Quantitatively Discriminating Alcohol Molecules by Thermally Modulating NiO-Based Sensor Arrays

Device Fabrication and Sensing Test: First, four kinds of NiO-based powder were dispersed in an appropriate amount of ethanol and continuously stirred for more than 1 h. The paste that formed was then coated onto the sensing ceramic substrates (with heater and test electrodes) as the sensing layer. All sensors were aged in air at ~ 220 °C for one week before testing. Three types of alcohol group gas of methanol, ethanol and IPA (with quite similar properties, see Table S2) were selected as the target analytes. Different concentrations of alcohol group gases were obtained by mixing standard gas and dry air in any ratio, controlled by mass flow controllers (CS-200, Sevenstar Electronics Co. Ltd). The total gas flow rate was set at 500 sccm. The gas sensing tests were carried out in a four-channel gas sensitive measurement system (SD101, Huachuang Ruike Technology Wuhan Co. Ltd.), which allows to measure the resistance ranging from 100 Ω to 1 GΩ with a verified accuracy of ±5.0% and an accuracy of ±18.3% for 10 GΩ, as seen in Figure S14. The temperature of the sensor was modulated from the peak temperature (50 °C, 100 °C, 150 °C, 200 °C, 250 °C or 300 °C) to a fixed low temperature of 50 °C by tuning the heating voltage. The alcohol group vapors were injected into the test chamber in the temperature drop range. After the temperature modulation, raising the heater temperature and maintaining a constant peak temperature (~ 120 s) facilitated the desorption of the tested analytes and recovery of the sensor baseline resistances. Before the sensors resistance under the target alcohol group gases (Rgas) was measured, the air resistance (Rair) was measured in the first temperature modulation cycle. The response of the alcohol group gases was defined as Rgas/Rair. Methanol, ethanol, and IPA, each at 25 different concentration levels (40 – 1000 ppm, 40 ppm ascending step), 75 analytes in total, were tested once in sequence from the second temperature modulation cycle.Statistical Analysis: The features used to distinguish alcohol group gases were extracted through preprocessing steps. For each temperature modulation test, the cycle duration time was fixed, and the response value was obtained from Rgas divided by Rair, both with cycle time dimension.S(t) = Rgas(t) / Rair(t)Next, the gas responses of different concentrations obtained by each temperature modulation were normalized.Snor(t) = (S(t) – Smin) / (Smax – Smin)where Smax and Smin are the maximum and minimum values of the response in the cooling range, respectively, Snor(t) represents the normalized response, and t is the time. Data processing was completed in Python 3.6. The features of the cycle time dimension were reduced to a two-dimensional visual space using PCA. In addition, LDA classified the samples based on the input features. LDA used leave-one-out cross validation, and each example was classified by functions derived from all examples except that example. A multilayer perceptron is a type of feedforward neural network proposed to solve nonlinear problems that cannot be solved by a single-layer perceptron. MLP is a simple neural network, which allows to quantitively evaluate the features of electrical responses obtained under different peak temperatures. The features of the four sensors were input to the MLP with a hidden layer, and the number of units in the output layer was 3 or 15 according to species or concentration classification. Softmax and tanh were used as the activation functions of the output and hidden layers, respectively. The cross entropy with faster convergence was selected as the cost function to evaluate the neural network model by minimizing the cost function. There were three training modes of the MLP, two of which were species classification, while one was concentration classification. For the first mode, 80% of the samples (75 samples in total) were used as the training set, and the rest were used as the test set. A five-fold cross validation can make full use of the sample data. For the second mode, 4% of the samples were used as the training set, and the rest were used as the test set. In other words, only one sample of each species (25 samples in total) was used as the training set, and samples of the same concentration of the three alcohol group gases were selected. Then, a 25-fold cross validation was performed to avoid accidental errors. For the third mode, gas samples of 0 – 1000 ppm were divided into five concentration ranges according to a segment of 200 ppm. 80% of the samples (four samples) of a concentration range were used as the training set, and the rest were used as the test set. Five-fold cross validation was used. The algorithms (PCA, LDA, and MLP) were implemented using IBM SPSS Statistics 26.