# Closing the Loop: A Six-Agent Framework for Aerodynamic, Acoustic, and Manufacturable Car Design

**Samarjith Biswas, PhD · samarjithbiswas.com**

## Abstract

Recent work has shown that a team of specialized AI agents can compress weeks of conceptual automotive design into minutes by orchestrating styling diffusion models, geometric deep learning, and computational fluid dynamics (CFD) tooling. The four-agent framework of Elrefaie et al. (arXiv:2503.23315, 2025), built on AutoGen and the DrivAerNet++ dataset, is the strongest published demonstration of this idea. It stops, however, at mesh quality metrics and surrogate or database-retrieved aerodynamics, and it couples aesthetics to engineering only loosely. This article presents a framework that extends that baseline into a six-agent, LLM-orchestrated system and removes its central limitation. The new system runs a real OpenFOAM `simpleFoam` k-omega-SST external-aero solve on freshly generated geometry, returning a measured drag coefficient (a validated example gives Cd = 0.292 for a DrivAer notchback in roughly 90 seconds). It grounds the styling render in the actual simulated geometry through a ControlNet pass over a clean clay render of the real part-labelled mesh, so the rendered concept and the simulated shape are the same object. It adds a dedicated Inspection Agent that gates geometry on watertightness and manifoldness before meshing, and an NVH agent that maps the aerodynamic result to a cabin wind-noise spectrum and designs a graded locally-resonant phononic metamaterial panel to attenuate the dominant band. Surrogate models trained on real DrivAerML CFD data reach validation R-squared of about 0.49 for drag and about 0.69 for the surface-pressure field. The contribution is a single design loop that couples aerodynamics, acoustic comfort, and manufacturability rather than treating them as separate downstream concerns.

## 1. Introduction

Car design balances measurable engineering performance against subjective styling and brand identity. The traditional workflow moves from sketch to CAD to physics-based simulation to wind-tunnel validation, with iteration cycles measured in weeks. Generative AI and geometric deep learning have made it possible to automate large parts of this chain, and the most complete published demonstration to date is the multi-agent framework of Elrefaie et al. (2025). That work introduced the idea of "Design Agents": four specialized agents (Styling, CAD, Meshing, Simulation) orchestrated by AutoGen and grounded in DrivAerNet++, a dataset of 8,000 industry-standard car designs with high-fidelity CFD labels. The Styling Agent turns sketches into renderings with SDXL and ControlNet. The CAD Agent retrieves the top similar designs from DrivAerNet++ and interpolates new shapes in a DeepSDF latent space. The Meshing Agent drives `blockMesh`, `snappyHexMesh`, and `checkMesh` in OpenFOAM. The Simulation Agent predicts drag with surrogate models (RegDGCNN, TripNet) or retrieves CFD results from the database.

The authors are explicit about where their framework stops. In their Section 6 they state that mesh evaluation focuses primarily on geometric and topological quality metrics from utilities like `checkMesh`, and that accurate evaluation should also incorporate validation through actual CFD simulation outcomes. In practice the loop is never closed: the framework reports whether a mesh passes quality checks, then leans on surrogates or database retrieval for the aerodynamic number. Three gaps follow from this. First, a newly generated shape never receives a ground-truth solve, so the surrogate is used without an on-the-shape check against it. Second, the styling render and the engineering geometry are produced by separate models conditioned on the same sketch, so the photorealistic concept is not guaranteed to be the shape that gets simulated. Third, the framework addresses aerodynamics and aesthetics but says nothing about acoustic comfort or printability, both of which matter for a real vehicle program.

This article describes a framework that keeps the multi-agent structure of the baseline and closes these three gaps with concrete mechanisms.

## 2. Related Work and Positioning

The direct predecessor is Elrefaie et al. (2025), and this work is framed as an extension of it. The same authors produced the underlying data: DrivAerNet (Elrefaie, Dai, Ahmed, 2024) and DrivAerNet++ (Elrefaie, Morar, Dai, Ahmed, 2025), released under CC BY-NC, which provide the parametric geometries, multimodal representations, and CFD labels used here for training and validation. RegDGCNN and TripNet originate from the same group as aerodynamic surrogates. Beyond that line, Ananthan et al. (2024) connected early-stage design to aerodynamic outcomes through a learned latent space, and Aréchiga et al. (2023) used drag-guided diffusion for vehicle images. None of these closes a real CFD loop on generated geometry, and none addresses cabin aeroacoustics.

