[Demo] Explainable 3D Mode Shape Recognition Using Region-Aware Graph Neural Networks

Can AI learn to reason about 3D structural dynamics the way experienced NVH engineers do? This is the topic of our latest work: Explainable 3D Mode Shape Recognition Using Region-Aware Graph Neural Networks

For decades, identifying structural vibration mode shapes has remained largely an expert-driven task. Engineers inspect simulation results, compare deformation patterns, and assign semantic labels such as bending, torsion, or local modes. While Physics AI has become increasingly capable of predicting engineering quantities, little work has focused on teaching AI to understand these engineering concepts.

In this work, we proposes a Region-Aware Graph Neural Network that learns from physically meaningful structural regions rather than numerical fields. The approach is validated using industrial CAE and test data from four vehicle platforms.

Key principles guiding our development of practical 3D Engineering AI:

✅ Transferability across different vehicle platforms, meshes, and experimental setups

✅ Learning from limited labelled data, where expert annotations are expensive

✅ Explainability, allowing engineers to understand the structural reasoning behind each prediction

✅ Industrial practicality, enabling AI to work consistently across simulation, testing, and historical engineering data

📄 Paper: https://lnkd.in/e9i5T-Xh (will be presented in ISMA2026 conference: https://www.isma-isaac.be/).

🎥 1-minute demo (make sure you watch until 20sec to meet the interesting AI Agent 😊)

This work was made possible through a close collaboration between Industrial AI researchers and highly experienced experts in NVH, Body Engineering, and 3D CAE. Their domain expertise was essential in shaping the problem formulation, engineering representation, and interpretation of the AI models. Thanks to my co-authors and collaborators: Marc Brughmans, Andrey Hense, Kohta Sugiura, Sebastian Ciceo, @Paolo di Carlo, and Theo Geluk.

It is a great example of how combining deep domain knowledge with AI can lead to practical and useful industrial AI.

Mode shape is just one example. The same principles could extend to CFD flow structures, crash deformation, thermal patterns, and other CAD-CAE domains where engineers reason using semantic concepts rather than raw numerical fields. Ultimately, this is a step toward AI that reasons like an engineer providing interpretable, trustworthy CoPilot rather than black-box predictions.

Do you think this is one of the potential future topics of 3D Engineering AI?