Classification of Tea Aromas Using Multi-Nanoparticle Based Chemiresistor Arrays

Nanoparticle based chemical sensor arrays with four types of organo-functionalized gold nanoparticles (AuNPs) were introduced to classify 35 different teas, including black teas, green teas, and herbal teas. Integrated sensor arrays were made using microfabrication methods including photolithography and lift-off processing. Different types of nanoparticle solutions were drop-cast on separate active regions of each sensor chip. Sensor responses, expressed as the ratio of resistance change to baseline resistance (ΔR/R0), were used as input data to discriminate different aromas by statistical analysis using multivariate techniques and machine learning algorithms. With five-fold cross validation, linear discriminant analysis (LDA) gave 99% accuracy for classification of all 35 teas, and 98% and 100% accuracy for separate datasets of herbal teas, and black and green teas, respectively. We find that classification accuracy improves significantly by using multiple types of nanoparticles compared to single type nanoparticle arrays. The results suggest a promising approach to monitor the freshness and quality of tea products.


Introduction
As one of the most popular beverages, tea is consumed by hundreds of millions of people worldwide [1]. Known for its health benefits such as energizing, relieving stress, strengthening the immune system, and alleviating digestive problems [2][3][4], tea has been intensively studied for its effects related to human health [5][6][7]. Green tea and black tea are harvested directly from the leaves of tea plants, Camellia sinensis, which contain many polyphenols that are natural antioxidants [1]. Herbal teas, on the other hand, are mixtures of caffeine-free leaves, seeds, spices, and plant roots that give a unique aroma and taste. Some of the most common ingredients include chamomile, cinnamon, ginseng, ginger, hibiscus, peppermint, and rose hip [2]. Similar to green tea and black tea, herbal tea also provides many medicinal properties, such as destressing and certain disease prevention [2, 8,9].
One key property that distinguishes tea products is aroma, which is also a key factor for consumer taste. Depending on the interactions between constituted ingredients and aroma compounds, each tea product has a complex odor and smells differently [10]. Furthermore, aroma is also related to freshness Four types of monolayer protected gold nanoparticles (AuNPs) were used for sensing experiments. DMAP-AuNPs were synthesized following a method reported by Hubble and coworkers with an average size of 5 nm [20]. Gold nanoparticles were first synthesized using the Brust method [32] and followed by phase transfer into an aqueous phase [33]. ODA-AuNP, MPA-AuNP, and ATP-AuNP with approximate size range 10-20 nm were prepared using an approach developed by Chen et al. [34] with minor modifications. ODA (4 mmol) and HAuCl 4 ·3H 2 O (0.05 mmol) were added into a 25-mL three-neck flask with a magnetic stir bar. The system was degassed by nitrogen gas to remove dissolved oxygen. The temperature was increased to 140 • C and kept for 20 min. Once the reaction was complete, a mixture of ethanol and toluene was added to remove excess reactants and surfactants. The nanoparticles were dissolved in ethanol for further use. For MPA-AuNP and ATP-AuNP synthesis, an equimolar of MPA and ATP were added to the reaction mixture, respectively, to complete ligand exchange reactions from ODA. Thiol-based chemistry is commonly used for monolayer formation on AuNPs and provides good sensing performance with long-term stability [35,36]. Other possible linker chemistries include diazonium compounds, which also provide robust core-shell structures [37][38][39]. Transmission electron microscopic imaging (Talos F200X TEM, Thermo Fisher Scientific, Waltham, MA, USA) was performed to measure the morphology and sizes of as-synthesized gold nanoparticles.

Fabrication of Sensor Arrays
Chemiresistor devices with 5 µm wide electrodes and 2 µm electrode separations were fabricated using standard photolithography and liftoff processes on 4 inch silicon wafer substrates, following a previously reported method [30]. Each sensor chip had 48 individually addressable sensor elements that were separated into four groups at each corner, consisting of a 4960 µm 2 total active sensing area. Photolithography was performed using a maskless aligner (Heidelberg Instruments, Heidelberg, Germany) that defined the devices and their electrical contacts. Gold microelectrodes and electrical contacts were obtained by electron-beam evaporation of 200 nm Au with a 10 nm Ti adhesion layer. The electrodes were obtained after liftoff processing. Subsequently, wafers were diced into individual chips  15.9 × 15.9 mm 2 in size. Before nanoparticle deposition, sensor chips were immersed in a 3-MPTES solution (2% v/v in toluene) overnight at room temperature for thiol-decoration on SiO 2 surface areas. The chip was then washed with toluene and dried by nitrogen, followed by a 2-hour bake at 110 • C to improve the coupling capability of 3-MPTES [40].

