Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

# A small climate-amplifying effect of climate-carbon cycle feedback

## Abstract

The climate-carbon cycle feedback is one of the most important climate-amplifying feedbacks of the Earth system, and is quantified as a function of carbon-concentration feedback parameter (β) and carbon-climate feedback parameter (γ). However, the global climate-amplifying effect from this feedback loop (determined by the gain factor, g) has not been quantified from observations. Here we apply a Fourier analysis-based carbon cycle feedback framework to the reconstructed records from 1850 to 2017 and 1000 to 1850 to estimate β and γ. We show that the β-feedback varies by less than 10% with an average of 3.22 ± 0.32 GtC ppm−1 for 1880–2017, whereas the γ-feedback increases from −33 ± 14 GtC K−1 on a decadal scale to −122 ± 60 GtC K−1 on a centennial scale for 1000–1850. Feedback analysis further reveals that the current amplification effect from the carbon cycle feedback is small (g is 0.01 ± 0.05), which is much lower than the estimates by the advanced Earth system models (g is 0.09 ± 0.04 for the historical period and is 0.15 ± 0.08 for the RCP8.5 scenario), implying that the future allowable CO2 emissions could be 9 ± 7% more. Therefore, our findings provide new insights about the strength of climate-carbon cycle feedback and about observational constraints on models for projecting future climate.

## Introduction

With increasing atmospheric CO2 and a warming climate during the industrial era, land and ocean reservoirs have together absorbed >50% of anthropogenic CO2 emissions1, playing a significant role in reducing anthropogenic warming2. Whether the ocean and land sinks will continue to take up a similar fraction in the future remains uncertain, and one of the major causes for that uncertainty is the feedback between the carbon cycle and the physical climate system3,4,5. Global climate change affects carbon uptake by land and oceans, which impacts the rate of increase in atmospheric CO2 and, in turn, climate change. This feedback loop between the physical climate system and the global carbon cycle of the Earth system was quantified using a modeling approach6,7,8,9.

Previous studies quantified the climate–carbon cycle feedback as a function of the carbon-concentration feedback response parameter (β) and the carbon–climate feedback response parameter (γ)7,10. The β (GtC ppm−1) and γ (GtC K−1) are also defined as the rates of change in land and ocean carbon storages relative to a fixed reference time to atmospheric CO2 concentration increase and to global climate change that is often quantified by the global-mean surface temperature change, respectively. From the perspective of the land and ocean reservoirs, β is positive, and γ is negative. Therefore, β-feedback reduces the impact of CO2 emissions on atmospheric CO2 concentrations and then global warming, while γ-feedback amplifies global warming. The combined effects of the β-feedback and the γ-feedback and the nonlinear interaction between them determine the strength of the climate–carbon cycle feedback loop which is known as the feedback gain factor (g).

Both β and γ were previously quantified using physical climate models coupled with the global carbon cycle under idealized experiments7,10,11. Based on the experiments by the advanced Earth system models, the last two successive assessments by the International Panel on Climate Change (IPCC) found that the uncertainty of the climate–carbon cycle feedback was dominated by β12,13. These Earth system models under idealized experiments did not account for the biophysical effect of land-use change on climate and ecological effects on the residence time of carbon in the land biosphere. Modeling experiments showed that in a high CO2-induced warming climate system, the nonlinear carbon–climate feedback can reduce the ocean carbon uptake by 3.6–10.6% based on the simulations of seven Earth system models14.

On the other hand, observations can be used to constrain the estimated climate–carbon cycle feedbacks. Applying an emerging constraint-based approach to instrumental records, Cox et al.15 estimated γ for tropical land to be −53 ± 17 GtC K−1 by 2100. Using data from three ice cores and multiple temperature reconstructions from tree ring data over 1050–1800, Frank et al.16 reported a ΔCATA (defined as η in this study) of 7.7 ppm K−1 with a likely range of 1.7–21.4 ppm K−1, which was much lower than previous estimates (e.g., ~40 ppm K−1)17,18. However, the η is not truly comparable with γ, because their relationship also depends on β (see Eq. (3) below). In quantifying the response of climate to an increase in atmospheric CO2 concentration or anthropogenic emissions, two other quantities are frequently used in the studies of carbon–climate interactions: the sensitivity of climate to atmospheric CO2 (α ≡ ΔTACA, in K GtC−1, which is equivalent to 1/(2.12 η) in this study)7,19, and the transient climate response to cumulative CO2 emission (TCRE ≡ ΔTACE, in K GtC−1)20. The values of β, γ, α, and TCRE are related to each other through one equation (see Eq. (7))10,21, and that equation is applicable to any individual frequency (or timescale) (see “Methods”).

## Results

### Estimates of β and γ across timescales

Based on the previous studies7,10,14,19, the climate–carbon cycle feedback framework with a nonlinear feedback term as a function of β and γ parameters, i.e., f(β,γ) in a unit of GtC ppm−1 K−1 or GtC GtC−1 K−1, in a CO2 emission-driven coupled climate–carbon cycle system, can be expressed as:

$$\Delta C_E = \Delta C_A + \beta \Delta C_A + \gamma \Delta T_A + f(\beta ,\gamma )\Delta C_A\Delta T_A$$
(1)

where ΔCE, ΔCA, and ΔTA are changes in cumulative CO2 emissions, atmospheric CO2, and global surface temperature, respectively, during a time interval (Δt). Here, we applied a Fourier analysis-based spectral decomposition approach to observations or reconstructed records of global surface temperature and atmospheric CO2 to quantify how the two feedback parameters vary over different periods of time, or across different timescales (1/frequencies) over the same time period. As shown in the “Methods” section, the relationship among the four parameters at any given timescale (k) is given by

$$({\mathrm{TCRE}}^{ - 1})_k = (1 + \beta _k)(\alpha ^{ - 1})_k + \gamma _k {\!\!}\ast$$
(2)

where $$\gamma \ast = \gamma + f\left( {\beta ,\gamma } \right)\Delta C_A$$ represents the combined carbon–climate feedback from the linear γ-feedback and the nonlinear impact of atmospheric CO2 on the carbon–climate feedback (see “Methods”). Equation (2) also shows that variations of TCRE−1 and α−1 for a given timescale are linearly related, with a slope of 1 + βk and an intercept of $$\gamma _k {\!\!}\ast$$. Our result from modeling experiments (see Supplementary Table 4) showed that the nonlinear feedback term f(β, γ) is relatively small, and its contribution (f(β, γCA) to the γ* is 15 ± 23%, while its contribution (f(β, γTA) to the β-feedback is negligible (3 ± 3%). Thus, we indicate that the nonlinear feedback term has negligible effect on the estimate of the slope of 1 + βk.

