Comparing two methods to measure oxidative pyrolysis gases in a wind tunnel and in prescribed burns

Keywords: compositional data analysis, Fourier transform infrared spectroscopy, FTIR, gas chromatography/flame ionisation detector, gas composition, GC/FID, log-ratio, Pinus palustris.

Wildland fire is a complex phenomenon of chemical and physical processes. Two of the chemical processes which are key to wildland fire are pyrolysis and combustion (Shafizadeh 1984; Ward 2001). During pyrolysis, a solid wildland fuel is heated and breaks down into constituent parts consisting of gases, tars and a solid material called char (Shafizadeh and Fu 1973; Shafizadeh 1982; Di Blasi 2008). During combustion, gaseous pyrolysis products react with oxygen releasing energy (heat) and a large assortment of gaseous and solid chemical compounds (e.g. Andreae and Merlet 2001; Akagi et al. 2011; May et al. 2014). Oxidation reactions involving atmospheric gases such as the oxides of nitrogen occur (Lobert et al. 1990; Crutzen and Brauch 2016).

The primary components of wildland fuels are cellulose, hemicellulose and lignin (Shafizadeh 1982; Collard and Blin 2014). These components decompose thermally by a variety of pathways which are (a) temperature dependent (Shafizadeh 1982; Neves et al. 2011; Sekimoto et al. 2018) and (b) affected by the presence/absence of O₂ (e.g. Senneca et al. 2007). During thermal decomposition, mass is conserved, meaning that the total mass of pyrolysis products will not exceed the sum of the beginning biomass and any O₂ which reacts with the pyrolysis products. An increase in the production of one product results in a concomitant decrease in one or more of the other products. Earlier works describing pyrolysis products from the foliar component of wildland fuels typically focused on determining the amount of combustible gas produced by heating and did not report the chemical composition (e.g. Susott et al. 1975, 1979; Susott 1982a, 1982b; Rogers et al. 1986). Other early works focused on other parameters such as moisture content, crude fat content and fibre content (e.g. Richards 1940; Philpot 1969, 1971). More recent flammability-related work isolated particular components of the mixtures (e.g. Alessio et al. 2008; Chetehouna et al. 2009). In fact, many gases comprise the pyrolytic mixture and both the plant species and heating rate/mode affect the compositional makeup (e.g. Tihay and Gillard 2010; Safdari et al. 2018; Amini et al. 2019a, 2019b; Safdari et al. 2019, 2020; Weise et al. 2022b). These mixtures reported in our recent work contained CO, CO₂, CH₄, H₂ and phenol in relatively large amounts; most of the other gases occurred in relatively small amounts as trace gases.

A wide variety of methods and instruments are available to identify and quantify the hundreds of gaseous and particulate compounds produced by the heating and combustion of wildland fuels. The methods and instrumentation have been compared in several studies (e.g. Fehsenfeld et al. 1987; Christian et al. 2004; Warneke et al. 2011; Haase et al. 2012; Kajos et al. 2015; Tasoglou et al. 2018; Li et al. 2019; Pistone et al. 2019). They have also been used in a complementary fashion to increase the detection and quantification of as large a suite of compounds as possible (e.g. Yokelson et al. 2007, 2013; Burling et al. 2010). Ward and Radke (1993) compared and contrasted different methods of describing smoke emissions from bench-scale measurement to aircraft and satellite measurement.

Linking laboratory results to field-scale phenomena has long been a topic of interest and the field of wildland fire and smoke emissions is no exception (Ward and Radke 1993; Yokelson et al. 2013). One potential measure to link results across scales is combustion efficiency η, defined as follows:

where CO₂a and CO₂t are the actual amount released by a fire and the theoretical amount if all carbon was converted to CO₂, respectively, and ‘C’ denotes the amount of carbon contained in each of the smoke components (HC, hydrocarbons; PM, particulate matter). The η parameter based on CO₂ was closely correlated with η based on the ratio of the amount of heat released by oxidation and the heat of combustion (Ward and Radke 1993). Further development of this concept led to the use of the CO/CO₂ ratio and modified combustion efficiency (MCE, ΔCO₂/[ΔCO₂ + ΔCO]) as descriptors of a fire and as ‘independent’ predictors of other smoke components (Ward and Hao 1991; Yokelson et al. 1996) because MCE was often ‘well-correlated’ with other emission components. The problematic use of correlation measures on proportional relative data has long been known (Pearson 1897; Aitchison 1982). Thus, the widely-assumed convention that CO/CO₂, and thus MCE, is a completely independent predictor for other smoke components has been shown to be incorrect (Weise et al. 2020b). While a suitable replacement for MCE has yet to be developed, Weise et al. (2020b) used a compositional linear trend to estimate trace gas composition over a range of combustion efficiency and reduced prediction error by 4% compared to linear regressions using MCE.

Much of the work describing the composition of pyrolysis products has occurred in bench-scale experiments and the results have been assumed to be applicable to wildland fires (e.g. Susott et al. 1979; Dimitrakopoulos 2001). However, we found no literature supporting the assumption that bench-scale experiments are applicable to the wildland fire setting. As a result, the U.S. Strategic Environmental Research and Development Program funded a series of studies designed to examine the linkages between bench-scale experiments and the wildland fire setting (Weise et al. 2022a). To date, we have used compositional data analysis (CoDA) techniques that consider the multivariate relative nature of these data to analyse the effects of plant species, heating mode and moisture content on non-oxidative pyrolysis products at bench-scale (Weise et al. 2022b). Using CoDA, we compared gases produced during flaming combustion and oxidative pyrolysis in a series of wind tunnel and field fires that were quantified using a gas chromatograph equipped with a flame ionisation detector (GC/FID) instrument (Weise et al. 2022c). The objective of the present manuscript is to compare the composition of oxidative pyrolysis gases measured in these same fires using GC/FID and Fourier transform infrared (FTIR) spectroscopy to determine if the methods produce comparable results. Both techniques have been used extensively in bench-scale pyrolysis experiments and are accepted methods to quantify the quantity and composition of pyrolysis gases.

