Physical admissibility, predictive accuracy, calibration, and cross-source transfer need not improve together in small materials datasets. We evaluate 511 Ca–P–O-containing structures with composition-disjoint partitions, repeated initializations, and validation-only selection. Clipping and a trained nonnegative head both remove negative gaps, but the small observed error difference does not establish a reliable advantage of ReLU over clipping across training runs. Analytical density supplies a deterministic reference. Shared-head ablations and limited graph and forest searches probe composition fusion, while fully unweighted training and validation remove the deterioration seen with weighted checkpoint selection. Classification remains modest relative to trivial decision rules. The main transfer finding is a reversal in gap-error ordering: descriptor trees lead internally, whereas variants of one CGCNN-inspired graph family transfer better to unseen OQMD compositions and composition-informed neural models perform worse than the fit-mean reference. Averaged equally over the tested model variants, the structure-only group has lower OQMD gap MAE than descriptor trees by 0.278 eV (family-adjusted composition-bootstrap interval 0.090–0.456 eV). This exploratory interval conditions on these fitted variants, not architecture-level uncertainty; all model-pair and group contrasts share one external correction family. The tiny unseen JARVIS cohort remains descriptive. Physical output validity and chemically meaningful descriptors therefore require assessment separately from predictive utility and transfer.
Copyrights © 2024