Dataset for: Selecting Descriptor Models and Graph Neural Networks For Structural Response Prediction Of Varying Geometric Complexity

Abstract

This paper compares descriptor models and geometric point graph neural networks for structural response prediction in procedurally generated L-shaped brackets with zero to eight holes. Complexity is quantified using hole count, removed-area fraction, boundary multiplier, inverse compactness and normalized minimum ligament. Finite element analysis and whole-group screening yield 8118 cases in 902 matched groups. Models are selected using training and validation groups and evaluated on 185 held-out groups. For maximum displacement and area-weighted 95th-percentile von Mises stress, the graph model reduces test mean absolute error by 46.0% and 50.5%, respectively, relative to the selected descriptor model. Paired group-bootstrap intervals support both improvements. Median combined input-construction and prediction times are 0.89 ms for descriptors and 6.21 ms for the graph model, versus 1.73 s for finite element analysis. These results demonstrate an accuracy–cost trade-off within the sampled domain without establishing a universal complexity threshold.