Oxidative pyrolysis gas composition was determined by two analytical methods (GC/FID and FTIR) in wind tunnel and field settings (Fig. 1). The complete GC/FID data have been analysed and presented elsewhere (Weise et al. 2022c). We first present an analysis of the FTIR data which includes reanalysis of the field fire data (Scharko et al. 2019a) and then examine the subcomposition (subset) of gases measured by both methods. The approach to analysing the collected gases consisted of several steps. The composition of oxidative pyrolysis gases measured using FTIR was compared between wind tunnel and field experiments. The relative amounts of the subcomposition of common gases in the GC/FID and FTIR datasets (CO, CO₂, CH₄, C₂H₂, C₂H₄, C₃H₆, C₄H₆ and isobutene) for the wind tunnel and field data were then compared to determine if the subcomposition was affected by the analytical method or the sampling location (wind tunnel and field).

Fig. 1.  Schema of oxidative pyrolysis gas sampling in wind tunnel and field. (a) Canister sampling for GC/FID and (b) FTIR and GC/FID sampling setup. Field sampling in canisters for GC/FID and FTIR and various analyses performed for GC/FID and FTIR samples.

Eighty-eight fuel beds 2.4 m long and approximately 0.8 m wide composed of longleaf pine (Pinus palustris Mill.) needles and various combinations of fetterbush (Lyonia lucida (Lam.) K. Koch), sparkleberry (Vaccinium arboreum L.), blueberry (V. darrowii Camp) and inkberry (Ilex glabra (L.) A. Gray) (Table 1) were burned in the Riverside Fire Laboratory (RFL) wind tunnel. Wind tunnel burns were carried out under ‘no wind’ and 1 m s⁻¹ conditions. Fuel beds were constructed by uniformly distributing 1000 g (wet weight) of needles over the entire fuel bed in sections based on a well-established procedure (Anderson and Rothermel 1965; Schuette 1965; Rothermel and Anderson 1966) and then placing the plants in predetermined locations. Composition and characteristics of the fuel bed components are available (Safdari et al. 2018; Matt et al. 2020). Fuel moisture content, fuel loading, ambient temperature and relative humidity in the wind tunnel varied between experiments. Pine needle dry mass ranged from 862 to 943 g, uniformly distributed over the fuel bed (Schuette 1965). The plants were placed in the fuel bed approximately 5 min before lighting the fire. Fuel beds were ignited with a line fire which spread the length of the fuel bed.

Table 1.  Summary statistics for pyrolysis gases mixing ratios in ppm determined by FTIR and GC/FID from fires in longleaf pine fuel beds in a wind tunnel at the U.S. Forest Service Riverside Fire Lab (RFL) or in prescribed burns at Ft. Jackson, SC.

Two separate oxidative pyrolysis gas sampling setups were used. An array of eight stainless-steel sample tubes along the length of the fuel bed filled a sample canister with multiple small aliquots of oxidative pyrolysis gases resulting from the volatile matter for the GC/FID analysis. A separate sample tube and canister upwind of the oxidative pyrolysis tubes captured flaming emissions for comparison. Further detail on the experiments and measurement methods can be found elsewhere (Weise et al. 2022c). The flaming combustion samples are not included in this analysis. For the FTIR samples, two approaches were used (Scharko et al. 2019a; Banach et al. 2021). In the wind tunnel, a fixed stainless-steel probe located in the vicinity of the sample tube array captured gases emitted by the pyrolysing plant and pine needle fuels prior to the arrival of any combustion gases or the flame front. The gases were pumped into a long-path White cell (White 1942) housed inside the sample compartment of a Bruker Tensor 37 FTIR spectrometer. Two different sampling modes (dynamic - continuous gas flow through the White cell; static - gas flow into the White cell for a fixed time which was then valved shut) were used. Only the static samples (Banach et al. 2021) were used in the present analysis. Analysis of the dynamic samples is currently incomplete and will be the focus of a future analysis. Canister samples for FTIR analysis were collected in the field fires as described below.

Five 0.1 ha experimental plots located at Ft. Jackson, SC in longleaf pine–slash pine forests (P. elliottii Engelm.) were burned in May 2018. The dominant understory shrub in these plots was sparkleberry (Hudak et al. 2020; Herzog et al. 2022). Further descriptions of the plots, fuel measurements and experimental burns are available (Scharko et al. 2019a; Hudak et al. 2020; Weise et al. 2022c). Real-time measurement of oxidative pyrolysis gases in an open-path environment using FTIR (Akagi et al. 2014) was not feasible so we collected canister samples. A 2.5 m gas sampling probe of 6 mm stainless-steel tubing with flexible stainless-steel tubing connecting to the sampling package captured the gases (Fig. 1), which we assume resulted from the volatilisation of the forest fuels. The field sampling package consisted of a swing Piston KNF Neuberger Pump,¹ 12-V gel cell rechargeable battery, stainless-steel tubing to a pressure relief valve and gauge. The flow rate to fill the canisters was 250 mL s⁻¹. There were two identical canister sampling packages: one for 0.85 L SUMMA^® canisters (172 kPa) for GC/FID analysis, and the second for 3 L SUMMA^® canisters (138 kPa) for FTIR analysis (Scharko et al. 2019a). The sampling strategy was to identify plants along the edge of the plot that had enough foliage to ignite and allow the chance of sampling oxidative pyrolysis. As the flame front advanced, the probe was positioned ahead of the flame front as it approached. Air samples were acquired in close proximity to the base of the leading edge of approaching flame, taking short interval sample aliquots when it was likely that oxidative pyrolysis was occurring.