We estimated (TCRE−1)k from the ratio of the variability in CE to the variability in TA, and (α−1)k from the ratio of the variability in CA to the variability in TA on a given timescale (see “Methods”). In Fourier analysis, the variabilities of CE, CA, and TA were represented by the amplitudes of their harmonics at a given frequency or period (Supplementary Fig. 1), which were then used to calculate TCRE−1 and α−1 across different timescales (see “Methods”). Different from the previous approaches based on a Taylor series expansion and modeling experiments (hereafter referred to as the FEA approach, see Eqs. (2225) in “Methods”)7,10,14, the estimated β and γ using Fourier analysis-based approach do not depend on a reference time that was used to compute changes in ΔCE, ΔCA, and ΔTA as in the FEA approach.

In the following, we apply the Fourier analysis-based approach to observed time series of CE, CA, and TA to estimate βk and γk for the industrial period (1850–2017). By ignoring the timescale dependence of β, and using the value of β for the industrial period, we estimated γk at different timescales using large ensembles based on combinations of ice-core atmospheric CO2 records (CA) and reconstructed surface air temperature (TA) datasets (“Methods”) over the period 1000–1850. Finally, we compare the observation-based estimates of β and γ, and corresponding feedback gain factor (g) with those estimated from the simulations from multiple Earth system models.

### Estimates of β and γ* for the industrial period (1850–2017)

Figure 1 shows the observed increase in atmospheric CO2CA) from ~285 ppm in 1850 to 405 ppm in 2017 in response to the increase in cumulative CO2 emissions (ΔCE), including fossil fuel combustion22 and land-use change23,24, which amounted to ~630 ± 42 GtC by 2017 since 1850, with respect to a concurrent increase in global-mean annual temperature (ΔTA) of ~1.1 K over the same period based on the averages of the four global temperature datasets. Using Fourier analysis, we found that the amplitudes of CE, CA, and TA for the industrial period increased nonlinearly with timescale (Fig. 2a and b), but the variations in (TCRE−1)k and (α−1)k with timescale were linearly related to each other (R2 = 0.99, P < 0.001) with a slope (mean ± 1 standard deviation) of 2.52 ± 0.15 (see Fig. 2c and Supplementary Table 1 for individual estimates). Across different timescales from inter-annual to multi-decadal (e.g., k from 2 to 90 years) for the 1880–2017 period, (α−1)k varied from 50 to 250 GtC K−1 and (TCRE−1)k from 125 to 600 GtC K−1 (Fig. 2c).

The slope of the linear regression between (TCRE−1)k and (α−1)k across different timescales can be used to estimate (1 + βk) based on Eq. (2). The result in Fig. 2c shows that the value of βk is approximately constant across different timescales, with an average value of 3.22 ± 0.32 GtC ppm−1 (or 1.52 ± 0.15 GtC GtC−1) throughout the industrial period in 1880–2017 (Supplementary Table 1). The uncertainty in β results from the uncertainties in CE from fossil fuel combustion and land-use change estimates and errors in the temperature datasets and the regression (see “Methods”). Using the estimated β for the industrial period, we estimated the corresponding γ* to be −10.9 ± 3.6 GtC K−1 for the period 1880–2017 based on Eq. (6).

To diagnose the constancy of β estimated from the industrial period and the reliability of our Fourier analysis-based approach, we used a box model to predict changes in temperature and CO2 concentration with CO2 emissions as forcing input (see “Methods”, Eqs. (29)–(30)). We ran the box model using the estimate of two parameters based on observations (i.e., β = 3.22 GtC ppm−1 and γ* = −10.9 GtC K−1) and the annual cumulative emissions over 1850–2017. The model predictions fitted very well to the observed values of atmospheric CO2 concentration and surface temperature (R2 = 0.99, RMSE = 3.5 ppm for CO2, and R2 = 0.96, RMSE = 0.17 K for temperature) (see Supplementary Fig. 2). The greater trend of the predicted global surface temperature relative to the observed after 1980s (Supplementary Fig. 2) is likely associated with γ* being treated as a constant, while in reality γ* could vary with internal climate variability and across different timescales. Using the box model predicted annual CO2 and temperature, we in return, estimated a nearly constant β (3.42 GtC ppm−1) and a γ* (−12.2 GtC K−1) for the period 1880–2017 based on our approach of Eqs. (1) and (2). Both values of β and γ* estimated from the predicted CO2 and temperature time series by the box model fall within the uncertainties of observation-based β and γ*, suggesting that the β for 1880–2017 being nearly constant is robust.

As the γ* for the industrial period consists of the γ-feedback and the nonlinear feedback effect of f(β, γCA, which results from the multi-decadal climate variability and anthropogenic CO2 emissions. Because of the relatively short records over the industrial period, we could not separate the γ-feedback parameter directly from the γ* using our analysis framework. The η (or α−1) for the industrial period includes a possibly significant contribution from an emissions-driven increase in atmospheric CO2 concentration to the climate–carbon feedback. As a result, the mean value of η for 1850–2017 is 109 ppm K−1, much higher than the estimates of 7.7–40 ppm K−1 during the preindustrial period before 1850s16,17 that is considered to be at quasi-equilibrium with small variation in greenhouse gas forcing. The reason for this large discrepancy in η between the preindustrial and industrial periods is likely a consequence of the nonlinear dependence of radiative forcing on atmospheric CO2 concentrations25, and the temperature change in response to the increase in atmospheric CO2 during the industrial era has not reached steady state, as large part of atmospheric CO2 increase during industrial period was driven by emissions not due to warming-induced CO2 release from land and ocean reservoirs. It is also well known that equilibrium climate sensitivity is often considerably larger than the transient climate sensitivity26. In the following, we used reconstructed records of atmospheric CO2 and surface temperature during the preindustrial last millennium from 1000 to 185016,17,18,27,28 to estimate the γ.