Aroma Sensing Experiments
Gold nanoparticle solutions were deposited onto the microelectrode regions of a sensor chip with direct drop-casting. A 2 µL droplet of each type of AuNP was placed at a corner of a sensor chip using a micropipette. The droplets covered the entire active sensing region of the devices. Once the solvent evaporated, the surface of the chip was cleaned by a stream of dry nitrogen gas (99.999% purity). Scanning electron microscope analysis (Verios 460L SEM, Thermo Fisher Scientific, Waltham, MA, USA) was conducted to verify nanoparticle assembly at the microelectrodes.
Room temperature sensing experiments were performed using a probe station (Rucker & Kolls, Mountain View, CA, USA) with a custom-designed probe card (Wentworth Laboratories, Brookfield, CT,  Figure 1 shows a schematic illustration of the electronic nose system. High speed data acquisition was made using a switch matrix/multimeter system (Keithley, Cleveland, OH, USA) that was connected to the probe card. Dry tea leaves and powders (average 0.40 ± 0.01 g) were extracted from their original packages and transferred into 20-mL syringes (BD, Franklin Lakes, NJ, USA). The tea samples were stored in syringes overnight to allow tea aromas to equilibrate. Tea aroma samples in air were delivered to a custom-made Teflon gas mixer by a syringe pump (Cole-Parmer, Vernon Hills, IL, USA). A dry nitrogen stream was used as a mixing/purge gas to combine with the aromas inside the gas mixer and to clear the sensor after each syringe pulse. The tea samples were stored in syringes overnight to allow tea aromas to equilibrate. Tea aroma samples in air were delivered to a custom-made Teflon gas mixer by a syringe pump (Cole-Parmer, Vernon Hills, IL). A dry nitrogen stream was used as a mixing/ purge gas to combine with the aromas inside the gas mixer and to clear the sensor after each syringe pulse. Real-time resistance data were obtained from the switch matrix / multimeter system, using a constant-current sourcing method for R0 ≤ 1 MΩ or a ratiometric method for R0 > 1 MΩ. To extract sensing data, a custom MATLAB program was used to process raw data and calculate sensor responses as ΔR/R0, where R0 is the baseline resistance and ΔR is the deviation from baseline resistance due to tea aroma interactions with sensor elements.

Data Analysis
For each tea aroma sample, a total of eight replicate measurements (pulses of vapor) were performed on all 48 devices simultaneously. The sensor responses were extracted into a 280 × 48 data matrix that was used for statistical analysis. The rows represent individual measurements, and the columns represent the sensor element responses (ΔR/R0). Principal component analysis (PCA) was conducted to visualize grouping within each aroma and separation between aromas. To evaluate the discrimination performance of the electronic nose array, cross-validation was conducted along with linear discrimination analysis (LDA), support vector machine (SVM), k-nearest neighbors (KNN), and random forest (RF) in order to calculate the classification accuracy. Data analysis and pattern recognition were done using the MATLAB Classification Learner Toolbox for LDA, SVM, and KNN, and coding with the programming language Julia for RF. Figure 2 shows scanning electron microscope images of micro-junctions assembled with MPAfunctionalized gold nanoparticles. The average baseline resistance for MPA-AuNP (15-nm average size) chemiresistors was 41.0 ± 21.8 MΩ. By using direct drop-casting, clusters of AuNP sensing elements covered adjacent electrodes to form a closed electrical circuit. The cluster formation ensured that electron transport occurs between nanoparticles. The low coverage of the electrodes in the figure was typical of the experiments, and the curvature of the microelectrodes is due to the resolution limit of the maskless writer. Real-time resistance data were obtained from the switch matrix/multimeter system, using a constant-current sourcing method for R 0 ≤ 1 MΩ or a ratiometric method for R 0 > 1 MΩ. To extract sensing data, a custom MATLAB program was used to process raw data and calculate sensor responses as ∆R/R 0 , where R 0 is the baseline resistance and ∆R is the deviation from baseline resistance due to tea aroma interactions with sensor elements.

