Gas Sensing with Nanoporous In₂O₃ under Cyclic Optical Activation: Machine Learning-Aided Classification of H₂ and H₂O

Abstract

Clean hydrogen is a key aspect of carbon neutrality, necessitating robust methods for monitoring hydrogen concentration in critical infrastructures like pipelines or power plants. While semiconducting metal oxides such as In₂O₃ can monitor gas concentrations down to the ppm range, they often exhibit cross-sensitivity to other gases like H₂O. In this study, we investigated whether cyclic optical illumination of a gas-sensitive In₂O₃ layer creates identifiable changes in a gas sensor’s electronic resistance that can be linked to H₂ and H₂O concentrations via machine learning. We exposed nanostructured In₂O₃ with a large surface area of 95 m² g⁻¹ to H₂ concentrations (0–800 ppm) and relative humidity (0–70%) under cyclic activation utilizing blue light. The sensors were tested for 20 classes of gas combinations. A support vector machine achieved classification rates up to 92.0%, with reliable reproducibility (88.2 ± 2.7%) across five individual sensors using 10-fold cross-validation. Our findings suggest that cyclic optical activation can be used as a tool to classify H₂ and H₂O concentrations.

1. Introduction

Over the past decade, a growing emphasis on sustainable energy solutions has led to a surge in research and development in the field of hydrogen-based technologies. As part of this trend, the European Union’s hydrogen strategy for a climate-neutral Europe has highlighted the increasing need for reliable and efficient hydrogen detection technologies [1]. At the core of these technologies are semiconducting metal oxides, which have the remarkable property of changing their electrical resistance upon exposure to reactive gas species. A detailed description of the resistance changes of semiconductors (in particular In₂O₃) upon gas exposure is described in the literature [2,3]. Metal oxides with a large surface area are attractive for gas sensing applications, such as mesoporous Indium(III) oxide (In₂O₃), due to the large number of surface adsorption sites that amplify the resistance change compared to non-porous materials. When the metal oxide is synthesized by a structure replication process on the nanometer scale (i.e., nanocasting), it possesses a well-defined structure enabling specific, target-oriented applications [4]. In nanocasting, ordered mesoporous silica is used as a mold, infiltrated with an indium oxide precursor (indium nitrate in this case), that is thermally converted to the oxide. After selective removal of the silica mold, nanostructured In₂O₃ remains [5].

Indium(III) oxide is a well-known n-type semiconductor with a surprisingly high charge carrier concentration of up to 10¹⁰ cm⁻³ [6]. Its large electrical conductivity is attributed to oxygen deficiency in the lattice and many intrinsic defects under a reducing atmosphere [7]. Therefore, In₂O₃ is a great candidate for low-temperature gas measurements. Illumination with sub-bandgap energy (e.g., blue LED, 472 nm, 2.6 eV) can be utilized because of the above-mentioned defects (In₂O₃ bandgap values reported in the literature range from 2.1 eV to 3.7 eV [8,9,10]) to increase sensor response and kinetics. Upon illumination, desorption of oxygen adsorbates and photo-reduction are two of several reasons why the electronic resistance of In₂O₃ changes [6,11,12,13]. These effects can also be used in photocatalytic applications [14,15].

Despite these advantages, resistive gas sensors might suffer from a lack of selectivity, meaning their inability to differentiate between different gases simultaneously [16,17]. To increase selectivity, various approaches can be pursued, such as adding barrier (filter) layers [18], catalysts [19,20,21], variation of measurement voltage [22], or temperature cycling [23,24,25,26,27] to create multi-signals [28,29,30,31]. Temperature cycling can be conducted in sub-second cycles when using appropriate sensor systems [32]. It should be noted that the sensor’s chemical response may be the rate-determining step [23,33,34] when using optical activation. Optical activation might not affect the whole sensing layer because the penetration depth of, e.g., UV light in dense In₂O₃ is around 10 nm [35,36]. This fact emphasizes the need for nanostructured metal oxides for optical activation in gas sensing because the radiation can penetrate deeper into the material due to the porosity. Materials synthesized by nanocasting are particularly interesting because their wall thickness is in the range of small mesopores (2–15 nm) and, therefore, in the same dimensions as the penetration depth of (UV-)light. Also, particle size and the sensing layer thickness need to be considered [37]. Constant optical activation, i.e., permanent illumination of a gas sensor by a light source, is already well-investigated and summarized in review articles [2,12,38]. However, transient optical activation is a novel trend in the resistive gas sensor community [39,40,41,42].

