Improving the Accuracy of Carbon Dot Temperature Sensing Using Multi-Dimensional Machine Learning

Abstract

Optical sensing methods offer a convenient noncontact approach to monitor different environmental parameters with a high spatial resolution and fast response times. Temperature monitoring can benefit from optical sensing using luminescent nanoprobes, but many of those substances are toxic or expensive. Carbon dots are a class of luminescent colloidal nanoparticles that have recently gained recognition as optical probes, which are easy to produce by environmentally friendly synthesis, nontoxic, and stable. While carbon dots show temperature-dependent optical properties, their broad emission profiles may constitute a challenge for optical sensing. In this study, three types of carbon dots with different emission profiles were tested as optical probes for intensity-, spectral-shift-, intensity-ratio-, bandwidth-, and lifetime-based temperature sensing. Depending on the optical characteristics of the specific probe, either intensity-or lifetime-based sensing was shown to be the most accurate, with accuracies of up to 1.65 and 0.70 K, respectively. Employing Gaussian fits improved accuracies of the intensity-ratio-based sensing to 1.24 K, with the additional benefit of greater stability against excitation fluctuations. Finally, a multiple linear regression model combining steady-state and time-resolved luminescence data of carbon dots has been applied to further increase the sensing accuracies with carbon dots to 0.54 K. Our study demonstrates how multidimensional machine learning methods can greatly improve temperature sensing with optical probes.