Data Analysis
For each tea aroma sample, a total of eight replicate measurements (pulses of vapor) were performed on all 48 devices simultaneously. The sensor responses were extracted into a 280 × 48 data matrix that was used for statistical analysis. The rows represent individual measurements, and the columns represent the sensor element responses (∆R/R 0 ). Principal component analysis (PCA) was conducted to visualize grouping within each aroma and separation between aromas. To evaluate the discrimination performance of the electronic nose array, cross-validation was conducted along with linear discrimination analysis (LDA), support vector machine (SVM), k-nearest neighbors (KNN), and random forest (RF) in order to calculate the classification accuracy. Data analysis and pattern recognition were done using the MATLAB Classification Learner Toolbox for LDA, SVM, and KNN, and coding with the programming language Julia for RF. Figure 2 shows scanning electron microscope images of micro-junctions assembled with MPA-functionalized gold nanoparticles. The average baseline resistance for MPA-AuNP (15-nm average size) chemiresistors was 41.0 ± 21.8 MΩ. By using direct drop-casting, clusters of AuNP sensing elements covered adjacent electrodes to form a closed electrical circuit. The cluster formation ensured that electron transport occurs between nanoparticles. The low coverage of the electrodes in the figure was typical of the experiments, and the curvature of the microelectrodes is due to the resolution limit of the maskless writer. Before conducting sensing experiments, the baseline resistance of each of the 48 individual devices was examined. For devices with baseline resistance R0 < 10 3 Ω, it is likely that large agglomerations of nanoparticles formed and covered the electrode junctions, or that layers contain partially sintered particles [20,30,41]. These devices usually have less than 0.01% change of resistance for interactions with vapor analytes and are not useful sensor elements. On the other hand, devices with high baseline resistance (R0 > 10 8 Ω) are also removed from the analysis. These high resistance chemiresistors may have incomplete nanoparticle films, and sensing responses are typically noisy (low S/N ratio). Therefore, only devices with baseline resistance in the range of 10 3 Ω ≤ R0 ≤ 10 8 Ω were included in sensor performance evaluation.

Sensor Response Profiles
The sensor chip was first measured under 0.5 SCFH (236 SCCM) of purge gas flow to reach a stable baseline resistance and then exposed to eight sequential 2.5 s pulses (3.3 mL per pulse) of a tea aroma. The total experimental duration was 170 s for each tea analyte. Each peak represents a sampling period that includes a pulse phase and a purge/recovery phase. The ratio between the maximum change of resistance (ΔR) and the baseline resistance (R0), after baseline correction with a Savitzky-Golay filter, was processed as the output. Figure 3 shows sample data of a DMAP-AuNP chemiresistor responding to three different tea aromas, as well as baseline resistance variation among the four investigated monolayer protected nanoparticles. As shown in Figure 3A for a selected sensor element assembled with DMAP-AuNP, the sensor responses are strongest toward Earl Grey (2.57 ± 0.07%), followed by Bengal Spice ® (1.75 ± 0.08%) and Matcha (0.83 ± 0.05%). Fast and reversible detection was demonstrated with a sampling rate of 20 ms between data points. A feature of nanoparticle-based chemiresistor arrays is the capability of the system to respond to a variety of analytes with different signal strengths. This is achieved by variations in the chemistry of the monolayer protected nanoparticles as well as nanoparticle assembly morphology [20,30,42]. From a microscopic view, tea aromas full of many chemical ingredients Before conducting sensing experiments, the baseline resistance of each of the 48 individual devices was examined. For devices with baseline resistance R 0 < 10 3 Ω, it is likely that large agglomerations of nanoparticles formed and covered the electrode junctions, or that layers contain partially sintered particles [20,30,41]. These devices usually have less than 0.01% change of resistance for interactions with vapor analytes and are not useful sensor elements. On the other hand, devices with high baseline resistance (R 0 > 10 8 Ω) are also removed from the analysis. These high resistance chemiresistors may have incomplete nanoparticle films, and sensing responses are typically noisy (low S/N ratio). Therefore, only devices with baseline resistance in the range of 10 3 Ω ≤ R 0 ≤ 10 8 Ω were included in sensor performance evaluation.
