CO 2 Surface Variability, from the Stratosphere or Not?

. Fluctuations in atmospheric CO 2 can be measured with great precision and are used to identify human-driven sources as well as natural cycles of ocean and land carbon. One source of variability is the 10 stratosphere, where the influx of aged CO 2 -depleted air can produce fluctuations at the surface. This process has been speculated a potential source of interannual variability (IAV) in CO 2 that might obscure the quantification of other sources of IAV. Given the recent success in demonstrating that the stratospheric influx of N 2 O-and chlorofluorocarbon-depleted air is a dominant source of their surface IAV in the southern hemisphere, we here apply the same model and measurement analysis to CO 2 . Using 15 chemistry-transport modeling or scaling of the observed N 2 O variability, we find that the stratosphere-driven surface variability in CO 2 is at most 10% of the observed IAV and is not an important source. Diagnosing the amplitude of the CO 2 annual cycle and its increase from 1985 to 2021 through the annual variance gives rates similar to traditional methods in the northern hemisphere (BRW, MLO), but can identify the emergence of small trends (0.08 ppm decade -1 ) in the southern hemisphere (SMO, CGO).


Introduction
The surface abundance of CO2, a.k.a. the Keeling Curve (Figure 1a), is used as the prime example of the human-driven increases in greenhouse gases. It also used to demonstrate control of CO2 by the land biosphere and the oceans through its annual cycles and interannual variations (Le Quéré et al., 2016;25 2018). The inverse modeling of surface sources based on these CO2 observations is used to infer regional sources of fossil fuel emissions as well as year-to-year changes in primary productivity of the biosphere or oceanic degassing (e.g., Gurney et al., 2002;Baker et al., 2006;Engelen et al., 2006;Nassar et al., 2011;Peylin et al., 2013;Frankenberg et al., 2016;Pandey et al., 2016;Nakazawa, 2020). There is concern that atmospheric variations in CO2, and hence the net sources derived from them, may be affected 30 by interannual variations (IAV) in tropospheric mixing or stratosphere-troposphere exchange (STE) (Gaubert et al., 2019), but there are no definitive studies. For example, Nazakawa's (2020) review of greenhouse gas studies mentions the stratosphere only in connection with CH4 and N2O, not with CO2.
The possibility of a true STE-driven IAV CO2 signal, raised by Gaubert et al. (2019), has not been 35 seriously investigated. For the most part, when studies investigate the stratospheric influence on CO2 source inversions, they are not concerned about STE fluxes directly but other factors that degrade the results: e.g., gradients across the tropopause, the effective tropospheric air mass diluting surface emissions, or the inclusion of CO2-depleted stratospheric air in column CO2 calculations (Nassar et al., 2011;Deng et al., 2015;Frankenberg et al., 2016;Pandey et al., 2016). For example, Le Quéré et al. 40 (2018) are concerned how emissions will mix throughout the troposphere and the stratosphere, but not how stratospheric air will come back down to the surface. Only studies of the CO2 triple-oxygen isotope signature (Δ 17 O) are concerned with accurate STE fluxes, recognizing its importance in the seasonal isotopic signals (Liang et al., 2017;Koren et al., 2019;Laskar et al., 2019).
Recent work has shown that the stratospheric quasi-biennial oscillation (QBO) modulates the STE and drives much of the IAV observed in surface N2O through the stratospheric influx of N2O-depleted air Ruiz and Prather, 2021). It also drives some surface variability in CFCl3 (Ray et al., 2020). Here, we use the N2O studies of Ruiz et al. (2021) with parallel model simulations of CO2 to place constraints on the CO2 IAV caused by atmospheric circulation, finding that this effect is a clear but minor 50 perturbation in driving the observed IAV.

Methods and Analysis
We investigate the CO2 IAV through the historical period 1990-2017 with UC Irvine chemistry-transport 55 model (CTM) simulations and also with surface N2O observations, which are known to be influenced by STE fluxes.
