Work

AI-Based Interfacial Contact Area Prediction in 3D Granular Media: Benchmarking Volumetric CNN, GNN, and Classical Surface Meshing

Year: 2026
Type: preprint
Abstract: Interfacial contact area between contiguous grains governs macroscopic elastic moduli, stress transmission chains, effective thermal and electrical conductivity, and solid-state diffusion rates in three-dimensional (3D) granular media. Standard characterization of X-ray micro-computed tomography (µCT) volumes relies on 3D Euclidean distance transform seeding and marker-controlled watershed segmentation followed by polygonal surface extraction via the Marching-Cubes algorithm (Lorensen & Cline, 1987). However, Cartesian voxelization introduces severe staircase terracing along curved inter-particle boundaries, inflating surface area measurements by 40% to 80%, while local mesh triangulation scales poorly to assemblies containing thousands of grains. Conducted under the supervision of Prof. Mayank Goswami , this study establishes a comparative benchmark of four data-driven surrogate architectures—an interpretable Random Forest ensemble, a 4-block 3D Convolutional Neural Network (3D CNN) operating on 48 3 binary dual-particle patches, a 3-layer bidirectional GraphSAGE Graph Neural Network (GNN) operating on the global contact graph G = (V, E) , and a multimodal Hybrid GNN + CNN fusion architecture—across 250 synthetic 3D granular phantoms (50,800 particles and 48,750 contact interfaces). Across 38 held-out test specimens (7,466 contact interfaces), we evaluate models under both DEM-coupled (known sphere geometry) and blind tomographic (voxel-only, zero metadata) operating regimes. When coupled with exact inter-particle distance and radius attributes, the 3-layer bidirectional GraphSAGE network achieves a mean absolute error (MAE) of 0.158 vx 2 , a root-mean-square error (RMSE) of 0.217 vx 2 , a mean absolute percentage error (MAPE) of 1.91% , and R 2 = 0.99999 at an inference latency of 2.89 ms per phantom , while the Random Forest baseline achieves R 2 = 0.99992 and MAE = 0.409 vx 2 . Crucially, an explicit feature-circularity ablation study demonstrates that removing pairwise distance and overlap features collapses tabular/graph accuracy to R 2 = 0.4387 (MAE = 44.68 vx 2 ), proving that inter-centroid geometry is physically indispensable; however, estimating these quantities 100% blindly from watershed-segmented voxels via principal inertia axes ( R̂ i = L maj,i /2, MAE = 0.048 vx) restores R 2 = 0.9959 (MAE = 3.35 vx 2 ) without ground-truth metadata. Furthermore, the purely image-based 4-block 3D CNN —which requires zero tabular or metadata inputs—achieves R 2 = 0.99788 and MAE = 2.434 vx 2 (166.23 ms/phantom), outperforming classical Marching-Cubes surface meshing (MAE = 42.750 vx 2 , MAPE = 72.24%, R 2 = 0.50471, 3,555.60 ms/phantom) by 17.6× in accuracy and 21.4× in speed. Scaling experiments up to N = 2,000 particles, overlap-depth stratification, per-particle contact area aggregation ( n = 6,879 active test grains), and domain adaptation to mildly non-spherical superellipsoids (3D CNN R 2 = 0.9979, MAE = 2.50 vx 2 ) establish practical guidelines, methodological limitations, and ethical safeguards for deploying AI surrogates in granular mechanics.
Language: English
Open Access status: green
OpenAlex ID: W7225892017
Cites: 21
Cited by:
Related to: 10