### Timescale dependence of γ over the preindustrial last millennium (1000–1850)

The preindustrial last millennium (1000–1850) climate could be considered as a quasi-equilibrium state with very little change in CO2 emissions, therefore the ΔCA during 1000–1850 was largely driven by carbon–climate feedback without the complication of concurrent increase in CE as the industrial period. We ignored the possible influences of CO2 emissions from the early land-use29 on global surface temperature during 1000–1850. As the atmospheric CO2 was relatively stable (280 ± 8 ppm) during the 1000–1850 period, i.e., ΔCA was only ~3% of atmospheric CO2 concentration (Fig. 1a), we assumed that its effects on β-feedback and γ-feedback were small (e.g., f(β, γCAΔTA ≈ 0). For the 1000–1850 period with ΔCE ≈ 0, Eq. (2) can be further simplified as

$$\gamma _k = - m\left( {1 + \beta } \right)\eta _k - f\left( {\beta ,\gamma } \right)\Delta C_A \approx - m\left( {1 + \beta } \right)\eta _k$$
(3)

where m = 2.12 GtC ppm−1 is a factor for converting atmospheric CO2 in ppm to GtC. As the β is found to be nearly constant during the industrial period, we assumed that the constant value is applicable to the 1000–1850 period. This is reasonable, as the error sensitivity analysis showed that the uncertainty in the estimated γ for 1000–1850 induced from the uncertainty in β is still smaller than those from the difference in the estimated γ from ice-core CO2 records and reconstructed temperatures (Supplementary Fig. 5). Thus, using 15 combinations of 3 ice-core CO2 records28,30,31 and 5 temperature reconstructions16,32,33,34,35 as shown in Fig. 1 and Eq. (3) with β of 3.22 ± 0.32 GtC ppm−1, we estimated that the γ at a 100-year timescale (γ100yr) for 1000–1850 was −122.8 ± 60.2 GtC K−1 (Fig. 3b and Supplementary Table 2). Furthermore, γk increases with timescale, varying from −33 ± 14 GtC K−1 on timescales of 10–70 years to −110 ± 40 GtC K−1 over timescales of 200–800 years (Supplementary Fig. 6).

Using another set of >1500 combinations of 521 reconstructed temperature records from 1000 to 185016 and 3 ice-core CO2 records (see “Methods”, Supplementary Fig. 4), we calculated much larger ensemble (>1500) estimates (EnOBS) of ηk and γk using the Fourier analysis-based approach. Our results confirmed that ηk systematically increased with timescale, varying from 6 ± 6 ppm K−1 on timescales of 10–70 years to 20 ± 8 ppm K−1 over timescales of 200–800 years due to the increased γk on magnitude with timescale (Supplementary Fig. 6). The range of the EnOBS-based η100yr for 1000–1850 was 7–23 ppm K−1 with a median of 9 ppm K−1, consistent with the estimate of η (1.7–21.4 ppm K−1) by Frank et al.16. The resulting γ100yr of EnOBS for 1000–1850 was −82 ± 42 GtC K−1, varying from a minimum of −260 GtC K−1 to a maximum of −15 GtC K−1 (Fig. 3b). The EnOBS-based γ100yr over timescales of 200–800 years was −80 ± 50 GtC K−1 (Supplementary Fig. 6), which is 30% smaller than the estimate of γ100yr from 3 ice-core CO2 records × 5 temperature reconstructions. These results suggest that the timescale or temporal dependance of η over the 1000–185016 is largely driven by the positive feedback of terrestrial and oceanic carbon pools to climate (i.e., the γ feedback), implying that on longer timescales, warming of the climate would cause more release of CO2 into the atmosphere and in return, amplify warming.

The estimate of γ also depends on the mean climate state (quasi-equilibrium or transient). Here we compared estimates of γ between the warmer period of the Medieval Climate Anomaly (1000–1300) with a mean temperature anomaly of −0.18 ± 0.09 K and the cooler period of the Little Ice Age (1400–1700) with a mean temperature anomaly of −0.41 ± 0.03 K, and both anomalies were calculated relative to the mean global surface temperature from 1961 to 1990. We further analyzed probability distributions of the EnOBS-based γ100yr for three different periods: 1000–1850, 1000–1300, and 1400–1700. The estimates of γ100yr for the cooler period of 1400–1700 had a comparable probability distribution with the 1000–1850 period (Fig. 3b). However, the γ100yr for the warmer period of 1000–1300 showed a much narrower distribution with a less negative mean (−38 ± 17 GtC K−1, or 50% smaller) than that for 1400–1700 (−72 ± 56 GtC K−1), suggesting that γ100yr feedback is sensitive to the state of the mean climate. The higher γ100yr for 1400–1700 resulted from a colder climate and some drastic fluctuations in CO2 (e.g., larger ΔCA leads to more nonlinear feedback contribution from f(β, γCA), especially the strong dip at ~1600 AD, while the lower γ100yr for 1000–1300 is associated with smaller variations in CO2 and a warmer climate (Fig. 1). The CO2 drop in the Little Ice Age (1400–1700) may have been driven not only by natural disturbances (e.g., volcanic eruptions) but also probably by human-induced land-use effects36, which could have led to a more negative estimate of γ100yr for 1400–1700.

### Comparisons of observation-based and model-based estimates of feedback parameters

Previous studies have noted the timescale dependence of η (i.e., ΔCATA) and the relationship between η and the climate–carbon cycle feedback parameters (β and γ)10,16,37. Here we further quantified the different timescale dependence of β and γ. Using an observation-based constraint18, Cox and Jones estimated that β was between 3 and 5 GtC ppm−1 and γ fell between −250 and −50 GtC K−1. For the β-feedback of land, recent studies estimated that the CO2 fertilization effect on global plant biomass carbon only was 25 ± 4 GtC year−1 for a 100 ppm ∆CO2 over 1980–201038, and for the global terrestrial C sink was 3.5 ± 1.9 GtC year−1 per 100 ppm ΔCO2 over 1959 to 201039, which is equivalent to a land β-feedback of 1.75 ± 0.95 GtC ppm−1 over the same period. Our findings indicate that observation-based (land + ocean) β is approximately constant across different timescales at 3.22 ± 0.32 GtC ppm−1 for 1880–2017. While magnitudes for γ increased with timescale from decadal to multi-centennial, and our estimate of γ100yr had a narrower uncertainty than previous estimates, with an average of −122.82 ± 60.16 GtC K−1 from the 3 ice-core CO2 records × 5 temperature reconstructions and of −81.91 ± 41.90 GtC K−1 from EnOBS for 1000–1850. Uncertainties related to the estimated γ100yr from the two datasets (15 members versus >1500 members of EnOBS) did not overlap, which may suggest the uncertainty in γ100yr was from −180 to −40 GtC K−1.

