A Highly Sensitive and Selective Laser-Based BTEX Sensor for Occupational and Environmental Monitoring

Mhanna, Mhanna; Sy, Mohamed; Arfaj, Ayman; Llamas, Jose; Farooq, Aamir

doi:https://doi.org/10.5194/egusphere-2023-514

Preprints

https://doi.org/10.5194/egusphere-2023-514

Preprints

24 Apr 2023

| 24 Apr 2023

A Highly Sensitive and Selective Laser-Based BTEX Sensor for Occupational and Environmental Monitoring

Mhanna Mhanna, Mohamed Sy, Ayman Arfaj, Jose Llamas, and Aamir Farooq

Abstract. A mid-infrared laser-based sensor is designed and demonstrated for trace detection of benzene, toluene, ethylbenzene, and xylene isomers at ambient conditions. The sensor is based on a distributed feedback inter-band cascade laser emitting near 3.29 μm and an off-axis cavity-enhanced absorption spectroscopy configuration with an optical gain of ~2800. Wavelength tuning and a deep neural networks (DNN) model were employed to enable simultaneous and selective BTEX measurements. The sensor performance was demonstrated by measuring BTEX mole fractions in various mixtures. At an integration time of 10 seconds, minimum detection limits of 11.4, 9.7, 9.1, 10, 15.6, and 12.9 ppb were achieved for benzene, toluene, ethylbenzene, m-xylene, o-xylene, and p-xylene, respectively. The sensor can be used to detect tiny BTEX leaks in petrochemical facilities and to monitor air quality in residential and industrial areas for workplace pollution.

Received: 20 Mar 2023 – Discussion started: 24 Apr 2023

Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this preprint. The responsibility to include appropriate place names lies with the authors.

Download & links

Mhanna Mhanna, Mohamed Sy, Ayman Arfaj, Jose Llamas, and Aamir Farooq

Status: closed

CC1:
'Comment on egusphere-2023-514', Dean Venables, 25 Apr 2023

This is a valuable spectroscopic contribution to realtime detection of aromatics. The authors should cite and compare their system to the recent, directly-relevant paper by Wang et al.:
Meng Wang, Ravi Varma, Dean S Venables, Wu Zhou, Jun Chen. A Demonstration of Broadband Cavity-Enhanced Absorption Spectroscopy at Deep-Ultraviolet Wavelengths: Application to Sensitive Real-Time Detection of the Aromatic Pollutants Benzene, Toluene, and Xylene, Anal. Chem., 94(10):4286-4293, 2022.
doi: 10.1021/acs.analchem.1c04940.

Citation: https://doi.org/10.5194/egusphere-2023-514-CC1
- AC1: 'Reply on CC1', Mhanna Mhanna, 08 May 2023
  
  Thank you for suggesting this relevant paper, we will add a comparison in our manuscript.
  
  Citation: https://doi.org/10.5194/egusphere-2023-514-AC1
RC1:
'Comment on egusphere-2023-514', Anonymous Referee #1, 15 May 2023
Although the authors have an interesting subject for a scientific study, the manuscript is not convincing and lacks scientifically too many flaws to be accepted for publication.
The discussion on the cavity enhanced absorption (line 87 to 99) is confusing and not according the generally agreed formula’s. I suggest to follow the formulas in the article: Cavity-enhanced absorption spectroscopy of molecular oxygen by the Gianfrani et al., Journal of the Optical Society of America B Vol. 16, Issue 12, pp. 2247-2254 (1999) https://tf.nist.gov/general/pdf/1324.pdf There is also shown that the increase in path length (or absorption) is equal to π.sqrt(R)/(1-R)

Figure 1 is not convincing concerning an interference free and selective sensing. I do not see the advantage compared to the UV. Where are the fingerprint spectra of the gases next to their absorption strengths? How orthogonal are the gases to each other?