To study the circulation-driven IAV of CO2, including STE, we focus on the southern hemisphere (SH) because small fluctuations in the large biosphere-driven seasonality in the northern hemisphere (NH) 60 ( Figure 1ab) will obscure any stratosphere-driven IAV. The UCI CTM uses ECMWF Integrated Forecast Fields at 1.1° horizontal resolution and has proven quite successful in simulating the historical IAV of surface N2O, ozone columns, and the Antarctic ozone hole Ruiz and Prather, 2021;Tang et al., 2021). For CO2, we develop two model scenarios to highlight the impacts of atmospheric transport. First, we define an emissions-driven, eCO2 scenario, in which the total atmosphere increases at 65 a constant rate of 2 ppm (parts per million, dry-air mole fraction) y -1 , forced by uniform surface emission over all land and ocean areas from 20° N -60° N. This eCO2 scenario is a simple experiment to test how atmospheric circulation driving a NH-SH gradient might affect the seasonal and interannual variability of SH surface CO2; it is obviously not realistic, lacking the large NH seasonality. A second, stratosphericdriven, sCO2 scenario, is forced with a net stratospheric flux of CO2-depleted air being transported into 70 the troposphere and down to the surface. This STE flux is calculated as the equivalent of the aging of stratospheric CO2 relative to the troposphere (2 ppm y -1 ), yielding an apparent negative CO2 flux of 0.8 PgC y -1 . This forcing flux is placed in the uppermost model layer (~80 km altitude) and transported to the surface. In both of these cases, CO2 is changing linearly with a known trend, and we subtract that trend to get the modeled anomalies. A third, independent method uses the observed SH surface N2O signal, 75 driven by stratospheric photochemical loss of 13 TgN (as N2O) y -1 , as a measure of STE influence. In this case we scale the results to CO2 using the ratio of the STE fluxes, i.e., 0.15 ppm CO2 per ppb N2O.
The monthly CO2 surface observations are gathered from NOAA ESRL (Dlugokencky et al., 2021a). We use 5 sites: BRW = Barrow AK, 71° N, 156° W; MLO = Mauna Loa HI, 20° N, 156° W; SMO = Tutuila, 80 Am. Samoa, 14° S, 171° W; CGO = Cape Grim, Tasmania, Australia, 41° S, 144° E; and SPO = South Pole, 90° S. Monthly average in situ observations are used with gaps filled by flask data at the same site. CGO is flask only. We have a continuous monthly record from 1985 through 2020 ( Figure 1a). We convert these to a stationary series of residuals by fitting polynomials: using 2nd, 3rd, 4th, and 5th order polynomials produced are almost identical results for any site (not shown), and we averaged the residuals 85 from the 3rd and 4th order fits. Results for the average of SPO+CGO are shown in Figure 1c as the red line, which shows a clear annual cycle plus equally large variability on decadal scales. The annual cycle (min to max) in CO2 and its rate of change is a critical metric used to evaluate the carbon cycle in Earth system models (Graven et al., 2013;Zhao and Zeng, 2014;Wenzel et al., 2016). Here, we calculate the cycle simply by averaging each calendar month of the year as shown in Figure 1b These are consistent with those previous studies, although SH cycles remain understudied and not well evaluated. In some months SMO at 14°S can be north of the South Pacific Convergence Zone and thus influenced by NH air, explaining its non-sinusoidal cycle when compared with SPO and CGO. Also shown is the annual cycle for the modeled eCO2 scenario (~0.18 ppm, dotted lines for SMO and SPO), 95 while that for the sCO2 scenario is even smaller (~0.06 ppm) and not shown. It is interesting that the SH annual cycles in eCO2 are similar in shape to those observed, even catching the double peak at SMO, but the magnitude is much smaller. There is no evidence in our direct modeling or analysis that stratospheretroposphere exchange, which does have an annual cycle in N2O, can produce a detectable annual cycle in CO2 above the large observed cycle. 100 Because of the clearly observed QBO signal in SH surface N2O , we look for a CO2 IAV signal in the 2-5 year range, removing the large 6-10 year IAV with a high-pass filter after removing the annual cycle with a low-pass filter (see Figure 1c for details). The resulting SPO+CGO IAV signal is shown as a thick black line in Figures 1c & d. In the SH where the annual cycle is small, 105 the low-pass filter clearly removes any vestige of an annual cycle; but in the NH, especially BRW, the annual cycle is large and slightly irregular in amplitude and thus the low-pass filter is unable to remove all of it.
