Towards accurate real-time luminescence thermometry: an automated machine learning approach

TL;DR

This paper introduces an automated machine learning approach to enhance the accuracy and simplicity of luminescence thermometry, enabling real-time temperature measurements with high resilience to spectral variances.

Contribution

The study presents a novel ML pipeline that significantly improves photoluminescence thermometry accuracy without complex physical modeling or deconvolution.

Findings

Accuracy improved by over 5.5 times compared to traditional methods

Method is simple and does not require physical mechanism knowledge

Resilient to spectral variance across measurements

Abstract

Luminescence thermometry has been extensively exploited in the last decades both from the fundamental and applied point of views. The application of photoluminescent nanoparticles on the microscopic level based on rare-earth doped (RED) nanostructures is yet a challenge. Distinct underlying physical mechanisms in the RED nanomaterials have been exploited, such as intensity ratio between radiative transitions associated with thermally coupled energy levels, energy peak and lifetime of an excited state variations with the temperature. The drawbacks of such systems are the relatively low thermal sensitivity (Sr), and the large temperature uncertainty. To overcome that, several research groups have been seeking new functionalized materials. The majority of the efforts have been directed towards increasing Sr with record around 10 %{\deg}C-1, which is, however, considered unsatisfactory. We propose the use of an automated machine learning tool to retrieve an ideal pipeline improving the response of photoluminescence thermometers. As a proof-of-concept, we used Nd3+-doped YAG nanoparticles, excited at 760 nm, and the photoluminescence spectra in the range from 860 nm to 960 nm as input parameters. In addition to the improvement in the accuracy (> 5.5x over traditional methods), the implementation is very simple, without the requirement of any deconvolution procedure or knowledge of any underlying physical mechanism. Our findings demonstrate that this approach is resilient to natural variances across various spectral acquisitions, which may otherwise lead to an inaccurate estimation of temperature, opening the door for real-time applications. Our open-source code is designed to be accessible to individuals without programming experience.