Observational Evidence of Large Contribution from Primary Sources for Carbon Monoxide in the South Asian Outflow

South Asian air is among the most polluted in the world, causing premature death of millions and asserting a strong perturbation of the regional climate. A central component is carbon monoxide (CO), which is a key modulator of the oxidizing capacity of the atmosphere and a potent indirect greenhouse gas. While CO concentrations are declining elsewhere, South Asia exhibits an increasing trend for unresolved reasons. In this paper, we use dual-isotope (δ13C and δ18O) fingerprinting of CO intercepted in the South Asian outflow to constrain the relative contributions from primary and secondary CO sources. Results show that combustion-derived primary sources dominate the wintertime continental CO fingerprint (fprimary ∼ 79 ± 4%), significantly higher than the global estimate (fprimary ∼ 55 ± 5%). Satellite-based inventory estimates match isotope-constrained fprimary-CO, suggesting observational convergence in source characterization and a prospect for model–observation reconciliation. This “ground-truthing” emphasizes the pressing need to mitigate incomplete combustion activities for climate/air quality benefits in South Asia.


Supplementary Notes Note S1. A discussion on the isotopic fractionation of CO produced in vehicular emissions
The isotopic signatures of CO emitted from fossil fuel-based traffic emissions show are quite variable in both the  13 C and  18 O dimensions (see Figure S7). Central to these variations are petrol vs. diesel emissions as well as catalytic processes [1][2][3] . The vehicle-to-vehicle variation of emitted CO and catalytic efficiency are judged using, for instance, CO:CO2 and H2:CO ratios, respectively 3 . While diesel engines have a higher combustion efficiency, thereby largely oxidizing CO to CO2, petrol engines run at the stoichiometric point with just enough oxygen to burn all fuel, resulting in high CO emissions. A three-way catalytic converter (TWC) is used in petrol engines, in part, to oxidize this CO. The TWC performs sub-optimally when it is not sufficiently heated (referred to as cold start emissions) and during lack of enough O2. These conditions affect the isotopic signatures of CO produced in the engine 3 .

Variations in δ 13 C
A clear distinction is found in CO sampled from individual stationary vehicles tested by varying parameters such as engine status (e.g., idling, revving), load, and speed (δ 13 C=-26±12‰; δ 18 O=25±7‰) vs. from a fleet of moving vehicles (δ 13 C= -27±2‰; δ 18 O=19±5‰) (see Figure S7). It is noteworthy that mean δ 13 C of a fleet in different urban locations, highways and tunnels are similar, whereas the spread in δ 18 O is larger. The spread in δ 13 C also gets muted to a large extent when comparing fleet with individual vehicles. High-emitting vehicles (e.g., cold petrol engines) with a large spread in CO:CO2 show a gradual enrichment in δ 13 C (relative to that of the fuel) with the oxidation of CO and therefore contribute significantly to the spread in the isotopic signatures when tested individually in stationary conditions 1,3 . Likewise, some vehicles with extremely low CO:CO2 have also shown a depleted δ 13 C relative to that of the fuel, indicating complexity of emission systems 3 . However, this effect is completely subdued in parking garage (low-speed cold-engines) vs. highway (high-speed hot-engines) comparison of a fleet of moving vehicles, implying i) the difference in driving conditions does not result in a significant difference in the integrated 13 CO/ 12 CO composition, ii) despite the different regimes, the overall isotopic signatures of CO in traffic possibly are dominated by the isotopic signatures of CO from the high-emitting vehicles, iii) the low CO:CO2 scenarios (depleted δ 13 C) of vehicular emissions do not affect the overall traffic signature 3 . Since nearly all carbon leaves the vehicle as CO2, for a large range of moderate CO:CO2 ratios, it is reasonable to assume that the 13 CO is closer to or slightly enriched than the C-isotopic signature of the fuel (~ -30 to -26‰) 4 .

