A demonstration of AI-in-the-loop engineering design — where an AI co-pilot reasons over physics results, proposes geometry and material changes, and iterates toward a converged solution alongside the engineer.
⚡ Human + AI Design Loop
1
🧑💻 ENGINEER sets parameters
You configure geometry (wall, top plate, channels), material, and cooling via the sidebar. Starting point is yours to own.
↓
2
🤖 AI runs FEA + diagnoses
The co-pilot meshes the geometry, solves stress / thermal / bow physics, and reads every number — diagnosing why a target fails, not just that it failed.
↓
3
🤖 AI proposes next design
Based on the cost function score and failure mode, the AI proposes specific parameter changes — with a written rationale you can read and critique.
↓
4
🧑💻 ENGINEER decides: Accept or Modify
You review the proposal and either accept it (AI drives the next iteration) or override values in the sidebar. You stay in control at every step.
↓ repeat until converged
🔄 Two-Phase Optimisation Strategy
The co-pilot uses a structured two-phase strategy to guarantee convergence — it never tries to cut cost while constraints are still violated.
PH
1
FEASIBILITY — Bring design in-bounds
Active when any physics target is violated (bow >3µm, stress >180MPa, temp >120°C, mass >3500g). The AI ignores cost entirely and focuses on the fastest path to a feasible region:
→Bow failing on high-CTE material? Escalate to Invar/SiC immediately — geometry alone can't save it.
→Bow failing on low-CTE material? Maximise cooling (HTC + channel count) to reduce ΔT.
→Stress failing? Increase wall thickness and top plate to reduce peak thermal stress.
→Temperature failing? Boost HTC and add channels; escalate heater management if needed.
Goal: all 4 targets GREEN. Cost score is not evaluated until Phase 1 exits.
▼ ALL CONSTRAINTS SATISFIED ▼
PH
2
OPTIMISATION — Minimise cost function
Active once all targets are in-bounds. The AI now holds feasibility as a hard constraint and works solely to reduce the composite cost score — trimming mass, easing cooling demands, and preferring cheaper materials where the physics margin allows:
→Large cost improvement available? Reduce wall thickness or channel count to cut mass score.
→On Invar/SiC with margin to spare? Try downgrading material if stress & bow headroom is sufficient.
Goal: lowest possible cost score while every constraint stays green.
What this shows
A 200mm semiconductor wafer chuck must stay flat to within 3µm under a 130°C process heater. The co-pilot iterates over geometry (wall thickness, top plate, cooling channels), material (AL6061, AL7075, Invar, SiC), and cooling physics (HTC, channel count) — diagnosing failures and reducing cost each iteration until all targets converge.
The actual project
The real system runs FreeCAD for parametric geometry, Gmsh for adaptive meshing, and CalculiX for full finite element analysis — converging in 5–8 FEA iterations in ~2.5 minutes. This demo uses analytical physics models calibrated from those real FEA runs to replicate the same convergence behaviour in the browser.
How it works here
Default run — physics models calibrated from real FEA runs. The reasoning engine reads live results and proposes parameter changes using the same logic as the real co-pilot.
With API key — a live language model receives the actual physics numbers and full iteration history, writes genuine engineering diagnosis, and proposes next parameters. Each run is unique.
Custom parameters — adjust any slider before starting. The analytical engine responds to your specific inputs and converges from wherever you begin.
Initial design configuration
200mm
chuck dia
130°C
heater temp
AL6061
material
~30µm
bow — iter 1
~116MPa
stress — iter 1
3.0µm
bow target
📬 Engineering IT Bridge Newsletter
Enjoyed this demo? Get more deep-dives on AI-in-the-loop engineering, industrial automation, and practical IT/OT integration — straight to your LinkedIn feed.
This analytical demo runs up to 5 iterations. For full multi-iteration optimisation with real FEA (FreeCAD + Gmsh + CalculiX), contact us to book a live session.
ConvergenceLIVE
Iteration log
📬 Engineering-IT Bridge Newsletter
More AI-in-the-loop engineering deep-dives, straight to your LinkedIn feed.
All physics targets satisfied. The optimiser has found a feasible, cost-minimised design.
This is an analytical demonstration. Results use calibrated physics models (bow, thermal stress, mass) that closely approximate real FEA. For production use, the same optimisation loop runs with full FreeCAD geometry, Gmsh meshing, and CalculiX finite element analysis — typically converging in 5–8 iterations with identical reasoning logic.
📬 Engineering-IT Bridge Newsletter
More AI-in-the-loop engineering deep-dives, straight to your LinkedIn feed.