In the static mode used for the field data, the gas was brought to the laboratory (typically on the same day), pumped from the SUMMA^® canister and isolated in the White cell; the inlet/outlet valves were simultaneously closed such that the emitted gases were isolated in the cell at a desired pressure (~98.7–93.3 kPa for high pressures, and 57.3–53.3 kPa for lower pressure measurements). The goal of the static mode was to obtain a higher fidelity spectral ‘snapshot’ for a given point in time of the burn; more interferograms were averaged at higher spectral resolution thus allowing for detection of more gaseous species with higher sensitivity (Scharko et al. 2019a).

Twenty FTIR static samples were thus obtained (10 each from the wind tunnel and the field). The static experiment spectra were recorded using the full resolution of the spectrometer (0.6 cm⁻¹), a 2 mm Jacquinot stop and double sided, forward–backward acquisition. The Tensor 37 FTIR was outfitted with a Ge/KBr beamsplitter, glow bar source and mercury cadmium telluride detector, providing data acquisition yielding spectra from 7500 to 500 cm⁻¹. Due to the higher resolution requiring a smaller aperture and thereby lower light throughput, acquisition time was extended by averaging multiple scans for a full 30 min, resulting in vastly improved signal/noise ratios. For analysis of such complicated gas-phase mixtures, infrared spectral resolution of 1.0 cm⁻¹ or better has been demonstrated to be advantageous (Burling et al. 2010; Akagi et al. 2014; Scharko et al. 2019a). The gas cell path-length was verified using purified isopropanol with a total of 10 spectra with neat gas pressures (MKS KF15 pressure transducer) ranging from 0.6 to 10.5 Torr. Resulting spectra were integrated from 3290 to 3515 cm⁻¹ and using known cross-sections a Beer–Lambert plot was created from which the cell path-length was determined to be 6.5 ± 0.2 m. All infrared data were reduced using MALT5 software (Griffith 1996; Griffith et al. 2012) in which the experimental spectrum is fitted to a simulated spectrum; the mean-squared residual between the two is minimised, resulting in calculated absorption coefficients of each input species. These are quantified by comparing the cross sections to quantitative databases, such as the HITRAN line-by-line dataset (Gordon et al. 2017) or the PNNL 50°C gas-phase reference spectra (Sharpe et al. 2004; Kochanov et al. 2019) as input libraries. Using different spectral ‘microwindows’ for each species and taking care that none of the measured bands were saturated, the software is used to calculate a vapour-phase mixing ratio for each gas (Scharko et al. 2019b). In addition to the static samples, 10 dynamic samples were collected for the wind tunnel also; some results from the dynamic samples are presented elsewhere (Banach et al. 2021).

Background air samples were taken during the wind tunnel and field experiments in 0.85 L canisters and analysed, so the reported gas concentrations are excess mixing ratios (Ward and Radke 1993; Urbanski et al. 2008). The canister emission samples were analysed using EPA method TO-14A (EPA 1999) for CO₂, CO, CH₄ and C₂ to C₅ hydrocarbon gases with an Agilent model 7890 GC configured with two columns running simultaneously. A 3.175 mm diameter, 2 m Carbosphere™ packed column with a nickel catalyst methaniser was used for analysis of CO₂ and CO using FIDs. The second column, a 0.53 mm diameter × 50 m length Agilent Al/S column, separated hydrocarbons and methane. Both columns went to FIDs and were run simultaneously. Chromatogram data were collected and processed using Agilent OpenLab software. A multipoint set of three standards bracketing the sample concentrations was analysed with each set of samples to construct a standard curve for each compound. Based on the integrated peak areas, the sample concentrations were calculated from the standard curves and written into a spreadsheet for analysis. Chromatograms of National Institute of Standards and Technology standard reference material gas standards for CO, CO₂, CH₄ and propane were acquired each day to validate the standard curve. Duplicate GC runs of canisters were performed for each sixth sample. A Trace Analytical RGA3 reduction gas analyser was used to measure H₂ concentrations – this is a chromatographic instrument with a molecular sieve column and a UV mercuric oxide detector that provides highly sensitive precise measurement of trace level H₂. The target range is 0–10 ppm H₂; most samples were diluted with ultra-high purity nitrogen to be in this range. An H₂ standard (Scott Specialty Gases) was used for calibration. Chromatograms of H₂ from this instrument were collected and integrated with Agilent OpenLab software interfaced to the instrument. On a subset of canisters with significant detectable oxidative pyrolysis gases, GC/MS analysis was performed using an Agilent 6890 GC with a HP-5 0.320 mm × 30 m column, helium carrier gas and Agilent 4590N MS detector. The composition of the gases in 204 canisters was analysed for the wind tunnel and field burns.

