Anal. Chem.

2000, 72, 206-215

Precision Trace Gas Analysis by FT-IR

Spectroscopy. 1. Simultaneous Analysis of CO2,
CH4, N2O, and CO in Air
Michael B. Esler,*,† David W. T. Griffith,† Stephen R. Wilson,† and L. Paul Steele‡

Department of Chemistry, University of Wollongong, Wollongong NSW 2522, Australia, and CSIRO Atmospheric Research,
Aspendale Vic 3195, Australia

We report the development of a method of trace gas et al.3 still predominates4 for background atmospheric CO2 mon-
analysis based on 1-cm-1 resolution Fourier transform itoring. This method offers a short-term precision (one standard
infrared (FT-IR) spectroscopy, deployable in both labora- deviation) as good as (0.01 in 360 µmol mol-1 (mole fraction
tory and field applications. Carbon dioxide, methane, mixing ratio) or (0.003%, for an hourly mean measurement. Alter-
nitrous oxide, and carbon monoxide may be analyzed natively, CO2 can be separated by gas chromatography (GC), con-
simultaneously in a single air sample using this method. verted to methane over a catalyst, and detected with a flame ioni-
We have demonstrated that the method can provide zation detector (FID). This GC-methanizer-FID approach can offer
analytical precision of the order of (0.15 µmol mol-1 for a short-term precision level of (0.02%.5 The other three gases
CO2, (0.9 nmol mol-1 for CH4, (0.3 nmol mol-1 for N2O, mentioned are usually analyzed using gas chromatography. Meth-
and (0.3 nmol mol-1 for CO, expressed as mole fractions ane in air lends itself readily to analysis by GC-FID as described
in dry air. The analytical precision is in all cases competi- by Steele et al.,6 where a precision level of (2.5 nmol mol-1 in
tive with or superior to that of the more usual methods of atmospheric background levels of 1720 nmol mol-1 CH4, or
analysis for these trace gases, namely, nondispersive (0.15%, is attainable.5,7 For N2O in air, GC with electron capture
infrared spectroscopy for CO2 and gas chromatography- detection (ECD) remains the standard technique. Precision of the
based techniques for CH4, N2O, and CO. The novel FT- order of (0.3 in 300 nmol mol-1 N2O, or (0.1%, is attainable.5,8
IR method relies on calibration using synthetically calcu- Analysis of CO at background atmospheric levels is usually done
lated absorbance spectra and a chemometric multivariate by GC separation followed by the reduction of hot HgO to Hg by
calibration algorithm, classical least squares. CO, the Hg vapor being measured by UV absorption.5,9,10 This
ensemble is known as a mercuric oxide reduction gas detector
The three most important anthropogenically influenced green- (RGD). Precision of ∼(1% can be obtained at 100 nmol mol-1
house gases are carbon dioxide, methane, and nitrous oxide.1 CO.
While not directly a greenhouse gas, carbon monoxide also plays Using the orthodox analytical techniques just described to
an important role in atmospheric chemistry as one of the most measure or routinely monitor the four species CO2, CH4, N2O,
important factors determining the concentration of atmospheric and CO, simultaneously at a single site currently requires an
oxidants.2 Measurements of the spatial distribution and temporal ensemble of four separate instruments: NDIR (or GC-methanizer-
trends of these species in the atmosphere have played, and will (3) Komhyr, W. D.; Harris, T. B.; Waterman, L. S.; Chin, J. F. S.; Thoning, K.
continue to play, a key role in the elucidation of their atmospheric W. J. Geophys. Res. 1989, 94, 8533-8547.
budgets. New measurement technologies have the potential to (4) Steele, L. P.; Beardsmore, D. J.; Pearman, G. I.; Da Costa, G. A. In Baseline
Atmospheric Program Australia 1994-95; Francey, R. J., Dick, A. L., Derek,
offer improvements over current strategies, as well as providing N., Eds.; Bureau of Meteorology and CSIRO Division of Atmospheric
a means of independently asessing well-established technologies. Research: Melbourne, 1996; p 103.
There is a high degree of consensus on the choice of instru- (5) Francey, R. J.; Steele, L. P.; Langenfelds, R. L.; Lucarelli, M. P.; Allison, C.
E.; Beardsmore, D. J.; Coram, S. A.; Derek, N.; de Silva, F. R.; Etheridge,
mental techniques employed for point sampling in situ measure- D. M.; Fraser, P. J.; Henry, R. J.; Turner, B.; Welch, E. D.; Spencer, D. A.;
ment of atmospheric trace gases. The nondispersive infrared Cooper, L. N. In Baseline Atmospheric Program (Australia) 1993; Francey,
(NDIR) spectroscopy method described by, for example, Komhyr R. J., Dick, A. L., Derek, N., Eds.; Bureau of Meterology and CSIRO Division
of Atmospheric Research: Melbourne, 1996; pp 8-29.
* Corresponding author: (e-mail) mesler@uow.edu.au. (fax) +61-2-4221-4287. (6) Steele, L. P.; Fraser, P. J.; Rasmussen, R. A.; Khalil, M. A. K.; Conway, T. J.;
†
University of Wollongong. Crawford, A. J.; Gammon, R. H.; Masarie, K. A.; Thoning, K. W. J. Atmos.
‡ CSIRO Atmospheric Research.
Chem. 1987, 5, 125-171.
(1) Schimel, D.; Alves, D.; Enting, I.; Heimann, M.; Joos, F.; Rayneud, D.; Wigley, (7) Dlugokencky, E. J.; Steele, L. P.; Lang, P. M.; Masarie, K. A. J. Geophys.
T.; Prather, M.; Derwent, R.; Ehhalt, D.; Fraser, P.; Sanhueza, E.; Zhou, X.; Res. 1995, 100, 23103-23113.
Jonas, P.; Charlson, R.; Rodhe, H.; Sadasivan, S.; Shine, K. P.; Fouquart, Y.; (8) Prinn, R.; Cunnold, D.; Rasmussen, R.; Simmonds, P.; Alyea, F.; Crawford,
Ramaswamy, V.; Solomon, S.; Srinivasan, J.; Albritton, D.; Derwent, R.; A.; Fraser, P.; Rosen, R. J. Geophys. Res. 1990, 95, 18369-18385.
Isaksen, I.; Wuebbles, D. In Climate Change 1995: The Science of Climate (9) Crill, P. M.; Butler, J. H.; Cooper, D. J.; Novelli, P. C. In Biogenic Trace
Change (Intergovernmental Panel on Climate Change); Haughton, J. T., Meira Gases: Measuring Emissions from Soil and Water; Matson, P.A., Harriss, R.
Filho, L. G., Callander, B. A., Harris, N., Kattenberg, A., Maskell, K., Eds.; C., Eds.; Blackwell: Oxford, 1995; pp 164-205.
Cambridge University Press: Cambridge, 1996; pp 65-131. (10) Novelli, P. C.; Elkins, J. W.; Steele, L. P. J. Geophys. Res. 1991, 96, 13109-
(2) Thompson, A. M. Science 1992, 256, 1157-1165. 13121.

