Measuring Hydrogen in Indoor Air with a Selective Metal Oxide Semiconductor Sensor

: Hydrogen is a ubiquitous but often neglected gas. In analytical measurements hydrogen—as a harmless gas—often is not considered so no studies on hydrogen in indoor air can be found. For metal oxide semiconductor (MOS) gas sensors that are increasingly pushed into the application as TVOC (total volatile organic compounds) sensors, hydrogen is a severe disturbance. On the other hand, hydrogen can be an intentional choice as indicator for human presence similar to carbon dioxide. We present a ﬁeld-study on hydrogen in indoor air using selective MOS sensors accompanied by an analytical reference device for hydrogen with an accuracy of 10 ppb. Selectivity is achieved by siloxane treatment combined with temperature cycled operation and training with a complex lab calibration using randomized gas mixtures, yielding an uncertainty of 40–60 ppb. The feasibility is demonstrated by release tests with several gases inside a room and by comparison to the reference device. The results show that selective MOS sensors can function as cheap and available hydrogen detectors. Fluctuations in hydrogen concentration without human presence are measured over several days to gain insight in this highly relevant parameter for indoor air quality. The results indicate that the topic needs further attention and that the usage of hydrogen as indicator for human presence might be precluded by other sources and ﬂuctuations. the of of ideal uniform and no corresponding to 0.16 mL. Here, an increase of the H 2 concentration 2–6 h after the VOC release is observed.


Introduction
Indoor air quality (IAQ) is a general term comprising a multitude of parameters like temperature, humidity, air flow or the concentration of many trace gases [1]. The term is known since 1970 and of increasing interest during the last decades [2]. Among these trace gases are specific harmful pollutants like nitrogen dioxide (NO 2 ); formaldehyde or benzene [3,4]; carbon dioxide (CO 2 ), which is often used as indicator for human presence [5,6]; many organic compounds, often simply referred to as VOCs (volatile organic compounds) [7]; and others. Despite this high complexity inexpensive sensors are wanted for the purpose of IAQ monitoring. Nondispersive infrared sensors (NDIR) for CO 2 monitoring are state of the art for demand-controlled ventilation if concentrations are above 1000 ppm, although this level is not harmful: Pettenkofer, who introduced this guideline value, already stated in 1858 that CO 2 is only an indicator for other harmful and smelly organic substances emitted by human beings [5] and recent results from aerospace research support this assessment [8]. In today's world VOCs are not only emitted by humans but also from other sources like furniture, personal care products or cleaning agents, so this approach of indirect VOC measurement might be erroneous [9].
Up-and-coming for real-time and direct VOC measurements are metal oxide semiconductor (MOS) gas sensors, which can already be found in many available IAQ monitors. These sensors show high sensitivity towards a wide range of reducing gases. Their selectivity-e.g., to a group of gases like VOCs-can be optimized by cyclically changing the temperature of the sensing layer (TCO, temperature cycled operation) [10,11]. This type