For the FTIR data, a combination of software was used for the post-acquisition spectral analysis and confirmation of the species observed during the campaign. The original MALT software (Multiple Atmospheric Layer Transmission) used classic least squares to create linear combinations of single-component spectra that minimised the difference between the measured and simulated spectra by adjusting mixing ratios (Griffith 1996). The MALT5 version uses a nonlinear least-squares algorithm (Griffith et al. 2012) and both HITRAN line-by-line data (Gordon et al. 2017) and the PNNL 50°C gas-phase reference spectra (Johnson et al. 2006, 2010; Kochanov et al. 2019) as input libraries to identify and quantify vapour-phase chemicals in the spectra. Spectra were compiled into parameter files and analysed using parameters including pressure, temperature, pathlength, resolution, as well as estimated initial values for chemical mixing ratios. To confirm that the chemical species were actually present, each spectrum generated by MALT5 was input to OPUS (Bruker Corporation LLC 2020) and subtracted from the measured spectrum; the target compound was purposefully omitted from the subtraction process to visually inspect if the omitted compound was present in the spectrum.

Atmospheric chemistry data which include smoke and pyrolysis gases are compositional data (e.g. Bandeen-Roche 1994; Engle et al. 2011; Jarauta-Bragulat et al. 2016; Gibergans-Baguena et al. 2020; Weise et al. 2020a, 2020b) and were analysed as such using the R system for statistical computing v4.1.3 (https://www.R-project.org/). The CoDA relies on log-ratio coordinate representations of the data and, hence, missing and zero values in the data set are problematic. The data were preprocessed using the multLN function in the zCompositions R package to impute random values below the minimum observed concentrations (Palarea-Albaladejo and Martín-Fernández 2013, 2015) for those instances where the gas concentration was below the detection limit (BDL) of the measuring instrument because this method provided results comparable to more complex imputation methods. Summary measures are defined for compositional data such as the centre (mean), total variance, log-ratio mean and variance (van den Boogaart and Tolosana-Delgado 2013). We calculated the log-ratio means for all gas pairs because the log-ratio can be used to examine the proportionality between each pair of gases (Lovell et al. 2015; Egozcue et al. 2018).

A proportionality measure ρ for compositional data, similar to the more familiar Pearson correlation coefficient r, ranges over [−1, 1] with one indicating perfect proportionality (Erb and Notredame 2016). Proportionality between each pair of gases was estimated for the FTIR and GC/FID data using the propr R package (Lovell et al. 2015; Erb and Notredame 2016; Quinn et al. 2017, 2018, 2019). Because ρ is the same for absolute (original) and relative forms of the gas compositions, it is a precise measure of correlation in the original data. Only positive values of ρ were considered to indicate proportionality based on the approach proposed by Egozcue et al. (2018). Since proportionality between D(D − 1)/2 gas pairs, where D is the number of gases, was tested for significance, the value of ρ was adjusted using the procedure updateCutoffs in propr to keep the false discovery rate (α) less than 0.05 (Benjamini and Hochberg 1995).

Seventeen scientifically meaningful comparisons between subgroups of gases were represented by so-called log-ratio balances, or simply balances, which are a special class of log-ratio coordinates where the log-ratio term compares the geometric mean of r_k parts in one subset (in the numerator, indicated with + symbol) with the geometric mean of s_k parts in another subset (in the denominator, indicated with − symbol). The general expression of the kth balance is given as follows:

The preceding multiplicative factor is a normalisation factor formally required to ensure comparability of the balances and ensure orthonormality of the resulting coordinate system. A simpler expression as the normalised difference between the means of log-transformed data is given as follows:

Scientifically meaningful balances were formulated by defining an adequate sequential binary partition of the original composition (Egozcue and Pawlowsky-Glahn 2005) (Table A1).

The FTIR dataset consisted of 27 gases identified by FTIR and 27 observations (22 wind tunnel and five field), which yielded a dataset of 27 observations and 26 balances. The assumption of multivariate normality is required for many multivariate methods and several different tests have been developed. Similarly, many univariate tests require univariate normality. We applied several multivariate normality tests to see if there was agreement between the different tests. The Mardia, Henze–Zirkler, Royston and Energy tests for multivariate normality as implemented in the MVN R package (Korkmaz et al. 2014) indicated that distribution of these balances was not multivariate normal (Table A2). The univariate Shapiro–Wilk (Royston 1982) and Kolmogorov–Smirnov (Conover 1999) tests indicated that several of the distributions of the individual balances were not normal (Table A3). Tests of equality of the covariance matrices for the wind tunnel and field as implemented in the covTestR R package (Barnard and Young 2018) yielded contradictory results: two tests rejected the null hypothesis of equality (Schott 2007; Srivastava et al. 2014) and two did not (Srivastava 2007; Srivastava and Yanagihara 2010). For these reasons, permutational multivariate analysis of variance (PERMANOVA) (Anderson 2001, 2017) was used to statistically test the potential effect of sampling location (wind tunnel and field) on the mean relative composition of gases using the adonis function in the vegan R package (Oksanen et al. 2020). The Kruskal–Wallis rank sum test (Hollander et al. 2014), a nonparametric analogue of univariate ANOVA, compared the average balances between wind tunnel and field. The resulting P-values were adjusted to account for the multiple testing using the method of Benjamini and Hochberg (1995).

Seven balances were derived for the subcomposition of common gases. The univariate tests indicated that six of the seven balances were not normally distributed. Tests of equality of the covariance matrices for the subcomposition were also contradictory as above. As a group, the seven balances were not multivariate normal. The univariate Shapiro–Wilk test showed that six of seven balances were not univariate normal (Table A4). The PERMANOVA tested the effects of analytical technique, location (wind tunnel, field) and their interaction on the subcomposition. The Kruskal–Wallis test compared the average balances between location and analytical technique (FTIR and GC/FID) individually. The resulting P-values were also adjusted to account for the multiple testing (Benjamini and Hochberg 1995).