206 Analytical Chemistry, Vol. 72, No. 1, January 1, 2000 10.1021/ac9905625 CCC: $19.00 © 2000 American Chemical Society
Published on Web 11/30/1999
Figure 1. Simple schematic of the instrument.

FID), GC-FID, GC-ECD, and GC-RGD. In turn, it is necessary to

characterize four separate instrument response functions and
implement four calibration protocols. We describe here a single
FT-IR spectrometer configured to provide the same analyses,
simultaneously and with the same level of precision. In the
following paper, we describe the application of FT-IR spectroscopy
to the determination of the stable isotope ratio 13C/12C in
atmospheric CO2.11

EXPERIMENTAL SECTION
The configuration of the instrument used in this work is
illustrated schematically in Figure 1. Briefly, a commercially
available Bomem MB100 FT-IR spectrometer, with 1-cm-1 maxi-
mum resolution, globar source, KBr beam splitter, and a liquid
nitrogen-cooled InSb infrared detector, was used to analyze air
samples introduced to a multipass (White) cell. The instrument
and White cell were enclosed in a sealed, thermostated box which
was purged with dry nitrogen.
This section also considers the main experimental factors in
obtaining good-quality spectra: namely, spectrum signal-to-noise
ratio, detector linearity, Beer-Lambert law linearity, temperature
and pressure control, instrument environment and sample pro-
Figure 2. Regions of an infrared absorbance spectrum of air
cessing. showing (a) 13CO2, N2O, and CO; (b) 12CO2, 13CO2, and N2O; and
Spectrum Signal-to-Noise Ratio. Spectra were collected at (c) CH4 absorbance features.
the spectrometer’s maximum resolution, 1 cm-1. Of the four
molecular species of interest in this work, some of the rotational (SNR) was the limiting factor on the analytical precision achieved,
structure of three of them is resolved at this resolution in the while for CO2 it was departure from Beer-Lambert law linearity.
2000-3000-cm-1 region, as illustrated in Figure 2. The rotational The White cell serves to increase the absorbance relative to
transitions are separated by less than 1 cm-1 and remain noise in the absorbance spectrum by increasing the path length
unresolved only for N2O. (The approximate rotational separations without proportional loss of signal.12-14 Two different White cells
in the most intense absorption features are as follows: CO2 ∼1.6 were used as described in Table 1. Although the 22-m cell provides
cm-1, CH4 ∼9.0 cm-1, CO ∼3.8 cm-1, and N2O ∼0.8 cm-1). For the best result in terms of absorbance relative to noise, consid-
the three less abundant trace gases, CH4, CO, and N2O, with the eration of the scarcity and expense of calibration gases will favor
correspondingly weak absorption features, signal-to-noise ratio
(12) White, J. U. J. Opt. Soc. Am. 1942, 32, 285-288.
(11) Esler, M. B.; Griffith, D. W. T.; Wilson, S. R.; Steele, L. P. Anal. Chem. 1999, (13) White, J. U., J. Opt. Soc. Am. 1976, 66, 411-416.
71, 216-221 (this issue). (14) Horn, D.; Pimentel, G. C. Appl. Opt. 1971, 10, 1892-1898.

Analytical Chemistry, Vol. 72, No. 1, January 1, 2000 207

Table 1. Specifications of the Two Multipass White
Cells Used

22 m cell 9.8 m cell

base path length, m 1.1 0.25
total path length, m 22.1 9.8
no. of passes 20 40
volume, L 8 0.5
transmitted energy, % ∼20 ∼12

