On the Detection of COVID-Driven Changes in Atmospheric Carbon Dioxide.

We assess the detectability of COVID-like emissions reductions in global atmospheric CO2 concentrations using a suite of large ensembles conducted with an Earth system model. We find a unique fingerprint of COVID in the simulated growth rate of CO2 sampled at the locations of surface measurement sites. Negative anomalies in growth rates persist from January 2020 through December 2021, reaching a maximum in February 2021. However, this fingerprint is not formally detectable unless we force the model with unrealistically large emissions reductions (2 or 4 times the observed reductions). Internal variability and carbon-concentration feedbacks obscure the detectability of short-term emission reductions in atmospheric CO2. COVID-driven changes in the simulated, column-averaged dry air mole fractions of CO2 are eclipsed by large internal variability. Carbon-concentration feedbacks begin to operate almost immediately after the emissions reduction; these feedbacks reduce the emissions-driven signal in the atmosphere carbon reservoir and further confound signal detection.

Introduction

Falling energy demand during the COVID‐19 pandemic led to rapid decreases in energy‐related carbon dioxide (CO2) emissions. In 2020, global annual CO2 emissions fell by 7% to 2011 levels (9.3 Pg C yr−1; Friedlingstein et al., 2020), and the rapid decline in emissions during the first half of 2020 surpassed the rate of emission declines during any previous economic recession or World War II (Forster et al., 2020; Friedlingstein et al., 2020; Le Quéré et al., 2020; Liu et al., 2020). Global annual CO2 emissions are forecast to remain below 2019 levels through 2021, and subsequent recovery of emissions is expected within a few years (International Energy Agency, 2021; Le Quéré et al., 2021). The precipitous and short‐lived drop in emissions during the COVID pandemic offers a unique opportunity to assess the detection of these types of emissions declines in observations of the global carbon cycle.

While the COVID‐related CO2 emissions reductions had a measurable impact on regional atmospheric CO2 concentrations (Buchwitz et al., 2021; Chevallier et al., 2020; Liu et al., 2021; Tohjima et al., 2020; Turner et al., 2020; Weir et al., 2020; Wu et al., 2021), as of this writing, there is no indication of a global‐scale decrease in the atmospheric CO2 mixing ratio or its growth rate due to the emissions reductions (NOAA Global Monitoring Laboratory, 2021; World Meteorological Organization, 2020). Even with a robust global measurement system, the detection of COVID‐related emissions reductions in global CO2 or its growth rate is challenging due to two factors: (1) internal variability in the climate system and (2) carbon‐concentration feedbacks. Internal variability is unforced climate variability that arises from the coupled interactions of the atmosphere and ocean (e.g., El Niño‐Southern Oscillation (ENSO); Deser, Phillips, et al., 2012). The role of internal variability in the growth rate of CO2 has been well documented in the literature (e.g., Frölicher et al., 2013; Keeling et al., 2001; Sarmiento & Gruber, 2002), and multiple studies implicate this variability in our inability to detect emissions changes in measurements of atmospheric CO2 (e.g., Peters et al., 2017). Carbon‐concentration feedbacks manifest from the sensitivity of the ocean and land carbon reservoirs to changing CO2 (Arora et al., 2013, 2020; Friedlingstein et al., 2006). Recent studies suggest that the ocean carbon reservoir rapidly responds to perturbations in CO2 (McKinley et al., 2020; Ridge & McKinley, 2021), and this can further confound the detection of emissions changes in measurements of atmospheric CO2. It is critical that we develop a deeper understanding of the role of internal variability and carbon feedbacks on the detectability of emissions changes to inform both near‐term (1–10 years) predictions of the carbon cycle (Ilyina et al., 2021) and the verification of future emissions reductions (Peters et al., 2017; Ridge & McKinley, 2021).

Initial‐condition large ensembles of Earth system models are a relatively new tool that provides a means to quantify the anthropogenic influence on the Earth system in the presence of internal climate variability (Deser et al., 2020). These large ensembles are a set of simulations with a single Earth system model: each simulation or ensemble member is initialized slightly differently to create diverging climate trajectories, while all ensemble members are externally forced with a common emission scenario (Deser, Knutti, et al., 2012). Multiple studies have used large ensembles to account for the role of internal variability in long‐term climate trends (e.g., Deser, Phillips, et al., 2012, 2016). Most recently, large ensembles have been used to estimate the anthropogenic influence on short‐term climate signals (such as for the COVID pandemic, see, e.g., Fyfe et al., 2021; Gettelman et al., 2021; Jones et al., 2021), and to make Earth system predictions over the near‐term (1–10 years; Yeager et al., 2018). However, no studies have used a large ensemble framework to assess the detectability of short‐term CO2 emissions reductions from atmospheric CO2 measurements.