To facilitate meaningful comparison of the two analytical techniques, all gases that were missing from more than 60% of the observations were removed from the GC/FID and FTIR compositions. This resulted in the removal of ethanol, N₂O, benzene, crotonaldehyde, isocyanic acid, NO, NO₂ and sulfur dioxide from the FTIR samples and hexane, 1-hexene, 1-heptane, benzene, toluene, furan and furfural from the GC/FID samples (Table 1). The wind tunnel and field fires yielded 88 and seven GC/FID samples containing 22 gases, respectively. These same fires yielded 22 wind tunnel and five field FTIR samples containing 27 gases, respectively. Water vapour and CO₂ were the two gases with the largest concentrations in the FTIR samples (Table 1); CO₂ also dominated the GC/FID samples followed by CO. The geometric mean concentrations estimated using FTIR for CO₂ were 15 272 and 46 229 ppm in the RFL wind tunnel and at Ft. Jackson, respectively. The GC/FID CO₂ concentrations were an order of magnitude smaller. The concentrations of the other gases measured by GC/FID were generally smaller than the concentrations measured by FTIR. The original concentrations were summarised as geometric mean ⋇ geometric standard deviation. The geometric counterparts to the ordinary ‘arithmetic’ measures are adequate when the data are ratios to some reference amount (Fleming and Wallace 1986). The geometric standard deviation is related to the geometric mean by multiplication and division (Butler et al. 2020) indicated by ‘⋇’.

The sign of a log-ratio indicates which part of the ratio is relatively larger: negative values indicate the numerator is smaller than the denominator. The relative dominance of H₂O, CO₂, CO and CH₄ in the wind tunnel and field compositions is indicated by their log-ratios with other gases being negative (Table 2) meaning that the relative amount of the other gas is smaller. In the CoDA context, perfect proportionality indicates that the log-ratio between two parts of a composition is constant, meaning that the relative amounts of the two parts are constant. The hypothesis of perfect proportionality between FTIR gas pairs was not rejected for 13 pairs in the wind tunnel data and a single pair for the Ft. Jackson samples, suggesting that these pairs may be proportional. Allene was proportional to acetylene in the field data, but not in the wind tunnel data. In the wind tunnel, HONO was proportional to both CO₂ and CO while CO was also proportional to acetylene and ethene. The last observation is consistent because the presence of CO is usually associated with incomplete combustion. Isoprene was proportional with propene and 1,3-butadiene; also understandable as all three alkenes are found in concomitant ratios before combustion occurs (Brauer et al. 2014). For comparison, in the GC/FID dataset, the test for perfect proportionality was not rejected for 59 pairs of gases (Table 3) in the wind tunnel; only a single test was not rejected in the field data. Acetylene and 1,3-butadiene were each proportional with nine other gases. As with ordinary correlation, the statistical test of the null hypothesis of perfect proportionality is affected by the variance of the log-ratio, so increased variation makes proportionality (and correlation) more difficult to detect.

Table 2.  Log-ratio means for pyrolysis gas composition measured by FTIR in (a) wind tunnel and (b) field burns.

Table 3.  Log-ratio means for pyrolysis gas composition measured using canisters and GC/FID in (a) wind tunnel and (b) field burns.

The PERMANOVA on log-ratio coordinates indicated a significant difference in mean FTIR gas compositions between the wind tunnel and the field (Table 4). Note that the results from this overall test are by construction independent of the log-ratio coordinate representation chosen for the original composition. However, to understand where the differences may arise, the 17 median values of the balances resulting from the sequential binary partition of the composition are shown in Fig. 2 and the details of the gases contained in the balances are found in Table A1. In the following sections the term ‘amount’ will be used for convenience – it is important to recall that the comparisons of what were originally mixing ratios are conducted in the space of log-ratio coordinates.

Fig. 2.  Comparison of median values for selected log-ratio balances of pyrolysis gases measured with FTIR spectroscopy by location using the Kruskal–Wallis rank sum test. P-values adjusted for false discovery rate. Grey band indicates P < 0.05. Light HC = light hydrocarbons (alkanes, alkenes and alkynes).

Table 4.  Summary of permutational analysis of variance testing effects of fire location (wind tunnel and Ft. Jackson) on relative composition of pyrolysis gases measured using FTIR.

The amount of H₂O and CO₂ relative to all other gases and the amount of H₂O relative to CO₂ were greater in the wind tunnel than the field. The reported concentrations were already adjusted to account for the ambient atmosphere, so this difference in H₂O is not due to differences in atmospheric humidity between the two locations, i.e. it is the excess H₂O due to the fire, similarly for CO and CO₂. Longleaf pine needle fuel moisture-content was comparable between wind tunnel and field fires and the live foliage moisture-content was approximately 100% and 200% in the wind tunnel and field fires, respectively (Weise et al. 2022c). The cause of the differences in the relative amounts may be proximity of the pyrolysing foliage to the sample probe (0.1 m in the wind tunnel versus 1 m in the field fires) or due to the size difference of the nursery and field plants.

More CO and light hydrocarbons were observed at Ft. Jackson relative to other pyrolysis gases. However, the relative amounts of CO and light hydrocarbons (alkanes, alkenes and alkynes) did not differ between the wind tunnel and Ft. Jackson, suggesting that the greater amount of CO and light hydrocarbons at Ft. Jackson is due to smaller relative amounts of the other gases. There were no significant differences in the relative amounts within the light hydrocarbon gases between the wind tunnel and Ft. Jackson fires.