The forced change in sensor conductivity by variable illumination helps overcome the conductivity drifts that are often observed in resistive gas sensors due to the equilibration of the sensor, which can take from minutes to days [13,43]. If a gas sensor is exposed to discrete gas concentrations, and these are used for learning and prediction, a classification algorithm is a common tool of choice. Therefore, a support vector machine (SVM; e.g., a classifier) was utilized in this study to classify the gas concentration by means of resistance value changes upon illumination [44]. Ten-fold cross-validation was utilized for learning and classification. The application of machine learning techniques is becoming more and more popular to overcome challenges in physical chemistry and gas sensor measurements in particular [13,45,46,47].

Recently, Schmuker et al. investigated the limitations of machine learning in gas sensing due to baseline drift [48,49]. They discussed baseline changes in the data that are limiting the usability of machine learning for concentration classification. For this reason, Ge et al. extracted the gas-specific features from the data and used them for gas identification. In this way, the aging effects of the sensor as well as changes in the environmental conditions could be suppressed in the training data [50]. Siadat et al. describe the utilization of sample set partitioning (SPXY) to select their training data and an SVM regression for monitoring air pollution [51]. Their calibration method proved to be quite efficient and reliable compared to classical direct standardization. SVM methods seem to address the quantification and differentiation between different gas species quite well. As shown by Yang et al., zinc oxide in different morphologies can create an array of gas sensors differentiating methanol, ethanol, propanol, acetone, formaldehyde, and ammonia [52]. A linear ridge classifier underperformed compared to the SVM classifier, which was correct in 99% of gas classifications. Furthermore, concentration classification for ethanol showed a strong performance of 98%.

Yoon and Park introduced transient illumination with UV light when using an In₂O₃ sensor decorated with gold nanoparticles to classify gas types (acetone, NO₂, ethanol, methanol, and air) and concentration (0–100 ppm). The pseudo-random change in illumination intensity led to electronic resistance changes that could be analyzed by a deep neural network. Gas-type classification resulted in an accuracy of 96% for a single gas sensor [13].

This study explores whether cyclic optical activation can be used to classify discrete H₂ and H₂O concentrations via an SVM. High surface area In₂O₃ without dopants is used as a gas-sensitive layer and is exposed to varied hydrogen gas (0–800 ppm) and water (0–70%) concentrations in synthetic air. Five individual sensors were utilized to test for classification reproducibility. Both H₂ and H₂O lower the electronic resistance of nanostructured In₂O₃, posing differentiation challenges that might occur in stationary power or an automotive context [53].

2. Materials and Methods

Synthesis of mesoporous silica: Ordered mesoporous silica number 6 from Korean Institute of Technology (KIT-6) [54] was synthesized by dissolving block co-polymer Pluronic P-123 (8.0 g, Sigma Aldrich, St. Louis, MO, USA, M_n~5800) in deionized water (240 mL), hydrochloric acid (28.6 g mL, 35%, Stockmeier, Bielefeld, Germany), and butan-1-ol (8.0 g, Stockmeier, Bielefeld, Germany) at 35 °C. After the addition of tetraethyl orthosilicate (16 g, 98%, Sigma-Aldrich, St. Louis, MO, USA), the solution is stirred for 24 h (35 °C), followed by hydrothermal treatment at 80 °C for 24 h in a glass-lined autoclave. After filtration and washing (three times) with water, the white precipitate is dried at 60 °C for 24 h. Finally, the polymer is thermally decomposed in a furnace at 550 °C for 5 h (2.5 °C min⁻¹).

