Topic 1 — CATIA V5/V6 sketch constraint automation covers the COM API, the AddMonoEltCst method most people miss, and an EKL snippet for 3DEXPERIENCE users, with a real warning about SP-level enum drift.
Topic 2 — SolidWorks generative design reality check cuts through the hype, explains the SW → nTopology → reimport workflow that actually ships, and includes a C# API snippet for driving topology study settings programmatically.
Topic 3 — Using LLMs to write CAD macros is probably the most practical thread — the "record first, then let AI clean it up" tip from Marcus is gold, and Carla's point about hallucinated CATIA method names is a real pitfall people hit constantly.
Topic 4 — Scan-to-CAD AI defect detection covers Geomagic's ML surface classification improvements and a full Open3D + YOLOv8 deviation-map-to-CNN pipeline with working Python code.
Topic 5 — Parametric optimization tools compares Dakota surrogates, Bayesian optimization via Optuna (with a live SW API driver loop in Python), and ANSYS optiSLang — with honest tradeoffs between each.

Owen_Whitfield
Senior CAE Analyst · Coventry
Feb 2026
Great thread. One practical thing I want to add for 2026 users — don't overlook ANSYS optiSLang if your company already has an ANSYS license, because it has pre-built connectors for both SolidWorks and CATIA geometry updates through a parametric interface, and its MOP (Metamodel of Optimal Prognosis) approach is specifically designed for engineering design spaces which tend to be lower-dimensional than ML problems but expensive to evaluate. The learning curve is steep but the certification courses on ANSYS Learning Hub are actually decent and free with a subscription. Where pure-Python tools like Optuna beat it is in flexibility and integration with custom solvers — optiSLang wins when your solver chain is all ANSYS (Mechanical + Fluent + optiSLang). For mixed toolchains, build your own loop as Leila showed above.

Leila_Sharma
Design Engineer · Hyderabad
Feb 2026
Adding to Farrukh's surrogate approach — for those more comfortable in Python than in Fortran-era tools like Dakota, the scikit-optimize and Optuna libraries give you Bayesian optimization out of the box and you can wrap your SolidWorks API calls inside the objective function. Here's the basic pattern for driving SW parametrically from a Bayesian loop:
Optuna's documentation at optuna.readthedocs.io is excellent and they have integration guides for engineering workflows specifically.

Farrukh_Nazarov ✓ PhD Aero
Aerodynamics / CFD · Toulouse
Jan 2026
The most underrated thing in AI-assisted parametric design right now is surrogate modelling — instead of running full FEA or CFD for every point in your parameter space, you build a fast ML proxy (Gaussian process, neural network, or polynomial response surface) trained on a small number of high-fidelity simulations, then run your optimization on the cheap surrogate. Tools like Dakota (from Sandia National Labs, free and open-source at dakota.sandia.gov) have been doing this for a decade but they've added ML-based surrogate backends in recent releases. The integration with CATIA or SW is through Python scripts that drive the CAD model via API, export geometry, run the solver, and pipe results back into Dakota's optimizer loop. On a recent turbine blade project we reduced CFD runs from 2,000 to about 80 while hitting the same Pareto front quality — the surrogate was a GPR trained on Latin Hypercube samples.