The discussion on the proper selection of the spectral wavelength is not convincing. Show the broadband spectra (between 3000 and 3100 cm-1), next to the spectra of water vapor and CO2. The latter two are important because their concentrations are orders of magnitude higher.

Since BTEX consist of 6 molecules, give all 6 bands.

Line 124: The focusing lens is not described. How is the beam divergence of the ICL? What was the type and performance of the oscilloscope?

Line 127-128 give a reference

Line 140: A BTEX mixture was used containing 6 gasses. What was the uncertainty in the concentration of each of the gases in the mixture? It seems here that al the gases are in one mixture, although line 153 states something else…?

A 99.97% reflectivity of the mirrors means an enhancement of 10470 in path length (see reference). With a cavity length of 5 cm, the total path length should be 520 meter. In the manuscript the Gain is 2800; can these two be compared? How is the overall NEAS (Noise Equivalent Absorption Sensitivity) of the system?

What was the averaged transmitted power through the cavity? How is this compared to the detector sensitivity and where is the performance of the latter in the Allan curve (Fig.3)?

The spectra leading to Fig.4 are not shown, while these are essential to convince the interested reader. Furthermore, I do not see any error bars in the figure.

Equation 11: Why is k limited from 1-4 and not extended to 6 (of higher, due to interfering species)?

While the uncertainty of alpha(CEAS) was determined to be 2% (line 160), the mixtures were calculated to be 3-digit accuracy (line 210-). There is no convincing performance of the DNN model to be justified.

Fig 7 cannot be properly judges due to all the raised questions above

Fig 8 is not convincing.
Citation: https://doi.org/10.5194/egusphere-2023-514-RC1
- AC2:
  'Reply on RC1', Mhanna Mhanna, 16 Jun 2023
  We thank the reviewer for his/her constructive comments which allowed us to improve our manuscript. Our replies to the comments are given below. The answers are also found in the attachment for better visualization.
  Although the authors have an interesting subject for a scientific study, the manuscript is not convincing and lacks scientifically too many flaws to be accepted for publication.
  The discussion on the cavity enhanced absorption (line 87 to 99) is confusing and not according the generally agreed formulae. I suggest to follow the formulas in the article: Cavity-enhanced absorption spectroscopy of molecular oxygen by the Gianfrani et al., Journal of the Optical Society of America B Vol. 16, Issue 12, pp. 2247-2254 (1999) https://tf.nist.gov/general/pdf/1324.pdf There is also shown that the increase in path length (or absorption) is equal to π.sqrt(R)/(1-R)
  
  While the paper suggested by the reviewer (published in 1999) is indeed a good reference on path-length amplification, our discussion is also well-established and follows the book by Gagliardi and Loock (2014) [1]. In fact, both references and discussions therein are similar and lead to the same results. Following Gianfrani’s equations [2]: F=pi.sqrt(r1.r2)/(1-r1.r2)=4336
  Thus, for the path-length of 5 cm given in manuscript, the equivalent pathlength is: Leq=(2/pi).FL=138.88 m
  This is same as the effective pathlength (139 m) given in line 157 of the manuscript based on our approach.
  Figure 1 is not convincing concerning an interference free and selective sensing. I do not see the advantage compared to the UV. Where are the fingerprint spectra of the gases next to their absorption strengths? How orthogonal are the gases to each other?
  
  In the UV, most hydrocarbons have broad overlapping spectra which makes it very challenging to isolate BTEX from other interfering molecules, whereas in the selected wavelength region, the main absorbers are only BTEX. Another advantage of the IR region versus UV is the availability of small-form semiconductor lasers in the IR region. Indeed, the goal of Fig. 1 is not to show the feasibility of interference-free and selective sensing of BTEX near 3.3 µm, but rather to show their overlapping spectra as an opportunity for their simultaneous sensing using DNN (deep neural networks). The absorbance spectra of BTEX are non-orthogonal in the selected wavelength range, which obfuscates their selective sensing using traditional fitting algorithms.
  The discussion on the proper selection of the spectral wavelength is not convincing. Show the broadband spectra (between 3000 and 3100 cm^-1), next to the spectra of water vapor and CO₂. The latter two are important because their concentrations are orders of magnitude higher.
  
  We have added 400 ppm CO₂ and 1.5 % water vapor (typical mole fractions in air) spectra to the selected wavelength range to show their low absorbance in comparison to the target BTEX species. We only showed these spectra in the narrow selected range rather than 3000 – 3100 cm^-1 because water absorbance is much higher than the target species in that range and it will overwhelm their spectra. The most important thing is that in the selected region (3039.25 – 3040.5 cm^-1), water and CO₂ absorbance is much smaller than BTEX.
  Since BTEX consist of 6 molecules, give all 6 bands.
  
  We have added these to the manuscript.
  Line 124: The focusing lens is not described. How is the beam divergence of the ICL? What was the type and performance of the oscilloscope?
  
  We have added these to the manuscript.
  Line 127-128 give a reference
  
  We have added a reference.
  Line 140: A BTEX mixture was used containing 6 gasses. What was the uncertainty in the concentration of each of the gases in the mixture? It seems here that all the gases are in one mixture, although line 153 states something else…?
  
  Uncertainties are now stated for each BTEX molecule. In Allan variance and sensor validation measurements, the target gases were mixed to prepare one mixture. However, in the cavity gain measurements, each species was diluted separately in nitrogen to characterize the gain with respect to the reference absorption cross-section of each species. This is because the magnitudes of absorption cross-sections are different for each molecule.
  A 99.97% reflectivity of the mirrors means an enhancement of 10470 in path length (see reference). With a cavity length of 5 cm, the total path length should be 520 meter. In the manuscript the Gain is 2800; can these two be compared? How is the overall NEAS (Noise Equivalent Absorption Sensitivity) of the system?
  
  Based on Eq. (5) in the reference provided by the reviewer [2], the enhancement factor is: (2/pi)F=2*sqrt(r1.r2)/(1-r1.r2)
  This means that a 99.97% reflectivity should give an enhancement of 3333 (not 10470). However, a reflectivity of 99.964% (as indicated in the manuscript) should give an enhancement factor of 2777 (according to the provided reference [2]). This is in perfect agreement with the gain value obtained in the manuscript using our approach (2789).
  The NEAS (Noise Equivalent Absorption Sensitivity) is 0.01%, as shown in the bottom panel of Fig. 9 of the revised manuscript.
  What was the averaged transmitted power through the cavity? How is this compared to the detector sensitivity and where is the performance of the latter in the Allan curve (Fig.3)?
  
  The average transmitted power through the cavity is calculated according to the following equation [3]: It=I0(1/2)(1-r)
  This is well above the noise-equivalent-power (NEP) of the photodetector, which is calculated as follows: NEP=sqrt(A*df)/D*
  where is the area of the photosensitive region, is the bandwidth, and is the specific detectivity of the photodetector.
  The Allan variance curve of the photodetector signal is given in Fig. R1. We have updated Figure 3 in the manuscript with this one.
  The spectra leading to Fig.4 are not shown, while these are essential to convince the interested reader. Furthermore, I do not see any error bars in the figure.
  
  Error bars are already included in the figure (new Fig. 5). They are clearer in the high absorbance region (top right corner) because of the higher error.
  The spectra leading to Fig. 4 are based on absorbance of single species, where each BTEX was diluted separately in nitrogen to obtain the cavity gain, as mentioned in the manuscript. A sample of these spectra is now added to the manuscript as Fig. 4.
  Equation 11: Why is k limited from 1-4 and not extended to 6 (of higher, due to interfering species)?
  
  Thank you for pointing this out, it was a typo and now we fixed it.
  While the uncertainty of alpha(CEAS) was determined to be 2% (line 160), the mixtures were calculated to be 3-digit accuracy (line 210). There is no convincing performance of the DNN model to be justified.
  
  Uncertainties are now added to the measured mole fractions (3.76%) based on combining the RMSE of the DNN model and the uncertainty in the measured cavity absorbance.
  Fig 7 cannot be properly judged due to all the raised questions above
  
  Now that the above concerns have been clarified, we hope that Fig. 7 is clear.
  Fig 8 is not convincing.
  
  We have now added the absorbance resulting from the fluctuation in the two incident laser intensity signals to make it more convincing.
  
  References:
  [1] Gagliardi, Gianluca, and Hans-Peter Loock, eds. Cavity-enhanced spectroscopy and sensing. Vol. 179. Berlin: Springer, 2014.
  [2] Gianfrani, Livio, Richard W. Fox, and Leo Hollberg. "Cavity-enhanced absorption spectroscopy of molecular oxygen." JOSA B 16.12 (1999): 2247-2254.
  [3] Moyer, E. J., et al. "Design considerations in high-sensitivity off-axis integrated cavity output spectroscopy." Applied Physics B 92 (2008): 467-474.
  
  Citation: https://doi.org/10.5194/egusphere-2023-514-AC2
RC2:
'Comment on egusphere-2023-514', Anonymous Referee #2, 02 Nov 2023

The authors present a machine learning approach to spectral deconvolution and quantitative laser absorption sensing of BTEX molecules in the mid-infrared. Cavity enhanced absorption spectroscopy is used to achieve adequate sensitivity. While the general approach is interesting, the viability and ultimate value of the method is not sufficiently demonstrated. Notably, the lack of distinguishing spectral features in the target domain (as shown in figure 6) suggests that very slight etalon fringes, misalignment, or other common distortions and noise in the baseline intensity will not be tolerable and will very likely lead to erroneous results. The authors do not provide an adequate description of how the method is or can be robust against such common concerns in laser absorption spectroscopy. Unfortunately, based upon the wavelength selection and limited tuning window, it seems unlikely such issues can be addressed with this approach.

Citation: https://doi.org/10.5194/egusphere-2023-514-RC2
- AC3: 'Reply on RC2', Mhanna Mhanna, 07 Nov 2023
  
  We thank the reviewer for his/her constructive comments. Our replies to the comments are given below.
  We have considered (and tackled) these issues while developing our sensor as we encountered these even in our simple laboratory setup. It is important to have a robust optical alignment as slight errors in alignment or etalon effects in windows can introduce small to moderate wavelength-dependent variations in laser intensity that could resemble the small differences in wavelength-dependent absorbance of BTEX species. We made sure that our alignment is sufficiently robust such that it is free of etalon fringes. In addition, such imperfections can be anticipated and included in the training set of the model so that the sensor gains immunity against them. In fact, a significant misalignment of the laser beam might lead to more serious issues with the sensor rather than etalon fringes, such as changing the cavity gain. In this case, the laser would need to be realigned, but the machine-learning model would be unaffected as the intensity of absorbance is impacted rather than its profile. It is worth mentioning that the normalization of the absorbed to non-absorbed laser intensity in Beer-Lambert law removes/minimizes the effect of etalons such that the resultant absorbance is nearly unaffected.
  Furthermore, careful selection of SNR in the training datasets is a key to enhance the predictive ability of the DNN model, which can fail if noise in the measured data is higher than that within the training datasets. There exist various noise profiles that can be included in the training datasets. For instance, in our models, we have included both white Gaussian noise and constant noise with varying SNR. This was based on actual characterization of noise in our experiments.
  We have now added this information to the manuscript to make it clearer.
  
  Citation: https://doi.org/10.5194/egusphere-2023-514-AC3

Status: closed

CC1:
'Comment on egusphere-2023-514', Dean Venables, 25 Apr 2023

This is a valuable spectroscopic contribution to realtime detection of aromatics. The authors should cite and compare their system to the recent, directly-relevant paper by Wang et al.:
Meng Wang, Ravi Varma, Dean S Venables, Wu Zhou, Jun Chen. A Demonstration of Broadband Cavity-Enhanced Absorption Spectroscopy at Deep-Ultraviolet Wavelengths: Application to Sensitive Real-Time Detection of the Aromatic Pollutants Benzene, Toluene, and Xylene, Anal. Chem., 94(10):4286-4293, 2022.
doi: 10.1021/acs.analchem.1c04940.

Citation: https://doi.org/10.5194/egusphere-2023-514-CC1
- AC1: 'Reply on CC1', Mhanna Mhanna, 08 May 2023
  
  Thank you for suggesting this relevant paper, we will add a comparison in our manuscript.
  
  Citation: https://doi.org/10.5194/egusphere-2023-514-AC1
RC1:
'Comment on egusphere-2023-514', Anonymous Referee #1, 15 May 2023
Although the authors have an interesting subject for a scientific study, the manuscript is not convincing and lacks scientifically too many flaws to be accepted for publication.
The discussion on the cavity enhanced absorption (line 87 to 99) is confusing and not according the generally agreed formula’s. I suggest to follow the formulas in the article: Cavity-enhanced absorption spectroscopy of molecular oxygen by the Gianfrani et al., Journal of the Optical Society of America B Vol. 16, Issue 12, pp. 2247-2254 (1999) https://tf.nist.gov/general/pdf/1324.pdf There is also shown that the increase in path length (or absorption) is equal to π.sqrt(R)/(1-R)

Figure 1 is not convincing concerning an interference free and selective sensing. I do not see the advantage compared to the UV. Where are the fingerprint spectra of the gases next to their absorption strengths? How orthogonal are the gases to each other?

The discussion on the proper selection of the spectral wavelength is not convincing. Show the broadband spectra (between 3000 and 3100 cm-1), next to the spectra of water vapor and CO2. The latter two are important because their concentrations are orders of magnitude higher.

Since BTEX consist of 6 molecules, give all 6 bands.

Line 124: The focusing lens is not described. How is the beam divergence of the ICL? What was the type and performance of the oscilloscope?

Line 127-128 give a reference

Line 140: A BTEX mixture was used containing 6 gasses. What was the uncertainty in the concentration of each of the gases in the mixture? It seems here that al the gases are in one mixture, although line 153 states something else…?

A 99.97% reflectivity of the mirrors means an enhancement of 10470 in path length (see reference). With a cavity length of 5 cm, the total path length should be 520 meter. In the manuscript the Gain is 2800; can these two be compared? How is the overall NEAS (Noise Equivalent Absorption Sensitivity) of the system?

What was the averaged transmitted power through the cavity? How is this compared to the detector sensitivity and where is the performance of the latter in the Allan curve (Fig.3)?

The spectra leading to Fig.4 are not shown, while these are essential to convince the interested reader. Furthermore, I do not see any error bars in the figure.

Equation 11: Why is k limited from 1-4 and not extended to 6 (of higher, due to interfering species)?

While the uncertainty of alpha(CEAS) was determined to be 2% (line 160), the mixtures were calculated to be 3-digit accuracy (line 210-). There is no convincing performance of the DNN model to be justified.

Fig 7 cannot be properly judges due to all the raised questions above

Fig 8 is not convincing.
Citation: https://doi.org/10.5194/egusphere-2023-514-RC1
- AC2:
  'Reply on RC1', Mhanna Mhanna, 16 Jun 2023
  We thank the reviewer for his/her constructive comments which allowed us to improve our manuscript. Our replies to the comments are given below. The answers are also found in the attachment for better visualization.
  Although the authors have an interesting subject for a scientific study, the manuscript is not convincing and lacks scientifically too many flaws to be accepted for publication.
  The discussion on the cavity enhanced absorption (line 87 to 99) is confusing and not according the generally agreed formulae. I suggest to follow the formulas in the article: Cavity-enhanced absorption spectroscopy of molecular oxygen by the Gianfrani et al., Journal of the Optical Society of America B Vol. 16, Issue 12, pp. 2247-2254 (1999) https://tf.nist.gov/general/pdf/1324.pdf There is also shown that the increase in path length (or absorption) is equal to π.sqrt(R)/(1-R)
  
  While the paper suggested by the reviewer (published in 1999) is indeed a good reference on path-length amplification, our discussion is also well-established and follows the book by Gagliardi and Loock (2014) [1]. In fact, both references and discussions therein are similar and lead to the same results. Following Gianfrani’s equations [2]: F=pi.sqrt(r1.r2)/(1-r1.r2)=4336
  Thus, for the path-length of 5 cm given in manuscript, the equivalent pathlength is: Leq=(2/pi).FL=138.88 m
  This is same as the effective pathlength (139 m) given in line 157 of the manuscript based on our approach.
  Figure 1 is not convincing concerning an interference free and selective sensing. I do not see the advantage compared to the UV. Where are the fingerprint spectra of the gases next to their absorption strengths? How orthogonal are the gases to each other?
  
  In the UV, most hydrocarbons have broad overlapping spectra which makes it very challenging to isolate BTEX from other interfering molecules, whereas in the selected wavelength region, the main absorbers are only BTEX. Another advantage of the IR region versus UV is the availability of small-form semiconductor lasers in the IR region. Indeed, the goal of Fig. 1 is not to show the feasibility of interference-free and selective sensing of BTEX near 3.3 µm, but rather to show their overlapping spectra as an opportunity for their simultaneous sensing using DNN (deep neural networks). The absorbance spectra of BTEX are non-orthogonal in the selected wavelength range, which obfuscates their selective sensing using traditional fitting algorithms.
  The discussion on the proper selection of the spectral wavelength is not convincing. Show the broadband spectra (between 3000 and 3100 cm^-1), next to the spectra of water vapor and CO₂. The latter two are important because their concentrations are orders of magnitude higher.
  
  We have added 400 ppm CO₂ and 1.5 % water vapor (typical mole fractions in air) spectra to the selected wavelength range to show their low absorbance in comparison to the target BTEX species. We only showed these spectra in the narrow selected range rather than 3000 – 3100 cm^-1 because water absorbance is much higher than the target species in that range and it will overwhelm their spectra. The most important thing is that in the selected region (3039.25 – 3040.5 cm^-1), water and CO₂ absorbance is much smaller than BTEX.
  Since BTEX consist of 6 molecules, give all 6 bands.
  
  We have added these to the manuscript.
  Line 124: The focusing lens is not described. How is the beam divergence of the ICL? What was the type and performance of the oscilloscope?
  
  We have added these to the manuscript.
  Line 127-128 give a reference
  
  We have added a reference.
  Line 140: A BTEX mixture was used containing 6 gasses. What was the uncertainty in the concentration of each of the gases in the mixture? It seems here that all the gases are in one mixture, although line 153 states something else…?
  
  Uncertainties are now stated for each BTEX molecule. In Allan variance and sensor validation measurements, the target gases were mixed to prepare one mixture. However, in the cavity gain measurements, each species was diluted separately in nitrogen to characterize the gain with respect to the reference absorption cross-section of each species. This is because the magnitudes of absorption cross-sections are different for each molecule.
  A 99.97% reflectivity of the mirrors means an enhancement of 10470 in path length (see reference). With a cavity length of 5 cm, the total path length should be 520 meter. In the manuscript the Gain is 2800; can these two be compared? How is the overall NEAS (Noise Equivalent Absorption Sensitivity) of the system?
  
  Based on Eq. (5) in the reference provided by the reviewer [2], the enhancement factor is: (2/pi)F=2*sqrt(r1.r2)/(1-r1.r2)
  This means that a 99.97% reflectivity should give an enhancement of 3333 (not 10470). However, a reflectivity of 99.964% (as indicated in the manuscript) should give an enhancement factor of 2777 (according to the provided reference [2]). This is in perfect agreement with the gain value obtained in the manuscript using our approach (2789).
  The NEAS (Noise Equivalent Absorption Sensitivity) is 0.01%, as shown in the bottom panel of Fig. 9 of the revised manuscript.
  What was the averaged transmitted power through the cavity? How is this compared to the detector sensitivity and where is the performance of the latter in the Allan curve (Fig.3)?
  
  The average transmitted power through the cavity is calculated according to the following equation [3]: It=I0(1/2)(1-r)
  This is well above the noise-equivalent-power (NEP) of the photodetector, which is calculated as follows: NEP=sqrt(A*df)/D*
  where is the area of the photosensitive region, is the bandwidth, and is the specific detectivity of the photodetector.
  The Allan variance curve of the photodetector signal is given in Fig. R1. We have updated Figure 3 in the manuscript with this one.
  The spectra leading to Fig.4 are not shown, while these are essential to convince the interested reader. Furthermore, I do not see any error bars in the figure.
  
  Error bars are already included in the figure (new Fig. 5). They are clearer in the high absorbance region (top right corner) because of the higher error.
  The spectra leading to Fig. 4 are based on absorbance of single species, where each BTEX was diluted separately in nitrogen to obtain the cavity gain, as mentioned in the manuscript. A sample of these spectra is now added to the manuscript as Fig. 4.
  Equation 11: Why is k limited from 1-4 and not extended to 6 (of higher, due to interfering species)?
  
  Thank you for pointing this out, it was a typo and now we fixed it.
  While the uncertainty of alpha(CEAS) was determined to be 2% (line 160), the mixtures were calculated to be 3-digit accuracy (line 210). There is no convincing performance of the DNN model to be justified.
  
  Uncertainties are now added to the measured mole fractions (3.76%) based on combining the RMSE of the DNN model and the uncertainty in the measured cavity absorbance.
  Fig 7 cannot be properly judged due to all the raised questions above
  
  Now that the above concerns have been clarified, we hope that Fig. 7 is clear.
  Fig 8 is not convincing.
  
  We have now added the absorbance resulting from the fluctuation in the two incident laser intensity signals to make it more convincing.
  
  References:
  [1] Gagliardi, Gianluca, and Hans-Peter Loock, eds. Cavity-enhanced spectroscopy and sensing. Vol. 179. Berlin: Springer, 2014.
  [2] Gianfrani, Livio, Richard W. Fox, and Leo Hollberg. "Cavity-enhanced absorption spectroscopy of molecular oxygen." JOSA B 16.12 (1999): 2247-2254.
  [3] Moyer, E. J., et al. "Design considerations in high-sensitivity off-axis integrated cavity output spectroscopy." Applied Physics B 92 (2008): 467-474.
  
  Citation: https://doi.org/10.5194/egusphere-2023-514-AC2
RC2:
'Comment on egusphere-2023-514', Anonymous Referee #2, 02 Nov 2023

The authors present a machine learning approach to spectral deconvolution and quantitative laser absorption sensing of BTEX molecules in the mid-infrared. Cavity enhanced absorption spectroscopy is used to achieve adequate sensitivity. While the general approach is interesting, the viability and ultimate value of the method is not sufficiently demonstrated. Notably, the lack of distinguishing spectral features in the target domain (as shown in figure 6) suggests that very slight etalon fringes, misalignment, or other common distortions and noise in the baseline intensity will not be tolerable and will very likely lead to erroneous results. The authors do not provide an adequate description of how the method is or can be robust against such common concerns in laser absorption spectroscopy. Unfortunately, based upon the wavelength selection and limited tuning window, it seems unlikely such issues can be addressed with this approach.

Citation: https://doi.org/10.5194/egusphere-2023-514-RC2
- AC3: 'Reply on RC2', Mhanna Mhanna, 07 Nov 2023
  
  We thank the reviewer for his/her constructive comments. Our replies to the comments are given below.
  We have considered (and tackled) these issues while developing our sensor as we encountered these even in our simple laboratory setup. It is important to have a robust optical alignment as slight errors in alignment or etalon effects in windows can introduce small to moderate wavelength-dependent variations in laser intensity that could resemble the small differences in wavelength-dependent absorbance of BTEX species. We made sure that our alignment is sufficiently robust such that it is free of etalon fringes. In addition, such imperfections can be anticipated and included in the training set of the model so that the sensor gains immunity against them. In fact, a significant misalignment of the laser beam might lead to more serious issues with the sensor rather than etalon fringes, such as changing the cavity gain. In this case, the laser would need to be realigned, but the machine-learning model would be unaffected as the intensity of absorbance is impacted rather than its profile. It is worth mentioning that the normalization of the absorbed to non-absorbed laser intensity in Beer-Lambert law removes/minimizes the effect of etalons such that the resultant absorbance is nearly unaffected.
  Furthermore, careful selection of SNR in the training datasets is a key to enhance the predictive ability of the DNN model, which can fail if noise in the measured data is higher than that within the training datasets. There exist various noise profiles that can be included in the training datasets. For instance, in our models, we have included both white Gaussian noise and constant noise with varying SNR. This was based on actual characterization of noise in our experiments.
  We have now added this information to the manuscript to make it clearer.
  
  Citation: https://doi.org/10.5194/egusphere-2023-514-AC3

Mhanna Mhanna, Mohamed Sy, Ayman Arfaj, Jose Llamas, and Aamir Farooq

Viewed

Total article views: 637 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	BibTeX	EndNote
413	193	31	637	24	15

HTML: 413
PDF: 193
XML: 31
Total: 637
BibTeX: 24
EndNote: 15

Views and downloads (calculated since 24 Apr 2023)

Month	HTML	PDF	XML	Total
Apr 2023	103	14	4	121
May 2023	84	13	3	100
Jun 2023	37	8	3	48
Jul 2023	18	15	3	36
Aug 2023	14	15	0	29
Sep 2023	17	22	1	40
Oct 2023	17	12	2	31
Nov 2023	52	18	6	76
Dec 2023	10	5	0	15
Jan 2024	8	5	1	14
Feb 2024	7	16	0	23
Mar 2024	6	18	1	25
Apr 2024	7	4	1	12
May 2024	3	9	1	13
Jun 2024	6	3	2	11
Jul 2024	7	5	3	15
Aug 2024	10	5	0	15
Sep 2024	2	3	0	5
Oct 2024	3	1	0	4
Nov 2024	2	2	0	4

Cumulative views and downloads (calculated since 24 Apr 2023)

Month	HTML	PDF	XML	Total
Apr 2023	103	14	4	121
May 2023	84	13	3	100
Jun 2023	37	8	3	48
Jul 2023	18	15	3	36
Aug 2023	14	15	0	29
Sep 2023	17	22	1	40
Oct 2023	17	12	2	31
Nov 2023	52	18	6	76
Dec 2023	10	5	0	15
Jan 2024	8	5	1	14
Feb 2024	7	16	0	23
Mar 2024	6	18	1	25
Apr 2024	7	4	1	12
May 2024	3	9	1	13
Jun 2024	6	3	2	11
Jul 2024	7	5	3	15
Aug 2024	10	5	0	15
Sep 2024	2	3	0	5
Oct 2024	3	1	0	4
Nov 2024	2	2	0	4

Viewed (geographical distribution)

Total article views: 634 (including HTML, PDF, and XML) Thereof 634 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 22 Nov 2024

Short summary

The recent advancement in machine learning models enabled scientists to use powerful tools in various applications. In the case of having similar molecules, it is very hard for lasers to distinguish between them; thus we used a special machine learning method to solve this problem. We have developed a highly sensitive laser sensor which can differentiate between very similar chemicals using machine learning.


Total:	0
HTML:	0
PDF:	0
XML:	0