Synthesis of mesoporous In₂O₃: Ordered mesoporous In₂O₃ was structurally replicated from cubic KIT-6 silica according to the procedure outlined in the literature [55]. A mortar, pistil, and silica were warmed to 100 °C in a drying cabinet. The amount of indium nitrate pentahydrate (99.99%, Sigma-Aldrich, St. Louis, MO, USA) was calculated by multiplying the silica mass with the pore volume of silica and the density of the molten precursor (around 2.4 mL g⁻¹), aiming to fill the pores completely. The indium nitrate was added to the preheated mortar as well as the preheated silica. Both components were homogenized and treated at 85 °C for 24 h. The composite was heated in an open vessel in a furnace under air to a temperature of 300 °C with a heating rate of 2 °C min⁻¹ for 2 h to convert the indium nitrate to the indium oxide.

The silica matrix is removed by leaching with a NaOH solution (5 mol L⁻¹, Stockmeier, Bielefeld, Germany) at a temperature of 60 °C for one hour. After centrifugation (4000 rpm for 15 min), the liquid phase is removed. This procedure was repeated three times. The remaining solid was washed with deionized water until a pH value of 7 was reached. The indium oxide was dried at a temperature of 60 °C for 24 h prior to further usage.

Material characterization: N₂ physisorption measurements were conducted at −196 °C after degassing the samples at 120 °C for 12 h in a vacuum using an Autosorb 6 (Quantachrome Instruments, Boca Raton, FL, USA). Surface areas were calculated using the Brunauer–Emmet–Teller (BET [56]) approach in a relative pressure region of 0.1–0.3, meeting the Rouquerol requirements [57]. Pore size distributions were calculated from the desorption branch, utilizing the Barrett–Joyner–Halenda model [58] as well as non-local density functional theory (NLDFT) kernel for cylindrical silica pores. The total pore volume was calculated at the penultimate adsorption point.

Diffraction data were collected on a AXS D8 advance (Bruker, Karlsruhe, Germany) with Cu K_α radiation (40 kV, 40 mA) with a step size of 2θ of 0.0075° (below 5°) and a step time of 300 s, and a step size of 0.02° (above 5°) with a step time of 3 s, respectively.

Particle size of In₂O₃ was determined by averaging over 25 particles collected on a Neon 40 scanning electron microscope (SEM, Zeiss, Oberkochen, Germany).

A Ultra Dry detector (Thermo Fisher Scientific, Waltham, MA, USA) determined residual silica content and In:O ratio from energy-dispersive X-ray spectroscopy.

The UV-vis spectrum was measured on a Lambda 650 (PerkinElmer, Waltham, MA, USA).

Gas sensor preparation and measurement: Mesoporous indium oxide (25 mg) was dispersed in deionized water (1 mL) and treated for 5 min in an ultrasonic bath. Then, 2 µL of the dispersion was dropped onto interdigitated electrodes (UST, sensor substrate 3 × 3 mm with IDS, 10 Ohm platinum heating element, mounted on T039 base, Geratal, Germany) and the sensor was dried at room temperature. Five individual sensors were prepared. Prior to the measurements, the sensors were first treated at 350 °C for 24 h and then at the measurement temperature of 150 °C for an additional 28 d. Resistance measurements (measured with 10 Hz) were performed under a constant gas flow of 500 mL min⁻¹ with different concentrations of hydrogen (0–800 ppm; Air Liquid) and relative humidity (0–70%) in synthetic air (20.5% O₂, 79.5% N₂, Air Liquid). Humidity values refer to carrier gas at a temperature of 22 °C. The sensor substrates were illuminated using a blue LED (InGaN, clear, 4.9 mm diameter, 8.6 mm height, peak wavelength 472 nm, FWHM = 40 nm, Luckylight, LL-504BC2E-B4-2GC) utilizing a square-wave voltage of 2.72 V. The distance of the LED to the gas sensor substrate was 10.7 mm with a radiation angle of 15°, leading to a light power of around 49 mW cm⁻². A so-called cycle contains an illumination period (5 s), followed by 5 s without illumination, and therefore lasts 10 s in total.

