Classification enhanced machine learning model for energetic stability of binary compounds

As contemporary computational technologies and machine learning methodologies rapidly evolve, machine learning (ML) models for predicting formation enthalpies of materials exhibited convincible numerical precision and remarkable predictive efficiency, thus establishing a solid foundation for materials thermodynamic design. Despite achieving numerically high global probability accuracy, current ML models for formation enthalpy nonetheless exhibit suboptimal local accuracy within specific physical domain, which can be attributed to the misalignment between the physical constraints of chemical bonds and the critical descriptors capturing class-specific traits. Herein, we propose a novel approach to improve the local precision of the ML model for predicting formation enthalpy by utilizing Miedema theory-based classification, which segments data into distinct categories according to the electronegativity difference, electron density discontinuity and atomic size difference. Utilizing ML algorithms to build surrogate models guided by the classification strategy significantly improves the local predictive accuracy of formation enthalpy for specific binary compounds, significantly raising the R² value from 0.4–0.9 to 0.8–0.9 compared to an unclassified method. Furthermore, feature importance analysis demonstrates that the pivotal factors for each category vary in some manner, highlighting the insufficiency of a sole ML model in classifying large-dimensional data, which can be addressed by adopting a physics-informed classification strategy. Our results suggest that employing physical-informed classification scheme into ML equips the models with broad applicability and local accuracy, which also shed light for other material properties predication.