Here, we develop an understanding of the role of internal variability and carbon feedbacks on the detectability of a short‐lived CO2 emissions reduction in the atmospheric mixing ratio of CO2 (χCO2) using output from an initial‐condition large ensemble of an Earth system model. This 30‐member ensemble evolves the Earth system under three, short‐term emissions reduction scenarios of differing magnitudes. We investigate the detectability of the emissions reduction using several modeled parameters that characterize atmospheric CO2: the χCO2 growth rate, the column‐averaged CO2(XCO2), and the integrated atmospheric carbon reservoir.

Methods

We utilize the Canadian Earth System Model version 5 (CanESM5), which consists of coupled atmosphere, ocean, sea‐ice, land surface, and land/ocean carbon cycle model components (Swart et al., 2019). The atmospheric model in CanESM5 is version five of the Canadian Atmospheric Model (CanAM5) that has an approximate 2.8° horizontal resolution and 49 vertical levels of varying thickness on a hybrid sigma‐pressure vertical grid and similar physical parameterizations as its predecessor (CanAM4; Swart et al., 2019). The land component of CanESM5 consists of the Canadian Land Surface Scheme (CLASS) and the Canadian Terrestrial Ecosystem Model (CTEM) that produce fluxes of energy, water, and carbon dioxide at the land‐atmosphere interface via the simulation of physical and biogeochemical processes, including the CO2 fertilization of photosynthesis (Swart et al., 2019). The ocean physical and biogeochemical components of CanESM5 used in this study are the CanNEMO physical model coupled to the Canadian Model of Ocean Carbon (CMOC), which simulates ocean carbon and its exchange with the atmosphere at approximately 1° horizontal resolution (Swart et al., 2019). In our study, the concentration of CO2 in the CanESM5 atmosphere is modeled as a three‐dimensional, prognostic passive tracer that responds to air‐sea and air‐land CO2 fluxes from the coupled land and ocean carbon cycle components, and to specified CO2 emissions.

We analyze output from five ensembles of CanESM5. In each case, ensemble members are initialized with slightly perturbed climate states to simulate a range of internal variability, but each member in a given ensemble experiences identical external forcing. The first ensemble (the historical ensemble) covers the period from 1850 to 2014 and consists of nine ensemble members of CanESM5 forced with a global historical emission data set of CO2 and other climate‐relevant gases and aerosols—this was devised for emissions‐driven historical simulations in Phase 6 of the Coupled Model Intercomparison Project (CMIP6; Figure S1a in Supporting Information S1; Hoesly et al., 2018). The second ensemble (the control ensemble) covers the period 2015–2100 and consists of 30 ensemble members of CanESM5 integrated under the ESM‐SSP2‐4.5 emissions scenario (Figure S1b in Supporting Information S1; O'Neill et al., 2016). The remaining 3 ensembles span 2019–2040 and consist of 30 members each that are forced with COVID‐like CO2 emissions reductions beginning in December 2019 and resolving in December 2021; peak emissions reductions of 25% (COVID‐like), 50% (2 × COVID‐like), and 100% (4 × COVID‐like) occur in May 2020 (Figure S1b in Supporting Information S1). Hereafter, we refer to these later three ensembles collectively as the CanESM5‐COVID ensemble, as described in Fyfe et al. (2021) and Lovenduski et al. (2021). The CO2 emissions in the historical and control ensembles have spatial and seasonal variability; emissions are highest near urban centers in the Northern Hemisphere (Figure S2a in Supporting Information S1) and peak in boreal winter when energy consumption in the Northern Hemisphere is at a maximum (Figure S2b in Supporting Information S1). Emissions are scaled uniformly for the COVID ensembles to maintain this spatial and seasonal variability. In the CanESM5‐COVID ensemble output we analyze here, emissions of CO2 from other sources (e.g., land use change) and emissions of other climate relevant gases and aerosols are prescribed from the ESM‐SSP2‐4.5 scenario, that is, these emissions do not change due to COVID.