We then compared our observation-based estimates of β and γ using Fourier analysis-based approach with those from Earth system models. Estimates of β and γ from Earth system models were calculated from three sets of model simulations: the biogeochemically-coupled (BGC), radiatively coupled (RAD), and fully coupled (COU) simulations based on the FEA approach7,14 (see “Methods”). Eleven first-generation coupled climate–carbon cycle models (C4MIP) were driven with the prescribed CO2 emissions from the historical period (1860–2005) and the future period (2006–2100) under the IPCC SRES A2 scenario without land-use change7. Only COU and BGC simulations were conducted for the C4MIP models. The subsequent analysis used nine models from phase 5 of the Coupled Model Inter-comparison Project (CMIP5) used 1% year−1 increasing CO2 for 140 years, no land-use change (or the 1pctCO2 experiment)11. β and γ were estimated from three simulations (COU, BGC, and RAD) by each participating model. Because of the nonlinear feedback, the changes in size of carbon pools (land, ocean, and atmosphere) in the COU simulation were not equal to the sum of the simulated changes of those pool sizes in the RAD and BGC simulations14,40,41. According to “Methods”, we calculated the direct β-feedback from the COU-BGC simulations ($$\beta ^{{\mathrm{GC}}} \approx \Delta C_B^{{\mathrm{BGC}}}/\Delta C_A^{{\mathrm{BGC}}}$$) and the direct γ-feedback from the COU-RAD simulations ($$\gamma ^{{\mathrm{RAD}}} = \Delta C_B^{{\mathrm{RAD}}}/\Delta T_A^{{\mathrm{RAD}}}$$) and the total γ-feedback (direct plus indirect) from the COU-BGC simulations ($$\gamma ^{{\mathrm{COU}} - {\mathrm{BGC}}} \approx \left( {\Delta C_B^{{\mathrm{COU}}} - \Delta C_B^{{\mathrm{BGC}}}} \right)/\Delta T_A^{{\mathrm{COU}}}$$, in theory this is the γ*) using the FEA approach, for the observation-overlapped period of 1880–2017 and the future emission scenario of 1880–2100 for the C4MIP models, and the 1pctCO2 140-year period for the CMIP5 models, respectively. We also estimated the nonlinear feedback term from the difference between COU simulations and the BGC and RAD simulations ($$f\left( {\beta ,\gamma } \right) \approx [\Delta C_B^{{\mathrm{COU}}} - \left( {\Delta C_B^{{\mathrm{BGC}}} + \Delta C_B^{{\mathrm{RAD}}}} \right)]/\Delta C_A^{{\mathrm{COU}}}\Delta T_A^{{\mathrm{COU}}}$$) and its contribution to γ-feedback ($$f\left( {\beta ,\gamma } \right)\Delta C_A^{{\mathrm{COU}}}$$) for the CMIP5 models (“Methods”).

The estimated βBGC from 11 C4MIP models for 1880–2017 using the FEA approach was 3.07 ± 0.68 GtC ppm−1 which is close to the observation-based β (3.22 ± 0.32 GtC ppm−1) for the same period using historical CO2 emissions as forcing. Previous studies demonstrated that the carbon-concentration feedback was strongly dependent on the growth rate of atmospheric CO2 and hence on emission scenarios10. When extending the calculation to the future high emission scenario of the IPCC SRES A2 (close to the RCP8.5 pathway, the mean CO2 growth rate was about 0.72% year−1), we found that βBGC from C4MIP models was largely reduced by 19.5% to 2.47 ± 0.60 GtC ppm−1 for the period of 1880–2100 (Fig. 3a, Supplementary Table 3). As the 1pcCO2 experiments of CMIP5 models were configured under a higher emission scenario (1% year−1 of the CO2 growth rate), we found that the estimated βBGC from nine CMIP5 models using the FEA approach was 1.71 ± 0.44 GtC ppm−1, which is 45% smaller than the observation-based or C4MIP-based β for the historical period 1880–2017 (Fig. 3a), indicating that carbon-concentration feedbacks became smaller under higher growth rates of CO2 in the climate–carbon cycle system.

The γCOU-BGC is theoretically the γ* in this study, as the γCOU-BGC was calculated from all feedback effects in COU simulations minus the direct β-feedback in BGC simulations. We found that the estimated γCOU-BGC from 11 C4MIP models were on magnitude increased from −27.52 ± 11.93 GtC K−1 for 1880–2017 to −52.18 ± 26.54 GtC K−1 for 1880–2100, when the βBGC was decreased for 1880–2100 (Fig. 3a, b, Supplementary Table 3). The estimated γCOU-BGC from nine CMIP5 models was −70.14 ± 32.43 GtC K−1, which theoretically came from direct feedback $$\gamma ^{{\mathrm{RAD}}}$$ (−65.08 ± 30.74 GtC K−1) and the nonlinear feedback contribution $$f(\beta ,\gamma )\Delta C_A$$ (−9.6 ± 10.03 GtC K−1, ~15 ± 17% of the γ-feedback) under the high emission scenario of the 1pctCO2 experiments (Fig. 3b, Supplementary Table 4). Without CO2 impact, the estimated nonlinear feedback parameter $$f(\beta ,\gamma )$$ was −11.22 ± 11.72 × 10−3 GtC ppm−1 K−1 (Supplementary Table 4). On the other hand, as the observation-based γk for 1000–1850 and the model-based γCOU-BGC refer to different periods and timescales which exaggerates their differences owing to the timescale dependency of γ-feedback, it should be noted that there is limited comparability between them.

To further assess whether β varies with timescale, we estimated $$\beta _k^{{\mathrm{BGC}}}$$ and $${\gamma}_k^{{\mathrm{RAD}}}$$ by applying the FEA approach to the simulations from the nine CMIP5 models from inter-annual to centennial timescales (see “Methods” and Supplementary Texts 1, 2). The results supported that β-feedback was approximately constant across different timescales while γ-feedback increased significantly with timescale for most models (Supplementary Fig. 7a, b). Similar timescale dependences were also found for $$\beta _k^{{\mathrm{BGC}}}$$ and $${\gamma}_k^{{\mathrm{RAD}}}$$ by applying the FEA approach to the simulations from C4MIP models (Supplementary Fig. 8).

### Estimates of feedback gain (g) for 1880–2017 and high CO2 emissions scenarios

For further comparisons, we estimated the observational and model-based airborne fraction (AF) of cumulative CO2 emissions ($${\mathrm{AF}}= \Delta C_A/\Delta C_E$$) and climate–carbon cycle feedback gain factor (g) across timescales ($$g = 1 - 1/[{\mathrm{AF}}(1 + \beta )]$$, see “Methods”). We estimated $${\mathrm{AF}}_k$$ over different timescales from amplitudes of atmospheric CO2 concentration and cumulative CO2 emissions using Fourier analysis (Fig. 2a), then calculated the average and standard deviation of AF from AFk over timescales (Supplementary Fig. 7). We showed that the observation-based AF for the period 1880–2017 was nearly constant across timescales with an average of 0.40 ± 0.05 (Supplementary Fig. 7c), suggesting that about 60% of CO2 emission was taken up by land and ocean. Previous studies found the relationship between AF and the two feedback parameters: AF = 1/(1 + β + αγ*)10. As for the period 1880–2017, we found a relatively smaller contribution of αγ* (~0.05 GtC GtC−1) from α (0.005 K GtC−1) and γ* (−10.9 GtC K−1) compared to β (~1.52 GtC GtC−1) from the observational estimates, i.e., βαγ*, suggesting that the carbon-concentration feedback (β) dominated the relatively stable cumulative airborne fraction over the industrial period.

