Machine-Learning-Assisted Discovery of Cr³⁺-Based Near-Infrared Phosphors

Cr³⁺-substituted inorganic phosphors exhibit three distinct near-infrared (NIR) photoluminescence emission peak shapes that typically fall between 650 and 950 nm. The exact position and shape are governed by the (weak, intermediate, or strong) crystal field splitting environment of the octahedrally coordinated Cr³⁺ ions. These emission characteristics are commonly quantified by the Dq/<i>B</i> ratio, where Dq represents the crystal field splitting parameter and <i>B</i> is the Racah parameter. Precise knowledge of this ratio is therefore critical for designing Cr³⁺-based NIR phosphors for applications like biomedical imaging, night vision, food quality analysis, and luminescence thermometry. Unfortunately, targeting specific Dq/<i>B</i> values in the solid state remains nontrivial due to the complex interplay between the composition, structure, and local coordination environment. To address this challenge, we developed a machine-learned regression model capable of predicting Dq/<i>B</i> trained on 193 experimentally determined Dq/<i>B</i> values and their associated compositional and structural features. We then applied it to estimate the Dq/<i>B</i> values of over 6060 known inorganic structures with potential octahedral Cr³⁺ substitution sites. Eight phosphor hosts, Y₂Mg₃Ge₃O₁₂, YInGe₂O₇, LiInW₂O₆, Gd₃SbO₇, Ba₂ScTaO₆, Ba₂MgWO₆, LiLaMgWO₆, and Ca₃MgSi₂O₈, representing a range of crystal field environments were selected from this list for synthesis and characterization. Their measured Dq/<i>B</i> values closely match model predictions, demonstrating the utility of this machine-learning framework for accelerating the discovery of application-specific Cr³⁺-substituted NIR phosphors.