The sensor chip was first measured under 0.5 SCFH (236 SCCM) of purge gas flow to reach a stable baseline resistance and then exposed to eight sequential 2.5 s pulses (3.3 mL per pulse) of a tea aroma. The total experimental duration was 170 s for each tea analyte. Each peak represents a sampling period that includes a pulse phase and a purge/recovery phase. The ratio between the maximum change of resistance (∆R) and the baseline resistance (R 0 ), after baseline correction with a Savitzky-Golay filter, was processed as the output. Figure 3 shows sample data of a DMAP-AuNP chemiresistor responding to three different tea aromas, as well as baseline resistance variation among the four investigated monolayer protected nanoparticles. Before conducting sensing experiments, the baseline resistance of each of the 48 individual devices was examined. For devices with baseline resistance R0 < 10 3 Ω, it is likely that large agglomerations of nanoparticles formed and covered the electrode junctions, or that layers contain partially sintered particles [20,30,41]. These devices usually have less than 0.01% change of resistance for interactions with vapor analytes and are not useful sensor elements. On the other hand, devices with high baseline resistance (R0 > 10 8 Ω) are also removed from the analysis. These high resistance chemiresistors may have incomplete nanoparticle films, and sensing responses are typically noisy (low S/N ratio). Therefore, only devices with baseline resistance in the range of 10 3 Ω ≤ R0 ≤ 10 8 Ω were included in sensor performance evaluation.
The sensor chip was first measured under 0.5 SCFH (236 SCCM) of purge gas flow to reach a stable baseline resistance and then exposed to eight sequential 2.5 s pulses (3.3 mL per pulse) of a tea aroma. The total experimental duration was 170 s for each tea analyte. Each peak represents a sampling period that includes a pulse phase and a purge/recovery phase. The ratio between the maximum change of resistance (ΔR) and the baseline resistance (R0), after baseline correction with a Savitzky-Golay filter, was processed as the output. Figure 3 shows sample data of a DMAP-AuNP chemiresistor responding to three different tea aromas, as well as baseline resistance variation among the four investigated monolayer protected nanoparticles. As shown in Figure 3A for a selected sensor element assembled with DMAP-AuNP, the sensor responses are strongest toward Earl Grey (2.57 ± 0.07%), followed by Bengal Spice ® (1.75 ± 0.08%) and Matcha (0.83 ± 0.05%). Fast and reversible detection was demonstrated with a sampling rate of 20 ms between data points. A feature of nanoparticle-based chemiresistor arrays is the capability of the system to respond to a variety of analytes with different signal strengths. This is achieved by variations in the chemistry of the monolayer protected nanoparticles as well as nanoparticle assembly morphology [20,30,42]. From a microscopic view, tea aromas full of many chemical ingredients As shown in Figure 3A for a selected sensor element assembled with DMAP-AuNP, the sensor responses are strongest toward Earl Grey (2.57 ± 0.07%), followed by Bengal Spice ® (1.75 ± 0.08%) and Matcha (0.83 ± 0.05%). Fast and reversible detection was demonstrated with a sampling rate of 20 ms between data points. A feature of nanoparticle-based chemiresistor arrays is the capability of the system to respond to a variety of analytes with different signal strengths. This is achieved by variations in the chemistry of the monolayer protected nanoparticles as well as nanoparticle assembly morphology [20,30,42]. From a microscopic view, tea aromas full of many chemical ingredients interact at the molecular scale with the monolayer protected nanoparticle organic shells, which induces a change in interparticle separation that affects electron transport through the nanoparticle films [18,30,31]. Figure 3B illustrates baseline resistances and sensor responses to Black Cherry Berry for four different monolayer protected nanoparticles on a single chip. After eliminating devices with out-of-criterion baseline resistance, 38 of 48 devices remained valid for analysis. For each monolayer protected nanoparticle, sensor response measurements were carried out with at least eight individual sensing elements. Error bars reflect the spread of R 0 and ∆R/R 0 for each grouping of monolayer protected nanoparticles. Sensor elements assembled with aqueous-phase DMAP-AuNPs possess the lowest baseline resistance (6.8 ± 4.8 × 10 5 Ω), followed by ODA-AuNPs (4.1 ± 2.4 × 10 6 Ω), MPA-AuNPs (4.1 ± 2.2 ×10 7 Ω), and ATP-AuNPs (4.9 ± 2.2 × 10 7 Ω). Sensor responses follow the same trend as baseline resistance up to 10 6 Ω range, after which responses stabilize around ∆R/R 0 = 5%.
