AI Agents in Engineering Design: A Multi-Agent Framework for Aesthetic and Aerodynamic Car Design (2503.23315v1)

Abstract: We introduce the concept of "Design Agents" for engineering applications, particularly focusing on the automotive design process, while emphasizing that our approach can be readily extended to other engineering and design domains. Our framework integrates AI-driven design agents into the traditional engineering workflow, demonstrating how these specialized computational agents interact seamlessly with engineers and designers to augment creativity, enhance efficiency, and significantly accelerate the overall design cycle. By automating and streamlining tasks traditionally performed manually, such as conceptual sketching, styling enhancements, 3D shape retrieval and generative modeling, computational fluid dynamics (CFD) meshing, and aerodynamic simulations, our approach reduces certain aspects of the conventional workflow from weeks and days down to minutes. These agents leverage state-of-the-art vision-LLMs (VLMs), LLMs, and geometric deep learning techniques, providing rapid iteration and comprehensive design exploration capabilities. We ground our methodology in industry-standard benchmarks, encompassing a wide variety of conventional automotive designs, and utilize high-fidelity aerodynamic simulations to ensure practical and applicable outcomes. Furthermore, we present design agents that can swiftly and accurately predict simulation outcomes, empowering engineers and designers to engage in more informed design optimization and exploration. This research underscores the transformative potential of integrating advanced generative AI techniques into complex engineering tasks, paving the way for broader adoption and innovation across multiple engineering disciplines.

• The paper introduces a multi-agent framework that automates critical car design stages using specialized AI agents for styling, CAD, meshing, and simulation.
• It leverages the diverse DrivAerNet++ dataset to achieve top-1% retrieval accuracy and highly precise drag coefficient predictions with a discrepancy of 0.00023.
• The framework reduces the design cycle from weeks to minutes, enabling rapid, data-driven exploration of both creative aesthetics and aerodynamic performance.

AI-Driven Multi-Agent Framework for Automotive Design: Integration of Aesthetics and Aerodynamics

Introduction

This paper presents a comprehensive multi-agent framework for automotive design that integrates vision-LLMs (VLMs), LLMs, and geometric deep learning to automate and accelerate the car design process. The approach is grounded in the DrivAerNet++ dataset, which provides a large-scale, multimodal benchmark of 8,000 industry-standard car designs with high-fidelity computational fluid dynamics (CFD) simulations. The framework introduces specialized "Design Agents"—Styling, CAD, Meshing, and Simulation Agents—each automating a critical stage of the design pipeline, from conceptual sketching to real-time aerodynamic evaluation. The orchestration of these agents is managed via the AutoGen framework, enabling seamless communication and workflow automation.

Figure 1: The multi-agent framework integrates VLMs, geometric deep learning, and LLMs to automate and coordinate the car design process, interfacing with engineering tools and enabling complex workflow automation.

Multi-Agent System Architecture

The system is structured around four specialized agents:

• Styling Agent: Utilizes SDXL and ControlNet to generate high-resolution, photorealistic renderings from 2D sketches and text prompts, supporting rapid exploration of aesthetic variations.
• CAD Agent: Employs geometric deep learning (DeepSDF, PointNet, RegDGCNN, TripNet) for 3D shape retrieval, generative modeling, and latent space interpolation, bridging 2D sketches and 3D geometries.
• Meshing Agent: Automates the generation and refinement of CFD-ready meshes using OpenFOAM’s snappyHexMesh, with iterative quality checks and optimization.
• Simulation Agent: Provides real-time aerodynamic predictions using surrogate models (TripNet) and retrieves simulation data from DrivAerNet++, enabling rapid performance evaluation.

The agents interact via AutoGen, supporting both sequential and hybrid workflows, and can execute Python commands to interface with engineering software such as Blender, OpenFOAM, and ParaView.

Figure 2: The multi-agent system enables iterative collaboration between designers and engineers, automating styling, CAD retrieval, meshing, and simulation for efficient design exploration and optimization.

Data Infrastructure: DrivAerNet++

DrivAerNet++ is a multimodal dataset comprising 8,000 car designs with associated 3D CAD models, meshes, point clouds, voxel grids, depth maps, part annotations, SDF representations, multi-view images, and sketches. This diversity supports a wide range of generative and retrieval tasks, enabling robust training and evaluation of AI models for both aesthetic and aerodynamic objectives.

Figure 4: DrivAerNet++ provides diverse data modalities, supporting retrieval, 3D reconstruction, styling, and aerodynamic simulation tasks.

