How the Discreteness of the Periodic Table Undermines Machine Learning Predictions

The discovery of novel optoelectronic materials relies on accurate bandgap characterization, enabling efficient exploration of uncharted materials space. Machine learning (ML) is a promising approach to accelerate materials discovery, yet accurate bandgap prediction remains a persistent challenge. This study finds that poor ML performance is not solely attributable to insufficient data or model robustness but rather to the intrinsic complexity and discontinuous nature of chemical space. The lack of a continuous trend in the periodic table fundamentally limits ML models, preventing them from leveraging systematic patterns. While smooth trends exist for properties like total and formation energies, bandgaps exhibit inherent fragmentation, leading models to memorize localized correlations rather than learn universal physical laws. Consequently, increasing dataset size fails to improve predictions. Thus, the core challenge of bandgap prediction lies not in model scale or complexity, but in bridging the discretized foundations of the periodic table itself.