The amount of nitrogen-bearing gases (methyl nitrite, NH₃, HCN and HONO) relative to the other volatile organic compounds (VOC, not including light hydrocarbons) was smaller at Ft. Jackson than in the wind tunnel. More HONO and NH₃ relative to HCN and methyl nitrite were observed in the wind tunnel fires. While we previously speculated that methyl nitrite’s presence at Ft. Jackson may have been associated with munitions (Scharko et al. 2019a), the amount of methyl nitrite relative to HCN did not differ significantly between the wind tunnel and field. It is important to note that 40% of the observations of methyl nitrite were BDL in contrast to 6% of HCN observations (Table 1). The relative amount of nitrogen present in the fuels may have differed because nursery plants were used in the wind tunnel as opposed to the native vegetation combusted at Ft. Jackson; this is even more likely if the nursery plants were fertilised. Similarly, the nitrogen content of the longleaf pine needles may have varied between the locations where the pine needles were harvested and the stands at Ft. Jackson. Fertilisation with nitrogen is a practice used in longleaf pine sites that are managed for pine straw production (Ludovici et al. 2018). No elemental analysis of the different fuels was performed so this is purely speculative. More HONO relative to NH₃ was observed at Ft. Jackson. Sekimoto et al. (2018) associated HONO with a high-temperature pyrolysis profile and NH₃ with a low-temperature pyrolysis profile. Based on this, our results suggest that more of the Ft. Jackson data resulted from high-temperature oxidative pyrolysis than the wind tunnel data. Maximum fuel temperatures observed below the FTIR probe in the wind tunnel ranged within 132–660°C and leaf temperatures at Ft. Jackson were 231–648°C (Weise et al. 2022c), which does not support differing pyrolysis temperature profiles in these data.

The amount of phenol relative to the other VOCs was greater in the wind tunnel (Banach et al. 2021). We previously reported difficulty in observing phenol and phenolic compounds at Ft. Jackson due to their infrared bands being somewhat weak and obscured by a number of other species, namely acetic acid, carbon dioxide, acetylene and hydrogen cyanide (Scharko et al. 2019a). This difficulty was reflected in the small balance value for Ft. Jackson. Results from bench-scale measurements of phenol produced in pyrolysis by different heating modes showed that after CO₂, H₂ and CO, more phenol relative to the remaining 84 gases was measured and plant species and moisture content affected the amount of tars relative to phenol (Weise et al. 2022b). More furan relative to naphthalene and more acetaldehyde relative to formaldehyde was observed in the oxidative pyrolysis gases at Ft. Jackson compared to the wind tunnel. The ratio of furan to acetylene has also been suggested as a possible measure to distinguish between high- and low-temperature pyrolysis factors (Sekimoto et al. 2018). Sekimoto et al. defined two factors that accounted for ‘much of the observed variability in VOCs’ as high- and low-temperature pyrolysis factors; air temperature was measured by an FTIR instrument at a sampling inlet after cooling and mixing with ambient air occurred. Our earlier analysis of this ratio suggested that high-temperature pyrolysis was the dominant process associated with the generation of gases at Ft. Jackson (Scharko et al. 2019a). That work and the work of Sekimoto et al. (2018) did not analyse the gas measurements as compositional data. Testing the furan to acetylene ratio could occur in the context of a different set of balances. We did not test whether this specific ratio differed between the wind tunnel and field. The log-ratios were very similar (Table 2), suggesting the possibility of no difference between the wind tunnel and Ft. Jackson. It may be possible to further explore this using the wind tunnel dynamic FTIR data coupled with heat flux and temperature data as in Banach et al. (2021).

The gases common to the GC/FID and FTIR samples were CO, CO₂, CH₄, C₂H₂, C₂H₄, C₃H₆, C₄H₆ and isobutene (C₄H₈). Bearing in mind that the aliquots of gases were collected close in time, but not from the same aliquot of air, the trends in the relative amounts of the gases by the two methods were quite similar (Fig. 3). However, the relative amount of CO measured by FTIR at Ft. Jackson was an order of magnitude larger than the wind tunnel or GC/FID measurements. The PERMANOVA showed that the GC/FID and FTIR methods had a significant effect on the relative composition of the common gases (Table 5) but the location and interaction effects were not significant.

Fig. 3.  Geometric mean relative amount of pyrolysis gases common to FTIR spectroscopy and GC/FID samples in wind tunnel (RFL) and prescribed fires in Ft. Jackson (FJ).

Table 5.  Summary of permutation analysis of variance testing effects of analytical method (FTIR and GC/FID) and fire location (wind tunnel and Ft. Jackson) on relative composition of pyrolysis gases measured by both analytical methods.

The median values of the four combinations of location and method and the overall median were calculated for each balance (Fig. 4) for the common gases. The analytical method significantly affected six of the seven balances as indicated by the Kruskal–Wallis test. The first two balances containing CO₂ were smaller for the FTIR measurements at Ft. Jackson; relatively less CO₂ and more CO and CH₄ were measured, resulting in smaller balances. The FTIR measured more CO relative to CH₄ than GC/FID; however, relatively more CH₄ was measured by GC/FID in the field burns. The FTIR balance values were smaller than GC/FID for alkenes vs acetylene. The amount of ethene relative to the other alkenes did not differ between analytical methods. However, the amount of 1,3-butadiene relative to the C₄ alkenes was greater in the FTIR samples. The propene vs isobutene balance suggested that more propene relative to isobutene was present in the FTIR samples compared to the GC/FID.

