Yb3+ Luminescence Thermometry Revived: Full‐Spectrum Machine Learning for Sub‐Kelvin Accuracy and Precision

Abstract

In this manuscript, we redefine the operational limits of Yb3+‐based near‐infrared luminescence thermometry. Using Ce3+/Yb3+ co‐doped Y3Al5O12 (YAG) nanoparticles, we established a full‐spectrum, data‐driven thermometric framework that transcends the constraints of the conventional Boltzmann‐based luminescence intensity ratio method. By systematically optimizing Yb3+ concentration and leveraging both unsupervised (Principal Component Analysis) and supervised (Random Forest and Gaussian Process Regression) machine learning approaches, we extracted robust, temperature‐encoded signatures from highly overlapped Yb3+ Stark‐level emissions. This approach enables accurate, full‐spectrum thermometry across the entire 100–700 K range, far surpassing the conventional Boltzmann limit (> 420 K). Gaussian Process Regression yields an exceptional peak performance with an average accuracy of 0.02 K and a precision of 0.11 K. Furthermore, when subjected to a rigorous temperature‐block splitting strategy to evaluate genuine generalization, the framework retains sub‐kelvin predictive power, demonstrating an absolute accuracy of 0.45 K and a precision of 0.13 K, outperforming all conventional and alternative machine learning methods under disjoint conditions. Beyond the demonstrated performance in Ce3+/Yb3+‐doped YAG, these results establish Luminescence Thermometry 2.0 as a full‐spectrum, data‐driven thermometric framework in which spectral complexity is exploited rather than avoided, and in which temperature prediction is evaluated through both precision and genuine predictive generalization.