Element priors and target support shape chemical transfer in materials graph networks

Abstract

Materials graph neural networks must often transfer to chemical regions weakly represented in training data. Such transfer can rely on predefined relations among elements or supervised evidence from target-containing structures, but these pathways are usually entangled. Here, held-out-element splits and incremental target support separate their roles. Without target-containing training structures, formation-energy errors depend strongly on the element representation, particularly for H, O and F. Matched perturbations show that representation-induced sharing matters beyond input dimension or numerical form, while a label-free similarity-graph prior reduces selected zero-shot errors. Adding a few target-containing structures sharply lowers errors and contracts differences among one-hot, k-hot and continuous inputs across ALIGNN and CGCNN. Calibration explains only part of this recovery, and freezing the initial element projection preserves most gains in five of six ALIGNN splits. Target support therefore shifts chemical transfer from reliance on static element relations toward learning from target-containing environments.