the smaller volume 9.8-m cell for many applications, particularly

for CO2 analysis where precision is not limited predominantly by
SNR. This cell has the added advantages of being quicker to fill
and pump out and, being smaller, is more easily incorporated in Figure 3. Absorbance spectra of wet (top) and dried (bottom) air,
a field-deployable instrument than the 22-m cell. at 740 Torr and 305 K. The wet air contains ∼1% H2O, and the dried
Detector Noise and Linearity. Detector noise was minimized air contains <10 µmol mol-1 H2O.
with the use of a liquid nitrogen-cooled InSb detector, which had
a specific peak detectivity (D*) of 1.9 × 1011 cm Hz1/2 W-1 and is apodization functions across a range of possible experimental
ideally suited to the 2000-3000-cm-1 region of the spectrum where configurations. Below a certain threshold, absorbance will vary
CO2, CH4, CO, and N2O have strong absorbance features. A band- linearly with concentration. In the earlier stages of our work, all
pass filter (OCLI, serial number W03999-4) to select only the interferograms were processed using a triangular apodizing
region containing the main CO2, CH4, CO, and N2O transitions, function. Later, a Happ-Genzel apodizing function was employed.
i.e., 2000-3200 cm-1, was used to avoid detector saturation. The latter allowed improved analytical precision as well as
There is a tradeoff between spectrum averaging time and SNR extending the range of absorbances for which Beer-Lambert law
(see Griffiths and de Haseth15 for a full account of the trading linearity prevailed. In our instrumental configuration, we have
rules in FT-IR spectroscopy). To double the SNR, the averaging found that the strength of the peak absorbance features due to
time (or the number of coadded scans, in effect) must be CO, N2O, CH4, and 13CO2 were well within this region of Beer-
quadrupled. For averaging times of up to ∼8 min (∼256 scans), Lambert linearity. Some of the 12CO2 lines used for analysis were
the quality of the spectral fit (and hence the precision of the somewhat stronger and potentially nonlinear. However, we have
technique as measured by SEP) appears to be limited by high- found experimentally that the degree of nonlinearity was not a
frequency noise. This high-frequency noise is reduced by extend- source of systematic error for measured peak absorbances up to
ing the coaddition time. Low-frequency effects such as drifts in ∼0.4. Thus, regions of the spectrum with absorbance greater than
baseline are well modeled by our classical least-squares (CLS) 0.4 are not used in the calibration and quantitative analysis
approach within this domain. However, beyond ∼8 min, the quality strategies described below.
of the spectral fit appears to be limited by features in the spectrum Temperature and Pressure Control. Spectroscopy inher-
that are not so well modeled by the CLS calibration such as more ently measures concentration (mass or mole per volume) for a
complex features in the baseline or residual line shape effects. constant path length. To convert concentration to mixing ratio
So coadding for longer than 8 min brings no further gains in requires the measurement of the sample density, achieved by
precision. In any case, an 8-min or less measurement is practical measuring temperature and pressure. The sample pressure was
for many applications. When the SNR-sensitive CO, CH4, and N2O determined with a capacitance manometer mounted on the White
were the subject of analysis, ∼256 scans, or 8 min, was found to cell. The temperature was measured with platinum resistance
be optimal. This was sufficient to provide single-beam root-mean- temperature detectors (RTDs). The manometer and RTD output
squared (rms) SNR of ∼50 000; equivalently absorbance spectra was monitored with a 16-bit data acquisition card. Sample pressure
with rms noise levels of the order of 10-5 at 2000 cm-1. The was typically in the range 720-760 Torr. The precision of the
experimentally determined relationship between precision and pressure measurement in this range was of the order of (0.1 Torr,
averaging time is illustrated in the section on verification of the or (0.01% at 760 Torr. The precision of the RTDs was better than
method, below. (0.03 K, or (0.01% in 300 K. The pressure and temperature
Beer-Lambert Law, Linearity, Apodization, and Resolu- measurements should not, therefore, limit the precision of the
tion. The Beer-Lambert law may be stated as A ) Rcl, where A method.
is absorbance, R is absorption coefficient, c is concentration, and Instrument Environment. The spectrometer and White cell
l is path length. In FT-IR spectroscopy, there will be a deviation were enclosed in a thermostated box and purged with dried high-
from Beer-Lambert law behavior where the absorption features purity N2 at a rate of 500-1000 mL min-1. Purging removes
are narrower than can be resolved using the available instrumental absorbing species from the part of the optical path lying outside
resolution. Zhu and Griffiths16 have quantified the expected the White cell. Thermostating was required to minimize instru-
deviations from Beer-Lambert linearity for triangular and other ment drift which otherwise becomes limiting to final measurement
precision.
(15) Griffiths, P. R.; de Haseth, J. A. Fourier transform infrared spectrometry; John
Wiley and Sons: New York, 1986. Sample Processing. Figure 3 illustrates the difference
(16) Zhu, C.; Griffiths, P. R. Appl. Spectrosc. 1998, 52, 1403-1408. between the spectra of ambient air, dew point ∼7 °C, containing
208 Analytical Chemistry, Vol. 72, No. 1, January 1, 2000
Figure 4. Flow chart of calibration and prediction process using MALT and CLS.