Variations in δ 18 O
Oxygen leaves the vehicle as H2O, and since CO2 can undergo isotopic exchange with H2O, the δ 18 O deviates from atmospheric oxygen (23.9‰) 5 . A conspicuous observation from the comparison of individual stationery vehicular emissions and fleet emissions of CO (see Figure S7) is that diesel engines show isotopically depleted C 18 O compared to petrol engines. This could be attributed to the combustion efficiency of diesel engines. The kinetic isotope effect (KIE) in the CO+OH· reaction (i.e., destruction of CO) induces a negative enrichment of (upto ~-10‰) in 12 C 18 O, implying the residual CO will be depleted in δ 18 O 1,6 . Nonetheless the shift in δ 18 O is still uncertain in petrol engines, wherein a positive enrichment in 18 O (normal KIE) has also been found during the destruction of CO in heavier Page S4 engines as well as negative enrichment (inverse KIE) in certain smaller engines 1,3 . However, one aspect determining the shift is the presence/absence of a catalytic converter as well as the metal surface of the catalytic converter which have both shown to cause large variations in C 18 O, respectively 1,7 . Based on the observations (in Figure S7), it is found that cold diesel emissions often form a distinct isotopic cluster compared to cold petrol emissions. Taken together, the reasons for the spread in δ 18 O are not well known and possibly related to the several factors including the engine-size, fuel-type, vehicle age 1,3,7-9 . Overall, it is reasonable to assume that the C 18 O is often slightly depleted than atmospheric O2.
Page S5 Note S2. South Asia-specific endmember for CO from fossil fuel combustion Emissions from traffic constitute a major fraction of the fossil fuel usage in South Asia 10 . The vehicular fleet, in this region, can be grouped into 2-wheelers (2-W), 3-wheelers, 4-wheelers, low duty diesel (LDDV) and high duty diesel vehicles (HDDV). 2-W vehicles make up the largest market stock (~78%) 11 . Over the past decades, the share of the two-stroke engines (high CO emitters) in 2-W and 3-W vehicles has drastically reduced, from 80% in 1990s to <5% post-2010 12 . Thus, the overall traffic-CO signature is most likely dependent on other factors, for instance, vehicle age.
A category of vehicles across all vintages that contribute disproportionately to pollutant emissions, known as superemitters, is established as 20% of all vehicle fleet (as a function of vehicle age) for South Asia 13 . Given the size of the fleet in this region, the superemitters account for as much as 20 million in 2-W and 4-W each, ~1.5 million in LDDVs and HDDVs 10 . This is much higher than in some parts of the US and Europe. For 2-W a large fraction of the fleet (~50%) is found to be older than 10 years in terms of vehicle age. This is slightly lower for 4-W and HDDVs.
The overall average fleet age of vehicles in South Asia (~13 years) is much higher than in countries in Europe and N. America 10 . A higher CO:CO2 ratio is found with increasing fleet age 14 , implying that the superemitters might have a much larger role in defining the overall traffic CO isotopic fingerprint in South Asia.
The largest share of PM2.5 (upto 75%) and black carbon (BC; upto 95%) emissions is attributed to diesel vehicles in South Asia 10 . A major chunk of this share is indeed found to be from HDDVs and superemitters (up to 80%). Using a conservative BC/CO ratio (0.01 µg m -3 / µg m -3 ), we find the total CO emissions from HDDVs and superemitters alone to be as high as ~14 Tg/yr, an overwhelming portion of the total fossil fuel combustion-derived CO estimate (~15.5 Tg/yr) 15 . This implies that diesel vehicles likely dominate the South Asian CO fingerprint from traffic emissions and possibly from the overall combustion of fossil fuel, respectively. As discussed in Note S1, the diesel engines are found to have distinct isotopic signatures compared to petrol engines, related to a more complete combustion process. Given that there are no isotopic studies of vehicular CO emissions from South Asia, we here establish the fossil fuel combustion endmember by averaging the mean of all fleet-based vehicular emission studies worldwide: δ 13 C= -27.8±1.5‰; δ 18 O=19.2±4.9‰ (see Table S4). In this approach, we find the derived δ 18 O is indeed closer to the C 18 O found in diesel emissions, and thus representative of South Asian traffic emissions 1,3 . We do not take the individually tested stationary vehicles (including testbench emissions) into account due to the wide range of isotopic signatures, inconsistent behaviour among vehicles and a lack of established systematic drivers for the observed emission signatures reported in literature.

