Machine learning-enhanced temperature sensing of spectrally featureless fluorophores

Current efforts in fluorescence-based optical thermometry have concentrated on probe optimization to minimize prediction errors, typically requiring fluorophores with multiple emission bands for ratiometric analysis or large emission peak shifts for reliable calibration. However, fluorophores exhibiting a single emission band with a modest peak shift - here termed spectrally featureless - are excluded from conventional thermometric approaches, as single-band intensity measurements are susceptible to excitation source fluctuations and photobleaching, while their peak shifts alone are insufficient for minimizing prediction errors in conventional temperature estimation methods based on calibration curves. Integration of machine learning (ML) with spectral analysis offers a promising route to overcome these limitations, yet has not been widely explored in this field. In this study, we aimed to optimize ML algorithms for fluorescence-based temperature sensing using resveratrone - a spectrally featureless organic fluorophore exhibiting a single emission band and a modest peak shift under temperature variation - as the sensing probe. We tested four different ML models: multiple linear regression, support vector regression, linear random forest, and multilayer perceptron algorithms as the primary regression frameworks, alongside conventional chemometric comparators for comparison. Our comparison showed that the pyramid-structure multilayer perceptron worked best at capturing temperature-dependent spectral changes and outperformed the other models in testing, achieving a relative performance index of 170 (with the single-feature polynomial regression baseline defined as 100), reflecting the lowest test MAE among all models evaluated. This advantage persisted under an independently measured condition involving a different day, concentration, and solvent matrix, where the pyramid-structure multilayer perceptron outperformed the other models by more than 1.2 °C in test MAE. This approach may help expand the range of viable materials for temperature sensing by enabling accurate temperature predictions from steady-state photoluminescence spectra of resveratrone, supporting potential applications of fluorescence nanothermometry within the temperature range investigated in this study (25-70 °C), which encompasses biologically and biomedically relevant conditions.