Machine learning analysis: The machine learning analysis serves the purpose of analyzing resistance changes of the gas sensor upon illumination that then can be correlated to experimental conditions, i.e., discrete gas concentrations. The experimental data (i.e., electronic resistance and gas concentrations) were used without any data clean-up or preprocessing for the discussion in Section 3.3. Subsets of the data were used as described in Section 3.4. and Section 3.5. Additionally, the normalization of data cycles was investigated as discussed in Section 3.4. Data analysis was performed using the Scikit-learn package [59]. A support vector machine with a linear kernel was used (regularization parameter C = 0.01). An unequal sample size was dealt with by undersampling. Ten-fold cross-validation was used for training and learning.

3. Results and Discussion

3.1. Characterization of Ordered Mesoporous In₂O₃

In₂O₃ is nanocasted from mesoporous silica (for characterization see Figures S1 and S2; physisorption raw data are available in .aif format [60] in the Electronic Supplementary Information Section) and characterized by N₂ physisorption, as depicted in Figure 1. The metal oxide exhibits a surface area of 95 m² g⁻¹ and a total pore volume of 0.62 cm³ g⁻¹ (see

Table 1. N₂ physisorption derived values of the silica matrix and the nanostructured In₂O₃.

Sample	Pore Size (a) nm	Pore Size (b) nm	Surface Area (c) m2 g−1	Pore Volume (d) mL g−1
silica	5.1	6.6	550	0.83
In2O3	6.0; 12.3	8.1; 14.5	95	0.62

^(a) derived from BJH model (desorption branch); ^(b) derived from NLDFT model (desorption branch), ^(c) BET surface area; ^(d) total pore volume.

). These values are larger compared to bulk metal oxides, especially when considering the apparent density of In₂O₃.

A type IV isotherm with two steps in a relative pressure region between 0.6 and 0.92 and small hysteresis can be observed. The isotherm and hysteresis shape indicate a successful nanocasting procedure and a mesoporous material with two maxima in the pore size distribution (Figure 1b). The smaller pore mode (6.0 nm BJH, 8.1 nm NLDFT) originates from the replication of the silica matrix. The larger mesopore (12.3 nm BJH, 14.5 nm NLDFT) is attributed to a partial displacement of the two interpenetrating pore networks, as described in the literature [55,61].

The displacement results in a lower symmetry of the metal oxide ($I{4}_1/a$ or lower) compared to the silica ($Ia\overline{3}d$), which is supported by the appearance of 110 in the low-angle diffractogram (Figure 2a), which is symmetry forbidden in $Ia\overline{3}d$ [61]. Reflections at larger angles are not visible anymore, which is due to the reduced degree of pore ordering or can be attributed to accidental extinction depending on the diameter of the metal oxide rods. The wide-angle diffraction pattern in Figure 2b confirms a successful synthesis of polycrystalline (cubic bixbyite structure, $Ia\overline{3}$) In₂O₃. The crystallite size of around 11 nm was determined via the Scherrer method, based on the 220 reflection. This fits well with the determined pore size, which restricts the unimpaired crystal growth of In₂O₃, thereby limiting the crystallite size [62].

The In₂O₃ particles exhibit a spherical shape with an average particle size of around 95 ± 5 nm (Figure 3), and the cubic pore structure is clearly visible. The atomic ratio of In to O is 0.59 ± 0.08: 1 which, with respect to the error of the measurement, resembles In₂O₃. The remaining silica content is around 3.4 ± 0.94 wt.%, which might be one reason for the atomic ratio of In to O showing a slight excess of oxygen. The synthesized In₂O₃ exhibits a direct bandgap of 3.83 ± 0.002 eV (see Figure S3).

3.2. Resistance Changes during Cyclic Optical Activation and Gas Exposure

Cyclic optical activation was performed by illuminating a gas-sensitive In₂O₃ layer with light from a blue LED for five seconds, leading to a decrease in electronic resistance that bounces back after the LED is switched off for five seconds, which is repeated throughout the entire measurement and therefore referred to as cyclic optical activation.

