"Design Agents" for the car design process
Car design takes months of iteration between stylists, CAD engineers, and aerodynamicists. This work introduces a multi-agent framework that compresses that loop: vision-language and large language model agents collaborate to take a concept from a hand-drawn sketch through photorealistic rendering, 3D CAD modeling, meshing, and CFD simulation — automatically.
The key design decision: the agents don't just talk, they use tools. Each specialist is wired to real engineering software — Blender for geometry and rendering, OpenFOAM for CFD, ParaView for post-processing — and grounded in the DrivAerNet parametric design space so every concept it produces is manufacturable and simulation-ready, not just a pretty picture. The result is a loop where a sketch drawn in the morning can have converged aerodynamic numbers by the afternoon, with the human designer steering taste and intent instead of shepherding file formats between departments.
Specialist agents, one design loop
Styling agents
VLM-driven agents interpret hand-drawn sketches and generate photorealistic concept renderings, keeping aesthetic intent in the loop.
Geometry agents
LLM agents translate concepts into parametric CAD, leveraging the DrivAerNet design space for manufacturable 3D geometry.
Simulation agents
Automated meshing and CFD evaluation close the loop, scoring every candidate on aerodynamic performance.
From hand-drawn sketch to photorealistic concept
The styling agents accept a rough 2D sketch and a text brief, and return photorealistic concept renderings that preserve the sketched proportions — bridging conceptual creativity and computational design in seconds instead of studio-weeks.
Aesthetics and aerodynamics, jointly optimized
By coupling generative styling with physics evaluation, the framework explores designs that look good and fly through air well — a step toward AI co-designers that handle the tedious handoffs between creative and engineering tools.
Citation
@proceedings{elrefaie2025designagents,
author = {Elrefaie, Mohamed and Qian, Janet and Wu, Raina and Chen, Qian
and Dai, Angela and Ahmed, Faez},
title = {AI Agents in Engineering Design: A Multi-Agent Framework for
Aesthetic and Aerodynamic Car Design},
series = {IDETC-CIE},
pages = {V03BT03A048},
year = {2025}
}