Fig. 4.  Comparison of median values for selected log-ratio balances of pyrolysis gases by method, location and overall median. P-values from Kruskal–Wallis test of difference between methods adjusted for false discovery rate. Grey band indicates P < 0.05. No significant differences between balance values between locations. FJ, Ft. Jackson; RFL, wind tunnel.

In the air pollution and smoke communities, emission ratios (ERs) are commonly analysed measures, perhaps starting with the work of Darley et al. (1966). Early air pollution and biomass smoke studies used linear regression to adjust laboratory and field gas measurements to common CO₂ values in order to compare other gas concentrations (Boubel et al. 1969). The use of linear regression in this instance assumed a correlative relationship between the gases. More recent work has used ERs with CO to link laboratory and field experiments (Yokelson et al. 2013). In seminal work, Crutzen et al. (1979) examined CO₂ ERs and reported ‘uniform CO/CO₂’ ratios and that other ERs did not deviate by ‘more than a factor of six’. A ‘factor of six’ suggests that the ER may not be constant, i.e. the concentrations of the two gases may not be proportional or correlated. At the base of the wealth of literature that has examined ERs is the assumption that correlation is an appropriate statistical measure to use on these data and that ratios are an appropriate transformation to apply to the original data. The use of ERs comes close to the statistically appropriate handling of these data. If an entire composition of gases (x) is expressed as ER relative to CO₂, i.e. , taking the logarithm of these ratios is equivalent to applying the so-called additive log-ratio (alr) transformation (Aitchison 1986; Weise et al. 2022b) and has been used in other atmospheric chemistry studies (Jarauta-Bragulat et al. 2016; Gibergans-Baguena et al. 2020). Although useful in parametric statistical modelling, the alr is not particularly adequate when distances, angles or shapes are involved (e.g. when performing clustering analysis or multidimensional scaling) as it deforms these (Aitchison 1986; Aitchison et al. 2000; van den Boogaart and Tolosana-Delgado 2013). An ER can be derived from any log-ratio x contained in Tables 2 and 3 by exp(x); for instance ER_CH4(C₂H₂) = exp(−1.5) = 0.22 for FTIR in the wind tunnel (Table 2).

Correlation and linear regression are frequently used to examine the internal relationships between different gases or between the gases and external variables (such as fire phase, fuel type and fuel moisture). Ideally, the relative relationships between pyrolysis gases should not be affected by the method used to identify the gases. Gases that are proportional in one method should be proportional in another method. If the proportionality between a pair of gases is known, it would be possible to only measure one of the pair of gases. In the present study, only a small fraction (13/351) of the possible pairs of gases measured in the wind tunnel using FTIR were identified as proportional; only five gases were proportional with CO or CO₂. This may be a surprising result based on previous studies that correlated many gases with CO or CO₂. It is important to recall that in those previous studies, the use of correlation to examine the relationships ignored the relative nature of the data. The data from those studies could be reanalysed using proportionality to examine the relationships of combustion gases with CO and CO₂. In the field measurements at Ft. Jackson, only one of 351 pairs was proportional. For the GC/FID oxidative pyrolysis data, 25% (59/231) and 1/231 of the pairs were proportional in the wind tunnel and Ft. Jackson data, respectively. Of the 59 pairs, 36 were proportional in the wind tunnel GC/FID flaming combustion data reported previously (Weise et al. 2022c). Proportionality is a measure based on the variability of the log-ratio variance; relatively large variance does not support proportionality. Fewer proportional pairs might suggest that the data measured using FTIR were more variable than the GC/FID data and that the Ft. Jackson field measurements were more variable than the wind tunnel measurements. However, this observation is based on gas compositions that have only nine gases in common. The variability could also be a natural characteristic of the gases emitted during the oxidative pyrolysis process. The potential increase in variability from the wind tunnel setting to the field setting is not surprising. Dimitrakopoulos (2001) argued for the use of tightly controlled bench-scale non-oxidative pyrolysis measurements due to the chaotic variability present in the fire environment. It is currently not possible to compare the current data to other data from similar studies to assess the quality of the data as these are the first such data describing pyrolysis gases in a fire setting. We sought to control as much variability as possible through the use of common sampling methods, replication of experiments, simplification of fuels and similar environmental conditions to produce quality data. Future scientists can use these data as a point of comparison to determine what the natural variability of gas composition is. It is important that physically-based fire models also incorporate this variability in composition instead of assuming that the composition is constant. However, we have recently shown using CoDA techniques that measurements made in simplified and controlled wind tunnel experiments can be related to field measurements (Weise et al. 2022c).

The balances associated with the FTIR data from the wind tunnel and Ft. Jackson suggested some differences in the relative amounts of various gases. The cause for the differences between the wind tunnel and Ft. Jackson may be due to the sampling probe location differences, insufficient sample size or greater natural variability. The relative amount of nitrogen content in the fuels may have differed due to different cultural practices used in commercial nurseries, pine stands managed for pine straw production and pine stands managed by the Department of Defence.

