Bilkent University
Department of Computer Engineering

---

M.S.THESIS PRESENTATION

 

MULTIMODAL ATTENTION-BASED GEOMETRIC LEARNING FOR MACHINING PROCESS IDENTIFICATION FROM 3D CAD MODELS

 

Aytaç Akyıldız
Master Student
(Supervisor: Prof. Dr.Uğur Güdükbay)

Computer Engineering Department
Bilkent University

Abstract: Machining Process Identification (MPI) concerns the inference of manufacturing operations directly from 3D Computer-Aided Design (CAD) geometry and is central to automated manufacturability analysis and process planning. The problem can be formulated as supervised geometric representation learning, where heterogeneous geometric signals must be mapped to a discriminative latent space. A multimodal learning framework, MAFNetMPI, is developed by integrating three complementary representations: heat kernel signatures derived from the Laplace--Beltrami operator, orthographic 2D projections, and 3D point clouds. Each modality is encoded separately and projected into a shared 256-dimensional embedding. An intermediate multi-head self-attention module fuses features by modeling cross-modal dependencies before classification. A parametric synthetic dataset of CAD models with controlled geometric variations was constructed for training and evaluation. The method was further assessed on publicly available benchmarks with increasing class granularity and imbalance. The model achieved 99.19% accuracy on the synthetic dataset, 99.85% on FeatureNet, and 97.4% on the 43-class CADNet dataset. Ablation experiments quantified the contribution of each modality and showed that attention improves cross-modal feature integration. The framework provides a learning-based approach to integrating machine learning into CAD/CAM pipelines, enabling automated inference of machining processes from geometric data.

 

DATE: June 11, Thursday @ 10:00 Place: EA 409