Experiments
SGP30 multilayer MOS sensors (four sensitive layers on one hotplate, Sensirion AG, Stäfa, Switzerland) were used for this study. One sensor device was installed as delivered and another one was pre-treated with Octamethylcyclotetrasiloxane at 2 ppm over 18 h yielding a H 2 selective sensor (compare [26,28]). The temperature cycle is specifically designed to achieve high sensitivity and selectivity by using the differential surface reduction (DSR) method, which is described in detail in previous publications [32][33][34]. For this purpose, the temperature cycle of the untreated sensor consists of several steps from high to low temperature, always starting for 5 s at 400 • C followed by 7 s each at 100, 125, . . . , 325 • C resulting in a total cycle duration of 120 s. For the pre-treated sensor, this cycle was adjusted as the sensor reaction is slower after siloxane treatment. Longer temperature Atmosphere 2021, 12, 366 3 of 13 plateaus and increased temperatures are required to compensate this effect: 10 s at 400 • C are here followed by 20 s each at 275, 300, 325, and 350 • C again resulting in a total duration of 120 s. Temperature control and sensor read-out are done using the SGP30 s integrated ASIC. Via a tiny 4.0 microcontroller board the temperatures from the TCO are controlled and the data is transferred to a computer. The sampling rate is 25 Hz.
The machine learning is performed with our open source toolbox DAV 3 E [35]. The sensor cycles are divided into 120 sections of 1 s duration and in each of these sections the linear slope of the logarithm of the sensor conductance is computed. Due to the four layers in each sensor 480 features are obtained from every individual sensor. For each sensor, feature selection is performed by recursive feature elimination and PLSR (partial least squares regression) models predicting the concentration of H 2 are trained [36]. Each model is validated by 10-fold cross-validation to define an appropriate number of PLSR components yielding the root mean squared error for validation (RMSE V ) [37]. During hold-out all cycles of one random gas mixture are always left out from the training to avoid vulnerability to over-fitting [37]. Moreover, 20% of the data is held out during the ML optimization process and applied as test data in the end yielding RMSE T as a final step to guarantee proper model functionality [37].
In a first step, the sensors were lab-calibrated inside a gas mixing system [38,39] using randomized gas mixtures [40], which means that all gases connected to the system, in this case six, are applied at once and their concentrations are randomly chosen from defined ranges (acetone 17-1000 ppb, carbon monoxide 150-2000 ppb, ethanol 4-1000 ppb, formaldehyde 1-400 ppb, hydrogen 400-4000 ppb, and toluene 4-1000 ppb). For the generation of this randomized mixtures Latin hypercube sampling is performed to ensure proper scanning of the full measurement range and the obtained concentration values are stored for training of the sensor models [41]. This method yields perfectly suited data for the training of machine learning algorithms as described before [40]. Each gas exposure has a duration of 20 min, 400 independent mixtures were measured during the calibration.
The sensors were then brought to the field-a normal office-where they were installed for two months interrupted by an additional lab measurement after four weeks. This measurement has the identical structure as the initial calibration and is used to identify and compensate drift in the sensor signals. For this purpose, both measurements-initial and repeated lab-calibration-were used to determine the ML models and evaluate the data presented in this manuscript.
During this time in the field several release tests were conducted inside the room: acetone, ethanol, isopropanol, limonene, and toluene were released by evaporating liquid in the middle of the room (the full amount of liquid of 0.1 to 0.16 mL-depending on the substance-inside a petri dish that was placed on a desk). Carbon monoxide was released by burning a tea candle. Hydrogen was released from a pressure cylinder with a concentration of 2000 ppm in air at a rate of 500 mL/min using a mass flow controller. During several days, a gas chromatograph with a reducing compound photometer as detector (Peak Performer 1, PeakLaboratories Inc., Mountain View, California, USA, detection limit and accuracy 10 ppb H 2 ) was installed in the room in parallel to the sensor systems as reference for the H 2 concentration. The instrument was provided on free loan and in that short period of time we were not be able to perform a proper recalibration which would have been needed according to the manual.
A plan of the field test room is shown in Figure 1. The area is 21.91 m 2 , the total volume of the room is 61.35 m 3 . The releases were performed close to the centre of the room and a fan from the corner was installed to produce some air flow and improve the (ideally uniform) distribution of released substances. From a CO 2 release and the decay curve measured by NDIR CO 2 sensors (SCD40, Sensirion AG, Stäfa, Switzerland) the air exchange rate with closed window and door is approximately 0.17 h −1 .

Lab Calibration
From the lab calibration the PLSR models shown in Figure 2 were obtained for both sensors. Each data point represents one cycle. The obtained values from the calibration are RMSEV = 44.8 ppb and RMSET = 40.5 ppb for the untreated sensor and RMSEV = 56.0 ppb and RMSET = 60.0 ppb for the pre-treated sensor. Since the pre-treatment enhances selectivity by reducing the sensitivity towards other gases more than towards H2, it is not surprising that quantification is somewhat worse for the pre-treated sensor. On the other hand, this sensor is chemically selective, whereas the untreated sensor might fail if gas compounds occur which were not part of the calibration dataset. Therefore, in the following we will always show both sensors, to check for synchronicity.

