Selective identification and quantification of VOCs using metal nanoparticles decorated SnO2 hollow-spheres based sensor array and machine learning

Abstract

Accurate and selective detection of target gas/volatile organic compounds (VOCs) is of utmost importance. The chemiresistive gas sensors have been a desirable candidate due to their compact footprint and ease of fabrication, but they show poor selectivity. This work presents a combination of nanomaterials-based chemiresistive gas sensors with machine learning (ML) techniques to achieve sensitive, selective, and quantified detection of tested VOCs. The sensor array consists of four separate sensing layers over interdigitated electrodes-based platform. The sensing materials were comprised of silver, gold, palladium, and platinum nanoparticles decorated on tin oxide hollow-sphere structures which were successfully synthesized through chemical routes and characterized accordingly. Surface decoration of different metal nanoparticles has produced sensitive and diverse sensing patterns among the tested VOCs. The sensing mechanism and related gas sensing kinetics were then analyzed to explain high sensitivity and diverse sensing phenomena. The subsequent incorporation of ML models has resulted in qualitative and quantitative detection of VOCs. A comparative analysis was carried out among different types of applied features and ML models with reasoning. Particularly, a deep neural network (DNN) model with time series (TS) response sequence as input information, delivered the best performance. The DNN_TS model presented an average classification accuracy of 98.33 %, in conjunction with excellent concentration prediction. The DNN_TS model showed a very fast prediction time of 2.74 µs with adaptive learning while utilizing minimum computing resources, which favors the real-time sensing capability. The reported results promote the development of an autonomous, smart, and selective gas sensor system for real-time applications.