Structural Representation in Crystal Property Prediction: A Controlled Benchmark on Band Gaps

Abstract

This Zenodo record contains the complete reproducibility package for the study Structural Representation in Crystal Property Prediction: A Controlled Benchmark on Band Gaps. The benchmark evaluates how structural representation, model capacity, and training-data availability affect crystal band-gap prediction under a controlled experimental design. Four lightweight model families are compared: a centered Cartesian baseline, a connectivity-only graph neural network, a distance-aware CGCNN-style model, and a SchNet-style continuous-filter model. Each representation is evaluated across approximately matched parameter budgets of 10k, 25k, 100k, and 400k trainable parameters, and across training-data fractions of 1%, 5%, 10%, 25%, and 50%, with repeated random train/validation/test splits. The archive includes the full source code, model implementations, configuration files, the exact Materials Project-derived dataset snapshot used in the experiments, complete run outputs, result logs, model checkpoints, and analysis materials required to reproduce the reported figures and benchmark statistics. The dataset contains crystal structural information, lattice parameters, atom-site information, and band-gap targets derived from the Materials Project. Because the original extraction script is no longer available, the archived dataset should be treated as the exact dataset snapshot used in the study. This record is intended to provide a complete and persistent reference for reproducing the benchmark, auditing the reported results, and reusing the experimental framework for related studies of structural representation in atomistic machine learning.