within memory limits on a typical 32 GB workstation.

Bottom Line

CADFit offers a compelling blend of transparency, speed, and precision for engineers who need to reverse‑engineer mesh data into editable CAD models without incurring licensing fees. Its open‑source nature and academic pedigree make it a strong alternative to commercial AI‑driven tools, especially for research labs and small‑scale manufacturers. As the feature set expands beyond basic solids, CADFit could become the de‑facto standard for mesh‑to‑CAD conversion in the CNC turning and broader manufacturing ecosystem.