Machine learning-based predictions of the 4fn-4fn-15d1 UV absorption for Pr3+ and Ce3+ ions in fluoride and oxide compounds

Abstract

The development of UV-C (100–280 nm) emitting phosphors utilizing the upconverted 4f^n-15 d¹→4fⁿ interconfigurational broad band emission transition of lanthanide ions requires the fundamental understanding of the host-impurity interaction, which determines the positions of energy levels of the 4f^n-15 d¹ electronic configuration. In this work, we developed an ML Random Forest–based prediction algorithm utilizing a set of 22 dataset descriptors, such as structural characteristics of the host materials, electronic band gaps, ionic radii, and other material properties. The model was successfully tested across 46 impurity centers in various fluoride (chloride) and oxide compounds doped with the Pr³⁺ or Ce³⁺ ions. As an outcome of the developed model, predictions of the 4f-5d absorption wavelengths of Pr³⁺ and Ce³⁺ ions in the investigated hosts, were generated, demonstrating good correlation with the available literature data used initially for training. The obtained results indicate that host materials characterized by (i) a higher polyhedral volume (defined as the volume of the polyhedron formed by the impurity ion and its nearest neighbors), (ii) a higher coordination number (8 or 9), and (iii) high anion electronegativity—particularly fluorides doped with Pr³⁺ and Ce³⁺—may be of interest for applications in the UV-C domain. The resulting prediction model is characterized by the mean absolute errors between the predicted and literature data on the absorption wavelength of less than 3 nm only, which confirms the model's robust and effective performance. The approach developed in this work can be readily extended to other families of materials and/or dopant ions.