Note S3. A discussion on the isotopic fractionation of CO produced in biomass burning
Only a handful of studies so far have investigated the isotopic composition of CO in biomass burning emissions 7,[17][18][19] . Results from controlled burn experiments and ambient samples have differed in magnitude of isotopic fractionation. Typically, fractionation effects are different for two main type of plants with distinct isotopic compositions i.e., C3 and C4 plants. δ 13 C is on average -27.1±0.2‰ for C3 plants and -13±1.2‰ for C4 plants 16  A strong decreasing trend has been observed in both isotopes over time from ignition. With a 30% drop in MCE the difference/depletion in δ 13 C and δ 18 O can be up to ~ -18‰ and ~21‰, respectively. Two clear groups of isotopic clusters are evident for CO based on burning phases: flaming and smoldering 17,18 . The flaming phase (defined as ~96±4% MCE) is accompanied by an enrichment of ~6‰ in 13 C relative to the δ 13 C of the fuel (~25.5‰) . The isotopic fractionation is weaker in the smoldering phase (defined as ~87±6% MCE) with a depletion of ~2‰ in 13 C relative to the fuel. A similar trend is found in δ 18 O. However, compared to atmospheric O2 (δ 18 18 O is found to be ~5‰ enriched in the flaming phase and ~9‰ depleted in the smoldering phase 17 . While the 13 C is close to that of the plant material in initial starting phase, the 18 O partly depends on oxygen isotopic composition in plant cellulose which is determined by that of meteoric water and highly dependent on relative humidity 18  Weak correlations between fuel moisture content and isotopic ratios have been found in indoor burn experiments, implying that dry and wet fuel-type would not show any difference in MCE. When taken together, the average isotopic composition during a whole combustion process is found to be closer to the smoldering phase emitted CO (i.e., depleted in δ 13 C and δ 18 O). This is also supported to ambient observations e.g., during wildfires in USA 7 and biomass burning influenced winter season in Europe 19 .
Both C3 and C4 plants have shown similar characteristics during combustion. The derived δ 13 C in CO from C4 plant burning is clearly more isotopically enriched than from C3 plant burning, however, the δ 18 O in CO are found to be overlapping (see Table S4). The C 18 O emitted in flaming phase of C4 plant burning are found to be similar to C 18 O from smoldering phase of C3 plant burning. Given this complexity it is not always possible to distinguish C3 and C4 burning by CO isotopes.
In South Asia, the biomass burnt can be classified into two sectors: open burning (crop residue, forest fires, garbage), domestic burning (agricultural residues, firewood, dung cake) 20 . While C3 and C4 plant burning would contribute with certain proportions, it is challenging to estimate the isotopic signatures of CO emitted from each of these activities in their respective sectors. A lack of CO isotopic studies in this region further complicates the matter.
This way we are able to account for most biomass types. This endmember is within the range of CO isotopes in the smoldering phase (characterized in indoor burn experiments) and thus representative for overall combustion process.
Page S8 Note S4. A discussion on the Keeling-plot approach A special case of the isotope mass balance is a gas sample taken as the mixture of two sources, namely background air (bgd, mbgd), and a pollutant (p, mp). The isotope ratio of the gas mixture is then given by (1) This equation can be rearranged as: Taking the background concentration, the background isotope ratio, and the isotopic composition of the pollutant as constant, Eq. (2) is linear in 1/mmixture with y-intercept p: This equation is very useful because it enables deriving the isotope value of the "pure" pollutant (p) from a regression of the measured isotope ratios of air samples as a function of the inverse of the measured concentrations, without any further knowledge required about mixing ratios or isotope ratio of the background 21 . With some limitations, Eq. (3) is also applicable for more complex mixtures, assuming e.g. the pollutant to be itself a mixture of two pollutants. The y-intercept can then be interpreted as the isotope ratio of the pollutant mixture, but the linearity of Eq. (3) does not strictly hold when the pollutant composition is variable.
Page S9 Note S5. A discussion on the background CO levels at MCOH By definition, "background" CO signal at MCOH would refer to the [CO] encountered in periods devoid of continental influence and/or long-range transport. Given that MCOH receives air masses spanning a large geographical domain, the [CObackground] would vary both temporally and seasonally. Apart from this, the isotopic composition of the CObackground would also vary with the kinetic isotope effect (KIE) induced due to atmospheric oxidation by hydroxyl radicals (OH·). It is thus important to establish the [CObackground] first for the current period of interest and then discuss the implications of scavenging process on the signal (details in Note S6).
The histogram of [CO] at MCOH for the winter campaign (see Figure S3 b) shows that the lowest [CO] is estimated to be ~70 to 75 ppb (as established by 5% percentile). This also corresponds to periods with the lowest black carbon (BC) concentrations as well lowest particle count suggesting limited influence from the continental outflow (see to the overall mixing ratios at MCOH or when apportioned for the contribution to the S Asiasource (see Fig. 3 in main manuscript). Hence, we exercise caution in choosing the background, and based on observed changes in the ancillary aerosol parameters (in line with previous observations in the region) find it reasonable to assume the value of ~70 to Page S10 Note S6. A theoretical model accounting for the effect of scavenging process on the background CO signal at MCOH CO is scavenged from the atmosphere mainly by homogeneous gas phase oxidation reaction with OH· radicals. This reaction also induces a kinetic isotopic fractionation (KIE) which plays an important role in altering the isotopic signature of CO emitted from various sources. Here we develop a relationship between the isotopic composition of background CO (COback,MCOH; intercepted at MCOH) and the background source signature (COback,source ; the initial starting signal of the background) by accounting for the fractionation effect (KIE) of the scavenging processes, in particular the reaction with OH·. Assuming steady-state, we have: J = CO flux (e.g., g m -3 s -1 ); k = reaction rate coefficient (e.g., s -1 ); [CO] = CO concentration (e.g., g m -3 ).
The isotope-ratio (R) can be expressed as: The kinetic isotope effect (KIE) is defined as: The isotope-ratio of the source can be estimated as the ratio of the fluxes: The observed isotope-ratio then relates to the source ratio and KIE as: Introducing the -scale (in per mill, where std is the isotope-ratio of the standard): Setting source is the CO isotopic signature of the background source (in this case COback,source) and obs is the measured CO isotopic signature in ambient air at the sampling site (in this case COback,MCOH) Page S11 and rearranging:  Figure S8.
It should be noted that the background signal would move along this line based on the extent of KIE.
Page S12

