Machine learning-enabled prediction and interpretation of spectroscopic properties in Nd and Y co-doped fluorite platform via local coordination engineering

Abstract

The design of advanced laser materials demands precise control over the relationship between local structure and spectroscopic performance. However, conventional approaches struggle to quantitatively link complex coordination environments with spectral properties in rare-earth-doped systems, particularly for correlated 4f-electron states. Herein, we present a machine learning-enabled compositional optimization strategy for tailoring the spectral characteristics of neodymium and yttrium-co-doped fluorites, a platform featuring a wealth of structures and spectroscopic properties, via local coordination engineering. By integrating experimental spectroscopy with structural data, we train interpretable models—including Linear Regression (LR), Support Vector Regression (SVR), Back Propagation Neural Network (BPNN), and Extreme Gradient Boosting (XGBoost)—to predict key spectroscopic parameters from the relative fractions of three dominant polyhedral motifs. The resultant models reveal, through Shapley value analysis, distinct roles of each structural unit in governing absorption, emission, and lifetime behavior. This data-driven framework not only overcomes limitations of traditional Density Functional Theory (DFT) in treating localized 4f states but also provides a scalable pathway for the intelligent compositional optimization of broadband, high-repetition-rate laser materials. This work highlights the effectiveness of combining local structural descriptors with interpretable machine learning to elucidate structure–property relationships in rare-earth-doped fluorite systems, providing guidance for the design of advanced laser materials.