Utilizing Deep Learning to Enhance Optical Sensing of Ethanol Content Based on Luminescent Carbon Dots

Abstract

Optical sensing methods have shown their potential in measuring environmental properties such as temperatures and sample concentrations, increasing the convenience and speed of such measurements. Carbon dots (CDs) are readily available, easy-to-synthesize, and non-toxic luminescent colloidal nanoparticles that have gained prominence as optical probes in recent years. A disadvantage of CDs used for optical sensing is their broad emission profile, leading to unspecific sensing and a potential overlap of their luminescence with the autofluorescence of the samples. Machine learning approaches can address these shortcomings and greatly enhance sensing accuracy. In this study, CDs were employed as optical probes, while machine learning methods were applied to optimize ethanol content determination in ethanol/water mixtures as well as in alcohol-containing beverages. A simple neural network was used to understand the importance of different optical parameters on sensing, while a modular deep learning model was developed for increased generalizability towards samples with strong autofluorescence. Drawing from multiple input channels, the deep learning model was able to predict ethanol concentrations with a mean absolute error of 0.4 vol % in pure solvents and 6.7 vol % in beverages (beers, wines, and spirits). While CDs are excellent candidates to demonstrate deep learning for optical sensing, the methods discussed in this work are promising for improving chemical sensing using various luminescent materials.