Machine Learning Assisted Discrimination of 20 Volatile Organic Compounds with Response Prediction Via UNIFAC Modeling

Abstract

Accurate detection of volatile organic compounds (VOCs) in complex background plays an important role in both natural and human-related activities, such as environmental protection, medical screening, food safety, and homeland security. Reliable monitoring of atmospheric VOCs is vital because vapor exposure can accelerate air-borne toxic generation and lead to tropospheric air pollution. While conventional analytical instruments such as gas chromatography-mass spectrometry (GC-MS) are largely used for identification of unknown VOCs, it is desirable for sensing technology to transition into a low cost, energy-efficient, and miniaturized media that can be implemented into personal devices.As one of the fastest growing types of chemical sensors, chemiresistors possess many desirable characteristics such as straightforward fabrication and easy to integrate. Sensing responses for nanoparticle-based chemiresistor are based on the sorbent-phase interaction of an analyte with the sensing materials, which can be explained by a thermally activated electron tunneling mechanism. The morphology of the gold-nanoparticle (AuNP) assembly (i.e. the interparticle distance) between the electrodes determines the baseline resistance of that individual sensor element.In this work, we demonstrate the detection and classification of 20 VOC vapors at room temperature using a multi-AuNP chemiresistor array fabricated with gold nanoparticles with chemically diverse ligands, including 4-(dimethylamino)pyridine (DMAP), tetradecylamine (TDA), and 4-aminothiophenol (4-ATP). The castellated electrode structures were fabricated using standard photolithography and liftoff processes. Classification accuracy was determined using a machine learning algorithm on a matrix of sensing response data. Supervised classifiers applied in this work include linear discriminant analysis (LDA), support vector machine (SVM), k-nearest neighbors (KNN), and random forest (RF). The results demonstrated that pattern recognition accuracy increases when a combination of ligands was used, as compared to sensors with one type of ligand. This proof-of-concept study illustrated that we were able to improve classification accuracy by choosing ideal sensing elements with optimized built-in redundancy. Meanwhile, a vapor sorption model was established to predict the chemiresistive sensor responses. One of the biggest challenges of AuNP-based chemiresistive sensor technology is finding the best candidates of materials that provide sufficient resolving power to differentiate chemicals in complex surroundings for molecular pattern recognition. To evaluate the sensor performance of ligand-analyte interactions, a UNIFAC (UNIQUAC Functional-group Activity Coefficients) model was established in this study to predict the sorption between a certain ligand material and VOC vapors. Therefore, a more performance-efficient sensing element library can be established for different types of analytes. Specifically, partition coefficients were determined from the UNIFAC method, and the sorbent phase concentration was obtained from modified Raoult’s Law. To validate our model, simulated UNIFAC partition coefficients were compared with experimental results and data reported from literature. Besides sensor responses, other characteristic parameters of chemiresistor performance are also studied, including sensitivity, sensitivity ratio, volumetric sensitivity, and molar sensitivity. This work illustrated the versatility of gold nanoparticle-based chemiresistor array systems for VOC detection, discrimination, and sensor signal prediction. Figure 1