A representative part of the resistance over time measurement at a constant atmosphere is shown in Figure 4. Illumination of the metal oxide layer on top of the interdigital electrode leads to oxygen desorption, photoreduction, and therefore a drop in resistance [6,11,12]. The electronic resistance depends on the atmosphere during the measurement and lies in a range between 100 kΩ and 340 kΩ for this particular gas sensor.

The sensor exhibits an electric resistance of around 230 kΩ at the start of the measurement that drops down to about 80 kΩ at the beginning of the first illumination cycle (as shown in Figure 5). The sensor’s resistance increases afterward because the dry synthetic air removes adsorbed water from the surface of the metal oxide.

Resistive gas sensors can take from minutes to days to achieve a stable baseline [43]. Fifteen minutes after starting the measurement, hydrogen (200 ppm) is added even though the baseline is still changing (i.e., the sensor is operated under non-equilibrium conditions for a fast concentration readout). A total of five sensors were prepared, four of which are shown in the Supplementary Information Section (Figure S4).

Adding a reducing gas leads to a decrease in electric resistance in an n-type semiconductor. Here, the hydrogen concentration was increased every 5 min in steps of 200 ppm until 800 ppm was reached. Afterward, the hydrogen addition was stopped, increasing resistance up to around 310 kΩ. Increasing the relative humidity to 30% reduces the resistance due to the adsorption of water on the sensor’s surface [63]. This general trend can be observed for all humidity levels. The sensor was exposed to five different concentrations of hydrogen (0 ppm, 200 ppm, 400 ppm, 600 ppm, and 800 ppm) and four relative humidity values (0%, 30%, 50%, and 70%), leading to a total of 20 different classes of gas concentrations.

Classical approaches to interpreting a gas sensor’s resistance typically wait for a stable baseline and a stable signal after a change of atmosphere. The change in resistance R (or response, which is resistance divided by base resistance R₀) is then associated with specific gas concentrations. As mentioned above, these processes can take a long time, especially at low temperatures. Therefore, we refrain from long waiting times and utilize machine learning to classify the resistance data, even though the sensors are far from reaching equilibrium.

3.3. Gas Concentration Classification

The gas sensor was exposed to discrete amounts of hydrogen (five concentrations: 0 ppm, 200 ppm, 400 ppm, 600 ppm, and 800 ppm) and relative humidity (four concentrations: 0%, 30%, 50%, and 70%) and the changes in the electronic resistance upon gas and light exposure (full cycle for a total of 10 s with a resistance measurement frequency of 10 Hz) were used for learning and evaluation of the classes. No gas concentrations between the aforementioned classes were added or evaluated, so we chose a classification algorithm. The actual concentrations (i.e., classes) versus the classification made by the machine learning algorithm are depicted in a confusion matrix in Figure 6. Cycles located on the diagonal (from bottom left to top right) are correctly classified sets, while deviations from the diagonal are wrongly associated. The best sensor (out of five) exhibits a correct classification of 92.0%, which is comparable to the literature values (see Table 2).

Four more sensors were prepared, measured, and evaluated in the same way to estimate the method’s reliability. It turns out that the accuracy of the gas concentration classification and discrimination between hydrogen and humidity was 88.2 ± 2.7% over five individual sensors, which is a little below the literature data (see Table 2). However, not many studies compare several sensors in parallel as we did in this study. Static and dynamic temperatures can lead to higher classification rates than our data; however, one must consider the semiconductors and gases investigated. The results presented by Yoon and Park et al. are the closest to our study (utilizing In₂O₃ decorated with Au particles), where a single sensor was investigated, leading to a classification rate of 96.53% using transient illumination and deep-learning [13].