Note S7. Source apportionment based on a hierarchical Bayesian statistical model
The hierarchical Bayesian model accounts for Keeling fit, isotope mass balance and primary endmember distribution.
Here, each is discussed in detail and the relevant steps in the formulation of the model are outlined:

Keeling fit
The good correlations of both isotope signatures ( 18 O: R 2 = 0.95 and  13 C: R 2 =0.81; Figure 2 in the main manuscript) with the inverse CO concentrations (i.e., the Keeling relation), suggests that the isotope variability at MCOH may be described by a two-state mixture: a constant background and a temporally variable source. While the background is likely to be affected by kinetic isotope effects (KIE), this is not expected to be the case for the isotope signature in the limit of [CO]→∞, as this would correspond to the "source" signature. As such, the values in this limit ( 18 OSAsia;

Isotopic mass balance
In this study, our aim is to estimate the fractional (f) contributions from primary and secondary CO to the South Asian continental CO emission. Assuming isotopic mass balance, we have:

Primary endmember distribution
The primary endmember distribution is here represented as a mix of three different primary source components.
However, the relative contributions of these are uncertain. But we do have prior information, e.g., from bottom-up emission inventories of primary CO emissions. This information may then be used as a prior in a Bayesian framework. Bayesian calculations of the posterior of fractional, a Dirichlet prior is often employed: Where, f are the fractional source contributions (here n = 3, representing C3, C4 and fossil) and, N() is the normalizing function: Where  is the gamma function.
The degree of prior information is encoded in the exponents, i. If all alphas are set to one, we have a flat prior, essentially no prior information. The mean relative contribution for a component i (i), is related to the exponents as: Thus, one may directly connect the exponents to the prior knowledge, e.g., from the bottom-up emission inventory.
Page S14 However, Eq. (17) is an under-determined system in itself: we cannot un-ambiguously solve for I using only known means. To estimate the alphas, we therefore need additional constraints.
The variance ( 2 ) for the dimensions i of a Dirichlet distribution is given by: We note that the variance for each individual dimension, i, as well as the total variance (∑

