Machine learning assisted prediction of the phonon cutoff frequency of ABO3 perovskite materials

One of the phonon properties, the phonon cutoff frequency, pertains to the vibration frequency of the strongest bond in a material, and it has a direct impact on the dielectric breakdown strength. In this study, the accurate prediction of the phonon cutoff frequency was achieved using the Light Gradient Boosting Machine (LightGBM) methodology, utilizing only 15 features related to the structural and elemental information of materials. The performance of the LightGBM model yielded R2 of 0.973, RMSE of 2.214, and MAE of 1.289, surpassing other models by a significant margin. Feature analysis revealed a close correlation between the phonon cutoff frequency and the minimum of atomic number among the elements in the composition through SHapley Additive exPlanations (SHAP). Notably, this research successfully predicted the phonon cutoff frequency of ABO3 perovskite materials accurately for bypassing the time-consuming first principles calculations and reveals the correlation between the phonon cutoff frequency and the physical and chemical information of the materials simultaneous.