Prior to the detailed analysis of the data using SVM, we conducted a brief screening utilizing other methods to see which approach potentially leads to the best classification rates. We investigated decision tree (62.0 ± 11.6%), k-nearest neighbor (67.8 ± 10.0%), multilayer perceptron (50 hidden layers, 21.0 ± 5.7%), and Gaussian process model analysis (39.0 ± 11.1%) for the presented five individual gas sensors. Furthermore, we conducted a principal component analysis. A full illumination–darkness cycle was split into half, and then both parts separately fitted against a power function to reduce dimensionality (from 100 to 4; e.g., analysis of slope and curvature for both half cycles) followed by classification. The classification rate was 50.75 ± 7.5%. When performed effectively, target-oriented dimensionality reduction can lead to enhanced sensor performance, which can be an interesting direction for future data evaluation [64]. A regression-based assessment is shown in the Supplementary Information section (Figure S5) for comparison. Due to the classification rate success of utilizing SVM, we decided to evaluate the data further.

Table 2. Classification rate based on literature results compared to this study.

	Faleh et al. [65]	Tonezzer, Van Duy et al. [27]	Kim, Shin et al. [66]	Ge et al. [50]	Yoon, Park et al. [13]	This Study
material	WO3	SnO2 + Ag, Pt	ZnO	commercial	In2O3 + Au	In2O3
no. sensor(s)	4	4	1	1	1	5
operation mode	static T a	static T a	cyclic T a	cyclic T a	time-variant illumination	cyclic illumination
data analysis	SVM	SVM	CNN b	SVM	CNN b	SVM
data preprocessing	yes	no	yes	yes	yes	no
classification rate/%	93.03	100	93.9	95.4	96.53	92.0

^a T = temperature, ^b CNN = convolutional neural networks.

Table 2. Classification rate based on literature results compared to this study.

	Faleh et al. [65]	Tonezzer, Van Duy et al. [27]	Kim, Shin et al. [66]	Ge et al. [50]	Yoon, Park et al. [13]	This Study
material	WO3	SnO2 + Ag, Pt	ZnO	commercial	In2O3 + Au	In2O3
no. sensor(s)	4	4	1	1	1	5
operation mode	static T a	static T a	cyclic T a	cyclic T a	time-variant illumination	cyclic illumination
data analysis	SVM	SVM	CNN b	SVM	CNN b	SVM
data preprocessing	yes	no	yes	yes	yes	no
classification rate/%	93.03	100	93.9	95.4	96.53	92.0

^a T = temperature, ^b CNN = convolutional neural networks.

3.4. Single-Gas Concentration Classification and Cycle Normalization

In the analysis so far, both offered gases were taken into account when it comes to the classification of concentrations. The group-based approach described in this subsection assesses the model’s robustness by testing its ability to generalize to new data combinations. When only using constant H₂ concentrations, e.g., 0 ppm and training with 0–75% r. H. (the second column in

Table 3. Classification rate of H₂O at constant concentrations of H₂.

constant H2 concentration [ppm]	0	200	400	600	800
H2O classification	79.8 ± 8.5	96.6 ± 1.7	96.2 ± 1.5	99.0 ± 1.3	99.4 ± 0.8

shows the classification rate from data collected during t = 0–15 min, 35–65 min, 85–115 min, 135–165 min, and 185–200 min), the classification rate of H₂O was around 79.8 ± 8.5%, which is comparably low compared to the other classes (see Table 3). This effect might be related to a substantial drift in electronic resistance as the setup, including heating and gasflow, was switched on at t = 0 min, which is usually uncommon for resistive gas measurements. It is quite widespread that resistive gas sensors are kept at constant conditions (temperature and atmosphere) for periods ranging from minutes up to several hours to avoid influencing data interpretation. This might be the reason why the classification rate is higher for constant H₂ concentrations of 200–800 ppm, as sensor drift might affect these regions less because the sensor had a longer operation period.

When only constant H₂O concentrations (e.g., 0% r. H. which is shown in column two in

Table 4. Classification rate of H₂ at constant concentrations of H₂O.

constant H2O concentration [%]	0	30	50	70
H2 classification	97.8 ± 1.7	96.4 ± 0.8	97.0 ± 1.7	95.1 ± 1.0

) were taken into account, classification levels were between around 95–98%, as summarized in Table 3.

