Bilkent University
Department of Computer Engineering
CS 590/690 SEMINAR
Multimodal Deep Learning for Machining Process Identification Using 3D Geometric Representations
Aytaç Akyıldız
Master Student
(Supervisor: Prof.Dr.Uğur Güdükbay)
(Co-Supervisor:Prof.Dr.Yiğit Karpat)
Computer Engineering Department
Bilkent University
Abstract: Machining process identification (MPI) is essential for automated manufacturability assessment and process planning. Traditional rule-based approaches struggle with complex geometries, necessitating data-driven methods. This study proposes a multimodal deep learning framework that determines whether a 3D part is machinable and, if so, identifies its machining process.We integrate three complementary modalities—heat kernel signature (HKS), 2D projections, and point clouds—to capture machining-relevant features. HKS, a spectral descriptor modeling heat diffusion over a surface, encodes intrinsic geometric properties, while 2D projections retain contour-based features, and point clouds provide spatial information. We employ CNNs for HKS and 2D projections and PointNet for point clouds, combining their latent features through intermediate fusion.To train our model, we generate a parametric dataset of 3D geometries with systematic variations. Experimental results demonstrate that our approach improves MPI accuracy over single-modality models, advancing automated feature recognition in CAD-based manufacturing.
DATE: March 17, Monday @ 13:00 Place: EA 502