Machine learning 5d-level centroid shift of Ce3+ inorganic phosphors

Information on the 5d level centroid shift (ɛc) of rare-earth ions is critical for determining the chemical shift and the Coulomb repulsion parameter as well as predicting the luminescence and thermal response of rare-earth substituted inorganic phosphors. The magnitude of ɛc depends on the binding strength between the rare-earth ion and its coordinating ligands, which is difficult to quantify a priori and makes phosphor design particularly challenging. In this work, a tree-based ensemble learning algorithm employing extreme gradient boosting is trained to predict ɛc by analyzing the optical properties of 160 Ce3+ substituted inorganic phosphors. The experimentally measured ɛc of these compounds was featurized using the materials' relative permittivity (ɛr), average electronegativity, average polarizability, and local geometry. Because the number of reported ɛr values is limited, it was necessary to utilize a predicted relative permittivity (ɛr,SVR) obtained from a support vector regressor trained on data from ∼2800 density functional theory calculations. The remaining features were compiled from open-source databases and by analyzing the rare-earth coordination environment from each Crystallographic Information File. The resulting ensemble model could reliably estimate ɛc and provide insight into the optical properties of Ce3+-activated inorganic phosphors.