For monthly N2O surface observations, also from NOAA ESRL (Dlugokencky et al., 2021b), we focus 110 only on SH mid-and polar latitudes, using SPO, CGO, plus 3 other sites: S30 = Western Pacific Cruise, 30° S, 168° E; USH = Tierra del Fuego, Ushauia, Argentina, 55° S, 68° W; and PSA = Palmer Station, Antarctica, 65° S, 64 °W. All five sites have nearly identical N2O records, and we average them to get our SH IAV signal with the same processing as for CO2. The QBO circulation is known to reach throughout the stratosphere and into the troposphere (Tung and Yang, 1994;Hamilton and Fan, 2000), 115 and multi-model studies have attributed the N2O IAV to the modeled STE flux . We can thus scale the surface N2O IAV with the ratio of STE fluxes (CO2:N2O) to give an observational estimate of the STE-driven IAV in CO2 (dashed blue line in Figure 1d). The IAV in SH (40° S -90° S) surface CO2 calculated from the sCO2 model scenario is also shown (dashed red line in Figure 1d). The modeled sCO2 and observed N2O-scaled IAVs are not always phase, but they are in strong agreement in 120 terms of amplitude: the STE IAV in CO2 is small, at most 10% of the observed IAV in the 2-5 year range. Similarly, the modeled eCO2 IAV shows that tropospheric circulation changes do generate IAV, but these are at most 12% of that observed.
We compare CO2 with other interannual cycles in the Earth system by plotting in Figure 1d (i) the QBO 125 phase change (from easterly to westerly zonal equatorial wind at 40 hPa, as thick gray vertical bars, see Newman, 2021) and (ii) the times of moderate to extreme El Niños (red stars) and La Niñas (blue stars) (Trenberth, 2021). From this analysis, we expect minimal contribution of the QBO-driven circulation to the CO2 IAV, and find no obvious additional correlation between the two. For the El Niño-Southern Oscillation (ENSO), this simple comparison is inadequate. At best it shows that the larger positive SH 130 IAV follows the El Niños; whereas we know that ENSO affects ocean upwelling and continental rainfall and the CO2 anomalies correlate very well with tropical ocean temperatures (Wang et al., 2021;Keeling and Graven, 2021). We have shown that the STE fluxes of old stratospheric air with "depleted" CO2 have little influence on the IAV or annual cycle. Similarly, the atmospheric transport of NH fossil fuel emissions into the SH produces signals that are 10-15% of what is observed. Thus the speculations of Gaubert et al. (2016) regarding atmospheric transport can be dismissed. 140 The latitudinal pattern of N2O IAV provides evidence for causes: e.g., the STE-driven signal weakens in the tropics and changes phase in the NH; QBO composites show a clear separation of hemispheric sources . For CO2 the 2-5 year IAV has similar amplitude but some phase shifts within the SH (SPO+CGO vs. SMO, solid and dotted black lines in Figure 1d). To expand on this, we reprocessed the 145 monthly CO2 data with a simpler algorithm to avoid possibly spurious signals from the band-pass filtering. We recognize that many options are available (see review in Piao et al., 2020). First, a polynomial fit to each series as defined above is removed. With 432 equally-weighted monthly values, the polynomial fit is extremely well determined and gives residuals and a stationary time series like the red curve in Figure 1c. Second, a 12-month running mean removes most of the annual cycle and gives 150 421 monthly values from 1 July 1985 to 1 July 2020. This processing leaves the 2-10 year cycles, and the resulting IAVs for MLO, SMO, CGO, and SPO are shown in Figure 1e. The BRW signal is very noisy and not shown here. At first glance, the sites are similar. This agreement is expected because most of this signal occurs over time scales >2 years and the interhemispheric mixing time scale is ~1 year. On time scales <2 years we find many instances where the amplitude is distinct by hemisphere or the phase is 155 shifted by 6-12 months (the interhemispheric mixing time). When the four sites are in synch, one can only presume that the CO2 perturbation is tied to changes in the growth/decay of tropical biomass transported equally to both hemispheres (Keeling and Graven, 2021). When the IAV change is opposite (e.g., years 2008, 2010, 2018) the CO2 perturbation likely occurred distinctly in one hemisphere, but it does not identify which. In some periods the NH obviously leads (e.g., 1993, 1998, 2003) or lags (e.g., 160 1991, 2002, 2015) the SH by about a year, again implying a single hemispheric origin, but now identifying the hemisphere because the global multi-year trend is clear.