Further analysis showed that the observation-based feedback gain (g) was very small (0.01 ± 0.05) for 1880–2017 (Fig. 3c). We showed that C4MIP-based AF for 1880–2017 was also nearly constant (0.45 ± 0.06) across different timescales (Supplementary Fig. 7c), slightly larger than the observed estimate. However, the C4MIP-based g for the same period 1880–2017 was 0.09 ± 0.04, larger than the observational value by about an order of magnitude (Fig. 3c). As a result, the observation-based and C4MIP-based feedback amplification G (G = 1/(1−g), see “Methods”) are 1.01 ± 0.05 and 1.10 ± 0.04, respectively, suggesting the modeled amplification effect is about 9 ± 7% larger. Under high emission scenarios, the C4MIP-based g increased to 0.15 ± 0.08 with increased AF (0.56 ± 0.09) for 1880–2100, and the CMIP5-based g increased to 0.13 ± 0.08 for the 1pctCO2 (Fig. 3c), which are much higher than the observation-based g.

## Discussion

This study expanded the traditional climate–carbon cycle feedback framework7,10,14,19 by including a nonlinear feedback term (see Eq. (1)). Using CMIP5 modeling experiments under a scenario of high CO2 growth rate (1% yr−1), we found that the contribution of the nonlinear term to land and ocean β-feedback is relatively small (3 ± 3%), while its contribution to land and ocean γ-feedback is 15 ± 23%. These estimates are noticeably smaller than those simulated from previous modeling studies on land and ocean (20% for β and 45% for γ), or on ocean only (6% for β and 60% for γ)10,14 (see also Supplementary Text 3 for detailed discussion). However, the estimated nonlinear contributions vary significantly across CMIP5 models (0.2–9.6% for β and 0.8–45% for γ, see Supplementary Table 4), which suggests a large uncertainty in the modeled nonlinear feedback among the advanced Earth system models.

This study also stated the relationship among the four commonly used quantities (β, γ, α, and TCRE) which is consistent to some similar relationships as stated by previous carbon–climate feedback studies10,21. Using historical temperature and CO2 records over 1880–2017, we estimated that the carbon-concentration feedback parameter β is 3.22 ± 0.32 GtC ppm−1 and the climate–carbon cycle feedback gain factor g is 0.01 ± 0.05. These estimates were nearly constant across inter-annual to decadal timescales. On the other hand, we also found that the γ-feedback parameter increased with timescale from −33 ± 14 GtC K−1 on a decadal scale to −122 ± 60 GtC K−1 on a centennial scale based on reconstructions over 1000–1850. Furthermore, the estimated climate amplification from carbon–climate feedback based on observations in this study is much smaller than the previous estimates by IPCC reports based on model simulations under high emission scenarios12,13. Our results based on observations have significant implications for understanding the strength of the climate–carbon cycle feedback and the allowable CO2 emissions to mitigate future climate change. For example, the allowable emissions based on Earth system models for a 68% probability of limiting warming to 1.5 °C above the preindustrial level by ~2050 as stipulated in the Paris Agreement42, are ~115 GtC (or 420 Gt CO2)43, using the much smaller feedback gain of 0.01 ± 0.05 than the C4MIP-based estimates during the same period, we estimate that the allowable emissions would be 9 ± 7% more, or 125 ± 8 GtC.

In this study, the observation-based β and AF were found to be nearly constant (with <10% change), which together determined the nearly constant feedback gain (g). This can be explained by the linear system in response to exponential increase of forcing (LinExp) theory44,45, in which the carbon–climate system over the industrial period can be approximated as a linear system of the carbon cycle forced by exponentially growing CO2 emissions with y = 0.27e0.018t for the 1850–2017 (Supplementary Fig. 9a, b), then all ratios of responses to forcings are constant44. Using the simulations over the historical period (1901–2010) from 14 terrestrial ecosystem models forced by observational climate, land-use change, and atmospheric CO246, we showed that when excluding the effects of climate change and land use change, both annual global GPP and cumulative land carbon sink increased exponentially in response to exponentially growing atmospheric CO2 over 1901–2010 (Supplementary Fig. 9c, d). Therefore, the ratio of the exponential increase in carbon uptake and the exponential increase in CO2 sustained a nearly constant value of β over the historical period.

The timescale dependency of the γ-feedback can be used to reconcile the diverging estimates of γ parameters in the previous studies. The estimated γ at centennial timescales for the preindustrial period (−180 to −40 GtC K−1) was found to be much more negative than the value of γ* at multi-decadal timescales (−14 to −7 GtC K−1) for the industrial period, implying that the climate sensitivity of carbon cycle depends on the base climate state. If we allow the present transient climate system to reach an equilibrium state, the estimated climate sensitivity of the carbon cycle would be more negative. Furthermore, the γ (or γ*)-feedback also depends on CO2 emissions scenario. By comparing observation-based γ* and model-based γCOU-BGC from C4MIP models for the industrial period (1880–2017) with the model-based γCOU-BGC estimate for the future period with higher CO2 emissions, we showed that under the conditions of the higher CO2 growth and warmer climate states, the γ-feedback and the nonlinear feedback contribution to the γ* became more negative, while the β-feedback decreased (Fig. 3, Supplementary Table 3). The contribution of the nonlinear carbon–climate feedback to total carbon–climate feedback was estimated to be 15 ± 17% for the 1pctCO2 high emission scenario as used in the CMIP5 study. For the ocean, the nonlinear feedback may become greater due to the decreased downward carbon transport to the deep ocean owing to reduced overturning in a warmer climate14. The γ encapsulates both the direct effect of warming on plant and soil respiration47,48 and sea water solubility (γ gets more negative), but also the indirect effect of warming on biological productivity and phenology49 and on stratification/reduction of the overturning14,41 (γ gets less negative). How the γ-feedback will change in a warmer climate with high CO2 emissions depends on the contributions by these competing processes, and the uncertainties in the earth system models used for the analysis4,50. Overall, the γ-feedback at equilibrium under a high emissions scenario would become much more negative than the γ-feedback for the industrial or preindustrial periods, and likely play the dominant role in the positive net climate–carbon cycle feedback.

