Interpretable machine learning screening and analysis of pressure sensitivity and high-pressure emission line shifts in lanthanide and transition metal doped phosphors

Abstract

In the field of high pressure optical sensing, developing pressure-responsive luminescent materials that integrate high sensitivity, a wide dynamic range, and robust stability remains a critical bottleneck for achieving precise, real-time, and remote manometry under extreme conditions. A key scientific challenge is to establish interpretable structure–property relationships from coupled physicochemical factors, including crystal structure, dopant species, and local coordination environment. Here, we develop an interpretable hierarchical machine learning framework for screening and estimating pressure-induced emission line shifts in lanthanide and transition metal doped crystals. A specialized dataset was constructed from experimental reports in the literature, incorporating nine physically meaningful descriptors such as ionic radius, coordination number, lattice volume, emission centroid, and phonon energy. The framework first classifies materials into different sensitivity regimes using an XGBoost-based screening module and subsequently performs quantitative regression using multiple machine-learning models. Benchmark evaluations show that CatBoost and TabM achieve strong internal performance, with R² values exceeding 0.90. External validation on recently reported unseen systems further indicates that the model can capture major pressure-response trends, while also showing larger deviations for chemically underrepresented or structurally complex cases. These results suggest that the present framework is most suitable for descriptor-guided screening, candidate prioritization, and semi-quantitative estimation. Beyond numerical estimation, SHAP and Sobol sensitivity analyses identify ionic radius and coordination number as major physical drivers of pressure-induced spectral shifts, and reveal transition-type-dependent descriptor contributions. This work provides a practical data-driven strategy for extracting materials-chemistry rules and accelerating the discovery of pressure-responsive luminescent materials.