Convolutional neural networks open up horizons for luminescence thermometry

Abstract

Luminescence thermometry has emerged as an effective way to implement non-invasive thermal reading and finds promising applications in many fields. Nevertheless, traditional luminescence thermometric methods, such as the luminescence intensity ratio, are generally carried out based on one single temperature-dependent spectral feature. As a result, other thermal related spectral features were ignored, which radically restricts the sensing performance. Herein, we preliminarily propose a novel strategy for driving luminescence thermometry via convolutional neural network. The trained network is able to autonomously select and extract multiple temperature-dependent features for regression temperature, so the temperature-dependent spectral data can be fully utilized. Using Y3Al5O12: Cr3+ as the temperature sensing material, the proposed thermometry method exhibits high accuracy and strong generalization over the additional test set. The maximum measurement error (emax) in the range of 35–315 °C is about 0.77 °C, accompanied with the average error (eaver) of 0.20 °C. The accuracy of the proposed approach is significantly superior to those achieved by the classical ratiometric technology (emax≈4.05 °C, eaver ~ 1.31 °C) and the widely used multiple linear regression method (emax≈4.27 °C, eaver ~ 1.69 °C), indicating deep learning has the potential to open up a new era of luminescence thermometry.