Uncertainties in the estimates of β, γ, and corresponding g arise from the errors in observations, model uncertainties, methodology, and study periods chosen. First, we did not consider separately the contributions of non-CO2 greenhouse gases (e.g., CH4, N2O, O3) from those of CO2 on climate variation, which may have led to some biases in the estimates of β and γ. Second, the overestimation of C4MIP-based g compared to the observation-based g for the same period 1880–2017, might be due to several reasons in model descriptions, e.g., poor descriptions of terrestrial ecosystem processes (carbon pools, carbon–water coupling, soil respiration, vegetation phenology, and ecosystem climate adaptation, etc.), omission of land-use change, and biases in model initial climate/carbon pool base-states, etc. Third, errors could arise from the two approaches (the Fourier analysis approach and the FEA approach) for estimating observation-based β and γ* or model-based βBGC and γCOU-BGC and corresponding g. Comparative analysis showed that the ensemble means of C4MIP-based βBGC and γCOU-BGC using the FEA approach are consistent with those of C4MIP-based β and γ* using the Fourier analysis approach for the 1880–2017 period, indicating a relatively small difference of C4MIP-based g between using the two approaches (Supplementary Fig. 11).

In addition, although the β-feedback was found to be stable over the industrial period with the approximately exponential growth of CO2, applying the same β value to the preindustrial period may result in bias in estimated γ for the preindustrial period. Sensitivity analysis found that an overestimation (or underestimation) of 50% in β would result in an underestimation (or overestimation) of about 30% in γk at timescales from 10 to 1000 years for the 1000–1850 period (Supplementary Fig. 5). This change of 30% (~25 GtC K−1) in γ100yr is still smaller than the uncertainty of γ100yr (a standard deviation of 41.90 GtC K−1) that was mainly caused by the large divergences in the three ice-core CO2 records and reconstructed temperatures. Furthermore, this study ignored the contribution of early land-use change to climate change ($$\zeta _{100yr} \approx 0$$) during 1000–1850, therefore may have overestimated the γ100yr, e.g., for the Little Ice Age (1400–1700). The observed atmospheric CO2 drop in the 1400–1700 period occurred because of the cooling-induced increase in terrestrial C uptake27, but it can also be partly explained by enhanced forest restoration from reduced land-use change as a result of abandonment of agricultural land from the collapse of native population in the Americas during 1500–165051. Despite these uncertainties, we demonstrated the nature of the timescale dependency of γ on climate–carbon cycle feedback and important implications for observational constraints on earth system models for projecting future climate changes.

## Methods

### Anthropogenic CO2 emission data

We calculated the global annual total anthropogenic CO2 emission flux (FE) as the sum of annual emission fluxes from fossil fuel combustion and industrial processes (FFF) and from land use and land cover changes (FLUC) by human activity covering the period 1850–2017, both of which were obtained from the Global Carbon Project’s annual global carbon budget report1. The FFF (uncertainty of ±5% for a ±1σ (68%) confidence interval) was estimated by ref. 22 and the FLUC (uncertainty of ±0.7 GtC year−1 representing a ±1σ confidence interval) was averaged from emission estimates based on two bookkeeping models by ref. 23 and ref. 24. We then calculated the cumulative annual anthropogenic CO2 emission: $$C_{\mathrm{E}}\left( t \right) = {\int}_0^t {F_{\mathrm{E}}dt} = {\int}_0^t {(F_{{\mathrm{FF}}} + F_{{\mathrm{LUC}}})} dt$$. Units were converted from GtC to ppm by dividing m (=2.12 GtC ppm−1) for comparison with atmospheric CO2 concentration52. The conversion factor m was also used for unit conversions of β, γ, α, and η parameters.

### Instrumental and ice-core reconstructed atmospheric CO2 records

Global annual atmospheric CO2 concentration over 1850–2017 was reconstructed from a combination of ice-core CO2 records and instrumental CO2 measurements since 1950s53,54. Information about the inter-annual variability in the CO2 data covering 1850–1940s from ice-core CO2 would be lost, as ice cores smooth atmospheric CO2 records by firn diffusion, which may lead to some biases in the timescale analysis.

The highly resolved CO2 covering the preindustrial period of 1000–1850 was compiled from ice-core records in Antarctica including three datasets at Law Dome30,55, WAIS Divide28, and Dronning Maud Land31, respectively. As the three records have different resolutions, CO2 values were resampled to the nearest calendar year and then smoothed using 30-year splines (Fig. 1a). However, large variations remain among the three records of CO2 from ice-core (see Fig. 1a and Supplementary Fig. 3a).

### Instrumental and reconstructed temperature datasets

All available instrumental temperature datasets including HadCRUT456 and Berkeley Earth57 over 1850–2017, and GISTEMP58 and NOAA GlobalTemp59 over 1880–2017 were used. Global annual temperature anomalies were calculated by area-weighted averaging over both land (2 m air temperature) and ocean (sea surface temperature) minus the mean surface air temperature of land and ocean for the period from 1961 to 1990.

Five reconstructions of Northern Hemispheric temperature were obtained from four groups which were abbreviated as PAGES2k32, Frank201016, Mann200934, MannEIV33, and Moberg200535 in this study (Fig. 1b). All temperature reconstructions were smoothed with 5-year splines, and adjusted to removed-means with respect to 1961–1990. Northern Hemispheric-mean temperature had been shown to be highly representative for global-mean temperature variations16.

The ensemble estimates of 521 calibrated temperature reconstructions were obtained from the Frank201016, in which all amplitude and variability on >30-year timescales are well preserved for uncertainty analysis. The ensemble estimates were recalibrated by nine available datasets of Northern Hemispheric-mean temperature reconstructions using a reconstructing technique based on a state-space time series and Kalman filter algorithm60. The ensemble means (Fig. 1b) and individual temperature reconstructions (Supplementary Fig. 4a) by Frank2010 are used for a large ensemble of >1500 estimates (EnOBS) of the η and γ over the preindustrial period. Our estimate based on Fourier analysis shows that uncertainty in the η on centennial timescales (Supplementary Fig. 4b) was consistent with the estimate by Frank2010 which was based on a lag-regression method (searching highest correlations) of CO2 and temperature reconstructions which were smoothed with spline using a range of cutoff time (50, 75,…, 200 years)16.

