Dassault Systèmes Aura & Leo Agents: VW Golf → Tiguan Door CAD–FEA Redesign Loop
Adapting a VW Golf Mk8 door to a VW Tiguan SUV aperture: Nemotron agents query requirements in ENOVIA, morph B-rep geometry in CATIA, run SIMULIA (CUDA-X accelerated) to catch a $31.54\text{ Hz}$ vibration failure, and loop back to refine the CAD ($46.37\text{ Hz}$). Again: 1 nominal part in a design silo, not 5,000 toleranced parts across a real CM assembly line.
VW Golf Mk8 Baseline Door Inner Shell
Drag canvas to orbit 3D mesh • Harmonic Mode 1 displacement contour amplified 25×
Nemotron Agent Trace (Aura & Leo)
Querying ENOVIA knowledge graph for Tiguan front-left door structural & NVH targets. Mode 1 natural frequency must exceed 42.0 Hz to prevent lower-panel acoustic booming.
Silo vs. 5,000-Part CM Assembly Line (Inference)
Monte Carlo distribution across 5,000 toleranced BIW doors
What We Built: The GTC Berlin Silo Loop vs. 5,000-Part Factory Reality
Closed-Loop CAD → FEA Inside a Single-Vendor Design Silo
In the NVIDIA GTC Berlin keynote, Dassault Systèmes demonstrated two AI co-pilots—Aura (requirements & PLM orchestration) and Leo (generative CAD morphing & CAE simulation)—powered by NVIDIA Nemotron models inside the 3DEXPERIENCE platform:
- 1. ENOVIA Query (Aura): Pulls engineering targets for adapting the VW Golf Mk8 front door to the VW Tiguan SUV body-in-white aperture ($\text{Mode 1} \ge 42.0\text{ Hz}$, $\text{Beltline Stiffness} \ge 185\text{ N/mm}$, $\text{Mass} \le 15.20\text{ kg}$).
- 2. CATIA Morphing (Leo): Stretches the parametric B-rep surface from $1045 \times 1110\text{ mm}$ to $1120 \times 1245\text{ mm}$. The unbraced lower panel span grows $+19.8\%$ ($680\text{ mm} \to 815\text{ mm}$).
- 3. SIMULIA CUDA-X Modal Solve: Runs GPU-accelerated Lanczos eigenfrequency extraction (`cuDSS` FP64) in $1.42\text{ s}$ and catches a severe panel drumming resonance at **$31.54\text{ Hz}$** (failing the $42.0\text{ Hz}$ spec).
- 4. Closed-Loop CAD Refinement: Leo embosses $12.5\text{ mm}$ stiffening X-beads across the strain-energy hotspot, bumps inner gauge from $0.75\text{ mm}$ to $0.85\text{ mm}$, and adds $6$ spot welds—passing at **$46.37\text{ Hz}$** ($14.59\text{ kg}$).
Why 1 Nominal Part in a Silo ≠ 5,000 Parts on a Real CM Line
Passing a deterministic FEA check at $46.37\text{ Hz}$ on a single zero-tolerance CAD file inside CATIA/SIMULIA does not mean the door builds cleanly across $5,000$ units on an external Contract Manufacturer (CM) BIW line:
- • Elastic Stamping Springback (38.5% of variance): Deep-drawing a $12.5\text{ mm}$ X-bead into Dual-Phase / AHSS sheet metal exhibits elastic recovery upon die release, losing $1.10 \pm 0.55\text{ mm}$ of effective bead depth and thinning the bead wall.
- • RSW Electrode Tip Wear & Shunting (29.0%): Across a production shift, copper electrode caps mushroom and pit, causing $1\text{–}4$ degraded or shunted spot welds per door and weakening inner-to-outer flange torsional coupling.
- • 180°C E-Coat Oven Thermal Buckling (18.5%): Differential thermal expansion between the outer skin and stiffer inner reinforcement rail introduces residual compressive membrane stress ("oil-canning").