Lab Calibration
From the lab calibration the PLSR models shown in Figure 2 were obtained for both sensors. Each data point represents one cycle. The obtained values from the calibration are RMSE V = 44.8 ppb and RMSE T = 40.5 ppb for the untreated sensor and RMSE V = 56.0 ppb and RMSE T = 60.0 ppb for the pre-treated sensor. Since the pre-treatment enhances selectivity by reducing the sensitivity towards other gases more than towards H 2 , it is not surprising that quantification is somewhat worse for the pre-treated sensor. On the other hand, this sensor is chemically selective, whereas the untreated sensor might fail if gas compounds occur which were not part of the calibration dataset. Therefore, in the following we will always show both sensors, to check for synchronicity.

Validation in the Field
To check the functionality of the calibrated PLSR models directly in the field, we performed several H 2 release tests. During the first two release tests no reference was available. Two consecutive releases were performed, the first one with a calculated concentration increase of 1 ppm inside the room and the second one with a calculated concentration increase of 2 ppm. These concentrations would be obtained if the gas released from the cylinder is evenly distributed inside the room and if there is no loss through ventilation, adsorption, or reaction; on average, somewhat lower values due to unavoidable losses are expected. Figure 3a,b show the recorded concentration predictions based on the PLSR model during the first (1 ppm) and second release test (2 ppm), respectively; the start and end points of the releases are marked. The untreated sensor shows slightly (100-200 ppb) lower absolute values than the pre-treated sensor over the whole course. The H 2 concentration increase during the first release is 900 ppb for the untreated sensor and 890 ppb for the pre-treated sensor. During the second H 2 release the predicted increases are 1430 ppb and 1240 ppb for the untreated and pre-treated sensor, respectively.