### Theoretical analysis for the climate–carbon cycle feedback

Following previous analysis7,10, the increase from anthropogenic CO2 emissions (ΔCE) is the sum of changes in the three stores,

$$\Delta C_E = \Delta C_A + \Delta C_L + \Delta C_O$$
(4)

where ΔCA, ΔCL, and ΔCO are the changes in carbon storage on atmosphere, land, and ocean, respectively, over a time period (Δt) evolved from the reference climate state. For the analysis for industrial period, we calculated ΔCE, ΔCA, ΔCL, and ΔCO (in a unit of GtC) during Δt since 1880. Note that change in atmospheric CO2 concentration in a unit of ppm can be converted to GtC by a factor of m (~2.12 GtC ppm−1)52. Within the climate–carbon cycle feedback system, following the linear carbon cycle feedback framework by refs. 7,10,19 and the nonlinearity of carbon cycle feedback discovered by ref. 14, we consider the changes in land and ocean carbon are the combined effects of carbon-concentration feedback (β) and carbon–climate feedback (γ) and the nonlinear feedback between climate and atmospheric CO2,

$$\left\{ {\begin{array}{*{20}{c}} {\Delta C_L = \beta _L\Delta C_A + \gamma _L\Delta T_A + f_L\left( {\beta _L,\gamma _L} \right)\Delta C_A\Delta T_A} \\ {\Delta C_O = \beta _O\Delta C_A + \gamma _O\Delta T_A + f_O(\beta _O,\gamma _O)\Delta C_A\Delta T_A} \end{array}} \right.$$
(5)

where βL and βO are the β-feedback on land and ocean, respectively, γL and γO being the γ-feedback on land and ocean, respectively. βL and βO are in a unit of GtC GtC−1 and can be converted to a commonly used unit of GtC ppm−1 by multiplying m (=2.12 GtC ppm−1). γL and γO are in a unit of GtC K−1. $$f_L(\beta _L,\gamma _L)$$ and $$f_O(\beta _O,\gamma _O)$$ are the nonlinear feedback (GtC ppm−1 K−1) as a function of β and γ feedback parameters on land and ocean, respectively. Combining Eqs. (4) and (5), we have

$$\Delta C_E = \Delta C_A + \beta \Delta C_A + \gamma \Delta T_A + f\left( {\beta ,\gamma } \right)\Delta C_A\Delta T_A = (1 + \beta )\Delta C_A + \gamma \ast \Delta T_A$$
(6)

where the total β-feedback is $$\beta _L + \beta _O$$, γ-feedback is $$\gamma _L + \gamma _O$$, and f(β, γ)-nonlinear feedback is $$f_L\left( {\beta _L,\gamma _L} \right) + f_O(\beta _O,\gamma _O)$$. As defined by ref. 14, when assuming the carbon stock in biosphere ($$C_B = C_L + C_O$$) at the reference climate state as a function of climate and CO2: $$C_B = F(C_A,T_A)$$, then β, γ, and f(β, γ) can be expressed as the 1st order and 2nd order coefficients of the Taylor series of CB since the initial time (t = 0): $$\beta = \frac{{\partial F}}{{\partial C_A}}|_0$$, $$\gamma = \frac{{\partial F}}{{\partial T_A}}|_0$$, and $$f\left( {\beta ,\gamma } \right) {\!}= \frac{{\partial ^2F}}{{\partial C_A\partial T_A}}|_0 + \frac{1}{2}\frac{{\partial ^2F}}{{\partial C_A^2}}|_0\frac{{\Delta C_A}}{{\Delta T_A}} + \frac{1}{2}\frac{{\partial ^2F}}{{\partial T_A^2}}|_0\frac{{\Delta T_A}}{{\Delta C_A}} + R_3$$. The nonlinear feedback f(β, γ) in this study represents the 2nd and high-order terms of the Taylor expansion. As previous studies mainly focused on the nonlinearity of the carbon–climate (γ-) feedback10,14,41, in this study, we combined the γ-feedback and the atmospheric CO2 change’s impacts on the nonlinear feedback as $$\gamma \ast = \gamma + f\left( {\beta ,\gamma } \right)\Delta C_A$$. From this definition, the $$\gamma \ast \approx \gamma$$, when $$\Delta C_A \approx 0$$.

The Eq. (6) can be rewritten as by dividing both sides by ΔTA,

$$\frac{1}{{{\mathrm{TCRE}}}} = (1 + \beta )\frac{1}{\alpha } + \gamma \ast$$
(7)

where $$\alpha \equiv \Delta T_A/\Delta C_A$$ is the sensitivity of climate to atmospheric CO2 (in a unit of K GtC−1), which represents the change in temperature in response to a change in CO2 concentration7,19. The α is a useful measure to quantify the feedbacks between climate and carbon cycle over both preindustrial and industrial periods. The $${\mathrm{TCRE}} \equiv \Delta T_A/\Delta C_E$$ is transient climate response to cumulative CO2 emission (in a unit of K GtC−1)20, which quantifies the ratio of change in temperature to cumulative carbon emissions, providing another useful measure to estimate the total allowable emissions for a given temperature change, as there is a near-linear relationship between cumulative CO2 emissions and global temperature change20.

In this study, we notated $$\frac{1}{\alpha }$$ as η, i.e., $$\eta \equiv \frac{1}{\alpha }$$. In some previous studies, the η was also defined as the sensitivity of atmospheric CO2 to climate for cases of no anthropogenic CO2 emission involved in the climate–carbon cycle system, e.g., over the preindustrial last millennium16,17,18,37,61. But for the industrial period, this definition expressing the sensitivity of CO2 to climate could be physically meaningless, as the CO2 increase over this period is not only due to the climatic impact on carbon stores, but also is primarily driven by the increasing anthropogenic CO2 emissions. During the preindustrial period (1000–1850), the atmospheric CO2 remained very stable (280 ± 8 ppm), with only small anthropogenic CO2 emissions from the land-use change and negligible anthropogenic emissions from fossil fuels, and those anthropogenic emissions had little induced global warming, thus37

$$\gamma _k = - m\left( {1 + \beta } \right)\eta _k - f\left( {\beta ,\gamma } \right)\Delta C_A$$
(8)

where m = 2.12 GtC ppm−1 is a factor for converting units in ppm to GtC.

### Feedback gain factor g

The gain factor (g) of the climate–carbon cycle feedback7,10,19 is expressed as

$$g = \frac{{ - \gamma \ast }}{{(1 + \beta )}}\frac{{\Delta T_A}}{{\Delta C_A}} = \frac{{ - \gamma \ast \alpha }}{{(1 + \beta )}}$$
(9)

Substituting Eq. (7) into Eq. (9) gives

$$g = 1 - \frac{\alpha }{{{\mathrm{TCRE}}}}\frac{1}{{(1 + \beta )}} = 1 - \frac{1}{{{\mathrm{AF}}(1 + \beta )}}$$
(10)

where $${\mathrm{AF}} = \Delta C_A/\Delta C_E$$ is the airborne fraction of cumulative CO2 emissions. Thus, we can further estimate the amplification factor $$G = 1/(1 - g) = {\mathrm{AF}}(1 + \beta )$$. For a net positive climate–carbon cycle feedback, g > 0 and G > 1.

