Autoencoder–Graph Neural Network Framework for Robust 3D Pore Connectivity Classification in Noisy Volumetric Images

Pore connectivity is a critical property governing the transport of fluids, gases, and ions in porous materials. Connectivity is often inferred from micro-computed tomography (μCT) volumes through thresholding and connected-component analysis. However, this pipeline is sensitive to noise and limited resolution, producing both false connections and missed links. This study investigates whether graph-based learning, using geometric features and 3D image representations, can correct such errors. We generate 250 synthetic 128³-voxel phantoms with known ground-truth connectivity, degrade them using point-spread blur and noise, and evaluate ten models spanning classical machine learning, 3D CNNs, GCN, GraphSAGE, and autoencoder–GNN fusion. The best-performing models (GraphSAGE and AE+GraphSAGE) achieve F1 scores of 0.954 and 0.957, respectively, compared with 0.468 for the raw connected-component baseline. Improvements are driven primarily by geometric features, including neck width, centroid distance, and pore volume, while frozen autoencoder embeddings provide negligible benefit. However, improved edge classification does not consistently translate into better cluster agreement or percolation prediction.