AI Agents in Engineering Design: A Multi-Agent Framework for Aesthetic and Aerodynamic Car Design

Journal article

Mohamed Elrefaie, Janet Qian, Raina Wu, Qian Chen, Angela Dai, Faez Ahmed
International Design Engineering Technical Conferences and Computers and Information in Engineering Conference, 2025

DOI: https://doi.org/10.48550/arXiv.2503.23315

View PDF

Cite

APA Click to copy
Elrefaie, M., Qian, J., Wu, R., Chen, Q., Dai, A., & Ahmed, F. (2025). AI Agents in Engineering Design: A Multi-Agent Framework for Aesthetic and Aerodynamic Car Design. International Design Engineering Technical Conferences and Computers and Information in Engineering Conference. https://doi.org/ https://doi.org/10.48550/arXiv.2503.23315

Chicago/Turabian Click to copy
Elrefaie, Mohamed, Janet Qian, Raina Wu, Qian Chen, Angela Dai, and Faez Ahmed. “AI Agents in Engineering Design: A Multi-Agent Framework for Aesthetic and Aerodynamic Car Design.” International Design Engineering Technical Conferences and Computers and Information in Engineering Conference (2025).

MLA Click to copy
Elrefaie, Mohamed, et al. “AI Agents in Engineering Design: A Multi-Agent Framework for Aesthetic and Aerodynamic Car Design.” International Design Engineering Technical Conferences and Computers and Information in Engineering Conference, 2025, doi: https://doi.org/10.48550/arXiv.2503.23315.

BibTeX Click to copy

@article{elrefaie2025a,
  title = {AI Agents in Engineering Design: A Multi-Agent Framework for Aesthetic and Aerodynamic Car Design},
  year = {2025},
  journal = {International Design Engineering Technical Conferences and Computers and Information in Engineering Conference},
  doi = { https://doi.org/10.48550/arXiv.2503.23315},
  author = {Elrefaie, Mohamed and Qian, Janet and Wu, Raina and Chen, Qian and Dai, Angela and Ahmed, Faez}
}

AI Design Agents: Redefining How We Design Cars

Designing a car has always been a blend of engineering precision and aesthetic creativity. Traditionally, the process involves manual sketching, iterative refinements, CFD simulations, and physical wind tunnel testing — often stretching across weeks or months.

But what if you could go from initial sketch to aerodynamic analysis in just minutes?

That’s exactly what we propose with our AI Design Agent framework — a system of collaborative, intelligent agents that automate key tasks in car design using the latest in generative AI, vision-language models, and geometric deep learning.

[Picture]

🚗 What Are AI Design Agents?

Imagine a team of AI collaborators, each specializing in a different aspect of car design:

Styling Agent: Generates high-resolution car renderings based on sketches and design prompts.
CAD Agent: Retrieves or generates 3D car models that match your styling concept.
Meshing Agent: Automatically creates CFD-ready meshes using OpenFOAM.
Simulation Agent: Instantly predicts aerodynamic performance using deep learning, or retrieves full CFD results from a database of 8,000 car designs.

These agents work together using natural language and Python APIs, coordinated via AutoGen, to accelerate the design process — all while keeping engineers and designers in the loop.

🌟 Key Contributions

1. End-to-End Agentic System

We introduce the first agentic design system that covers multiple aspects of car design — from 2D sketching, to 3D shape retrieval/generation, to aerodynamic meshing and simulation — all driven by AI agents that collaborate through language and APIs.

2. New Stylized Data Modalities

We extend the DrivAerNet++ dataset with:

8,000+ high-resolution, photorealistic images using Stable Diffusion XL + ControlNet
Hand-drawn-style sketches using Clipasso and advanced computer vision pipelines
These additions enable vision-language modeling and cross-modal generation for the first time in aerodynamic design.

[Picture]

3. Shape Retrieval and Generation via Modified DeepSDF

We enhance DeepSDF to support two powerful modes:

Shape Retrieval: Search the full 8k DrivAerNet++ dataset in real-time using a 2D sketch or 3D shape input.
Shape Generation: Interpolate across the learned latent space to generate new car geometries from sketches or design attributes.

[Picture]

4. CFD Meshing Agent

Our Meshing Agent uses large language models to:

Automatically update and write snappyHexMeshDict files for OpenFOAM
Run meshing and check quality using OpenFOAM utilities
All based on a text prompt and a 3D car mesh

This enables full real-time simulation preparation, reducing CFD setup time from hours to minutes.

[Picture]

🔁 From Sketch to Simulation, Seamlessly

With just a hand-drawn sketch and a text prompt, designers can:

Generate styled concept renderings.
Retrieve or morph 3D shapes from the DrivAerNet++ dataset.
Automatically generate simulation-ready meshes.
Get aerodynamic predictions in real-time — no waiting for CFD runs.

This end-to-end pipeline not only speeds up iterations, but also enables cross-modal design exploration, letting you design by text, shape, or performance goals.

🧠 Built on the DrivAerNet++ Dataset

The agents are trained on DrivAerNet++, a large-scale open dataset with:

8,000 car designs
Full CFD simulation results
Multi-modal representations: meshes, SDFs, point clouds, sketches, and more

It’s the first large-scale multi-modal dataset to comprehensively bridge aesthetics and aerodynamics for data-driven car design.

✨ Why It Matters

Our system empowers engineers and designers to:

Rapidly visualize and compare new car concepts
Optimize performance and style simultaneously
Make design decisions informed by real aerodynamic insights

This is more than automation — it's about enabling co-creation between humans and AI, driving innovation in engineering design.

Want to explore more? Read the full technical paper here.
Or try the DrivAerNet++ dataset on GitHub.