Conference Agenda
Overview and details of the sessions of this conference. Please select a date or location to show only sessions at that day or location. Please select a single session for detailed view (with abstracts and downloads if available).
Please note that all times are shown in the time zone of the conference. The current conference time is: 18th Apr 2026, 04:06:12pm CEST
|
Agenda Overview |
| Session | ||
D452: AI METHODS FOR GEOMETRIC MODELS AND DESIGN DATA
| ||
    | ||
AI-enhanced computer-aided design: predictive modelling of operations 1Technical University of Munich, Germany; 2BMW Group, Germany; 3University of Passau, Germany This work introduces a graph-based CAD assistant that predicts the next modelling operation in parametric design sequences. Real CATIA V5 models from the automotive domain are converted into directed acyclic graphs capturing feature dependencies, enabling learning directly from structural design data. A four-layer Graph Attention Network achieved a top-5 prediction accuracy of 94%, outperforming a frequency-based non-parametric baseline. The results show that graph representations and attention-based message passing provide a strong foundation for context-aware modelling assistance. Characterizing geometric variability of industrial 3D models to guide preparation of synthetic datasets for machine learning applications 1University of Zagreb Faculty of Mechanical Engineering and Naval Architecture, Croatia; 2Neo Dens Ltd., Croatia This paper presents a characterization approach for analysing geometric variability in industrial 3D model datasets to support the preparation of synthetic datasets for machine-learning applications. By implementing pairwise Hausdorff distances and manifold-based embedding techniques, the study identifies variability ranges required for generating representative synthetic data and demonstrates how targeted augmentation can effectively reproduce real data's variability, ultimately leading to more reliable and robust NN model performance. Automatic feature recognition from imperfect models using a novel workflow of data surrogation 1University of Bristol, United Kingdom; 2Dresden University of Technology, Germany Imperfect CAD models with non-smooth features are common outputs of the latest digital tools. These are unsuitable for the feature recognition needed for end applications like computer-aided manufacturing. This paper proposes to recognise features from imperfect models by contributing a comprehensive dataset, a novel data surrogation method, and ML-based automated feature recognition model. Results show that the data surrogation method accurately replicates manual imperfections with voxel accuracy >0.9 and a Dice coefficient >0.6. Ultimately, feature recognition achieves 92.8% test accuracy. Entity matching for recurring engineering components: a bottom-up enabler for reference architecture reconstruction 1Fraunhofer IEM, Germany; 2Heinz Nixdorf Institute, Paderborn University, Germany Engineering organisations increasingly aim to reuse historical BOM, CAD, and requirements data to identify recurring components. A key prerequisite is Entity Matching (EM), whose performance on heterogeneous engineering data is unclear. This paper evaluates classical models, zero-shot LLMs, and hybrid EM on Amazon–Google and a multimodal engineering dataset. Random Forest and XGBoost achieve near–state-of-the-art results; LLMs perform well but are costly, hybrids add little. EM transfers under controlled conditions and forms a foundation for reference architecture reconstruction. | ||