The variations of baseline resistance and sensor response levels are essential for pattern recognition using machine learning algorithms. Depending on the effective chemical interactions between analytes and monolayer protected core-shell nanostructures, the AuNP layers swell at different magnitudes [43], which results in variations of sensor responses. The sorption of vapors into the sensing materials can be qualitatively understood in terms of non-covalent van der Waals interactions that vary with each analyte and sensor material [18,44,45].

Sensor Performance and Classification Accuracy
Sensor element responses (∆R/R 0 ) to the 35 different tea aromas were collected and processed as inputs for statistical analysis. Principal component analysis (PCA) was performed as an unsupervised method to examine the natural grouping and separations between the different teas. The analytes were analyzed in two separate groups. The first group consisted of the eight black teas and five green teas, and the second group comprised all 22 herbal teas. Figure 4 shows PCA plots of both groups to reveal successful separation and clustering of the data. In Figure 4A it is apparent that the green tea analytes (Honey Lemon Ginseng, Matcha, China Sencha, Young Hyson, and Gunpowder) all have PC1 scores <0, while most of the black teas have PC1 scores >0, with the exception of Formosa, Keemun congou, and Nilgiri. The latter, in particular, is located closer to some of the green teas. This can be partly explained by the texture of tea leaves, as the Nilgiri tea sample consists of fine powders, which was similar to both Honey Lemon Ginseng and Matcha. Overall, the PC1 separation suggests intrinsic differences between most of the green and black teas. Figure 4B shows more overlap of the herbal teas, which may be due to the similarity of ingredients. For instance, both Watermelon Lime Zinger ® and Wild Berry Zinger ® have hibiscus and rosehips listed as the top two ingredients, which can explain the overlap in PCA scores. The separation of clusters is limited by the 2D plot that only captures 63% of the variance in PCs 1 & 2.
Sensors 2018, 18, x FOR PEER REVIEW 7 of 12 interact at the molecular scale with the monolayer protected nanoparticle organic shells, which induces a change in interparticle separation that affects electron transport through the nanoparticle films [18,30,31]. Figure 3B illustrates baseline resistances and sensor responses to Black Cherry Berry for four different monolayer protected nanoparticles on a single chip. After eliminating devices with out-of-criterion baseline resistance, 38 of 48 devices remained valid for analysis. For each monolayer protected nanoparticle, sensor response measurements were carried out with at least eight individual sensing elements. Error bars reflect the spread of R0 and ΔR/R0 for each grouping of monolayer protected nanoparticles. Sensor elements assembled with aqueous-phase DMAP-AuNPs possess the lowest baseline resistance (6.8 ± 4.8 ×10 5 Ω), followed by ODA-AuNPs (4.1 ± 2.4 ×10 6 Ω), MPA-AuNPs (4.1 ± 2.2 ×10 7 Ω), and ATP-AuNPs (4.9 ± 2.2 ×10 7 Ω). Sensor responses follow the same trend as baseline resistance up to 10 6 Ω range, after which responses stabilize around ΔR/R0 = 5%. The variations of baseline resistance and sensor response levels are essential for pattern recognition using machine learning algorithms. Depending on the effective chemical interactions between analytes and monolayer protected core-shell nanostructures, the AuNP layers swell at different magnitudes [43], which results in variations of sensor responses. The sorption of vapors into the sensing materials can be qualitatively understood in terms of non-covalent van der Waals interactions that vary with each analyte and sensor material [18,44,45].