Validation in the Field
To check the functionality of the calibrated PLSR models directly in the field, we performed several H2 release tests. During the first two release tests no reference was available. Two consecutive releases were performed, the first one with a calculated concentration increase of 1 ppm inside the room and the second one with a calculated concentration increase of 2 ppm. These concentrations would be obtained if the gas released from the cylinder is evenly distributed inside the room and if there is no loss through ventilation, adsorption, or reaction; on average, somewhat lower values due to unavoidable losses are expected. Figure 3a,b show the recorded concentration predictions based on the PLSR model during the first (1 ppm) and second release test (2 ppm), respectively; the start and end points of the releases are marked. The untreated sensor shows slightly (100-200 ppb) lower absolute values than the pre-treated sensor over the whole course. The H2 concentration increase during the first release is 900 ppb for the untreated sensor and 890 ppb for the pre-treated sensor. During the second H2 release the predicted increases are 1430 ppb and 1240 ppb for the untreated and pre-treated sensor, respectively.
During two periods a reference instrument (Peak Performer 1) was available. Figure  4a shows the first of these periods, where data from the reference instrument was recorded in parallel with both sensors over one weekend. While only minimal concentration changes were observed during that period, good correlation between all three signals is nevertheless evident. The untreated sensor shows an offset of approximately 200 ppb and only slightly higher noise than the analytical reference instrument. The pre-treated sensor shows an even higher offset of 500 ppb and, in accordance with the calibration results, a significantly higher noise level with almost double amplitude.
The second test period including the reference instrument is shown in Figure 4b. Again, we measured over one weekend, but this time some significant changes in H2 concentration were observed. Both sensor signals show a very good correlation to the reference device. On the final day a release test with a calculated concentration of 2 ppm H2 was performed. The room was thoroughly ventilated through open window and door before this release test. The reference instrument indicated a concentration increase during the release of 1465 ppb. The two sensors indicate an increase of 1650 ppb (untreated sensor) and 1365 ppb (pre-treated sensor). Despite the higher offset, the relative PLSR signal of the chemically selective MOS sensor correlates better with the analytical measurement (7% deviation vs. 13% deviation in the observed concentration increase). A correlation analysis over the whole period when the reference device was measuring in parallel was performed. Due to the different sampling rates (2 min for the sensors and 3.6 min for the Peak Performer 1) the data were resampled to obtain data points every minute using linear interpolation, Figure 5. From the 9560 data points per device (covering 160 h) Pearson correlation coefficients between the Peak Performer 1 and the PLSR model predictions from the untreated sensor and the pre-treated sensor were obtained of 0.97 and 0.95, respectively.
In field measurements it is impossible to gain uncertainty levels from noise calculation that are free of doubt because one cannot distinguish between changes in concentration and pure signal noise. However, we chose a 2 h window on 1 November with tolerably constant concentration level. A Lilliefors test was applied to the data points of all devices indicating they are normally distributed (5% significance level for rejection of the null hypothesis) so the assumption of constant concentration is reasonable [42]. The standard deviation of the Peak Performer 1 on that time scale is 15.9 ppb, which is close to the given accuracy of the instrument of 10 ppb. The untreated sensor shows a standard deviation of 29.8 ppb and the pre-treated sensor 36.1 ppb.
