Detection of trace levels of volatile organic compounds (VOCs) has widespread applications, including wearable diagnostics, IoTs, and indoor air quality control. Although metal oxide semiconductors (MOS) arguably offer the best sensitivity for a wide range of VOCs, their poor selectivity limits their performance. Here, we demonstrate a machine learning (ML)-based analysis and framework using a single, non-selective MOS sensor made of RF-sputtered nickel oxide thin film with gold contacts, aiming to achieve VOC classification and concentration prediction with a high degree of accuracy (>90%) and eliminate biases. Both time-independent and time-dependent features were evaluated using classifiers and regressors, including ensemble methods, artificial neural networks, and recurrent architectures (LSTMs and GRUs). The features identified as excluding time reference (response, its gradient, and Laplacian) were highly effective for baseline classification, achieving near-ideal accuracies (98%) with ensemble models. On the other hand, the time-dependent features (continuous, discrete, and time-sliced) complement the analysis by capturing dynamic adsorption-desorption kinetics via sequential models, leading to accuracies of 94% and above. Regression analysis techniques enhance the predictive capabilities of ensemble and neural approaches, yielding higher R2 values and lower RMSE. Thus, the methods adopted in this work highlight the complementary approach of ML-based modeling with that of material innovation to achieve an important performance metric, namely, selectivity of MOS-based sensors, as a way forward for scalable, real-time VOC monitoring in a complex background of other gases. This approach is highly scalable for other toxic gases, pollutants, and biomarkers for relevant applications.
Savio et al. (Wed,) studied this question.