[Paper] Graph Learning for 3D Engineering AI: Explainable Workflows

• cost of data generation (simulation computation time or physical testing campaigns), labeling efforts by experts.

• generalization across different model variants, sensor configurations, and system structures.

• and importantly how to explain AI predictions to domain experts such as mechanical, electrical, and CFD engineers. At the end of the day, engineers need to understand whether the outputs from neural networks make physical and intuitive sense and feel confident applying AI to their work.

We explored some of these topics in our recent paper: “Toward Generalizable Graph Learning for 3D Engineering AI: Explainable Workflows”.  Read the paper on arXiv

The paper presents two industrial use cases: CAE Mode Shape Classification and CFD Field Prediction. The applications use both simulation and test data variants, and go beyond high prediction accuracy by providing physically meaningful explanations aligned with engineering terminologies.

The work will be presented at the JSAE Annual Congress in Japan later this month. Please feel free to join if you will be attending.

It has been a pleasure coordinating this collaboration with Siemens Digital Industries Software engineering colleagues across Europe, Asia, and the US, with co-authors from nine nationalities :)

Kohta Sugiura, Marc Brughmans, Andrey Hense, Zhihao Liu, Amirthalakshmi V, Ajinkya Bhave, Jay Masters, Paolo di Carlo, Theo Geluk

Hope you enjoy reading, and have a nice weekend.