Direct comparison of gas measurements in a fire environment produced with different instruments is challenging and influenced by many factors. In prior work and the current study, 57 gases associated with wildland fire (Johnson et al. 2010; Weise et al. 2015; Scharko et al. 2019b) were added to the PNNL spectral library. Although 57 is small compared to the hundreds of compounds that can be identified using other methods such as GC-FID, FTIR is a viable technique for wildland fire. Different instruments and analytical techniques may measure the same gases with differing resolution or are unable to detect the same compounds (Ward and Radke 1993). A wide variety of instrumentation is often deployed to measure a large suite of compounds (e.g. Yokelson et al. 2013). We previously noted that different methods influence the relative abundance of measured emissions (Weise et al. 2015) even though CoDA was not used in this earlier study.

The present study used two methods readily adapted to field use. The inherent spatial, temporal and compositional variability in fuels as well as the non-steady conditions under which the fuels were heated can affect the composition of the pyrolysis and combustion products. Some of this variability can be controlled by using a common sample line or assuming that sample collection points in proximity are true replicates. The wind tunnel fuel beds were constructed so that the amount and depth of longleaf pine needles was uniform and the density of the uniformly distributed plants was increased due to the plants’ small size in order to produce a detectable concentration of gases. The distribution of both pine needles and live plants in the field was not uniform and contained additional fuel types. Using more uniform field sites such as pine plantations instead of natural stands in future experiments may reduce variability. Using more sensitive instrumentation that is field-hardened and sampling closer to the live foliage may improve measurements; however, the mixture of gases sampled in the wind tunnel resulted from live and dead fuels in close proximity which may be difficult to achieve when measuring mature plants in a field setting. Performing elemental analysis to enable carbon and nitrogen balances is another improvement that could be used in future experiments. As the elemental makeup of plants is also compositional data, log-ratios of the different elements should be used.

In the present study, the assumption was made that samples taken in proximity to each other in wind tunnel fuel beds or in small prescribed burns in natural fuels each represented true replicates. In three of the four (location × method) combinations, samples were collected in canisters for subsequent analysis, either later in the day (FTIR) or several weeks after collection (GC/FID). The wind tunnel FTIR spectroscopic measurements were collected in real time. The SUMMA^® canisters used for sample collection and storage have been shown to be reliable storage containers for up to 32 weeks and are used routinely to collect fire gas samples (Evans et al. 1998; Austin et al. 2001; Wang et al. 2014; Strand et al. 2016).

While all these factors may affect the absolute values for the oxidative pyrolysis gases collected, it is the relative values that provide the relevant information since these are compositional data. Based on the subcomposition of gases measured by both FTIR and GC/FID, the analytical method did significantly affect relative composition of oxidative pyrolysis gases based on the PERMANOVA while the fire location (wind tunnel versus field) effect was not significant. The dominance of the primary pyrolysis gases (CO₂, CO and CH₄) relative to the light hydrocarbon trace gases differed between methods. Future experiments using instrumentation with a greater number of common gases could be used to further evaluate the applicability of wind-tunnel scale results to field fires. In pyrolysis models within physics-based fire models, these three primary gases are the ones often used (e.g. Westbrook and Dryer 1981; Zhou and Mahalingam 2001; Di Blasi 2008). The balance between CO₂ and the fuel gases (CO and CH₄) and the relative amount of CO to CH₄ did not differ between locations. This potential lack of difference for the dominant pyrolysis gases is encouraging and may support the use of relatively simple oxidative pyrolysis models in physics-based wildland fire models as in Borujerdi et al. (2022). Application of common statistical methods to meaningful log-ratio coordinate representations of compositional data conveying relative information ensures that the results are not an artifact of the presence or absence of a gas in the composition. A future manuscript will compare the composition of pyrolysis gases to determine if it is possible to link the non-oxidative bench-scale results with the oxidative pyrolysis fire-scale results. Lastly, another future manuscript will compare the changes in composition of gases measured over time using dynamic FTIR as described in Banach et al. (2021).

Oxidative pyrolysis gases were successfully measured in a wind tunnel and in small field prescribed burns using two different methods: FTIR-spectroscopy and GC/FID analysis. Using CoDA techniques for the FTIR data, the composition of wind tunnel and field oxidative pyrolysis gases differed. In the wind tunnel setting, use of FTIR enabled the measurement of instantaneous samples of gases that were not measured by the GC/FID setup used. Although not presented here, FTIR can also provide dynamic description of the changing composition of gases as fuels heat, pyrolyse and combust.

Gases potentially proportional to each other were identified for both methods; however, the number of proportional pairs differed markedly due to the observed variability in part. The subcomposition of gases measured by both methods showed that the compositions were affected by the measurement method but not by location. The subcomposition differed in the relative amounts of the trace gases. The relative amount of the two primary fuel gases (CO and CH₄) was significantly affected by the analytical method. Although the composition of trace gases differed between the analytical methods, from a fire modelling perspective, each method yielded comparable description of the primary fuel gases, providing independent verification of measurements. While the importance of the trace gases in combustion models remains to be determined, complementary measurement methods such as FTIR and GC/FID can provide characterisation of a larger set of compounds than each method separately. This study provided measurements of the compositional variability associated with oxidative pyrolysis gases that can be used in future comparisons.

Once approved for release by the sponsor of this study (SERDP), the data will be available through the Forest Service Research Data Archive https://www.fs.usda.gov/rds/archive/.

The authors declare no conflicts of interest.

This work was supported by the DOD/DOE/EPA Strategic Environmental Research and Development Program project RC-2640. JPA was supported by the project PID2021-123833OB-I00, funded by the Spanish Ministry of Science and Innovation (MCIN/AEI/10.13039/501100011033) and by ERDF A way of making Europe. USDA Forest Service annual research appropriations provided in-kind support from 2016 to 2022.