To judge the absolute accuracy of all three devices the ventilation event before the release test can be used. The atmosphere contains 500 ppb of H2 so lower levels are unrealistic-higher levels might be possible due to sources nearby (traffic or industry) [13,14]. The Peak Performer 1 shows a minimum concentration of 380 ppb during this ventilation, the untreated sensor 465 ppb and the pre-treated sensor 560 ppb. Thus, in this case the offset for the analytical reference device is at least −120 ppb and for the untreated sensor at least −35 ppb, the absolute accuracy of the pre-treated sensor cannot be determined.   During two periods a reference instrument (Peak Performer 1) was available. Figure 4a shows the first of these periods, where data from the reference instrument was recorded in parallel with both sensors over one weekend. While only minimal concentration changes were observed during that period, good correlation between all three signals is nevertheless evident. The untreated sensor shows an offset of approximately 200 ppb and only slightly Atmosphere 2021, 12, 366 6 of 13 higher noise than the analytical reference instrument. The pre-treated sensor shows an even higher offset of 500 ppb and, in accordance with the calibration results, a significantly higher noise level with almost double amplitude.

Results during Field Tests
The first 25 days of the field test before the interim calibration is shown in Figure 6, the second part after that calibration spanning 22 days in Figure 7. Indicated H2 concentrations are between 450 ppb and 2300 ppb. During most of the time, the indicated concentration inside the room is around 1000 ppb, which is high compared to the atmospheric background level of 500 ppb. Numbers along the line mark events or periods, a detailed description of these in given in Appendix A, Table A1. All events labelled below the traces of the PLSR model predictions indicate ventilation events, i.e., the window and/or door of the room were opened, and the room thoroughly ventilated to achieve a clean baseline before and after the conducted release tests.  The second test period including the reference instrument is shown in Figure 4b. Again, we measured over one weekend, but this time some significant changes in H 2 concentration were observed. Both sensor signals show a very good correlation to the reference device. On the final day a release test with a calculated concentration of 2 ppm H 2 was performed. The room was thoroughly ventilated through open window and door before this release test. The reference instrument indicated a concentration increase during the release of 1465 ppb. The two sensors indicate an increase of 1650 ppb (untreated sensor) and 1365 ppb (pre-treated sensor). Despite the higher offset, the relative PLSR signal of the chemically selective MOS sensor correlates better with the analytical measurement (7% deviation vs. 13% deviation in the observed concentration increase).
A correlation analysis over the whole period when the reference device was measuring in parallel was performed. Due to the different sampling rates (2 min for the sensors and 3.6 min for the Peak Performer 1) the data were resampled to obtain data points every minute using linear interpolation, Figure 5. From the 9560 data points per device (covering 160 h) Pearson correlation coefficients between the Peak Performer 1 and the PLSR model predictions from the untreated sensor and the pre-treated sensor were obtained of 0.97 and 0.95, respectively.
In field measurements it is impossible to gain uncertainty levels from noise calculation that are free of doubt because one cannot distinguish between changes in concentration and pure signal noise. However, we chose a 2 h window on 1 November with tolerably constant concentration level. A Lilliefors test was applied to the data points of all devices indicating they are normally distributed (5% significance level for rejection of the null hypothesis) so the assumption of constant concentration is reasonable [42]. The standard deviation of the Peak Performer 1 on that time scale is 15.9 ppb, which is close to the given accuracy of the instrument of 10 ppb. The untreated sensor shows a standard deviation of 29.8 ppb and the pre-treated sensor 36.1 ppb.  . PLSR model predictions for H2 for both sensors (untreated and pre-treated) compared to an analytical reference instrument (Peak Performer 1): (a) a weekend without additional events, the reference instrument was not connected to the pump for a short period of time and indicates 0 ppb during that period; (b) during a weekend followed by a release test with a calculated concentration increase of 2 ppm H2.