Conclusions, Speculations, and Digressions
The Samoan site SMO provides a valuable but very challenging record of CO2 and other trace gases with dominant NH emissions (e.g., N2O, SF6, chlorofluorocarbons). As seen in Figures 1de, sometimes SMO 165 is synchronous with the SH mid-and polar-latitude sites (CGO and SPO, which are almost always synchronous with each other) and at other times it links with MLO and the NH. Thus, to use SMO CO2 as a metric for carbon cycle models, one must recognize that SMO is not representative of the SH tropics. When evaluating carbon cycle models, one must be sure that the tracer transport in these models has the correct IAV for SMO. On interannual time scales there are clear times when SMO aligns as SH and is 170 distinct from MLO (e.g., 1986MLO (e.g., , 2002MLO (e.g., , 2010MLO (e.g., , 2018 and times when it aligns with MLO and is distinct from CGO+SPO (e.g., 1999, 2012, 2015. These interannual shifts provide an excellent test for CO2 historical simulations using weather forecasting systems (e.g., McNorton et al., 2020) and realistic sources and sinks (e.g., Piao et al., 2018;Wang et al., 2020). 175 A by-product of this alternate method is a cleaner, statistical measure of the annual cycle. Each of the 421 running-mean quantities in Figure 1e consists of 12 months, and each contains an annual cycle. If the cycle is sinusoidal, then the min-to-max amplitude is equal to twice the square root of 2 times the standard deviation. With this approximation, we calculate a mean amplitude of 17.0, 6.4, 1.4, 1.2, and 1.2 ppm for BRW, MLO, SMO, CGO, and SPO, respectively. These results are almost identical to those 180 from the composited annual cycles plotted above, but disagree at SMO as might be expected because of its double-peaked annual cycle. A linear fit to the standard deviations gives trends for the period 1985-https://doi.org/10.5194/esd-2021-98 Preprint. Discussion started: 1 December 2021 c Author(s) 2021. CC BY 4.0 License. 2020 of 0.11, 0.014, 0.008, 0.008, and 0.003 ppm y -1 for BRW, MLO, SMO, CGO, and SPO, respectively. In terms of % y -1 the increases in annual cycle amplitude are 0. 62, 0.22, 0.62, 0.67, and 0.25. Are these trends significant? Without trying to assess the number of degrees of freedom in the time 185 series, we simply compare the 36-year change in annual cycle amplitude (signal) with the standard deviation of the residuals after the fit (noise). This ratio is about 4:1 at BRW, making the change in annual cycle obvious to the eye; the ratio is 1:1 for MLO, SMO and CGO, making for a solid detection; but the ratio at SPO is much less because the trend in annual amplitude is much smaller. Our results for BRW and MLO agree well with other more extensive data analyses (Graven et al., 2013;Zhao and Zeng, 190 2014; Wenzel et al., 2016;Piao et al., 2018;Wang et al., 2020), but provide subtler tests of the seasonality and trends in the SH, which is not often used for model evaluation. A more serious uncertainty analysis focusing on the SH sources and sinks, the annual and IAV cycles, and their trends would help solidify our knowledge of the carbon cycle. 195 Code availability. The full chemistry-transport model (CTM) code used for the tracer simulations is included in the archived data. the emissions-driven scenario (eCO2) are shown as dotted lines for SMO and SPO (same color coding for each site) The annual cycle in eCO2 has similar phasing at SPO and is double-peaked at SMO, but the amplitudes (~0.18 ppm) are much smaller and barely detectable on this scale. The stratosphere driven scenario (sCO2) has an even smaller amplitude (~0.06 ppm). The atmosphere-only transport model uses historical meteorology from the ECMWF IFS system. The eCO2 scenario is driven with a uniform surface flux from 20°N to 60°N equivalent to 2 ppm annual increase in CO2 (~4.2 PgC/y). The sCO2 scenario is driven 335 by a negative CO2 flux from the mesosphere that is equivalent to 2 ppm/y loss over the stratosphere only (~0.8 PgC/y).
(c): Residual surface CO2 variability (ppm) averaged over two primary southern extra-tropical stations (SPO+CGO, red line) after removal of a polynomial fit (average of 3rd and 4th order fits). Polynomial fits of 3rd, 4th, and 5th-order produce nearly identical CO2 residuals. SMO, a tropical station, is not used because it has very different annual cycles, see (b) above. This 340 residual curve shows evidence of annual to decadal variability. The interannual variability (IAV, black line) is calculated by a sequence of high-pass (Matlab: highpass(series, 0.32, 12)) followed by low-pass (Matlab: lowpass(series, 0.30, 12) filters leaving variability with periods of 2-5 years to focus on the stratospheric quasi-biennial oscillation (QBO). The low-pass filter results for these SH sites are almost identical to a 12-month running mean series. https://doi.org/10.5194/esd-2021-98 Preprint. Discussion started: 1 December 2021 c Author(s) 2021. CC BY 4.0 License.