∼1% H2O, and dried air. Strong H2O absorption features occur in handling and to log temperature and pressure. A 16-bit ADC was
the same region as CH4, N2O, and CO. In principle, the quantitative used to ensure that measurement precision was not limited by
spectrum analysis algorithm employed (described below) is the ADC. In each cycle of the program, single-beam spectra were
capable of resolving individual species contributions to a spectrum collected, saved to disk, and ratioed with reference spectra to
of their mixture, even when the absorption lines obscure and produce absorbance spectra which were then fitted as described
overlap each other seriously, as in Figure 3. In practice, signifi- in the quantitative analysis section below. The parameters of the
cantly greater reproducibility is obtained for CH4, N2O, and CO fit provided the concentration of the trace gas constituents directly.
mixing ratio retrievals if the samples are well dried first. The H2O These are combined with pressure and temperature data to
absorption lines, as well as obscuring the usually much weaker provide mixing ratios.
CH4, N2O, and CO features, are sufficiently strong to behave in a
nonlinear Beer-Lambert manner. This can introduce serious QUANTITATIVE ANALYSIS OF SPECTRUM
“analytical cross-talk” into the fitting procedure, where the This section deals only with how the absorbance spectra of
apparent mixing ratios of CH4, N2O, and CO will be slightly air samples were analyzed. This may be considered as a three-
dependent on the level of H2O present. In ambient air, moisture phase process, as outlined in Figure 4. In the first phase, the
levels are typically in the range 0.5-2.5% and are quite variable multiple atmospheric layer transmission (MALT) program17 is
with time. Accordingly, all samples were dried to <10 µmol mol-1 used to calculate a set of absorbance spectra, A, that closely
H2O (-70 °C dew point) prior to analysis by passing them over simulate the spectra obtained on the actual instrument. In the
anhydrous magnesium perchlorate, a very efficient desiccant. A second phase, this set of spectra, A, and a matrix, C, containing
stainless steel cartridge, volume ∼200 mL, was filled with the trace gas concentration information for the calculated spectra
granulated anhydrous magnesium perchlorate and placed in the are used as the training set input to the multivariate calibration
sampling line. Sample residence time in the dryer is <1 s. One algorithm, in this case CLS.18 In practice, the entire calculated
such charge of desiccant was sufficient to dry ∼1600 L of ambient spectrum is not used for calibration. Rather, an optimized region
air. This desiccant is known not to perturb the concentrations of within the spectrum is chosen for each species. A procedure for
CO2, CH4, N2O, and CO in air.5 optimal spectral region selection is detailed below. In the second
As far as possible, the sampling manifold was constructed of phase, the CLS calibration algorithm produces as its output a
glass, copper, and stainless steel tubing, PTFE (Teflon), and Viton. matrix, K̂, the calibration matrix, which describes the relationship
The system was kept as free as possible of hydrocarbons. The between species concentration and absorbance strength for all
moving components in the solenoid valves were cleaned of species and regions of the spectrum included in the calibration.
lubricants. An activated alumina hydrocarbon trap was fitted to Absorbance is assumed to be proportional to concentration (Beer-
the rotary oil pump to prevent backstreaming of oil vapors to the Lambert law) at all wavenumbers ν̃. The third phase is the
manifold and White cell. prediction CLS step, in a sense the inverse of the calibration CLS
The random error in the analysis due to the inconsistency of step. The calibration matrix, K̂, is used to perform a linear least-
the human operator was effectively removed by automation. A squares fit of a synthetic spectrum to a real experimental
program in the GRAMS-Array Basic (Galactic Industries Inc.) spectrum, As. The parameters of the fit, the vector C, directly give
language was used to automate the FT-IR analysis. The program
(17) Griffith, D. W. T. Appl. Spectrosc. 1996, 50, 59-70.
addressed a data acquisition and control card (Strawberry Tree (18) Haaland, D. M.; Easterling, R. G.; Vopicka, D. A. Appl. Spectrosc. 1985, 39,
Industries, Mini-16) to activate the solenoid valves for sample 73-84.

