1. Sensor Fabrication

The fabrication process of nanocomposite hydrogen Sensor is illustrated in Fig. 1. First, Si wafer is prepared, and SiO$_{2}$ layer is deposited at 300 nm using wet oxidation. After that, MoS$_{2}$ is deposited on the wafer through chemical vapor deposition (CVD). CVD process conditions are maintained at 200 $^{\circ}$C in the upstream region and 720 $^{\circ}$C in the central heating region. Ar gas is injected with 200 Standard cubic centimeters per minute (SCCM). Molybdenum trioxide powder (2 mg) and sulfur powder (100 mg) are placed in a quartz tube in a 3-zone tube furnace. A combination of molybdenum trioxide powder (2 mg) and sulfur powder (100 mg) is placed in a quartz tube. Through this deposition process, MoS$_{2}$ accumulates to a thickness of about 1.4 nm and forms about eight layers. After MoS$_{2}$ deposition, a show mask is used to pattern sensor. Pd is deposited over the MoS$_{2}$ layer through DC sputtering. Pd is deposited for 4 s with an Ar flow rate of 30 SCCM and a power of 100 W to form discrete Pd particles of 4 nm thickness. This condition was set to effectively balance sensor sensitivity and response time. Thickness of Pd layer affects sensor performance. An extremely thin Pd layer, such as 1 nm, forms discontinuous Pd particles. As a result, contact area with gas is limited. Sensitivity and response time are relatively reduced. Conversely, as thickness increases, sensitivity and reactivity increase. However, the formation of continuous layers can degrade performance due to charge accumulation and reduced contact area. A layer of 10 nm Ti and a layer of 40 nm Au are deposited through DC sputtering ^[4].

Fig. 1. The fabrication process of nanocomposite hydrogen Sensors.

2. Sensor Mechanism

This section describes basic behavior and fundamental principles of nanocomposite hydrogen Sensor. Sensor detects presence of H$_{2}$ through interaction of the components. Detection of nanocomposite hydrogen Sensors takes advantage of unique characteristics of MoS$_{2}$ and Pd. MoS$_{2}$ performs two main functions to improve performance of sensor. First, MoS$_{2}$ acts as a channel for current flow. Second, MoS$_{2}$ adsorbs oxygen molecules on surface to form oxygen ion species. This adsorption process increases resistance of MoS$_{2}$ since adsorbed oxygen is formed as oxygen ions through electrons of MoS$_{2}$. Fig. 2 shows mechanism of nanocomposite hydrogen Sensor. In the absence of H$_{2}$, MoS$_{2}$ adsorbs oxygen molecules to form oxygen ion species that increase resistance of material. When exposed to H$_{2}$, Pd plays an important role in dissociating H$_{2}$ molecules and absorbing generated H$_{2}$ atoms. This leads to a "spillover effect" in which H$_{2}$atoms move from Pd layer to MoS$_{2}$. On the MoS$_{2}$ surface, H$_{2}$ atoms react with oxygen ion species to produce water and electrons. This electron is injected into the MoS$_{2}$ channel, resulting in a change in resistance. The sensor detects H$_{2}$ leakage through a change in resistance caused by interaction of H$_{2}$ ^[5].

Fig. 2. The Mechanism of nanocomposite hydrogen Sensors.

Fig. 3. Random forest model for environmental variable prediction.

3. Machine Learning Algorithm

The machine learning (ML) model determined for sensor application is random forest (RF) model. RF models derive reliable results as a way to ensemble them into decision trees. As shown in igure 3, RF models are the process of generating multiple decision trees using different training datasets and combining the predictions of these trees for final output. RF models use a bagging approach to reduce the model variance and prevent overfitting. Bagging generates multiple bootstrap samples from original data and trains each decision tree with each sample. In this process, each tree learns different data and reduces risk of overfitting individual trees. Furthermore, RF models compensate for the weaknesses of individual models and increase their stability by combining predictions from multiple models through ensemble learning ^[6].