Styling Agent: Generative Aesthetic Exploration

The Styling Agent leverages SDXL and ControlNet to transform hand-drawn sketches into high-fidelity renderings, conditioned on text prompts for stylistic diversity. Sketches are generated using Canny edge detection and CLIPasso, ensuring alignment with early-stage design intent. The agent supports rapid iteration over multiple aesthetic directions, facilitating creative exploration without sacrificing structural fidelity.

Figure 6: The Styling Agent generates diverse, photorealistic renderings from sketches and text prompts, enabling rapid exploration of styling options across car categories.

CAD Agent: Bridging 2D and 3D Design

The CAD Agent implements a DeepSDF-based pipeline for encoding 3D car shapes as continuous implicit functions, enabling efficient retrieval and generative modeling. A CNN predicts latent codes from sketches, supporting both retrieval of similar designs and interpolation between models for novel shape synthesis. The latent space is structured to allow smooth transitions and high-quality reconstructions, with modifications such as positional encoding and reduced latent dimensionality for improved performance.

Figure 3: The CAD Agent achieves high retrieval success rates, accurately identifying the closest 3D mesh from a given sketch across car configurations.

Figure 9: DeepSDF-based interpolation enables smooth transitions between car designs, supporting structured exploration of the design space for optimization.

Meshing Agent: Automated CFD Mesh Generation

The Meshing Agent automates the generation and refinement of CFD meshes using OpenFOAM’s snappyHexMesh, guided by natural language prompts. The agent iteratively improves mesh quality based on engineer feedback, executing domain partitioning, refinement, and quality checks. This automation significantly reduces the time and expertise required for mesh preparation, a traditional bottleneck in simulation workflows.

Figure 5: The Meshing Agent interacts with engineers to iteratively refine mesh quality, automating OpenFOAM commands and quality analysis.

Simulation Agent: Real-Time Aerodynamic Evaluation

The Simulation Agent enables rapid aerodynamic analysis by retrieving CFD data from DrivAerNet++ or predicting performance using TripNet, a triplane neural surrogate model. The agent supports both shape-based and performance-based retrieval, facilitating data-driven design optimization. On an unseen test set of 1,200 designs, the surrogate model closely matches ground truth drag coefficients, with minor oscillatory deviations but strong overall trend fidelity.

Figure 7: The Simulation Agent’s predictions closely track ground truth drag coefficients across diverse car designs, supporting reliable real-time evaluation.

Figure 8: The predicted and ground truth drag coefficient distributions exhibit strong overlap, indicating accurate statistical modeling of aerodynamic performance.

Results and Numerical Performance

• Retrieval Accuracy: The CAD Agent achieves top-1% retrieval rates across car categories, demonstrating robust mapping from sketches to 3D geometries.
• Surrogate Modeling: TripNet achieves high correlation with CFD ground truth on drag prediction, with the largest observed discrepancy in drag difference being 0.00023, which is within acceptable engineering tolerances.
• Workflow Acceleration: The multi-agent system reduces the end-to-end design cycle from weeks to minutes, automating tasks that traditionally require significant manual effort and domain expertise.

Implications and Future Directions

The integration of AI-driven agents into engineering design workflows has significant implications for both practice and theory:

• Practical Impact: The framework enables rapid, iterative exploration of the design space, supporting both creative and performance-driven objectives. The automation of meshing and simulation democratizes access to high-fidelity analysis, reducing reliance on specialized expertise.
• Theoretical Advances: The use of implicit representations and cross-modal retrieval bridges the gap between conceptual and detailed design, while the orchestration of agents via AutoGen demonstrates scalable, modular workflow automation.
• Limitations: Current mesh evaluation is limited to geometric/topological metrics; future work should incorporate simulation-based validation and user studies for aesthetics. The orchestration framework could be further generalized to other engineering domains.
• Future Work: Extending the agent framework to downstream tasks such as prototyping, manufacturing, and market analysis, and exploring alternative orchestration strategies for improved scalability and robustness.

Conclusion

This work demonstrates a scalable, modular multi-agent framework for automotive design, integrating state-of-the-art generative AI, geometric deep learning, and surrogate modeling. By automating the full pipeline from sketch-based concept generation to real-time aerodynamic evaluation, the system enables efficient, data-driven exploration of both aesthetic and engineering objectives. The approach is validated on a large-scale, high-fidelity dataset, with strong numerical results in retrieval accuracy and surrogate prediction. The framework is extensible to other engineering domains, providing a blueprint for the integration of AI agents into complex design and simulation workflows.