Analytical Chemistry, Vol. 72, No. 1, January 1, 2000 209

the concentrations of the trace gas components of the real fully elsewhere.17 Normally real experimental spectra are required
sample. as calibration input to the CLS calibration algorithm. For the
Müller et al.19 have recently shown the effects of wavenumber simultaneous analysis of several atmospheric trace gas species
scale instability on FT-IR trace gas measurements when using a in air, these calibration spectra would be obtained by analyzing a
multivariate analysis approach in the absorbance domain. To avoid suite of air samples of known composition. Furthermore, within
these effects, we iteratively fit the wavenumber scale simultaneous this suite it would be necessary to ensure that the concentration
to the CLS process of fitting the absorbance scale. This is done of the different species varied independently from each other and
rapidly in real time by computer and means that any wavenumber covered the range of concentrations anticipated in any unknown
instability will not affect the fitting of spectral absorbances and samples to be analyzed. MALT removes the need for a suite of
hence the retrieved component concentrations. real calibration gases since it calculates simulated spectra which
Normally, the first step, MALT simulation, and the second step, model very closely the real spectra obtained using a given FT-IR
CLS calibration, need to be performed once for a given type of spectrometer. A change in measurement conditions merely
analysis. The resulting calibration matrix K̂ is stored in the requires the rerunning of MALT to update the calibration. What
computer. The CLS prediction step is performed every time a follows is only a brief summary of MALT.
spectrum is fitted to derive its concentrations. The first and second The calculation of synthetic spectra is based on a compilation
steps can be completed in a few minutes. The third step takes of absorption line parameters. One suitable line parameter set is
less than 1 s of computer time and so may be performed on-line HITRAN,21 which includes line parameters for 35 individual
in real time following the collection of each spectrum. atmospheric gases, including the species of interest here. HITRAN
CLS-Multivariate Calibration and Prediction. CLS analysis includes, for each absorption line of each molecule, the line
is one of several chemometric techniques developed in the past frequency, the integrated line strength, and the pressure and
decade or so and is ideally suited to the retrieval of quantitative temperature-dependent Lorentzian half-width. For each absorption
information from spectra. The theory is fully described else- line of each molecule, MALT first calculates the contribution to
where.18 We used a CLS package originating from Galactic the absorption coefficient at each wavenumber. This includes
Industries Inc. which has been significantly customized for this temperature and pressure contributions to the strengths and line
application. shape due to both Doppler and pressure broadening. The
A general disadvantage of the CLS method is that all interfering absorption coefficient is then multiplied by the amount of the
chemical components in the spectral region of interest need to molecular species (i.e., its concentration times the path length)
be known and included in the calibration or the method will give to provide the monochromatic optical depth, τ, at any given
erroneous results. In many analytical situations, this renders CLS wavenumber. The total monochromatic optical depth spectrum
an inappropriate chemometric technique. Often, the conceptually is arrived at by summing over all absorption lines of all molecules.
similar but more complex partial least-squares (PLS) technique This spectrum is equivalently the spectrum measured on a perfect
is applied. The significant advantage of PLS over CLS is that PLS spectrometer of arbitrarily high resolution. However, any real
does not require that all absorbing components of the system be spectrometer convolves the true monochromatic transmission
known, as long as the variance of the training set adequately (transmission ) e-τ) spectrum with an instrument line shape to
represents the variance of the unknown spectra. The disadvantage produce the observed or measured transmission spectrum. MALT
of PLS is that much of the transparency of CLS and the valuable models the contributions to instrument line shape due to instru-
qualitative information it provides is lost. Indeed, Haaland et al.20 ment resolution (finite optical path difference), the apodization
have recently described the use of CLS to shed light on PLS function applied to the interferogram, and to the divergence of
calibrations of solution spectra. In the present work all significant the collimated beam in the spectrometer. When these effects are
absorbing components of the sample are known; they are H2O, convolved with the monochromatic transmission spectrum, the
CO2, CO, and N2O in the 2000-2500-cm-1 region and CH4 near result is the instrumentally degraded spectrum. MALT converts
3000 cm-1. Except for rare contamination events due to non- this into the required output units (for either transmittance or
methane hydrocarbons, the absorbance due to other trace gas absorbance), and it is saved in the standard GRAMS (Galactic
species is less than the noise level of the recorded spectra. This Industries) file format.
enables the system to be fully characterized, i.e., calibrated in all For a set of calibration spectra, the number of spectra required
absorbing species present, as required when the CLS technique and the range of concentrations for each absorber are input to
is employed. In the developmental stage of this project, a parallel MALT and a set of spectra with random concentrations within
study was performed, where both CLS and PLS techniques were the given ranges is calculated. Random noise may also be
applied to analysis of the same samples. The results indicated that optionally included in MALT calculated spectra. Finally, the
there was nothing to be gained by using PLS rather than the calibration set calculation also produces a list file of all calibration
simpler CLS. While it was observed that PLS typically gave better spectra and their species concentrations in a format suitable to
fits to experimental spectra than did CLS, this did not translate define the C matrix for direct input to the CLS software.
into higher analytical precision for trace gas analysis. In operational terms, one full run of MALT is equivalent to
MALT/HITRAN (Working Calibration). MALT, a program the generation of a full set of calibration spectra. Typically 40
for generating synthetic absorbance spectra, has been described
(21) Rothman, L. S.; Rinsland, C. P.; Goldman, A.; Massie, S. T.; Edwards, D. P.;
(19) Müller, U.; Heise, H. M.; Mosebach, H.; Gärtner, A. G.; Hausier, T. Field Flaud, J.-M.; Perrin, A.; Camy-Peyret, C.; Dana, V.; Mandin, J.-Y.; Schroeder,
Anal. Chem. Technol. 1999, 3, 141-159. J.; McCann, A.; Gamache, R. R.; Wattson, R. B.; Yoshino, K.; Chance, K. V.;
(20) Haaland, D. M.; Han, L.; Niemczyk, T. M. Appl. Spectrosc. 1999, 53, 390- Jucks, K. W.; Brown, L. R.; Nemtchinov, V.; Varanasi, P. J. Quant. Spectrosc.
395. Radiat. Transfer 1998, 60 (5), 665-710.

210 Analytical Chemistry, Vol. 72, No. 1, January 1, 2000

Figure 5. Example of two regions of a real spectrum of clean air, Figure 6. Spectrum of air sample and of the N2O component of
and the residual spectrum (difference between real and fitted spectra, the air, showing the optimal spectroscopic window for N2O analysis.
enlarged ×5) resulting from the CLS fit of MALT/HITRAN calculated
spectra to the real spectrum. The sharp features evident in the CH4
residual spectrum are most likely due to inadequate MALT modeling
of the actual instrumental line shape which may be asymmetric due
to misalignment and pressure shifts.