The locally-resonant metamaterial component draws on the phononic-crystal literature, where sub-wavelength resonators open band gaps well below the Bragg frequency. That distinction matters here: audible wind noise sits in a frequency range where Bragg scattering would demand impractically large lattices, so local resonance is the correct mechanism for a thin cabin panel. Coupling that acoustic design to an aerodynamic result inside the same agentic loop is, to the author's knowledge, new.

| Capability | Elrefaie et al. (2025) baseline | This framework |
|---|---|---|
| Agents | 4 (Styling, CAD, Meshing, Simulation) | 6 (adds Inspection, NVH) |
| Orchestration | AutoGen | LLM planner |
| Aerodynamic result | Surrogate or DB retrieval | Real `simpleFoam` solve, measured Cd |
| Styling vs. geometry | Decoupled concept art | Render forced onto the simulated mesh |
| Geometry source | DeepSDF interpolation, retrieval | FFD morph of part-labelled DrivAer mesh |
| Mesh gate | `checkMesh` quality only | Watertightness and manifold QC before meshing |
| Acoustic comfort | Not addressed | Wind-noise spectrum to phononic panel |

## 3. Methodology

The system is six agents driven by an LLM planner that decomposes a chat or sketch request into a task graph and routes geometry, mesh, simulation, and design artifacts between agents.

### 3.1 Styling Agent (geometry-grounded)

The styling render is forced to follow the actual 3D model. The pipeline takes a clean clay render of the real part-labelled mesh, extracts a Canny edge map, and conditions a Stable Diffusion ControlNet pass on that map. The output is a photorealistic image whose silhouette and proportions are those of the geometry that will be meshed and solved. This is the first true novelty relative to the baseline, whose styling images are concept renderings conditioned on a sketch and therefore decoupled from the engineering shape. Here the beautiful render and the simulated object are one shape.

### 3.2 CAD Agent (part-labelled realistic geometry)

Instead of decoding shapes from an implicit latent space, the CAD Agent morphs a real part-labelled DrivAer CAD mesh using free-form deformation. The source mesh carries 49 named parts, which are collapsed into 8 semantic regions: body, glass, lights, tyres, rims, grille, and underbody. A connectivity-based label denoising step cleans region boundaries so that each material region is contiguous and watertight, which matters both for rendering material assignment and for the downstream acoustic and manufacturability reasoning. The result is a realistic, semantically labelled vehicle rather than a smoothed latent reconstruction.

### 3.3 Meshing Agent and 3.4 Simulation Agent (closed-loop real CFD)

These two agents together close the loop the baseline left open. The path runs STL to `blockMesh` to `snappyHexMesh` to `checkMesh` to `simpleFoam` to `forceCoeffs`, executing a steady RANS external-aero solve with the k-omega-SST turbulence model and returning a measured drag coefficient with its force-coefficient history. A validated example produced Cd = 0.292 for a DrivAer notchback in roughly 90 seconds. This is the second true novelty: the baseline stops at mesh quality and uses surrogates or retrieval for the drag number, whereas this framework obtains ground-truth Cd on the newly generated geometry and uses it to validate the surrogate on the same shape.

The surrogates are retained for fast iteration, not as the final word. A RegDGCNN drag model and a DGCNN surface-pressure (Cp) field model are trained on real DrivAerML CFD data, and a Transolver implementation is available for the pressure field. During a design session the surrogate gives an instant estimate, and the closed-loop solve is invoked when a candidate is worth confirming, so the measured Cd both grounds the decision and provides a check on the surrogate.

### 3.5 Inspection Agent (geometry QC gate)

The Inspection Agent is new relative to the baseline. Before any geometry reaches the mesher it runs automated mesh QC: counting boundary edges (holes), non-manifold edges, connected components, and degenerate or duplicate faces. Geometry that fails is rejected or repaired before meshing rather than failing silently inside `snappyHexMesh` or producing a misleading solve. This gate is also the manufacturability and printability check the baseline lacks, since a watertight manifold surface with a single connected shell is the precondition for both reliable CFD and fabrication.

### 3.6 NVH Agent (aero-acoustic phononic co-design)

The NVH agent is the core novelty. It takes the aerodynamic result, maps it to a cabin wind-noise spectrum, identifies the dominant frequency band, and designs a graded locally-resonant phononic metamaterial panel to attenuate that band. Because the troublesome energy sits in the audible range, the design uses local resonance rather than ultrasonic Bragg scattering, with resonator masses and stiffnesses graded across the panel to broaden the attenuation band around the measured dominant frequency. This adds a comfort dimension that is entirely absent from the baseline and links it directly to the aerodynamic solve: the flow result drives the acoustic target, and the acoustic target drives the panel design.

### 3.7 Data

Training uses real DrivAerNet and DrivAerNet++ labels (CC BY-NC) and is scaled toward 10,000 or more samples with physically-grounded augmentation. Real and augmented samples are documented separately so that reported metrics can be attributed to genuine CFD labels.

## 4. Results

The headline result is the closed-loop solve. For a DrivAer notchback, the full STL-to-`forceCoeffs` path returned a measured Cd of 0.292 in approximately 90 seconds, demonstrating that ground-truth external aerodynamics on freshly generated geometry is attainable inside an interactive agent loop rather than only as an offline batch job.

