CuO–ZnO p-n junctions for accurate prediction of multiple volatile organic compounds aided by machine learning algorithms

Highlights

• CuO, ZnO nanoflakes and CuO–ZnO heterojunctions were synthesized using facile hydrothermal method.

• All the nanostructured metal oxides were tested for methanol, acetonitrile, isopropanol, and toluene.

• The highest sensitivity towards the VOCs were observed to be modulated with the composition of the heterostructures.

• Machine learning algorithms were used to get high classification and quantification accuracies.

• Only two sensors' response and response times could lead to excellent classification and quantification of all the four VOCs.

Abstract

Detection and quantification of multiple volatile organic compounds (VOCs) are emerging as critical requirements for several niche applications including healthcare. It is desirable to get multiple gases identified rapidly and using minimum number of sensors. Heterojunctions of metal oxides are still among the top-picks for efficient VOC sensing because they unfold exciting sensing characteristics in addition to enhanced response. This work reports the synthesis of nanostructures of CuO, ZnO, and three CuO–ZnO p-n junctions having different weight percentages (1–0.5, 1-1, and 0.5–1) of CuO and ZnO, using a facile one-pot hydrothermal method. The nanomaterials were characterized using X-ray diffraction, field emission scanning electron microscopy, and UV–Visible spectroscopy. Resistive sensors were fabricated of all five nanomaterials and were tested for 25–200 ppm of four VOCs – isopropanol, methanol, acetonitrile, and toluene. The CuO and CuO–ZnO (1–0.5) sensors showed the highest response for isopropanol (7.5–65.3% and 19–122%, respectively) at 250 °C, CuO–ZnO (1-1) and CuO–ZnO (0.5–1) exhibited the highest responses for methanol (9–60%) and isopropanol (15–120%), respectively at 350 °C, and the intrinsic ZnO showed maximum response to toluene (29–76%) at 400 °C. All the sensing layers were observed to exhibit finite responses to the other three VOCs so, an attempt to classify and quantify the four VOCs accurately was made using support vector machine (SVM) and multiple linear regression (MLR) algorithms. The response and response times of two sensors were observed to be sufficient as inputs to the machine learning algorithms for classifying and quantifying all the four VOCs. The combinations of (CuO–ZnO (1–0.5) & (1-1) and CuO–ZnO (1-1) & (0.5–1) demonstrated the highest classification accuracy of 98.13% with SVM. The combination of CuO–ZnO (1–0.5) & (1-1) demonstrated the best quantification of the four VOCs using MLR.

Introduction

In the past few decades, resistive sensors have been extensively utilized in the detection of volatile organic compounds (VOCs) for various applications such as environmental monitoring [1], explosive detection [2], disease diagnosis [3], precision agriculture for crop health monitoring and in food safety, etc [4]. Among all, disease diagnosis using resistive gas sensors is gaining rapt attention and is considered to be an absolute necessity [5]. This is predominantly because resistive sensors offer excellent sensitivity, exceptional stability, adaptability (simple to use and can be incorporated into portable devices), reduced power usage, and are inexpensive as compared to their competitors like optical sensors and chromatography-based sensors [6,7]. In diagnostics, vapor sensors leverage the fact that breath exhaled by humans comprises multiple VOCs, that are by-products of various metabolic activities occuring inside the body. At the onset of disease, either the concentrations of some of these VOCs change or in a few cases, new VOCs get introduced into the breath samples. These changes are very specific to certain metabolism which are altered due to the incidence of that particular disease [8]. For example, acetone, methanol, isoprene, toluene, isopropanol, acetonitrile, formaldehyde, 2-butanone, etc [9,10] were repeatedly identified as biomarkers of lung cancer which is one of the leading causes of deaths globally and its incidence is equally high in both males and females [11]. Hence, the detection and quantification of these biomarkers open avenues for non-invasive and rapid detection of the disease [12].