spectra were calculated. The number of calibration spectra

required depends in part on deviations from the ideal Beer-
Lambert law; in an ideal Beer-Lambert case, if there are N
components and no noise, only N spectra are required. However,
as the spectra are generated synthetically on a computer within
reasonable time limits, the cost in time of generating large
calibration sets is negligible. There is no advantage in generating
fewer spectra. When fitting real spectra with MALT/CLS-
calculated synthetic spectra, it is possible to achieve very good
fits where the residual spectrum (difference between the fitted
and the real spectra) is dominated by the random spectral noise.
For example, Figure 5 illustrates an example of the CLS best fit
to two regions of a real spectrum, based on a set of calibration
spectra calculated by MALT/HITRAN. The real spectrum is that Figure 7. MALT-modeled SEP surface for N2O. SEP of calibration
associated with a clean air sample. The residual spectrum is useful vs upper and lower edge of spectrum region used in calibration.
as a diagnostic indicator of the quality of the fit. In this case,
random noise dominates the CO and N2O region of the residual of 24 reference air tanks was analyzed by FT-IR. Each tank had
spectrum indicating a good fit, but the CH4 residual spectrum independently been assigned mixing ratios for one or all of the
shows evidence of an imperfect MALT modeling of the CH4 line species CO2, CH4, CO, and N2O by GASLAB (Global Atmospheric
shape. Sampling LABoratory, CSIRO Atmospheric Research, Australia).
Absolute Calibration. The process just described, of using The links between the GASLAB scales and existing calibration
the MALT-calculated spectra derived from the HITRAN database scales that are internationally recognized are described in Francey
and CLS analysis serves to carry out a working, or “MALT/CLS”, et al.5 The trace gas mixing ratios in the set of tanks more than
calibration of the instrument. This is sufficient to determine the spanned the range of mixing ratios typically found in the
analytical precision of the method, for example, as illustrated in atmosphere. For all four species, we found that the MALT/CLS
Figure 8. However if we wish to probe the accuracy of the method, calibration was linearly related to the GASLAB scale over the
this MALT/CLS calibration must be referenced to an independent entire range of mixing ratios tested, but that there can be a
“absolute” scale. An “absolute” calibration of the FT-IR instrument systematic difference of up to 5% between the MALT/CLS-derived
can be carried out by analyzing real samples from a suite of mixing ratios and those on the GASLAB scales. We have found
calibration tanks containing air that has been independently that in practice reference to one or two air standards is sufficient
calibrated relative to reference standards and then determining to relate the MALT/CLS calibration to some absolute scale to a
their apparent composition using the MALT/CLS calibration. high degree of accuracy. For example, Figure 9 (discussed below)
Absolute calibration allows for the effects of uncertainty in path illustrates five weeks of half-hourly FT-IR trace gas monitoring
length, instrumental line shape, and HITRAN line parameters to data which are accurate (i.e., in agreement with parallel measure-
be largely removed from the measurement. We have carried out ments using conventional instruments) to the order of 0.1% for
a study (which will be reported fully elsewhere) in which a suite CO2, CH4, N2O, and 1% for CO.
Analytical Chemistry, Vol. 72, No. 1, January 1, 2000 211
Figure 8. Experimental SEP vs number of coadded scans for CO2, CH4, N2O, and CO. Filled squares indicates the no-sample-handling
experiment results, and open triangles are the spectroscopy + sample handling results. (MALT parameters: resolution 1 cm-1, 740 Torr, 305
K, path length 9.8 m, 0 < H2O < 10 µmol mol-1, 340 < CO2 < 380 µmol mol-1, 1680 < CH4 < 1750 nmol mol-1, 300 < N2O < 320 nmol mol-1,
and 50 < CO < 100 nmol mol-1. CLS calibration windows: 12CO2 2020-2306 cm-1; 13CO2 2020-2290 cm-1; CH4 2805-3195 cm-1; CO
2020-2200 cm-1; N2O 2020-2250 cm-1.)

Optimal Spectral Window Selection. It is necessary to spectra and the CLS calibration procedure. The first step is to
decide which region(s) of the spectrum to fit for the analysis of generate a set of spectra using MALT that closely simulates the
each species. For example, Figure 6 illustrates a spectrum of a line shape and range of concentrations of the instrumentally
sample of air that includes 310 nmol mol-1 N2O. The calculated obtained spectra to be analyzed, including realistic levels of noise.
spectrum for N2O is also illustrated, offset in the y-axis for clarity. The wavenumber region of the MALT spectra should extend
Virtually all the infrared absorption features for the species N2O beyond the range of wavenumber regions to be considered for
occur in the region 2160-2260 cm-1, suggesting this as an the optimal calibration window. First a guess at the best left and
appropriate calibration window. However, there are significant right edges of the calibration window is made. Using these, along
gains to be made in precision by systematically rather than with the MALT-calculated spectra as input, the CLS calibration
intuitively determining the ideal upper and lower edges of the step is undertaken. As well as producing a calibration, the CLS
calibration window for each species. Note that for N2O in air the algorithm produces as output a statistical estimate of the precision
P-branch (2160-2225 cm-1) lies relatively clear of other absorption of that calibration. This is the standard error of prediction (SEP)
bands with just a few weak CO line overlapping with it. The of the calibration:
R-branch (2225-2260 cm-1) lies under considerably stronger

x
absorption lines due mainly to 13CO2. If a decision is made to fit
N2O only in the region not obscured by 13CO2, then this may ∑(y
n
actual - ypredicted)2
effectively discard as much as half of the N2O information which
SEP )
could potentially have made the measurement more precise. If, n
on the other hand, all of the N2O information under 13CO2 is
included, the risk of having the N2O measurement perturbed by
interference from the much stronger 13CO2 absorption features There are n (typically 40) spectra in the calibration training set,
is much higher. and yactual and ypredicted are the actual and CLS-predicted concentra-
For ideal spectra, the optimal CLS window for a species can tions, respectively, of a given species in a particular training
be systematically determined solely by use of MALT-calculated spectrum. In the calculation of the SEP, each spectrum in turn is
212 Analytical Chemistry, Vol. 72, No. 1, January 1, 2000
Figure 9. FT-IR, AGAGE-GC, and NDIR trace gas mixing ratios at Cape Grim (a) during the period 23 May-28 June, 1995 and (b) during an
extended period of clean air conditions, June 17-22, 1995.