- • 3-2-1 RPS Fixture & GD&T Stack-Up (14.0%): Combining coil gauge drift ($\pm 0.045\text{ mm}$), hemming roll variation, and hinge/latch datum drift pushes **$17.5\%$ of doors below $42.0\text{ Hz}$** and **$19.6\%$ into line rework**.
Reduced-Order Orthotropic Shell FEA & Monte Carlo Equations
Embossing diagonal stiffening ribs of depth $d_{\text{rib}}$ into an inner panel of thickness $t_{\text{inner}}$ increases the effective bending thickness $h_{\text{eq}}$ and orthotropic flexural rigidity $D_{\text{eff}}$:
Deff ∝ E · heq3 / [12(1 − ν2)]
In Stage 4, increasing $d_{\text{rib}}$ from $2.0\text{ mm}$ to $12.5\text{ mm}$ and $t_{\text{inner}}$ from $0.75\text{ mm}$ to $0.85\text{ mm}$ raises $h_{\text{eq}}$ from $1.13\text{ mm}$ to $3.225\text{ mm}$, restoring modal stiffness with minimal mass penalty ($+1.49\text{ kg}$).
The fundamental panel drumming / torsional resonance $f_1$ scales inversely with the unbraced lower panel span $L_{\text{span}}$ ($680\text{ mm} \to 815\text{ mm}$) and directly with spot-weld coupling $\eta_{\text{weld}}$:
Stretching $L_{\text{span}}$ to $815\text{ mm}$ in Stage 2 drops $(680/815)^{1.80} = 0.722$, pulling $f_1$ from $43.70\text{ Hz}$ down to $31.54\text{ Hz}$ (`FAIL < 42.0 Hz`).
For each production unit $k \in \{1, \dots, 5000\}$, Box-Muller Gaussian variates $Z_1, \dots, Z_4 \sim \mathcal{N}(0, \sigma^2)$ perturb the physical parameters prior to solving:
Even though the nominal design sits $+4.37\text{ Hz}$ above spec ($46.37\text{ Hz}$), asymmetric springback and weld shunting shift the population mean to $43.39\text{ Hz}$ with $17.5\%$ tail escapes.
Full Stack Comparison & Runnable Implementation
| Architecture Layer | Official GTC Berlin Demo (Dassault + NVIDIA) | Our Runnable Reconstruction & CM Line Audit |
|---|---|---|
| Agent Orchestration | Aura (PLM/Requirements) & Leo (CAD/CAE) powered by NVIDIA Llama-3.1-Nemotron-70B & Nemotron-4-340B NIMs inside 3DEXPERIENCE | AuraAndLeoOrchestrator in agents.py + interactive 5-stage state machine with structured tool-call and reasoning traces |
| PLM & Requirements | ENOVIA V6 Knowledge Graph & Specification Tree | EnoviaRequirements dataclass & live multi-constraint gate ($\ge 42.0\text{ Hz}$, $\ge 185\text{ N/mm}$, $\le 15.20\text{ kg}$, $\pm 1.20\text{ mm}$) |
| Parametric CAD Morph | CATIA Generative Shape Design / SFE Concept B-Rep Morphing | Parametric 3D quad-mesh generator (project3D, drawScene) with real-time aperture stretch, X-bead embossing, RSW nodes, and 3-2-1 RPS pins |
| FEA & Physics Solver | SIMULIA Abaqus Lanczos EigenSolver accelerated by NVIDIA CUDA-X (cuDSS FP64 direct sparse solver) |
Calibrated orthotropic stiffened-shell modal solver (run_simulia_cudax_fea in Python, Invoke-SimuliaCudaXFea in PowerShell, computePhysics in JS) |
| 5,000-Part CM Reality | Omitted — evaluates 1 nominal part in a deterministic design silo | Box-Muller Gaussian Monte Carlo Engine (run_5000_part_cm_line_simulation) modeling springback, gauge drift, RSW wear, e-coat bake, and GD&T stack-up |
| Hosting & Edge Delivery | Omniverse / 3DEXPERIENCE Enterprise Cloud | Cloudflare Workers Static Assets on aura-leo.kogenlabs.dev & gtc-berlin.kogenlabs.dev |