Luminescence Thermometry 2.0: Machine-Learning Temperature Read-Out with Luminescence of YOF: Yb3+, Er3+ Nanoparticles

Here, we evaluate a full-spectrum, data-driven readout framework, termed Luminescence Thermometry 2.0, in which normalized emission spectra are used directly as inputs to temperature regression. We benchmark this approach against conventional luminescence-intensity ratio thermometry using the same experimental dataset obtained from the upconversion and downshifting emissions of Er3+,Yb3+-co-doped yttrium oxyfluoride nanoparticles. Relative to the best-performing conventional LIR calibration, full-spectrum regression reduces the absolute prediction bias by up to approximately 2.5-fold and the prediction spread by 24% for visible green upconversion. For near-infrared emission, the corresponding reductions reach approximately 2.2-fold and 8.6-fold, respectively. The approach reduces feature-selection bias and exploits temperature-encoded information that conventional analysis discards. Shapley additive explanations for the best-performing Gaussian process regression model identify the Er3+ spectral regions that dominate temperature prediction, linking the improved performance to physically meaningful emission features. These results demonstrate improved temperature readout using full-spectrum regression for YOF:Er3+/Yb3+ under the investigated conditions. Matched synthetic-background tests showed that training-set augmentation reduced background-induced set-point bias, with effects that depended on the regression method and normalization.