MAFNetMPI: A multimodal attention fusion network for machining process identification in 3D CAD models

Machining process identification (MPI) is a critical step in automated manufacturability assessment and computer-aided process planning, enabling early decisions on feasible process chains directly from part geometry. Existing learning-based approaches to manufacturing-oriented shape analysis rely predominantly on a single geometric representation, such as voxels, point clouds, or multiview projections, each of which captures only a subset of the geometric cues relevant to machining. Furthermore, publicly available CAD datasets are primarily used for manufacturability analysis or feature recognition rather than for direct classification of machining processes. To address these gaps, we propose MAFNetMPI, a Multimodal Attention Fusion Network that classifies 3D CAD models according to their associated machining processes (turning, milling, or both). The framework leverages three complementary input modalities: heat kernel signatures (HKSs), 2D projection images, and 3D point clouds. Each modality is independently encoded using a domain-specific encoder. The resulting feature vectors are fused through a modality-level self-attention mechanism before classification. To support training and evaluation, we construct a parametric synthetic CAD dataset with controlled geometric variations and balanced process labels. Experimental results demonstrate that the multimodal approach outperforms single modality baselines, achieving 99.19% accuracy on the proposed dataset, 97.40% on CADNet, and 99.85% on FeatureNet, indicating that the learned representation transfers across part-level and feature-level classification tasks.