removed from the calibration set and is predicted by the n - 1 baseline regions. Only when the baseline is well characterized
others, to avoid self-prediction. are the nonbaseline absorption features well constrained by
For a given set of input spectra and a given calibration window reference to it, thus leading to improved measurement precision.
specified by its upper and lower edges, the CLS calibration step This is particularly important for a broad “featureless” spectrum
estimates how precisely it can retrieve each species concentration such as N2O. Similar precision surfaces for the molecules CO2,
from the spectroscopic information provided in that window. To CH4, and CO are generated in the same way.
systematically determine the optimal window for a given species,
the CLS calibration step is performed many times while the VERIFICATION OF METHOD AND APPLICATION TO
calibration window is systematically varied. The minimum SEP ATMOSPHERIC MONITORING
specifies the calibration window that will produce most precise A series of instrumental and modeling experiments was
retrievals of that species concentration from MALT spectra. The performed in an attempt to verify the validity of the approach
assumption is then made that the same SEP vs window edge described above. We also report results of a field trial of the
surface will give a good indication of achievable precision for real instrument where it was employed to continuously monitor trace
as opposed to synthetic spectra. gases.
Figure 7 illustrates such an SEP surface for calibration of N2O Experimental Precision. A sample of air was introduced to
measurements in the region of the spectrum bounded by 1990 the 9.8-m path length White cell from a calibration tank. This air
and 2290 cm-1. The lower limit of the lower edge was determined was known from independent gas chromatographic analysis by
by the cutoff of the optical filter we used. The position of the GASLAB to have the following approximate trace gas mixing
minimum on this surface suggests that an optimal choice for an ratios: 358 µmol mol-1 CO2, 1700 nmol mol-1 CH4, 90 nmol mol-1
N2O calibration window would be the region 2000-2240 cm-1. CO, and 310 nmol mol-1 N2O. The pressure and temperature of
This optimal window is illustrated in Figure 6. Choosing an upper the sample were 740 Torr and 305 K, respectively. A set of 30
edge much higher than 2240 cm-1 leads to a loss of measurement single-beam spectra of this sample was collected over a 4-h period,
precision due to the inclusion in the window of too much strong each spectrum consisting of 256 coadded scans and taking 8 min
13CO information which obscures or dilutes the N O information to collect. These were converted to absorbance spectra with
2 2
content. Somewhat counterintuitively, the optimal lower edge for reference to a 256-scan spectrum of the evacuated White cell
the N2O window is 2000 cm-1, resulting in the inclusion of a large collected earlier. In a similar manner, four further sets of 30
region where there is very little absorption, by N2O or any other absorbance spectra were collected, consisting of 128- (4 min
species, other than the sharp CO2 feature at 2075 cm-1. It is collection time), 64- (2 min), 32- (1 min), and 16-scan spectra (0.5
plausible that the inclusion of this CO2 feature in the calibration min), respectively. All of these spectra were analyzed using the
window helps to constrain N2O. However, on the basis of CLS prediction procedure. Thus, each set of spectra provided four
observations of the other species’ optimal calibration windows, sets of 30 trace gas concentrations, one set each for CO2, CH4,
we believe the main beneficial effect is due to the inclusion of the CO, and N2O.
Analytical Chemistry, Vol. 72, No. 1, January 1, 2000 213
 After accounting for a slow linear instrument drift due to a Table 2.a Modeled and Experimental Analytical
change in the laboratory temperature, the precision of the analyses Precision
(one standard deviation) for each of the four species was approx FT-IR-NDIR
mixing precision (σ ((σ%) or FT-IR-GC
determined for the 256-scan experiment, the 128-scan experiment,
SEP ratio modeled experimental (∆ ((∆%)
and so on. The results of this set of experiments are plotted as
solid squares separately for the four species concentrations in CO2 (µmol mol-1) 360 0.05 (0.014%) 0.15 (0.04%) 0.02 (0.006%)
CH4 (nmol mol-1) 1700 0.6 (0.04%) 0.9 (0.05%) 0.5 (0.03%)
Figure 8. Figure 8 illustrates analytical precision (expessed as CO (nmol mol-1) 50 0.3 (0.6%) 0.27 (0.5%) 0.6 (1.0%)
SEP) in both absolute and percentage terms as a function of the N2O (nmol mol-1) 310 0.1 (0.03%) 0.28 (0.1%) 0.4 (0.1%)
number of coadded scans in the spectra. a The third column lists for each species, the modeled precision
The same experiment was repeated, except that the sample for spectroscopic noise levels equivalent to a 256-scan spectrum; the
was changed between measurements to assess the effects of fourth column lists the experimental precision actually attained using
256 scans and 22 m. As an estimate of the accuracy of the FT-IR
sample handling on precision. The same tank of calibrated well- technique, the fifth column lists for each species the mean of the
characterized air was used. Sample pressure and temperature were differences, ∆, between parallel measurements of the CO2 mixing ratio
by FT-IR and NDIR, and CH4, CO, and N2O by FT-IR and AGAGE-
maintained at 760 Torr and 306 K, respectively. After each analysis, GC. The mean differences are expressed both in terms of molar fraction
the White cell was evacuated and then refilled from the tank via mixing ratios and as a percent of the mixing ratio of the species in
Southern Hemisphere baseline air, as listed in the second column. For
the sampling manifold, the whole process automated by a the fifth column instrument difference results, only data collected
computer program. A set of 32 consecutive 256-scan spectra was during baseline (clean air) conditions were considered.
collected in this way, as well as a set of 40 consecutive 64-scan
spectra and a set of 12 32-scan spectra. These were analyzed in
the same manner as described above. The level of analytical mixing ratio scales. The CH4, N2O, and CO GC data were provided
precision achieved in this experiment, involving sample handling, by the Advanced Global Atmospheric Gases Experiment (AGAGE)
plotted as open triangles in Figure 8, was essentially the same as program.22 Figure 9b shows a five-day period within this time
that achieved in the experiment not involving sample handling. characterized by continuous baseline (clean air) conditions. The
This confirms that the analytical precision is not currently limited FT-IR data exhibited a level of analytical precision and accuracy
by sample handling. similar to that attainable using the conventional techniques
Application to Atmospheric Monitoring. The spectrometer, employed by the AGAGE instrument. The levels of analytical
fitted with the 22-m White cell, was temporarily installed at the precision and accuracy attained using FT-IR spectroscopy in this
Cape Grim Baseline Air Pollution Station (CGBAPS), on the application are listed in Table 2. (For CH4, N2O, and CO these
northwest coast of Tasmania. The station is part of a global are slightly superior to those shown in Figure 8, where the 9.8-m
network of atmospheric monitoring stations. The instrument was White cell was used. The longer path length of the 22-m White
programmed to draw in a sample of the atmosphere, and record cell allows a somewhat greater absorbance relative to noise to be
an absorbance spectrum of the air sample, every 30 min over a obtained, which translates directly into better measurement
period of about five weeks. These spectra were analyzed in real precision.)
time to provide a record of CO2, CH4, N2O, and CO mixing ratios. MALT Modeling of Precision vs Noise. It is possible to
The calibrated mixing ratio ranges, and the calibration windows compare the experimentally obtained analytical precision levels
used were with purely MALT-modeled analytical precision levels. The mod-
eled precision can be used to indicate the best performance
340 < CO2 < 380 µmol mol-1, 2020-2310 cm-1 obtainable in an ideal experiment. The difference between the
modeled and the experimental analytical precision indicates how
1680 < CH4 < 1750 nmol mol-1, 2810-3150 cm-1 nearly the real measurement approaches the ideal.
The MALT program was used to calculate sets of validation
40 < CO < 100 nmol mol-1, 2050-2220 cm-1 spectra of air containing the trace gases in the mixing ratio ranges
and spectral regions as listed above for the Cape Grim experiment.
310 < N2O < 320 nmol mol-1, 2020-2265 cm-1 The MALT model also specified 1-cm-1 resolution, pressure of
760 Torr, temperature of 300 K, and path length of 22 m. The
MALT spectra were calculated with a peak-to-peak absorbance
These concentration ranges, although relatively narrow, were noise level of 3.5 × 10-5 in the region of the spectrum where CO2,
appropriate for the monitoring of Southern Hemisphere maritime CO, and N2O are analyzed and ∼7 × 10-5 in the CH4 region. This
boundary layer air at a baseline station. Few excursions beyond is the noise level observed in a 256-scan spectrum of air under
these ranges were observed. Every 6 h, a sample of air from a the conditions described above. A CLS calibration was generated
calibration tank containing well-characterized clean air (with CO2, with this set of spectra. For each species, the CLS calibration
CH4, CO, and N2O mixing ratios assigned by GASLAB) was procedure provides the SEP as a measure of the precision of the
analyzed, to provide an absolute calibration for the FT-IR analyses. calibration (see equation).
This calibration enabled the FT-IR data to be reported with The MALT-modeled SEP results determined here and the
reference to the GASLAB mixing ratio scales. Figure 9a illustrates corresponding experimental precision results, reported immedi-
the mixing ratio record for the four trace gases over a five-week
period. Also illustrated are the parallel mixing ratio data acquired (22) Steele, L. P.; Lucarelli, M. P.; Fraser, P. J.; Derek, N.; Porter, L. W. In Baseline
Atmospheric Program Australia 1994-95; Francey, R. J., Dick, A. L., Derek,
by routine instruments at CGBAPS, gas chromatography for CH4, N., Eds.; Bureau of Meteorology and CSIRO Division of Atmospheric
N2O, and CO and by NDIR for CO2, also reported on the GASLAB Research: Melbourne, 1996; pp 134-140.