Sensor chip inside the chamber was connected to an electronic circuit that collects sensor signals and sends them to PC via Inter-Integrated Circuit (I2C) interface. The output voltage of each sensor node was collected from a resistor connected in series with sensor node. An ATMega 2560 microcontroller is programmed to process an 8-bit digital-to-analog converter (MCP3008) and related electronic components and send voltage signals to PC. Therefore, sensor can measure voltage range from 0 to 5 VDC with an accuracy of 0.1 mV. In this process, we used MiCS-2714 sensor, due to compatibility issues with nanocomposite sensors. This approach was able to confirm future applicability of nanocomposite sensors and ML model ^[7]. Fig. 4 shows a flowchart of ML model generation and completion process. To build the ML model, input values were assigned as temperature, humidity, voltage, and current, and output values were assigned as H$_{2}$ concentrations. Dataset is divided into training and testing datasets at an 80:20 ratio to check predictive accuracy of algorithm ^[8].

Improve accuracy of H$_{2}$ concentration detection in various environmental conditions by incorporating the RF model with sensors. The approach ensures stability and reliability of nanocomposite hydrogen sensor system.

Fig. 4. Flow chart for generating Random forest model and completion process.

1. Measurement Results

Fig. 5 shows sensitivity of nanocomposite sensor at room temperature according to H$_{2}$ concentration. Fig. 5(a) illustrates sensor response to a 1% H$_{2}$ concentration, indicating an average sensitivity of 1.34. When H$_{2}$ concentration is increased to 3% (Fig. 5(b)), average sensitivity increases to 2.17. Finally, at 4% H$_{2}$ concentration (Fig. 5(c)), average sensitivity reaches 2.77. Response time was 4 to 5 seconds. Response time was defined as time required for current to change by 90%. Sensitivity is calculated as $S$ = Ig / I0. Ig is current after H$_{2}$ gas injection, and I0 is current without H$_{2}$ gas. Fast response time was caused by high carrier mobility and simple structure of sensor. Unlike typical MoS$_{2}$ based sensors, which often require high operating temperatures, these sensors have demonstrated reactivity to H$_{2}$ at room temperature. Pd aggregates H$_{2}$ and lowers activation energy with H$_{2}$. The sensitivity graph shows that sensitivity increases as H$_{2}$ concentration increases. The reason is that when concentration of H$_{2}$ increases, the number of H$_{2}$ molecules that react with Pd particles increases. Thus, the number of electrons injected into the channel increases, resulting in an increase in current.

Fig. 5. The sensitivity of nanocomposite hydrogen sensors using H2 concentration at room temperature: H2 concentration is (a) 1%; (b) 3%; (c) 4%.

2. Machine Learning Accuracy

Application of RF models to hydrogen sensors demonstrates high accuracy in predicting H$_{2}$ concentration and state. As shown in Fig. 6, the RF model achieved a mean square error (MSE) of 0.0506 and an R$^{2}$ score of 0.994 in H$_{2}$ concentration prediction. A high R$^{2}$ score indicates that model accounts for 99.4% of the data variance, indicating a nearly perfect fit between predicted and true concentrations. Strong correlation is shown in scatterplot where predicted concentration almost follows true concentration line. The RF model was trained from concentration data and state data to extract more accurate state predictions.

The confusion matrix for sensor state prediction shows an overall accuracy of 0.979. Among the total predictions, the RF model correctly categorized 318 instances as "Safe" and 661 instances as "Help". In the case of misclassification, 11 "Safe" instances were "Help" and 10 "Help" instances were "Safe", which were only a few. Results of the RF model are as follows. The RF model utilize ensembles of decision trees to help average errors and reduce overfitting. Ensemble approaches ensure stability against noise in the data. Bagging reduces variance through training on a random subset of data. The RF model predict from weights of environmental conditions such as temperature, humidity, voltage, and current. Therefore, the weighting of environmental conditions categorizes factors that directly affect response of sensors and determines importance of environmental conditions. The RF model are crucial for accurately predicting H$_{2}$ concentrations in different environments.

Fig. 6. (a) Confusion matrix for nanocomposite sensor status; (b) estimated concentration versus actual concentration for the tested data.