The global carbon cycle in CanESM5 compares well with observational metrics and is thus an appropriate tool for the study of the detectability of short‐term emissions reductions in atmospheric χCO2. Air‐sea and air‐land CO2 fluxes from the historical simulation of CanESM5 were previously evaluated in Swart et al. (2019). Briefly, Swart et al. (2019) illustrate high skill and low root mean square error between simulated and observed spatial patterns of Gross Primary Production (GPP) and air‐sea CO2 flux over 1981 to 2010. CanESM5 tends to overestimate GPP in sub‐saharan Africa and underestimate GPP in the Amazon rainforest, likely due to precipitation biases (Swart et al., 2019). Historical CanESM5 air‐sea CO2 fluxes are biased high in the North Atlantic and low in the Southern Ocean, such that the globally integrated air‐sea CO2 flux exhibits little bias as compared to observations (Swart et al., 2019). CanESM5 captures the broad features of the amplitude and phasing of the seasonal cycle of χCO2 measured at Barrow (BRW), Mauna Loa (MLO), and South Pole (SPO), though the seasonal drawdown of CO2 occurs too early at Point Barrow, and the amplitude is biased high at Mauna Loa (Figure S3 in Supporting Information S1; the model is sampled at the approximate latitude, longitude, and height of the flask sample in the real world). The de‐trended interhemispheric gradient in observed, annual mean χCO2 exhibits large annual‐to‐decadal fluctuations over 1960–2020 that are generally replicated by the model (Figure S4 in Supporting Information S1). The de‐trended interhemispheric difference in the CanESM5 historical ensemble members encapsulate the observations and the ensemble mean replicates the decadal variations in the observations, though the interannual variance of individual ensemble members is greater than that of the observational record (Figure S4 in Supporting Information S1). Finally, the CanESM5 control ensemble mean exhibits a similar growth rate in χCO2 (2.4 ppm yr−1 over 2015–2019; see Figure 1) as calculated from observations (2.57 ± 0.08 ppm yr−1 over 2015–2019; https://gml.noaa.gov/ccgg/trends/gl_gr.html). The actual growth rate derived from observations is slightly higher due to the impact of the 2015–2016 El Nio event on the carbon cycle (Chatterjee et al., 2017; Liu et al., 2017).

Figure 1

Temporal evolution of the growth rate of de‐seasoned, monthly χCO2 from the CanESM5 COVID ensemble sampled at (a) Mauna Loa, and (b) the average of 12 flask sites (as in Cadule et al., 2010) over 2020–2024. Growth rate is calculated as the difference in χCO2 for a given month relative to the same month in the previous year. Thin lines show individual ensemble members, and thick lines show the ensemble mean for each emissions scenario. Red dot and range illustrates the mean and 2σ (95%) confidence interval in February 2021 for the COVID‐like emissions scenario. Subplots show the temporal correlation coefficients of individual ensemble members with the ensemble mean over Jan 2020–Dec 2021 for each emissions scenario. Small circles show the correlation coefficients across the 30 ensemble members, large circles show the mean correlation coefficients, and dashes indicate 2σ (95%) confidence intervals.

Results

The 30‐member CanESM5 COVID ensemble predicts a decrease in the de‐seasoned, monthly growth rate of χCO2 from January 2020 through February 2021, followed by an increase in growth rate from February through December 2021 under all of the COVID emission scenarios when sampled at both Mauna Loa and 12 global flask sites (Figure 1; growth rate is calculated as the difference in χCO2 for a given month relative to the same month in the previous year; flask sites as in Cadule et al., 2010). The decrease in the growth rate is largest for the 4 × COVID‐like emissions scenario and smallest for the COVID‐like emissions scenario and peaks in February 2021 under all COVID scenarios (Figure 1). Meanwhile, the control ensemble exhibits little change in its growth rate over this period (Figure 1). This suggests that the χCO2 growth rate is highly sensitive to the magnitude of the emissions reduction, and that growth rate anomalies at Mauna Loa and across the global flask network tend to be largest ∼9 months after the peak emissions reduction (May 2020; Figure S1b in Supporting Information S1). This delay is consistent with the global mixing timescale for CO2 in the atmosphere (Forster et al., 2007). Under the 4 × COVID‐like scenario, the growth rate exceeds the control growth rate from 2022 through 2024 before returning to control values (Figure 1). Figure 1 also reveals that internal variability tends to obscure the emissions reduction signal in χCO2 at an individual site more than in the global average (cf. Figures 1a and 1b); averaging across multiple sites tends to dampen the effects of internal variability that manifest most strongly at local and regional scales (Hawkins & Sutton, 2009). As a result, the ensemble‐mean February 2021 COVID‐like χCO2 growth rate is significantly different from the control ensemble mean in the average of the 12 flask sites, but not at Mauna Loa (Figure 1; significance calculated using a 2σ (95%) confidence interval across the COVID‐like ensemble members). If we wish to detect a signal of the COVID‐driven emissions reduction in the real‐world growth rate of χCO2, our modeling study suggests that we are most likely to find it in early 2021 by averaging across measurements collected in the global flask network.