Sensor Performance and Classification Accuracy
Sensor element responses (ΔR/R0) to the 35 different tea aromas were collected and processed as inputs for statistical analysis. Principal component analysis (PCA) was performed as an unsupervised method to examine the natural grouping and separations between the different teas. The analytes were analyzed in two separate groups. The first group consisted of the eight black teas and five green teas, and the second group comprised all 22 herbal teas. Figure 4 shows PCA plots of both groups to reveal successful separation and clustering of the data. In Figure 4A it is apparent that the green tea analytes (Honey Lemon Ginseng, Matcha, China Sencha, Young Hyson, and Gunpowder) all have PC1 scores <0, while most of the black teas have PC1 scores >0, with the exception of Formosa, Keemun congou, and Nilgiri. The latter, in particular, is located closer to some of the green teas. This can be partly explained by the texture of tea leaves, as the Nilgiri tea sample consists of fine powders, which was similar to both Honey Lemon Ginseng and Matcha. Overall, the PC1 separation suggests intrinsic differences between most of the green and black teas. Figure 4B shows more overlap of the herbal teas, which may be due to the similarity of ingredients. For instance, both Watermelon Lime Zinger ® and Wild Berry Zinger ® have hibiscus and rosehips listed as the top two ingredients, which can explain the overlap in PCA scores. The separation of clusters is limited by the 2D plot that only captures 63% of the variance in PCs 1 & 2.  As PCA only captures a qualitative account of aroma classification, supervised multivariate and machine learning models were used to provide a quantitative analysis [46,47]. One of the most robust classification techniques is linear discriminant analysis (LDA), which characterizes the input data based on linear combinations of sensor responses, and maximizes the distance between different classes [48]. Five-fold cross-validation was applied to all classification analyses. Specifically, the data entries were randomly divided into five groups. Among the five groups, four were used in the training set, and the remaining group was tested. The process rotated five times so that every data entry had been used in the testing set. For the first group of analytes (black teas and green teas), out of 104 measurements, the training size was 83 and the testing size was 21 for each rotation. For the second group (herbal teas), out of 176 measurements, the training size was 141 and the testing size was 35 for each rotation. Figure 5 shows an LDA plot and confusion matrix for both groups of analytes. As PCA only captures a qualitative account of aroma classification, supervised multivariate and machine learning models were used to provide a quantitative analysis [46,47]. One of the most robust classification techniques is linear discriminant analysis (LDA), which characterizes the input data based on linear combinations of sensor responses, and maximizes the distance between different classes [48]. Five-fold cross-validation was applied to all classification analyses. Specifically, the data entries were randomly divided into five groups. Among the five groups, four were used in the training set, and the remaining group was tested. The process rotated five times so that every data entry had been used in the testing set. For the first group of analytes (black teas and green teas), out of 104 measurements, the training size was 83 and the testing size was 21 for each rotation. For the second group (herbal teas), out of 176 measurements, the training size was 141 and the testing size was 35 for each rotation. Figure 5 shows an LDA plot and confusion matrix for both groups of analytes. As shown in Figure 5, 100% classification accuracy was achieved for green and black teas, and 97.7 % accuracy was achieved for herbal teas, where confusion occurs three times. For example, one of the misclassifications is treating Raspberry Zinger ® as Lemon Levander Lane TM . This can be rationalized from the PCA plot where close clustering is observed for these two classes. If the teas are analyzed as a single group of 35, the LDA classification accuracy is 98.6%.  As shown in Figure 5, 100% classification accuracy was achieved for green and black teas, and 97.7% accuracy was achieved for herbal teas, where confusion occurs three times. For example, one of the misclassifications is treating Raspberry Zinger ® as Lemon Levander Lane TM . This can be rationalized from the PCA plot where close clustering is observed for these two classes. If the teas are analyzed as a single group of 35, the LDA classification accuracy is 98.6%. Table 2 summarizes the LDA classification accuracy for a full sensor chip, as well as a comparison of classification accuracy of the four separate monolayer protected nanoparticles used in this study. The full sensor chip with all 38 sensing elements is obviously better than any of the separate chemistries Sensors 2019, 19, 2547 9 of 12 alone, although the DMAP chemistry does very well with the green and black teas (97.1%). For a more even comparison, two devices from each sensing material were randomly selected to make a group of eight sensing elements for comparison with single particle data on an equal number basis (eight multi-particle vs. eight single particle). The result showed a 91.4 ± 3.3% accuracy for black and green teas, and 88.6 ± 3.7% for herbal teas. The combination of the four monolayer protected nanoparticles clearly performs better overall than any single type of nanoparticle, which supports a broader hypothesis that more chemical diversity can enhance sensor performance. We speculate that adding more of each type of monolayer protected nanoparticle will also improve overall sensor performance.

