Thermodynamic stability of ABX2-type ternary chalcogenides from machine learning

Ternary chalcogenide compounds (ABX₂) have shown great potential for optoelectronic and energy-related applications due to their outstanding functional properties. Nevertheless, their thermodynamic stability remains a major challenge that hinders practical implementation. In this study, a symbolic regression framework based on genetic programming, integrated with SHAP-driven feature analysis, was developed to accurately predict the formation energies of 293 ABX₂ compounds. Through the systematic evaluation of a wide range of atomic and chemical descriptors, the framework successfully identified the most significant variables influencing stability, while maintaining a balance between predictive performance and model interpretability. Among these, a newly defined polarity factor P = E 2 _ X / PE _ B —derived from the interaction between the second ionization energy of the X-site element and the Pauling electronegativity of the B-site cation—was identified as a key descriptor for thermodynamic stability. This descriptor achieves an accuracy of nearly 90 % and demonstrates clear stability thresholds across different A-site element categories. The proposed framework provides an interpretable and efficient approach for descriptor-guided inverse design, facilitating the systematic discovery and rational development of stable ABX₂ chalcogenide materials. • A thermodynamic stability descriptor (polarity factor) for ABX₂ chalcogenides has been proposed. • SHAP-based feature selection and symbolic regression algorithm were combined to identify the key descriptor. • The stability descriptor effectively captures the thermodynamic stability trends of ABX₂ chalcogenides.