Is it possible to detect the change in the de‐seasoned, monthly χCO2 growth rate from flask observations in the real world, where we have only a single “ensemble member”? To answer this question, we turn to a formal statistical detection framework, where we use the unique ensemble mean “fingerprint” of the growth rate in the model sampled at flask sites (i.e., the V‐shaped dip and recovery in the ensemble‐mean growth rate over January 2020 to December 2021 in Figure 1) and quantify the correlation of each individual ensemble member with this fingerprint for each emissions scenario. This statistical detection approach for hypothetical observations (we haven't yet measured the growth rate in December 2021, for example) is identical to the one outlined in Lovenduski et al. (2021) and mimics the approach for the detection of a climate change signal in real‐world observations (Bindoff & Stott, 2013). The resulting correlation coefficients are shown in the subplots of Figure 1, where small circles show the set of 30 Pearson's correlation coefficients (r) with the ensemble mean fingerprint across the 30 ensemble members, and large circles show the mean correlation coefficients (calculated using a Fisher's z transform; see Lovenduski et al., 2021) for each COVID‐like emissions scenario. For the model sampled at Mauna Loa, the mean correlation coefficient for the COVID‐like ensemble is 0.4 with a wide range; stronger emissions reductions increase the mean correlation coefficient and narrow the range (subplot in Figure 1a), suggesting a higher probability of detecting the fingerprint from a single ensemble member or hypothetical observational record under higher emissions reductions. Indeed, the range of correlation coefficients is only statistically different from zero (using a 2σ/95% confidence interval) under the 2 × COVID‐like and 4 × COVID‐like emission scenarios (subplot in Figure 1a), indicating that significant detection of the COVID fingerprint is only possible in cases with more extreme emissions reductions than those that occurred during the COVID pandemic. Similar patterns are observed when the model is sampled at 12 flask sites (Figure 1b), though the correlations are overall higher due to reduced internal variability.

Model‐estimated, column‐averaged dry‐air mole fraction of χCO2 (referred to as XCO2 by the satellite community; Crisp et al., 2004) averaged over the extratropical Northern and Southern Hemispheres (20°N–55°N and 20°S–55°S, respectively) shows only a small signal of COVID‐like emissions reductions amid large internal variability (Figure 2). While the emissions reduction signal is more pronounced in the Southern Hemisphere extratropics, there is large overlap of the various model ensemble members from the various emission scenarios (Figure 2b), and only the 4 × COVID‐like ensemble mean is significantly different from the control ensemble mean at the 2σ (95%) level (not shown). The vertical integration of the atmospheric column and the diffusive nature of atmospheric transport makes the modeled column concentrations less sensitive to changes in the surface emissions signal (Miller et al., 2007; Rayner & O'Brien, 2001), thus making the signal more difficult to detect in the column.

Figure 2

Temporal evolution of monthly, column‐averaged χCO2 over (a) the Northern Hemisphere, 20°N–55°N, and (b) the Southern Hemisphere, 20°S‐55°S, simulated with the CanESM5 COVID ensemble. Thin lines show individual ensemble members, and thick lines show the ensemble mean for each emissions scenario.

