Dataset for: Complexity-Aware Model Selection for Structural Simulation Surrogates

Machine-learning surrogates can approximate finite element analysis cheaply, but the choice of surrogate is usually made by convention rather than measured complexity. Using a parametric L-bracket extended into four part families of increasing geometric and topological complexity under a common linear-elastic pipeline, this study defines the boundary at which a descriptor-based surrogate ceases to be adequate and a geometry-aware graph neural network becomes justified. The boundary is a graded shift in representational effort rather than a single accuracy cliff: descriptor models are near-saturated at low complexity, whereas at rich topology a fixed-length encoding either becomes unstable or requires bespoke feature engineering. On the mesh-converged distributed targets, the compact graph model reduces mean absolute error by 30–42 % relative to the best descriptor, yielding a complexity-aware, representation-driven model-selection rule and a reproducible benchmark for structural simulation surrogates.