Sensor Stability
To evaluate sensor stability, a five-day experiment was performed for the classification of black and green tea aromas. Firstly, the baseline resistance of each working device was monitored over time. Figure 6A shows a gradual decrease of baseline resistance, which was larger during Days 1-3 and flattened after Day 4. All sensing materials underwent a decrease in baseline resistance. Sensing profiles on Figure 6B show that for DMAP-AuNP and ODA-AuNP, there was a trend of decreasing sensitivity, but not for MPA-AuNP or ATP-AuNP. This is likely due to the difference in particle morphology changes during the reversible sensing process.  Table 2 summarizes the LDA classification accuracy for a full sensor chip, as well as a comparison of classification accuracy of the four separate monolayer protected nanoparticles used in this study. The full sensor chip with all 38 sensing elements is obviously better than any of the separate chemistries alone, although the DMAP chemistry does very well with the green and black teas (97.1%). For a more even comparison, two devices from each sensing material were randomly selected to make a group of eight sensing elements for comparison with single particle data on an equal number basis (eight multi-particle vs. eight single particle). The result showed a 91.4 ± 3.3% accuracy for black and green teas, and 88.6 ± 3.7% for herbal teas. The combination of the four monolayer protected nanoparticles clearly performs better overall than any single type of nanoparticle, which supports a broader hypothesis that more chemical diversity can enhance sensor performance. We speculate that adding more of each type of monolayer protected nanoparticle will also improve overall sensor performance.

Sensor Stability
To evaluate sensor stability, a five-day experiment was performed for the classification of black and green tea aromas. Firstly, the baseline resistance of each working device was monitored over time. Figure 6A shows a gradual decrease of baseline resistance, which was larger during Days 1-3 and flattened after Day 4. All sensing materials underwent a decrease in baseline resistance. Sensing profiles on Figure 6B show that for DMAP-AuNP and ODA-AuNP, there was a trend of decreasing sensitivity, but not for MPA-AuNP or ATP-AuNP. This is likely due to the difference in particle morphology changes during the reversible sensing process. Besides linear discriminant analysis (LDA), two other types of classification algorithms were examined using MATLAB Classification Learner Toolbox, including support vector machine (SVM) and k-nearest neighbors (KNN). Again, five-fold cross-validation was applied for all methods. Table  3 shows the classification accuracy of black and green teas over five days for 38 sensing elements on a sensor chip. For all four methods, the classification accuracy declined at Day 3, however, it mostly Besides linear discriminant analysis (LDA), two other types of classification algorithms were examined using MATLAB Classification Learner Toolbox, including support vector machine (SVM) and k-nearest neighbors (KNN). Again, five-fold cross-validation was applied for all methods. Table 3 shows the classification accuracy of black and green teas over five days for 38 sensing elements on a sensor chip. For all four methods, the classification accuracy declined at Day 3, however, it mostly recovered by Day 5. This may be related to changes of baseline resistance for Days 1-3. The baseline resistance stabilized at Day 5, which resulted in an improved classification accuracy. The Random Forest (RF) method is a relatively newer machine learning approach that has been successful in many other studies, therefore, we also tested it on our data set. The RF method performed very well for all days with >98% accuracy, but it also shows a small dip at Days 3 & 4.

Conclusions
In this project, we demonstrated gold nanoparticle based chemiresistor arrays in order to classify 35 types of tea beverages, including eight black teas, five green teas, and 22 herbal teas. Sensor performance was evaluated based on classification accuracy using linear discriminant analysis (LDA) and several machine learning algorithms. A classification accuracy higher than 97% was achieved for both green and black teas (group 1), and herbal teas (group 2). The results demonstrate more robust performance for sensor arrays assembled with multiple types of monolayer protected nanoparticles compared to arrays of a single type, which supports a broader hypothesis that more diverse chemistry can enhance sensor performance for resolving complex odors, including teas. Short term sensor stability was assessed by performing a five-day experiment using teas from group 1. During the five-day experiment, a classification accuracy higher than 90% was obtained and sensor performance recovered to initial performance after the fifth day. The findings suggest that nanoparticle sensor arrays can be useful for detection and discrimination of complex odors, including tea aromas.