Carbon‐concentration feedbacks further obscure the detection of COVID emissions reductions in the atmospheric carbon reservoir. Figure 3 shows the anomaly in the ensemble‐mean cumulative change in the modeled atmosphere, ocean, and land carbon reservoirs from December 2019 to December 2040, where the anomaly is calculated relative to the control ensemble mean. In the atmosphere, the cumulative reservoir anomaly is negative for the duration of the simulations regardless of emissions scenario (Figure 3a), indicating that each of the COVID‐like emissions perturbations leads to a forced change in the cumulative atmospheric reservoir lasting well beyond the emissions recovery in 2022 (cf. Figures 3a and S1b). For both the 2 × COVID‐like and 4 × COVID‐like scenarios, the atmosphere reservoir anomaly falls outside of the ensemble spread due to internal variability (gray shading) for several years. Meanwhile, the ocean and land carbon reservoirs also respond to the COVID‐like emissions reductions, the ocean carbon sink immediately slows with decreasing χCO2 under all COVID‐like scenarios (Figure 3b; Lovenduski et al., 2021), and the land carbon sink also weakens, most noticeably under the 2 × COVID‐like and 4× COVID‐like scenarios (Figure 3c). The ocean reservoir anomaly falls outside of the spread due to internal variability only in the 2 × COVID‐like and 4 × COVID‐like scenarios (Figure 3b; Lovenduski et al., 2021). The land carbon sink anomaly is fully within the internal variability bounds (Figure 3c), due to high internal variability in the land‐air CO2 flux (Denman et al., 2007). Nevertheless, these results suggest that COVID emissions reductions cause both the ocean and land to absorb less carbon than usual in our model, thus reducing the perturbation in the atmosphere. To illustrate this point further, we estimate the ensemble‐mean, cumulative change in the atmospheric carbon reservoir due only to emissions changes and plot the resulting reservoir anomaly as dashed lines in Figure 3a. This illustration reveals a critical role for carbon‐concentration feedbacks in the detection of COVID‐driven emissions reductions: if not for the slowing ocean and land carbon sinks, the COVID‐like emissions reduction signal in the atmospheric carbon reservoir would have been detectable above the noise of internal variability for three consecutive years (2022–2025), and for longer durations with larger emission perturbations (Figure 3a).

Figure 3

Cumulative changes in the (a) atmosphere, (b) ocean, and (c) land carbon reservoirs from December 2019 onwards, as simulated by the CanESM5 COVID ensemble. Colored lines show the anomaly in the ensemble‐mean reservoir size relative to the control ensemble mean (SSP2‐4.5), and gray shading indicates the spread in the cumulative reservoir anomaly across the control ensemble. Dashed lines in (a) show the cumulative changes in atmospheric carbon due to anomalous emissions alone.

Conclusions and Discussion

We use an initial‐condition large ensemble of an Earth system model to assess the detectability of the COVID‐driven emissions reductions signal in measurements of atmospheric CO2 above the noise of internal variability and carbon‐concentration feedbacks. We find a unique fingerprint in atmospheric CO2 growth rates calculated from simulated χCO2 measurements under COVID‐like emissions reductions. The largest negative anomalies in the atmospheric χCO2 growth rate appear in February 2021, ∼9 months after the peak emissions reductions. This growth rate signal is more likely to be detected above the noise of internal variability when averaging over global flask network sites, rather than at an individual site. However, this unique fingerprint is not formally detectable using a climate signal detection statistical approach, unless we force the model with unrealistically large emissions reductions. Internal variability obscures the detection of change in the interhemispheric difference of simulated χCO2 from flask measurements and the simulated extra‐tropical column‐average XCO2 from satellite observations. Carbon‐concentration feedbacks further reduce the emissions signal in the atmospheric carbon reservoir. When we remove the effects of these feedbacks on the atmospheric carbon reservoir, the signal in the cumulative reservoir anomaly is detectable above the noise of internal variability over a three consecutive year period (2022–2025).

Our study illuminates the challenges associated with detecting brief CO2 emissions reductions in global‐scale atmospheric CO2 from our established observational measurement systems. In order to see the emergence of the signal of COVID‐driven emissions reductions in atmospheric CO2, one needs to first remove the influence of internal climate variability and carbon‐concentration feedbacks from the atmospheric CO2 measurements. While we are getting closer to quantifying the internal contribution to the total signal from our measurements and producing near‐real time estimates of this variability (e.g., Betts et al., 2016, 2020), we are not yet capable of quantifying carbon feedbacks from our current, exploratory observational system (e.g., Sellers et al., 2018). Further, the ocean and terrestrial carbon reservoirs are only sparsely observed and, with the exception of a few surface ocean buoys that measure ocean CO2 partial pressure (Sutton et al., 2019), the high‐quality estimates of changing air‐sea and air‐land CO2 fluxes that are available in a historical context are not yet available in near‐real time due to the high costs of fast data dissemination and other impediments. As we move into a world characterized by intentional emissions reductions associated with international climate change mitigation policies, we should consider this measurement infrastructure in the ocean and terrestrial biosphere to detect the signal and monitor the impact of intentional emissions reductions in atmospheric CO2.

Supporting Information S1

Click here for additional data file.