Toward Generalizable Graph Learning for 3D Engineering AI: Explainable Workflows for CAE Mode Shape Classification and CFD Field Prediction

A research team from academia and industry has published a new framework that applies Graph Neural Networks (GNNs) to core engineering simulation tasks, aiming to move beyond task-specific AI models. The system converts heterogeneous 3D engineering data—like finite element models, CAD geometry, and CFD meshes—into unified, physics-aware graph representations. This allows a single AI workflow to handle both classification (e.g., identifying vibration modes in a car's body-in-white) and regression tasks (e.g., predicting full aerodynamic pressure fields), addressing the industry's need for reusable and interpretable AI tools.

The framework was validated on two critical automotive applications. For Computer-Aided Engineering (CAE), it performs explainable vibration mode shape classification even with limited labeled data, helping engineers understand why a design vibrates a certain way. For Computational Fluid Dynamics (CFD), it acts as a physics-informed surrogate model, predicting pressure and wall shear stress across different vehicle body shapes. A key innovation is symmetry-preserving downsampling, which maintains accuracy while significantly reducing the computational graph size and associated costs.

Beyond just predictions, the framework includes data generation guidance that tells engineering teams which additional simulations or data labels would be most valuable to collect next, optimizing their R&D pipeline. This creates a practical, closed-loop workflow for accelerating design cycles. The result is a move toward generalizable engineering AI that provides faster, more explainable decision support for CAE and CFD, directly tackling pressures to shorten development timelines and improve performance.