ANALYTICAL MODE
IDLE
AI-in-the-loop engineering · live demo

Design with Copilot: Make Geometry, Physics and Material decisions for fastest path to passing design and then optimize it

⚠️ Disclaimer
🔧 Personal project — built on personal hardware, own time, using free open-source software. Not affiliated with any employer, client, or institution, past or present.
🤖 AI-generated use case — the wafer chuck scenario (physics, materials, thermal loads, targets) was proposed by an AI when prompted for a semiconductor machinery use case. It does not reflect the author's professional domain or any past/current employer knowledge.
📐 Synthetic data — all geometry, FEA results, and cost models are illustrative approximations for demonstration. No real product or proprietary design is represented.
🎯 Intent — explore AI-augmented engineering workflows and share findings openly with the community. Feedback welcome.
What is a wafer chuck?
A wafer chuck is the precision fixture that holds a silicon wafer flat against a heated platen during semiconductor lithography, deposition, or inspection. It must be dimensionally perfect — any bow in the wafer surface causes pattern misalignment at nanometre scale, ruining the die.
The physics at play
The chuck sits on a resistive heater at 80–150 °C while the fab room stays near 20 °C. That 100 °C+ gradient drives differential thermal expansion — the body tries to grow, the cooled top plate resists, and the whole assembly bows. The stiffer and thicker the top plate, and the lower the material's CTE, the less it bows. Cooling channels carved through the body extract heat, but they also remove material — reducing mass and stiffness simultaneously, so every change has a second-order effect somewhere else.
Design trade-offs that matter
Material CTE — Al6061 expands 23× more per °C than Invar. Switching material alone cuts bow by up to 20×, but Invar costs 4× more and is harder to machine.
Top-plate thickness — thicker plate increases bending stiffness but adds mass and raises the centre of gravity.
Cooling channels — more / deeper channels lower chuck temperature (reducing ΔT and bow), but thin walls risk stress failure under vacuum clamping loads.
Mass budget — light chuck = faster stage settling. Heavy Invar body costs throughput. Every gram matters at high-velocity scanning.
How this co-pilot reaches a passing design faster
Rather than sweeping parameters blindly, the co-pilot reads the FEA output, identifies the dominant failure mode (bow, stress, temp, or mass), and makes a causal move — switching material when CTE is the root cause, or stiffening geometry when it isn't. It runs a two-phase strategy: get every metric inside its limit first, then trim cost without re-breaking anything. When you push back with your own idea, it runs yours, explains the result, and only overrides when the evidence is clear.
29.9µm
BOW · START
→
2.0µm
BOW · FINAL
150s
5 ITERS
ITER 1 — FAIL bow 29.9µm ✕ 3mm plate AL6061 130°C heater → ITER 5 — CONVERGED bow 2.0µm ✓ 8mm plate Invar · CTE 1.2µm/m°C
Cross-section · iter 1 vs converged
The reasoning loop
YOU
Set parameters
Geometry · material · cooling
↓
FEA
Simulate physics
Stress · bow · temp · mass
↓
AI
Diagnose & propose
Root cause → written rationale → next params
↓
YOU
Accept or override
You stay in control — always
↻ repeat until converged
Two-phase optimisation — why this never thrashes
PH
1
Feasibility
Any target failing? AI ignores cost entirely. Gets the design inside bounds first. No exceptions.
→ bow fail + high CTE → switch material
→ stress fail → wall + cooling
→ temp fail → max HTC + channels
ALL
⇒
GREEN
PH
2
Optimisation
Feasibility locked. Now AI hunts cost — exploiting every margin without ever re-breaking the physics.
→ trim wall → cut mass score
→ reduce HTC → lower cooling cost
→ Δcost <1% → declare converged
The real system

Full FEA on a personal laptop. No cloud. No licences. Converges in ~150 seconds.

FreeCAD 1.0 Gmsh CalculiX 2.23 Python 3 Anthropic API
This demo · three modes
▶ Default run
Analytical physics calibrated from real FEA. Same convergence behaviour in the browser, instant.
⚡ With API key
Live LLM gets actual physics numbers. Writes real engineering diagnosis. Every run is unique.
⚙ Custom parameters
Adjust any slider. The engine responds to your specific starting point and converges from there.
Not a parameter sweep. A physics-aware reasoner.
Scripted sweep
Vary param ±Δ → re-solve → accept if better.
No causal understanding.
This co-pilot
Reads FEA output → diagnoses failure mode → escalates material / cooling → understands bow physics.
The material call
Bow 10× over limit → AI switches AL6061 to Invar. CTE drops 20×. Bow drops to 1.38µm instantly. No human needed to figure that out.
Mesh self-repair
Iter 4: geometry became unmeshable. AI diagnosed, repaired parameters, re-ran. Zero human input.
📬 Engineering IT Bridge

AI-in-the-loop engineering · industrial automation · IT/OT integration — straight to your LinkedIn.

Subscribe on LinkedIn
Cost = 0.35·mass + 0.25·cooling + 0.15·matCost + 0.15·mfg + 0.10·bow — each term normalised 0→1, lower is better