Results during Field Tests
The first 25 days of the field test before the interim calibration is shown in Figure 6, the second part after that calibration spanning 22 days in Figure 7. Indicated H2 concentrations are between 450 ppb and 2300 ppb. During most of the time, the indicated concentration inside the room is around 1000 ppb, which is high compared to the atmospheric background level of 500 ppb. Numbers along the line mark events or periods, a detailed description of these in given in Appendix A, Table A1. All events labelled below the traces of the PLSR model predictions indicate ventilation events, i.e., the window and/or door of the room were opened, and the room thoroughly ventilated to achieve a clean baseline before and after the conducted release tests. To judge the absolute accuracy of all three devices the ventilation event before the release test can be used. The atmosphere contains 500 ppb of H 2 so lower levels are unrealistic-higher levels might be possible due to sources nearby (traffic or industry) [13,14]. The Peak Performer 1 shows a minimum concentration of 380 ppb during this ventilation, the untreated sensor 465 ppb and the pre-treated sensor 560 ppb. Thus, in this case the offset for the analytical reference device is at least −120 ppb and for the untreated sensor at least −35 ppb, the absolute accuracy of the pre-treated sensor cannot be determined.

Results during Field Tests
The first 25 days of the field test before the interim calibration is shown in Figure 6, the second part after that calibration spanning 22 days in Figure 7. Indicated H 2 concentrations are between 450 ppb and 2300 ppb. During most of the time, the indicated concentration inside the room is around 1000 ppb, which is high compared to the atmospheric background level of 500 ppb. Numbers along the line mark events or periods, a detailed description of these in given in Appendix A, Table A1. All events labelled below the traces of the PLSR model predictions indicate ventilation events, i.e., the window and/or door of the room were opened, and the room thoroughly ventilated to achieve a clean baseline before and after the conducted release tests.
Events 11 and 12 indicate the H 2 release tests shown in Figure 3a,b, #25 indicates the weekend with the data of the reference instrument recorded in parallel (cf. Figure 4a), and #47-49 the weekend and H 2 release test with the second recording of the reference instrument in parallel (cf. Figure 4b).
Weekends are indicated by #7, 14, 23, 25, 36, and 47. During these days no one entered the room, and no release tests were conducted. Nevertheless, ascending as well as descending concentrations are observed, especially at #14, where oscillating values indicate a diurnal cycle. The maxima of these cycles are typically reached in the morning around 6-8 a.m., the corresponding minima in the evening around 6-8 p.m. During other weekends, this clear correlation with time is less obvious, but in general descending values are more likely in the afternoon.
During the field tests, release tests were also performed with different VOCs. The amount of evaporated liquid was always chosen to theoretically reach 600 ppb inside the room assuming ideal uniform distribution and no ventilation, corresponding to 0.1 to 0.16 mL. Here, an increase of the H 2 concentration 2-6 h after the VOC release is observed. This increase often starts around 6 p.m., which is the same time at which the minimum in the diurnal cycle is observed. Looking at those days with release tests, during which a higher H 2 concentration is observed than on the day before (#16, 20,22,27,29,31,40,42, and 46) a correlation to the substances acetone, isopropanol, toluene and xylene is found. Release of limonene (#35) results in an immediate increase of the PLSR output of the chemically selective pre-treated sensor, while the signal of the untreated sensor shows a less pronounced increase after 5-6 h. of limonene (#35) results in an immediate increase of the PLSR output of the chemically selective pre-treated sensor, while the signal of the untreated sensor shows a less pronounced increase after 5-6 h. Event 5 shows a strong increase in H2 concentration overnight after a day without any identified events or tests during that day. Event 33 shows a very strong increase of the H2 concentration prediction only from the untreated sensor while the signal obtained from the pre-treated sensor stays nearly constant. After ventilation (34) on the next day again a milder increase is observed with the same pattern. We were able to attribute these events to construction work inside the building using a binding agent, which was done on the same floor as the field test room on the first day and on the floor below on the second day.
Due to the Covid-19 pandemic no systematic investigation of the effect of human presence was possible. Only on some days (#9, 15, 29, and 45) individual persons sporadically entered the room and for every release test one person entered the room. However, in none of these cases could the human presence be observed from the H2 signal.    Table A1.

Discussion
Two questions arising from the results need to be discussed in more detail: (1) can MOS sensors function as reliable H2 detectors also in further research, and (2) what do we learn from the results concerning H2 in indoor environments?
Especially Figure 4-the PLSR predictions obtained from the MOS sensors recorded in parallel with an analytical reference instrument-strongly suggests, that the calibrated sensors can correctly measure relative fluctuations. For absolute measurements, an infield one-point calibration might be sufficient to eliminate the observed but constant offset between the three devices. Similar to low-cost CO2 sensors, this could be realized with an automatic baseline correction, i.e., setting the lowest value over a longer period to the  Table A1. Event 5 shows a strong increase in H 2 concentration overnight after a day without any identified events or tests during that day. Event 33 shows a very strong increase of the H 2 concentration prediction only from the untreated sensor while the signal obtained from the pre-treated sensor stays nearly constant. After ventilation (34) on the next day again a milder increase is observed with the same pattern. We were able to attribute these events to construction work inside the building using a binding agent, which was done on the same floor as the field test room on the first day and on the floor below on the second day.
Due to the Covid-19 pandemic no systematic investigation of the effect of human presence was possible. Only on some days (#9, 15, 29, and 45) individual persons sporadically entered the room and for every release test one person entered the room. However, in none of these cases could the human presence be observed from the H 2 signal.

