Answers are grounded in this paper and cite their sources.
Double perovskites (DP) (A 2 BB’X 6 ) represent a material suitable for photovoltaic and optoelectronic applications, yet their vast compositional space poses significant challenges for computational discovery. In this work, we conduct a systematic comparative analysis of ten machine learning (ML) methods, comprising five ensemble and kernel-based algorithms (CatBoost, XGBoost, LightGBM, Random Forest, and SVM) and five neural network models (SMLP, DMLP, CNN, RNN, and Transformer), to predict the bandgap and heat of formation of 540 DP compounds, using a set of 21 physics-informed descriptors derived from elemental properties. Model generalisation was assessed using 5-fold cross-validation with bootstrap confidence intervals, and accuracy is reported on held-out data. The ensemble methods, particularly CatBoost, attain the highest accuracy for heat of formation (test-set R 2 = 0.999), while the Transformer architecture gives the strongest generalisation for bandgap (test-set R 2 = 0.989). The performance of CatBoost can be ascribed to its ordered boosting and symmetric tree structure, which captures the relationships governing thermodynamic stability, whereas the Transformer uses an attention mechanism to learn the crystallographic dependencies associated with electronic structure. SHapley Additive exPlanations (SHAP) interpretability analysis, together with a holistic ranking approach, identifies Space Group (SG) as the dominant feature for bandgap prediction, followed by anion electronegativity and B’-site electronegativity, and anion electronegativity for heat of formation, with anion electron affinity and B’-site electronegativity also ranking highly, across all ten algorithms. Our framework provides practical guidelines for algorithm selection in computational materials science and demonstrates the relationship between predictive accuracy and physical interpretability achievable with ensemble learning methods and neural networks.
No takes yet. Share an insight, caveat, or question.
Akinpelu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: