Photoacoustic Multicomponent Gas Analysis Using A Levenberg-Marquardt Fitting Algorithm
Photoacoustic Multicomponent Gas Analysis Using A Levenberg-Marquardt Fitting Algorithm
Photoacoustic Multicomponent Gas Analysis Using A Levenberg-Marquardt Fitting Algorithm
Abstract. A mobile CO2 laser photoacoustic spectrometer radiation, which are both influenced by the amount of wa-
has been developed for the in-situ monitoring of atmospheric ter vapor and CO2 present in the gas sample. This is caused
trace gases in different environments. Numerous air pollu- by resonant energy transfer processes between vibrational en-
tants, such as ammonia, ethene, ozone and alcohols, can be ergy levels of excited molecules, the most important being
monitored with a time resolution of a few minutes. A new fit- the energy transfer between the first mode of the asymmetric
ting algorithm based on the numerical method of Levenberg– stretch vibration of CO2 and a nearby energy level of nitro-
Marquardt is discussed and applied to the derivation of indi- gen, the prevalent compound in ambient air.
vidual gas concentrations from measured photoacoustic sig- These effects were first described by Wood et al. [2] and
nal amplitudes and phases at selected CO2 laser wavelengths. are referred to in literature as kinetic cooling effects, since
The algorithm has been tested with artificially generated mul- they result in an initial cooling of the absorbing gas sample. In
ticomponent gas mixtures exhibiting gas concentrations in the previous analyses of multicomponent air samples containing
ppb to ppm range. Furthermore, its potential is demonstrated water vapor and CO2 we have used the in-phase part of the PA
with the analysis of an air sample from a fruit storage cham- signals in a weighted linear least squares fit of the measured
ber and with ambient air measurements during a field study in spectra [3]. Several difficulties are, however, imposed by this
a rural environment. approach: since only the in-phase part of the signal is used
for the analysis, half of the information entailed in the meas-
PACS: 82.80.Kq; 02.60.+y; 89.60.+x urement is lost in the calculations. In practice, an additional
phase offset resulting from the experiment can be superposed
to the theoretical photoacoustic phase. Since this parameter
In photoacoustic (PA) spectroscopy the radiation absorbed can vary with the measurement conditions it is difficult to
by gas molecules in an absorption cell is detected purely determine in advance. This problem has been accounted for
acoustically. Molecular energy transfer processes convert en- by restricting the analysis to phase lags of π in the case of
ergy stored in excited vibrational modes of the molecules dominant CO2 absorption and of 0◦ in other situations and
into acoustic waves that can be detected with microphones. by determining the phase offset manually [4]. Particularly at
The CO2 laser is a suitable radiation source, since it pro- larger CO2 concentrations, which typically occur in polluted
vides powerful radiation in the wavelength range between 9 ambient air, the photoacoustic phase can, however, assume
and 11 µm in which numerous molecules exhibit characteris- values between 0◦ and π. This is revealed by the examples of
tic strong absorption bands. By modulating the exciting laser multicomponent analyses presented below.
beam at an acoustic resonance frequency of the absorption Rooth et al. [5] have added ammonia as a further compon-
cell a high sensitivity can be achieved that allows the detec- ent to the idealized atmosphere in Wood’s model and calcu-
tion of absorption coefficients as low as 10−8 cm−1 [1]. This lated the resonant PA signal amplitude and phase at selected
technique thus allows the monitoring of numerous organic CO2 laser wavelengths. They considered an atmosphere that,
and inorganic compounds at ppb concentrations. One prob- besides oxygen and nitrogen, contained only CO2 , water va-
lem associated with resonant PA spectroscopy at CO2 laser por and ammonia, and they derived ambient ammonia con-
wavelengths concerns the amplitude of the detected acoustic centrations during a field study in the Netherlands. We have
signal and its phase with respect to the modulated exciting extended this model and developed a fitting algorithm with
the numerical method of Levenberg–Marquardt that derives
∗ Present address: ARITRON INSTRUMENTE AG, CH-8127 Forch, unknown gas concentrations not only for ammonia but simul-
Switzerland taneously for a number of trace components from the meas-
450
and
y
θ = arctan + θ0 . (12)
x
The cartesian components are
I0 τs αi ci + αH cH + αC cC
px = ωτs (13)
1 + (ωτs) 2 ωτs
ωτN ωτN
− βαC cC +1
1 + (ωτN ) 2 ωτs
and
I0 τs
py = ωτs − (αi ci + αH cH + αC cC ) (14)
1 + (ωτs)2
ωτN 1
+ βαC cC + ωτN .
1 + (ωτN )2 ωτs
The characteristic time τs for heat losses of the gas sample
is given by [6]
1 1 1
= + , (15)
τs τT τL
where τT denotes the thermal relaxation time of the gas at the
cell walls and τL is the characteristic time for gas exchange
from the cell to adjacent volumina. In practice τs is of the
order of 0.1 s to 1 s. For sufficiently high chopper frequen- Fig. 1. Photoacoustic signal amplitude R and phase θ at the main ethene
cies it can therefore be assumed that ωτs 1. This permits (C2 H4 ) absorption line 10P(14) of the 12 C 16 O2 laser of a gas sam-
the following simplifications ple containing 100 ppb ethene and variable CO2 and H2 O concentra-
tions at atmospheric pressure. Data calculated for a modulation frequency
f = ω/2π = 2860 Hz
1
≈0 (16)
ωτs
and
signal amplitude decreases due to the kinetic cooling ef-
1 + (ωτs )2 ≈ (ωτs )2 . (17) fect [2], passes a minimum at CO2 concentrations between
1000 and 2000 ppm and finally increases when CO2 absorp-
Introduction of these terms into (13) and (14) finally yields tion becomes dominant. Close to the minimum of the signal
the expressions for the measured resonant photoacoustic am- amplitude the signal phase rapidly changes and increases by
plitude R and phase θ: 180◦ . The phase depicted in Fig. 1 has been inverted for com-
" parison with the ethene, CO2 and water vapor measurement
I0 (αi ci + αH cH + αC cC )2 discussed further below. The inverse phase in the experiment
R= A
ω 1 + (ωτN )2 compared to the data calcutated with (20) is caused by the
#1/2 detection electronics.
(ωτN )2 αi ci + αH cH + (1 − β)αCcC ]2
+ (18)
1 + (ωτN)2
2 The Levenberg–Marquardt algorithm
αi ci + αH cH + αC cC 1 + (ωτN)2
θ = arctan − ωτN + θ0 .
βαC cC ωτN
2.1 Introduction
(19)
The proportionality factor A in (18) corresponds to the cell The Levenberg–Marquardt algorithm is a standard method for
constant of the photoacoustic cell that is determined by cali- determining the parameters of nonlinear functions in least-
bration measurements. It depends on several parameters such squares problems. A brief summary of its main principles is
as the geometry of the photoacoustic cell, its quality factor, given in the following. It is to our knowledge for the first time
the shape of the exciting laser beam and the pressure of the that this algorithm is applied to determine gas concentrations
gas sample. Figure 1 illustrates the variations of the photo- from resonant photoacoustic measurements of multicompo-
acoustic amplitude and phase at the main ethene absorption nent gas mixtures. This nonlinear approach takes also the
line 10P(14) in a gas sample containing 100 ppb ethene and quadratic dependence of the water vapor continuum absorp-
variable CO2 and water vapor concentrations. At constant ab- tion on its concentration into account [8]. An extensive dis-
solute humidity and with increasing CO2 concentration the cussion of the algorithm’s features and guidance for its prac-
452
tical implementation into computer programs can be found in and (20) and the parameters to be determined are the concen-
the literature [9, 10]. trations of the individual components of a gas mixture. The
index j represents the measurement at a specific laser line.
The known parameters xj are determined by the known ab-
2.2 Mathematical principles sorption coefficients of the individual components and further
known data such as the chopper frequency, the laser power,
In a least-squares regression problem with nonlinear model the air pressure, the molecular relaxation times presented in
functions yij ≡ yi (xj , a), i = 1, . . . , n, j = 1, . . . , m, of Table 1 and the energy ratio β.
known parameters xj and unknown parameters a, the set With the definitions
of parameters a ≡ {ak }, k = 1, . . . , p, should be determined τN
which minimizes the merit function x 0 ≡ βαC cC , a0 ≡ A ,
1 + (ωτN)2
X yijmeas − yi (xj , a) 2 1
χ 2 (a) ≡ . (20) x 1 ≡ x 0 ωτN − (αH cH + αC cC ) , a1 ≡ θ0 ,
∆ij ω
i, j 1
x n+1 ≡ αn (n = 2, . . . , m) , an+1 ≡ cn (n = 2, . . . , m) ,
ω
Here yijmeas denotes the experimentally determined value at (25)
xj and ∆ij denotes its standard deviation. At a minimum
of χ 2 (a), the gradient of χ 2 (a) is necessarily equal to zero. (18) and (19) can be rewritten as
Close to a minimum of the merit function, χ 2 (a) can be ap- !2 1/2
proximated by a quadratic form. It is X
n+1
R ≡ y1 (xj , a) = a0 x 02 + x 1 − ai x i (26)
1 i=2
χ 2 (a) ≈ γ − d · a + aDa (21)
2 and
with a suitable constant γ and a vector d of the dimension p.
P
n+1
D indicates the Hessian matrix of χ 2 (a) with the elements x1 − ai x i
X 1 ∂yi (xj , a) ∂yi (xj , a)
i=2
θ ≡ y2 (xj , a) = arctan + a1 . (27)
∂ 2 χ 2 (a) −x 0
Dkl ≡ =2
∂ak ∂al i=1,... ,n
∆2ij ∂ak ∂al
j=1,... ,m
The derivates of the two model functions are given by the
∂ 2 yi (xj , a) Jacobian matrix
− yijmeas − yi (xj , a) . (22)
∂ak ∂al
∂y1 ∂y1 ∂y1
··· ···
In order to simplify the calculations somewhat, the second dy ∂a0 ∂a1 ∂ak k≥2
= (28)
derivatives in the last term of the above equation are neglected da ∂y2 ∂y2 · · · ∂y2 ···
in the following [9]. ∂a0 ∂a1 ∂ak k≥2
From (21) the set of parameters a ≡ {ak }, k = 1, . . . , p,
which minimizes χ 2 (a) can be calculated in a single step. It with
is given by !2 1/2
∂y1 X
n+1
By introducing the above relations into (22) to (24), the merit controllers that can be operated individually or in a master–
function χ 2 (a) can be minimized. The scaling factor A as slave mode. Two mass flow controllers with a full scale range
well as the phase offset θ0 and the gas concentrations of the of 2 l/min, three with 10 cm3 /min, one with 100 cm3 /min,
individual components of the gas mixtures can then be deter- one with 1 cm3 /min for the reference gas nitrogen and fi-
mined with the Levenberg–Marquardt algorithm. nally one with a full-scale range of 10 cm3 /min for pure
CO2 enabled us to dilute certified gases from bottles. These
contained the individual compounds either in pure form or
3 Experimental section diluted in synthetic air at concentrations of 100 ppm. Due
to its strong chemical reactivity, ozone was generated sep-
3.1 The mobile PA spectrometer arately from pure synthetic air with a calibrated generator
(Environics, model S100) and was consecutively added to
Our PA spectrometer has been designed for the monitoring the gas mixture. The actual concentrations were calculated
of air pollutants at various measuring sites and is therefore from the mixing ratios of the gases and compared to the pho-
installed in a mobile trailer. A commercially available line- toacoustically derived concentrations. Since water vapor and
tunable cw CO2 laser provides radiation at approximately CO2 concentrations can easily be monitored with commer-
80 wavelengths between 9 and 11 µm. The laser beam is cial equipment, the concentrations of these two compounds
intensity-modulated by means of a mechanical chopper at were recorded in addition with a dew point mirror, a capac-
a frequency corresponding to the strongest acoustic reson- itive humidity meter and an infrared absorption monitor for
ance of the PA absorption cell at approximately 2860 Hz with CO2 (LI-COR, model LI-6262). The latter device operated
a Q value of 168. The small resonance width of 17 Hz re- at the strong CO2 absorption band at 4.26 µm and provided
quires accurate control of the resonance frequency. The cell a detection limit of a few ppm CO2 .
construction is based on the original design by Gerlach and
Amer [11], extended by two buffer volumes to further reduce
the noise induced by flowing gas when operated in the flow- 4 Results and discussion
through mode. It has a volume of 1.8 l and is usually operated
at flow rates between 0.5 and 2 l/min. The acoustic signal First we present two examples from a number of artificially
generated by the absorption of the gas sample is recorded by generated multicomponent gas mixtures with time-varying
two electret microphones and a lock-in amplifier. The sig- concentration profiles, followed by examples of an analysis of
nals are normalized with the laser power to account for power sampled air and a field study.
fluctuations in independent measurements. The system can
be operated fully automaticly and exhibits detection limits in
the ppb range for various compounds. A detailed technical 4.1 Three-component mixture containing ethene (C2 H4 ),
description of the mobile spectrometer and of its calibration water vapor and CO2
with gases of certified concentrations and known absorption
cross sections reported by various authors, can be found else- Figure 3 presents the analysis of a three-component mix-
where [4, 12]. ture with ethene, water vapor and CO2 . The concentration of
CO2 was increased stepwise from 0 ppm up to a final con-
centration of 4000 ppm while the absolute humidity and the
3.2 Generation of multicomponent mixtures ethene concentrations were kept at constant levels of 1.3%
and 100 ppb, respectively. Figure 3 shows the variation of the
Multicomponent mixtures containing water vapor and CO2 photoacoustic signal amplitude and phase with time at three
at typical ambient concentrations have been generated in laser transitions. The 10P(14) transition is characteristic for
order to test the fit algorithm. The experimental set-up for ethene absorption and the 10R(20) transition for water vapor
the generation of these gas mixtures is plotted in Fig. 2. The line absorption, while none of them show appreciable absorp-
commercial gas mixing unit (MKS, model 147C) consists tion at the 10P(12) transition. With increasing CO2 concen-
of a computerized control unit and up to eight mass flow tration the signal amplitudes at the 10P(14) and the 10R(20)
500
case of Fig. 1, the minimum signal amplitude is passed at CO2
PA amplitude concentrations slightly above 1000 ppm simultaneously with
signal amplitude [mV/W]
10P(12) measurement
400
10P(12) fit a rapid increase of the signal phase. The measured amplitudes
10P(14) measurement and phases are well approached by the Levenberg–Marquardt
300
10P(14) fit algorithm. In addition to the plotted data measurements at
10R(20) measurement the 10P(16) and the 10P(20) laser lines have been performed.
200
10R(20) fit
In agreement with the Levenberg–Marquardt fit, these show
100
a similar time dependence as the depicted data. They have,
however, been omitted for clarity in Fig. 3. The resulting
0 ethene concentration was derived by fitting the signal am-
120 plitude and phase at all five laser lines with equal weights.
90 PA phase 10P(12) measurement
As the bottom part of Fig. 3 shows, the ethene concentration
60 10P(12) fit remains at the expected value of 100 ppb according to the set-
signal phase [°]
30 10P(14) measurement ting of the gas mixing unit with slow variations on the order
0 10P(14) fit of 25 ppb. Variations can also be observed in the concentra-
-30 10R(20) measurement tions of the other two compounds. They are caused by thermal
-60
10R(20) fit drifts of the mass flow controller zero points due to tempera-
-90
ture fluctuations during the measurement. The CO2 and H2 O
-120
concentrations depicted in Fig. 3 have been obtained from
-150
the LI-COR infrared absorption monitor and the dew point
200
0 60 120 180 240 300 360 420 480 mirror, respectively. The data were used to determine intial
estimates for the fitting procedure.
150 ethene [ppb]
setting
100 4.2 Four-component mixture containing ethene (C2 H4 ),
methanol (CH3 OH), water vapor and CO2
50
100
9P(34) fit 10 4
0
150
120
PA phase 10P(16) measurement 10 3
90 10P(16) fit
180
signal phase [°]
60 10P(14) measurement
30 10P(14) fit measurement
0 10R(20) measurement 120 PA phase
fit
-180
50 900 950 1000 1050 1100
wavenumber [cm-1 ]
0
300 Fig. 5. Photoacoustic absorption spectrum of an air sample from a fruit stor-
250 age chamber. The measured photoacoustic amplitude and phase and the
200 ethene [ppb] data obtained with the Levenberg–Marquardt algorithm are plotted. The
gas concentrations
150 setting corresponding gas concentrations resulting from the fit are given in Table 2
100
50
0
3000
to the supposition that besides water vapor, CO2 , ethene and
2500
Institute for Forest, Snow and Landscape Research (WSL Figure 6 depicts the concentration profiles of the three
Birmensdorf/Zurich) in a mountain valley far from primary gases at the measuring site between July 5 and July 24, 1996.
air pollution sources. Between July 5 and July 24, 1996, the In comparison to the photoacoustic measurements, the ozone
mobile PA system was operated at the unattended meteoro- concentrations derived with the commercial ozone analyser
logical monitoring site “Erlenhöhe” at an altitude of 1200 m of the WSL Birmensdorf and the absolute humidity calcu-
above sea level. The monitoring site is situated on a slightly lated from the data of a capacitive humidity meter are plotted.
descending slope in a lightly wooded area. Neither agricul- The CO2 concentrations in Fig. 6 have been obtained from the
tural activities nor any traffic were present in the immedi- near-infrared gas sensor (Gas Card) of the mobile system. The
ate environment. Harsh meteorological conditions during the photoacoustic ozone concentrations show the expected varia-
measurement period and large diurnal temperature variations tions in the daytime in correlation to the data obtained with
with maximum temperatures exceeding 35 ◦ C and minimum the commercial monitor. Average concentrations are on the
temperatures close to 7 ◦ C posed a challenge to the perform- order of 40 ppb during the measuring period while maximum
ance of the automatic system, particularly in view of the fact
that the air conditioning in the trailer had to be switched ozone analyser WSL
off due to the low electrical power available at that site. mobile system
500
measurements at a nearby site revealed yearly average NH3
concentrations of only 0.20 ± 0.17 µg/m3 (≈ 0.3 ppb) [22],
which is also below the detection limit of 2 ppb of our mobile 400
that exhibits the strongest ozone absorption at 12 C 16 O2 laser Fig. 6. Ozone, water vapor and CO2 concentrations at the meteorological
frequencies. Additionally the 10R(18), 10R(20) and 10R(22), monitoring site Erlenhöhe (Brunni, Switzerland) between July 5 and July
as well as the 10P(18), 10P(20) and 10P(22) transitions have 24, 1996. The concentration profiles derived from the photoacoustic spec-
tra of ozone and of the absolute humidity are compared with the data from
been included since they are representative for H2 O and CO2 a commercial ozone analyser of the Swiss Federal Institute for Forest, Snow
absorption. Both signal amplitude and phase were fitted with and Landscape Research (WSL Birmensdorf) and with the absolute humid-
the same statistical weight. ity calculated from the data of a capacitive humidity meter
457
concentrations reach up to approximately 80 ppb. In contrast with a line-tunable CO2 laser. In contrast to previous ana-
to the data from the commercial monitor the photoacoustic lyses this algorithm calculates signal amplitudes and phases
concentrations come close to zero on July 16. This could separately, which extends the potential of the analysis sub-
most probably be ascribed to erroneous results of the fitting stantially, particularly with respect to multicomponent mix-
procedure when the ozone concentrations come close to our tures which often cause interfering spectra. Without any
detection limit of 13 ppb. The natural background concentra- pre-treatment air samples of different origin can thus be
tion of ozone in clean tropospheric air exceeds 20 µg/m3 , i.e. analyzed.
10 ppb [23]. The ambient ozone concentration is increased With a commercial gas mixing device the analysis proced-
above this level by ozone originating from local precursors at ure was tested for the first time with multicomponent mix-
distances between 10 and 50 km and by contributions from tures containing time-varying gas concentrations. The mix-
higher atmospheric layers where ozone concentrations re- tures were prepared with water vapor and CO2 at ambient
main at rather constant levels even in periods of bad weather. concentrations and additional compounds in the ppb range.
Ambient ozone concentrations below 10 ppb are therefore not Excellent agreement between measured photoacoustic ampli-
very probable. The mismatch between the concentrations de- tude and phase spectra and the data calculated from refer-
termined photoacoustically and the data from the commercial ence spectra of the individual compounds was obtained. The
system between July 22 and July 24 can be ascribed to ther- derived gas concentrations are in good agreement with the
mal detuning of laser transitions in the 9P laser branch. This expected values according to the settings of the gas mixing
resulted in larger measurement errors of the recorded spectra. unit. Deviations are on the order of 10% to 20%. Furthermore,
It is very unlikely that such deviations are caused by ozone re- the algorithm was applied to the analysis of an air sample
actions in the gas tubings because similar tube lengths have of unknown composition from an apple storage chamber. Ex-
been used for the calibration measurements and the field stud- cept for a few laser lines, excellent agreement between the
ies and, furthermore, the deviations occur in both directions. measured and the calculated data was obtained by including
The absolute humidity determined from the photoacoustic reference spectra of water vapor, CO2 , ethene and ethanol
data is in good agreement with the conventionally measured vapor into the fitting procedure. The resulting gas concentra-
concentration. It shows diurnal variations with amplitudes tions correspond to the expected levels according to earlier
on the order of 0.75% and an additional slow modulation investigations.
with time caused by the changing meteorological conditions. In a field campaign our mobile CO2 laser photoacoustic
The CO2 concentrations exhibit diurnal variations around system was operated unattended at a biological monitoring
350 ppm with maximum values after midnight on the order of site in a rural mountainous area at an elevation of 1200 m
400 ppm and minimum concentrations around 330 ppm in the above sea level. The ambient air was flown continuously
afternoons. No comparative data were available for the CO2 through the gas cell. Photoacoustic spectra were recorded
measurements depicted in Fig. 6. However, the CO2 data were over 20 days and analyzed after the campaign. Water vapor,
used to derive the ozone concentrations from the photoacous- CO2 and ozone could easily be identified from the spectra
tic spectra. The good agreement between the measured and and their diurnal variations were derived, but no evidence was
the calculated photoacoustic spectra sustain the assumption found for additional absorbing compounds. Independent ana-
that the correct CO2 concentrations were determined with lyses of air samples showed that the concentrations of volatile
the near infrared gas analyser. Diurnal and seasonal changes organic compounds were in fact below the detection limit of
in atmospheric CO2 concentrations in different environments our system. In contrast to laboratory measurements on gas
have recently been investigated by various authors [24–26]. mixtures containing only known compounds, field measure-
In these studies, diurnal variations of the CO2 concentra- ments have larger errors in derived trace concentrations in
tion with enhanced concentrations during nighttime, mini- ambient air due to potential absorption interferences of un-
mum concentrations in the afternoon and differences between known species.
maximum and minimum concentrations on a single day in With regard to future field measurements at sites with
summer of 50 up to 95 ppm were observed. The enormous low ambient pollutant concentrations, emphasis is thus put on
diurnal variation of CO2 is ascribed to the daily cycle of lowering these detection limits, e.g., by implementing a mul-
photosynthesis and respiration by the surrounding vegeta- tipass photoacoustic cell. On the other hand, an improvement
tion superimposed on the day–night contrast of atmospheric in detection selectivity, which is particularly important in
mixing [25]. These results agree well with our CO2 measure- highly polluted areas, is envisaged by increasing the number
ments. It should be noted that the diurnal CO2 variation is of operational wavelengths, e.g., by the simultaneous oper-
less pronounced during the first six days of the monitoring ation of two different laser isotopes. Ideally, a narrow-band
period, which were characterized by rather bad weather with continuously tunable laser with a broad wavelength range in
low solar radiation densities and hence reduced photosynthe- the mid-IR is employed and great efforts are currently being
sis activity. Our CO2 data are also supported by a previous put into the development of such sources that are suitable also
study performed in a rural area with enhanced CO2 concen- for field measurements.
trations during night time [1, 4].
Acknowledgements. We are grateful to the help provided by the Swiss Fed-
eral Research Station for Fruit-Growing, Viticulture and Horticulture in
5 Conclusion Wädenswil/Zurich and for the support and the comparative ozone data ob-
tained from the Swiss Federal Institute for Forest, Snow and Landscape
Research (WSL Birmensdorf/Zurich). We owe special thanks to Victor
On the basis of the numerical method of Levenberg–Marquardt Stadelmann (PSI Würenlingen) for the gas chromatography analysis of two
we developed a new fitting algorithm for the analysis of air samples. This project is supported by the Swiss National Science Foun-
photoacoustic spectra of multicomponent gas samples taken dation and the ETH Zurich.
458