214 Analytical Chemistry, Vol. 72, No. 1, January 1, 2000

ately above, are summarized in Table 2. Comparing the modeled a reference gas, the method offers accuracy to within its limits of
and experimental precision results, it is generally the case that precision. The analysis is carried out on a benchtop, 1-cm-1
the MALT/CLS model is a broad guide (to within a factor of 3) instrument which is deployable in field studies. It has also been
to what we have achieved in practice. For CO2 and N2O, the demonstrated that the MALT program allows the modeling an
experimental precision was never better than a third of that of FT-IR experiment without having to perform it. The influence of
the model. For CO2, we suspect that this is attributable to the various experimental parameters may be investigated and the
strong and slightly Beer-Lambert nonlinear absorbance of some experiment optimized first on a virtual platform, offering econo-
of the 12CO2 lines included in the calibration window. For N2O, mies in both time and resources. The results returned by the
we suspect that the discrepancy between model and experiment modeling exercise offer realistic estimates as to what may be
is due to the CLS prediction algorithm not being entirely success- achieved by actual experiment. The method described here for
ful in distinguishing between the broad featureless N2O band and trace gas mixing ratio analysis is, in the companion paper,11
low-frequency baseline fluctuations in the N2O region which are expanded and refined to allow the determination of the stable
a feature of experimental spectra but not of MALT spectra. isotope ratio 13C/12C in CO2. The potential for broader application
The implications of this predictive capability of the MALT/ to other geophysically important stable isotope ratio analyses is
CLS model are important. The MALT program combined with considered.
the CLS calibration and prediction processes as described above
can serve as a virtual experimental platform, allowing systematic ACKNOWLEDGMENT
optimization of the many experimental parameters and tradeoffs We thank the staff of GASLAB, and in particular Ray Langen-
to be performed before any samples are measured on a spec- felds, for the provision of calibration gases used in this work; L.
trometer. Porter and other CGBAPS staff for assistance at Cape Grim; and
the AGAGE program for early access to some of their data. M.B.E.
CONCLUSION thanks CSIRO Atmospheric Research and the Bureau of Meteo-
We have described a novel method of precise and accurate rology (through CGBAPS) for research and living support, and
trace gas analysis using FT-IR spectroscopy. The method is the Australian Academy of Science for a travel grant.
applied to whole air samples, the only pretreatment usually
required being drying. Simultaneous analysis of CO2, CH4, CO,
and N2O mixing ratios in air was achieved. The precision of the Received for review May 25, 1999. Accepted September
mixing ratio retrievals was of the order of 0.1% for CO2, CH4, and 21, 1999.
N2O and 0.5% for CO at clean air levels. With calibration against AC9905625

Analytical Chemistry, Vol. 72, No. 1, January 1, 2000 215