The surrogates, trained on real DrivAerML CFD data, reach the validation metrics below.

| Surrogate | Target | Validation R-squared |
|---|---|---|
| RegDGCNN | Drag coefficient (scalar) | about 0.49 |
| DGCNN | Surface-pressure field, Cp | about 0.69 |
| Transolver | Surface-pressure field, Cp | implemented |

Geometry quality is visible in the region cleanup. Connectivity-based label denoising on the 49-part DrivAer mesh yields clean glass, body, and underbody regions, removing the stray-face speckle that otherwise contaminates material assignment and acoustic region definition. The geometry-grounded render shows the second contribution working: the ControlNet pass over the clay render produces a photorealistic image whose silhouette matches the meshed and solved geometry, so the concept image and the engineering shape are the same object rather than two separately conditioned outputs.

## 5. Discussion

The advance is not the count of agents but the fact that aerodynamics, acoustics, and manufacturability now live in one loop. In the baseline these are either absent or deferred: drag comes from a surrogate or the database, acoustics is not modelled, and printability is implicit in a `checkMesh` pass. Here the chain is explicit and connected. The Inspection Agent guarantees a clean manifold shape, the Meshing and Simulation Agents return a measured Cd on that exact shape, and the NVH Agent turns that aerodynamic result into an acoustic comfort target and a physical panel design. Coupling these means a styling change propagates through to a measured drag number and a redesigned noise-attenuation panel in the same session, which is the kind of cross-domain feedback a real vehicle program needs and which a surrogate-only, aero-only pipeline cannot provide.

Grounding the render in the simulated geometry matters for the same reason. When the concept image and the engineering shape diverge, any aesthetic decision taken on the render is an estimate about a different object. Forcing the render onto the actual mesh removes that gap and makes styling and engineering decisions refer to one geometry.

## 6. Limitations and Future Work

The closed-loop solve is a steady RANS k-omega-SST computation, which is appropriate for mean drag but does not resolve the unsteady, broadband flow that actually drives wind noise. Bridging the aerodynamic result to the cabin spectrum currently uses a mapping rather than a scale-resolving aeroacoustic simulation, so a natural next step is a hybrid or detached-eddy approach feeding the NVH agent directly. The surrogate R-squared values, while useful for fast iteration, leave room for improvement and motivate the move toward a larger labelled and augmented training set. The phononic panel is designed against a target spectrum but has not yet been validated experimentally, and a fabricated panel measured in a flow-acoustic rig is the clearest way to close that loop too. Finally, aesthetic preference remains subjective, and structured human evaluation of the rendered concepts would quantify the styling side as the engineering side is already quantified.

## 7. Conclusion

This framework keeps the multi-agent structure introduced by Elrefaie et al. (2025) and removes its central limitation. By running a real OpenFOAM `simpleFoam` solve on freshly generated geometry, it replaces a surrogate-or-retrieval drag estimate with a measured coefficient (Cd = 0.292 on a validated notchback). By conditioning the styling render on a clay view of the actual part-labelled mesh, it makes the photorealistic concept and the simulated shape the same object. By adding an Inspection Agent and an aero-acoustic NVH agent, it folds manufacturability and cabin comfort into the same loop as aerodynamics and styling. The result is a single, coupled design loop across aerodynamics, acoustics, and manufacturability, built on the CC BY-NC DrivAerNet and DrivAerNet++ data and on surrogates trained against real DrivAerML CFD.

## References

1. M. Elrefaie, J. Qian, R. Wu, Q. Chen, A. Dai, F. Ahmed. "AI Agents in Engineering Design: A Multi-Agent Framework for Aesthetic and Aerodynamic Car Design." arXiv:2503.23315, 2025.
2. M. Elrefaie, A. Dai, F. Ahmed. "DrivAerNet: A Parametric Car Dataset for Data-Driven Aerodynamic Design and Graph-Based Drag Prediction." IDETC/CIE, arXiv:2403.08055, 2024. CC BY-NC.
3. M. Elrefaie, F. Morar, A. Dai, F. Ahmed. "DrivAerNet++: A Large-Scale Multimodal Car Dataset with Computational Fluid Dynamics Simulations and Deep Learning Benchmarks." Advances in Neural Information Processing Systems, Vol. 37, 2025. CC BY-NC.
4. Q. Chen, M. Elrefaie, A. Dai, F. Ahmed. "TripNet: Learning Large-scale High-fidelity 3D Car Aerodynamics with Triplane Networks." arXiv:2503.17400, 2025.
5. Y. Wang, Y. Sun, Z. Liu, S. E. Sarma, M. M. Bronstein, J. M. Solomon. "Dynamic Graph CNN for Learning on Point Clouds." ACM Transactions on Graphics, Vol. 38, No. 5, 2019.
6. L. Zhang, A. Rao, M. Agrawala. "Adding Conditional Control to Text-to-Image Diffusion Models." Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 3836-3847, 2023.
7. D. Podell et al. "SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis." arXiv:2307.01952, 2023.
8. C. Greenshields. OpenFOAM v11 User Guide. The OpenFOAM Foundation, 2023.
9. V. Ananthan et al. "Machine Learning for Road Vehicle Aerodynamics." SAE Technical Paper, 2024.
10. N. Aréchiga, F. Permenter, B. Song, C. Yuan. "Drag-Guided Diffusion Models for Vehicle Image Generation." arXiv:2306.09935, 2023.