=1
), depend on Since bottom-up emission inventories often report high uncertainties (e.g. 125% or higher) 47 , and since such uncertainty estimates are also often not available/estimated, we here use a prior with minimum prior constraints. The least informed prior, while still retaining prior information regarding mean contributions, should then maximize the variance; minimize the sum of i. We here assume that the prior distribution is mono-modal. This means i ≥1, for all i, while k >1 for at least one k. We then have the following optimization problem: we need to find the smallest sum of alphas, such that all values are equal or larger than one. The smallest sum is obtained when at least one of the alphas equals 1. Since all other alphas need to be equal or larger than 1, this means that the alpha that equals one (k = 1) need to correspond to the smallest mean (k = min(1, 2, …, n,)). We arrive at the following parametrization: We note that in this formulation with three dimensions, i = 1/3 for all i, equates to I = 1 for all i; the un-informed prior (unit simplex).

Posterior probability density function
The Bayesian model used here to compute the relative contributions from primary and secondary sources to CO in the South Asian continental emissions then relies on a number of modules, that are combined into the model. We can summarize the posterior probability density function as: Is assumed normal, fulfilling isotopic mass-balance. The constraints provided by the multiple data points (i) are implemented through the product, as in Eqn. (14a-b), The prior for the slopes: ( 18 ; 13 ) Was assumed normal, while the estimator for the variability inverse gamma (see above for details).
Page S15 The prior for the fractional source contributions to the primary endmember: ( 3 ; 4 ; ) Was assumed Dirichlet distributed (see above for details).
And the prior for the fractional contribution of primary: ( ) Was assumed uniform.
The MCMC simulations were implemented using a Metropolis-Hastings algorithm 50,51 , implemented in MATLAB version 2015b code with 1000.000 iterations with an initial burn-in phase of 10.000 (to assure proper annealing prior to estimation of the probability density functions) and a data thinning of 100 (to remove correlations between iterations) 25 .
Page S16

Note S8. Establishing an informed prior using emission estimates from bottom-up inventories
The main primary sources include burning of C3-biomass (e.g., wood; and agricultural waste burning of wheat and rice), burning of C4-biomass (e.g., agricultural waste burning of sugarcane, millet and maize) and fossil fuel combustion.
Biomass burning can be divided into two categories: crop residue burning and residential/domestic use. For crop residue burning, the fraction of C3-vs C4-plant residue generated, and residue burnt was first estimated and the CO produced from each was apportioned from estimates of total CO 48 . This was ~7 Tg/yr for C3 plants and ~2 Tg/yr for C4 plants.
The total CO estimated from an Asia-specific bottom-up emission inventory 47 ~62 Tg/yr. This was then divided into CO from various sectors and grouped into two categories: Biomass burning-CO and Fossil fuel combustion -CO.
The estimates for total biomass burning CO was ~43 Tg/yr and fossil fuel combustion was ~19 Tg/yr.  49 . However, the corresponding changes in the distribution of CO between various sectors is unclear and therefore in this study we relied on the 2008 estimates.

Supplementary Figures
Page S27  Table S3) and only the Mean ± SD are shown here. Also reported are the isotopic signatures for background CO (see Note S5; Figure S8). A discussion on fractionation effects can be found in Notes S1-S3.    Table S1), which is a censored version of the full dataset. The censoring is primarily done for fossil fuel combustion-and Biomass burning-CO isotopic signatures and is based on the arguments in Note S1-S3 (see also Fig. S7). The dataset with all "raw" data (complied from all published CO studies) will be available on the Bolin Centre Database (http://bolin.su.se/data/Dasari-2021). Note that the uncertainties in the isotopic signatures as well as the isotopic fractionation reported for the corresponding measurements is accounted in this compiled version.

Supplementary Tables
Note: The superscripts in the