Despite advancements in computational design, a gap persists between theoretical predictions and experimental synthesizability. While data-driven semisupervised models offer a promising alternative to thermodynamic stability estimates for synthesizability prediction, their inherent limitations remain underexplored. In this work, we systematically investigate the boundaries of these models by evaluating performance shifts under various data manipulations─specifically random, structurally similar, and dissimilar case removals. By comparing these results with thermodynamic-based estimations, we quantify how structural similarity dictates model effectiveness, particularly for newly synthesized materials. Our findings reveal a significant similarity dependency in current frameworks. To address this, we demonstrate that incorporating complementary material properties can partially alleviate this bias, significantly improving recall for structurally dissimilar materials and providing a more robust pathway for inorganic synthesizability prediction.
Kim et al. (Thu,) studied this question.