Discussion
Two questions arising from the results need to be discussed in more detail: (1) can MOS sensors function as reliable H 2 detectors also in further research, and (2) what do we learn from the results concerning H 2 in indoor environments?
Especially Figure 4-the PLSR predictions obtained from the MOS sensors recorded in parallel with an analytical reference instrument-strongly suggests, that the calibrated sensors can correctly measure relative fluctuations. For absolute measurements, an in-field one-point calibration might be sufficient to eliminate the observed but constant offset between the three devices. Similar to low-cost CO 2 sensors, this could be realized with an automatic baseline correction, i.e., setting the lowest value over a longer period to the atmospheric background concentration. It is remarkable, that applying this approach during a ventilation event yields, that the analytical reference shows the highest offset from the baseline level of 500 ppb. Keeping in mind that the analytical device is three orders of magnitude more expensive than MOS sensors the performance of the sensors is very good with approximately doubled noise level compared to the analytical device and correlation coefficients of 0.95 and 0.97, respectively.
As the sensors are running parallel in almost all situations and given the very different methods of achieving the selective H 2 signal, this also strongly implies the capability of MOS sensors to reliably measure H 2 at least at ppm and sub-ppm concentrations. There are two events where significant deviations between both models occur: #33, where an emission during construction work was classified as H 2 by the untreated sensor. In this situation, the chemically selective pre-treated sensor proves to be more reliable when untrained and unusual gases occur. During the release tests of limonene (#35), the opposite effect is observed: the model of the pre-treated sensor indicates an increase of H 2 , while the model of the untreated sensor remains at a constant value during the release and only indicates increasing values several hours after the release test. Since in contrast to all the other released gases we do not see any sensor reaction in the raw sensor data, we assume that the sealing membrane on the sensor prevents the large limonene molecules reaching the sensor surface. For the pre-treatment this membrane was removed, thus limonene can reach the hot sensor surface of the pre-treated sensor. We assume that limonene reacts on the hot surface releasing H 2 (cf. [43]), resulting in a sensor response and an increase of the PLSR model prediction. Reapplying the sealing membrane would help to resolve this issue. Thus, in terms of selectivity the pre-treated sensor would be more reliable. In return a lower signal to noise ratio needs to be accepted. In contrast to a pre-treated sensor at constant temperature, which can be found in literature and suffers from low response times [30,44], the DSR mode overcomes this issue.
From the field tests we observe that releasing VOCs like toluene and acetone is followed by an increase in H 2 concentration hours later. This might be due to decomposition processes such as photochemical oxidation [15]. However, this effect is also overlaid by the observed diurnal cycle with minimum H 2 concentrations in the evening. This diurnal cycle is not fully correlated with the time of day, because we do not observe it every day. Probably also incident solar radiation, which is higher in the afternoon because the window is directed westwards, or temperature play a role. Therefore, high VOC levels probably also cause high H 2 levels. During the field tests Tenax samples were drawn and analyzed by GC-MS, but no unusually high concentrations of any substance were found. Thus, the observed H 2 base level inside the room, which was confirmed by the reference instrument to be significantly higher than atmospheric level, cannot be explained and is not fully understood. It seems that an unknown H 2 source is inside the room as only lower concentrations were found in neighboring rooms and outside using the same sensors. Perhaps this high base level is caused by some unknown VOC source [16], e.g., from building materials or furniture, or by microorganisms, which can emit H 2 from their metabolism (in fact, they are the cause for human H 2 emissions).
Unfortunately, the observed changes of the H 2 concentration overshadow all moments with human presence, which were short and only included one person. Thus, it was not possible to detect any effect of human presence against this background. In former experiments we found an increase of 2500 ppb over a span of 1 h in a room of similar size while 13 people were inside (same building, same floor). This correlates to an increase of approximately 200 ppb per person and hour. In a normal indoor air environment this can represent a valid indicator for human presence, but effects of VOC release, i.e., from cleaning agents would have to be considered. In any case, further experiments to determine H 2 concentrations and fluctuations in different rooms under normal use (e.g., cooking, cleaning etc.) are required.

Conclusions
Two important conclusions can be drawn from the presented field test results: first, MOS sensors can reliably detect and quantify H 2 in indoor environments; second, H 2 is a dynamic parameter that needs more attention. More research is required on many open questions: Does H 2 emitted by humans dominate in normal rooms so that it can be used as indicator for human presence? Does H 2 represent a good indicator for other relevant and potentially harmful situations, e.g., high VOC values or microbial issues such as mold? Do many TVOC monitors based on MOS sensors, especially when driven at constant temperature, mainly measure H 2 ?

Acknowledgments:
We thank Rainer Lammertz Pure Gas Products for providing the Peak Performer 1 reference instrument for this study.

Conflicts of Interest:
The authors declare no conflict of interest.