Resistive sensors lead to the identification and quantification of the VOCs of interest by resulting in a change in the resistance of the sensing layer. Several materials have been explored as sensing layers for resistive sensors to detect a set of VOCs [13]. The advent of nanomaterials, however, has taken the capabilities of the resistive sensor to a different plane. This is because gas sensing is a surface phenomenon and the nanomaterials with at least one of their dimensions in the nanoscale, extend very high surface-to-volume ratios [14]. Of all, nanostructures of metal oxides have been synthesized and tested as VOC sensors, the most. This is because the semiconducting sensing layers exhibit ultra-high sensitivity, fast response, and recovery [15]. A number of metal oxide nanostructures, such as SnO₂ [16], TiO₂ [17], WO₃ [18], CuO [19], Fe₂O₃ [20], ZnO [21] etc. have been reported to function as vapor sensors. Despite of such outstanding properties, metal oxides offer extremely limited selectivity [22]. To address the problem of poor selectivity and render better sensitivity to the intrinsic metal oxides, various efforts were made on multiple fronts. One front is the surface modification of intrinsic metal oxides with functional activators such as noble metals (Au, Ag, Pt, Pd, etc.) [23,24], carbon nanomaterials (carbon nanotubes, graphene oxide, graphitic carbon nitride, activated carbon, etc.) [25,26], rare earth metals (Lanthanide series elements) [27,28], etc. Another excellent way to get a functionalized surface for vapor sensing is the controlled formation of metal oxide hetero junctions. In this technique, two dissimilar semiconductors are used to form either a p-n junction or a heterojunction. This helps in extending the type of space charge regions available in the sensing layer, thereby making it more favourable for vapor adsorption [29]. For example, Du et al. have demonstrated excellent formaldehyde sensing by hydrothermally synthesized SnO₂–In₂O₃ at 375 °C [30]. Similarly, Yan et al. prepared SnO₂–ZnO nanofibres by electrospinning method and showed a stable and excellent response towards ethanol at 300 °C [31]. Naik et al. fabricated a composite of WO₃ and ZnO by mechanical mixing and demonstrated ultrahigh sensitivity towards ethanol at 350 °C [32]. Another exciting possibility could be forming heterostructures with CuO which is a p-type semiconductor having a band gap between 1.2 and 2 eV and ZnO which is an n-type material having band gap of 3–3.5 eV [33,34]. Several reports are available that demonstrate sensing by CuO–ZnO heterostructures for detection of various gases such as acetone [35], formaldehyde [36], ethanol [37], CO₂ [38], NO₂ [39], etc. The collective finding of previous works is that the CuO–ZnO heterostructures form simple, rapid, and highly sensitive vapor sensors [40]. This inspired the authors to explore this combination for sensing VOCs that have diagnostic relevance.

Though the surface engineering techniques enhance the sensitivities of the intrinsic metal oxides and were observed to tune the selectivity, they failed to provide adequate selectivity to the semiconductors which would make them suitable for applications like breath analysis [41,42]. Hence, exploring machine learning (ML) algorithms on the data generated by the resistive sensors have begun in the past few years. Several ML algorithms have been seen to efficiently classify and quantify the VOCs of interest. These algorithms include linear regression [43], support vector machines (SVM) [44], random forest [45], k-Nearest neighbours [46], Naïve Bays classifier [47], logistic regression [48], artificial neural networks [49], etc. Among all, SVM can function with a limited amount of training data and provides a better generalization by maximizing the linearly separating margin resulting in reduced test errors. Since the scenario of limited data is true in the case of gas sensors, SVM was believed to be a better option for classification and similarly, linear regression (LR) was expected to be better for quantification of gases.

In this work, CuO–ZnO heterostructures with different weight ratios of the two metal oxides were synthesized following a cost-effective and simple one-pot hydrothermal route. Resistive sensors were developed and detailed studies were conducted with intrinsic CuO and ZnO nanoflakes and with three CuO–ZnO heterostructures (1–0.5, 1-1, and 0.5–1) for detecting four VOCs–methanol, acetonitrile, toluene, and isopropanol which also happen to be biomarkers of lung cancer. The sensing layers were exposed to ten cycles of four different concentrations (25–200 ppm) of all the VOCs at their respective optimum temperatures of operation. It was observed that none of the sensing layers exhibited high selectivity towards any particular VOC. Hence, in an attempt to improve the prediction capabilities of the sensors, their steady-state responses and response times were taken as input features for SVM and LR algorithms for the classification and quantification of the four gases, respectively.