Furthermore, a model was used where the training data consisted of relative humidity values except for 50%, which was then tested for. The classification rate was about 72.3 ± 4.0%. When doing the same for H₂ concentrations (learning with 0 ppm, 200 ppm, 600 ppm, and 800 ppm of H₂) and testing for 400 ppm, the algorithm was correct in 41.5 ± 10.3%. We tested how cycle normalization between 0 and 1 affected the results to elucidate the importance of absolute electric resistance on the classification rate. The classification rate after normalization (of each individual cycle) drops to 31.6 ± 5.7% over all five sensors. These results lead to the assumption that the absolute resistance plays a significant role in the classification process. This might also explain why it is not possible with our approach to learn with four individual sensors and classify the gas concentrations of the fifth sensor (see also ESI Figure S4.) and why PCA was less reliable when testing on the data set. The results indicate that overfitting by the algorithm is an issue that surfaces via the cross-validation. In this approach, the training data always contain similar gas concentration values compared to the data that is to be classified. Therefore, the algorithm struggles to interpret unknown data, which can be seen in the group-based approach.

3.5. Classification Success for Different Stages of the Optical Activation Process

We tried to test whether parts of the illumination (or darkness) cycle lead to similar classification rates. Therefore, the data were divided into subsets to identify the importance of certain parts of the measurement for the classification algorithm. Figure 7 shows four different kinds of (sub)datasets.

As already discussed, if the entire measurement is utilized (LED on and off for 5 s each, 10 s total) the gas concentration classification accuracy is 88.4% (Figure 7a). The accuracy reduces to 80.4 ± 4.5% if the data were utilized from when the LED is on for 5 s (b) and 81.2 ± 6.0% when the LED is off for 5 s (c), respectively. If the data for classification are chosen from the middle part of the darkness–illumination cycle (2.5 darkness, 2.5 s illumination) as depicted in (d), accuracy is about 86.6 ± 4.5%, as summarized in

Table 5. Classification rate depending on the utilized dataset.

Dataset	LED Status	Duration of ... s			Classification Rate %
		Illumination	Darkness	Dataset
(a)	on-off	5	5	10	88.4 ± 3.0
(b)	on	5	0	5	80.4 ± 4.5
(c)	off	0	5	5	81.2 ± 6.0
(d)	off-on	2.5	2.5	5	86.6 ± 4.5

.

The evaluation of the subset of data shows that a data reduction, i.e., using 5 s of a cycle, slightly decreases when the data is split into half (10 s total illumination–darkness cycle for (a) and only 5 s for every case form (b)–(d)). Therefore, it seems that the necessary information for the classification can be found in every subset investigated above, which again points towards the influence of absolute resistance on the classification rate. However, overfitting needs to be taken into account, as discussed previously.

4. Conclusions

This study successfully demonstrates a unique approach of integrating machine learning with resistive gas sensing, utilizing cyclic optical illumination to classify H₂ (0 ppm–800 ppm) and H₂O (0–70%) concentrations. Using a gas-sensitive In₂O₃ layer, cyclically illuminated with sub-bandgap energy, enabled us to employ a linear support vector machine for classification. This method achieved a classification rate of up to 92.0% using 10-fold cross-validation, maintaining high accuracy across five individual sensors (88.2 ± 2.7%).

However, we observed a decrease in accuracy when splitting learning data into half, and a strong decrease when normalizing individual cycles. While these aspects present limitations in our current model, they offer valuable insights for further investigations. Notably, our findings underline the substantial potential of combining cyclic optical activation and machine learning for gas sensing. A careful balance between the complexity of the model and the available data is required.

Increasing the amount of training data and reducing dimensionality could enhance classification rates in future studies. Furthermore, we would vary the illumination frequency, light intensity, and activation wavelength to investigate the classification rate and feature extraction, and potentially differentiate between other gases and gas concentrations at a time, especially with the rising interest in the global H₂ market. This aspect could be particularly interesting for industrial applications such as fuel cell monitoring or catalytic conversion of CO (e.g., PROX process). Moreover, unraveling the underlying mechanism of identifying and associating features to gas concentrations could improve the sensor system. The integration of artificial intelligence with traditional sensing technologies, as shown in our study, indicates a promising and exciting pathway for future research in the field of gas sensing.