Comparative assessment of composition- and structure-based surrogate models across 2D materials databases

Machine learning (ML) surrogate models are increasingly employed to accelerate materials discovery, yet their transferability across heterogeneous databases remains unclear. In this study, we benchmark the accuracy and transferability of composition-based and structure-based surrogate models using three widely adopted two-dimensional (2D) materials databases: Computational 2D Materials Database (C2DB), 2D Materials Encyclopedia (2DMatPedia), and Joint Automated Repository for Various Integrated Simulations (JARVIS-2D). We evaluate predictive performance for energy per atom and bandgap under database-to-database transfer, probe the effects of dataset size and coverage through down-sampling, and analyze error correlations across surrogate models used in this study. Energy-related properties were predicted robustly, whereas bandgap prediction proved to be substantially more difficult due to data imbalance and inconsistencies in DFT parameters across databases. Composition-based models generally exhibited more stable cross-database performance than structure-based models, underscoring that incorporating structural features does not necessarily lead to better generalization. Through down-sampling and error correlation analyses, we demonstrate that cross-database performance is primarily determined by the coverage of the training dataset and that distinct error patterns emerge from differences in the models' feature representation and architecture. Together, this study provides a systematic characterization of surrogate model robustness across 2D materials databases, offering insights into the factors that determine their transferability in materials discovery.