### Estimating β and γ across different timescales from observations

Over the industrial CO2 emission forcing period (1850–2017), as both changes in anthropogenic CO2 emission (ΔCE) and atmospheric CO2 concentration (ΔCA) are accumulated over time intervals (Δt). From Eq. (6) with the three unknows β, γ, and f(β, γ), we can have

$$\frac{{\Delta C_E}}{{\Delta t}} = \frac{{\Delta C_A}}{{\Delta t}} + \beta \frac{{\Delta C_A}}{{\Delta t}} + \gamma \frac{{\Delta T_A}}{{\Delta t}} + f(\beta ,\gamma )\Delta C_A\frac{{\Delta T_A}}{{\Delta t}}$$
(11)

when Δt → 0, then ΔCA → 0, the nonlinear feedback contribution is close to zero, i.e., $$f\left( {\beta ,\gamma } \right)\Delta C_A \to 0,$$ hence $$\gamma \ast \approx \gamma$$. Therefore, rewriting Eq. (11) into partial differential form (Δt → 0) is

$$\frac{{\partial C_E}}{{\partial t}} = (1 + \beta )\frac{{\partial C_A}}{{\partial t}} + \gamma \ast \frac{{\partial T_A}}{{\partial t}}$$
(12)

As the year-to-year or decade-to-decade variations of CE, CA, and TA in the climate system are mostly driven by more than one factor (e.g., the El Niño-Southern Oscillation) at different timescales62,63. Based on the theory of Fourier analysis, fluctuations of these variables with time (including variability and trend) thus can be seen as the wave superposition over different frequencies64.

In this study, we then apply the Fourier analysis to the nonlinear climate–carbon cycle feedback framework to estimate β and γ across different timescales from observations. Expressing the CE, CA, and TA as functions of time t, writing them as the sum of periodic basis functions at different frequencies (Supplementary Fig. 1); this can be written as

$$\left\{ {\begin{array}{*{20}{c}} {C_E\left( t \right) = \mathop {\sum }\limits_k a_k\sin ( {\omega _kt + \varphi _{1,k}} )} \\ {C_A\left( t \right) = \mathop {\sum }\limits_k b_k\sin ( {\omega _kt + \varphi _{2,k}} )} \\ {T_A\left( t \right) = \mathop {\sum }\limits_k c_k\sin ( {\omega _kt + \varphi _{3,k}} )} \end{array}} \right.$$
(13)

where k is wavenumber (or with respect to timescale), $$\omega _k = 2\pi \frac{k}{N}$$ is the angular frequency, and ak, bk, and ak are amplitudes of CE, CA, and TA, respectively, and N is the time period in years. The time variable t varies from 0 to N. Then,

$$\left\{ {\begin{array}{*{20}{c}} {\frac{{\partial C_E}}{{\partial t}} = \mathop {\sum }\limits_k a_k\omega _k\cos ( {\omega _kt + \varphi _{1,k}} )} \\ {\frac{{\partial C_A}}{{\partial t}} = \mathop {\sum }\limits_k b_k\omega _k\cos ( {\omega _kt + \varphi _{2,k}} )} \\ {\frac{{\partial T_A}}{{\partial t}} = \mathop {\sum }\limits_k c_k\omega _k\cos ( {\omega _kt + \varphi _{3,k}} )} \end{array}} \right.$$
(14)

We can substitute Eq. (14) into Eq. (12), which yields,

$$\mathop {\sum }\limits_k h_k( {a_k - \left( {1 + \beta } \right)b_k - \gamma \ast c_k} )\omega _k\sin \left( {\omega _kt + \xi _k} \right) = 0$$
(15)

where $$\xi _k = \arctan \left( - \frac{{\cos \varphi _{1,k} + \cos \varphi _{2,k} + \cos \varphi _{3,k}}}{{\sin \varphi _{1,k} + \sin \varphi _{2,k} + \sin \varphi _{3,k}}}\right)$$, and $$h_k = \sqrt {( {\sin \varphi _{1,k} + \sin \varphi _{2,k} + \sin \varphi _{3,k}} )^2 + ( {\cos \varphi _{1,k} + \cos \varphi _{2,k} + \cos \varphi _{3,k}} )^2} {\,}\ne{\,} 0$$. We find that Eq. (15) only holds when

$$a_k - \left( {1 + \beta } \right)b_k - \gamma \ast c_k = 0$$
(16)

Comparing Eq. (16) with Eq. (7), we notate $$\frac{1}{{{\mathrm{TCRE}}}}$$ as ζ, i.e., $$\zeta \equiv \frac{1}{{{\mathrm{TCRE}}}}$$, and $$\frac{1}{\alpha }$$ as η, then have the solutions of ζ and η over timescales,

$$\zeta _k = \frac{{a_k}}{{c_k}}$$
(17)
$$\eta _k = \frac{{b_k}}{{c_k}}$$
(18)

The η can be estimated from ratios of the amplitudes in CE to the amplitudes in TA at any given timescale. Similarly, the ζ can be estimated from the amplitudes in CE and TA. Fourier analysis based on fast Fourier transform (FFT)65 is used for estimating amplitudes for each time series of CE, CA, and TA over industrial or preindustrial periods (see Supplementary Fig. 1). From Eqs. (16)–(18), we have

$$\zeta _k = \left( {1 + \beta } \right)\eta _k + \gamma \ast$$
(19)

From Eq. (19) or Eq. (2) and Fig. 2c, we find that the β can be estimated from a linear regression of ζk and ηk with residual errors ()

$$\zeta _k = p\eta _k + {\it{\epsilon }}$$
(20)

where $$p= \frac{{{\mathrm{cov}}(\zeta _k,\,\eta _k)}}{{{\mathrm{var}}(\eta _k)}}$$, then $$\beta = p - 1$$. The γ* on different timescales ($$\gamma _k \ast$$) for the industrial period can be calculated by Eq. (20) with input from estimates of ζk (or $${\mathrm{TCRE}}_k^{ - 1}$$) and ηk (or $$\alpha _k^{ - 1}$$) for a given timescale. If considering that the β-feedback of the climate–carbon cycle system is the same over both the industrial and preindustrial periods on long-term timescales, e.g., of 100 year, assuming $$\zeta _{100yr} \approx 0$$, from Eq. (8) we have

$$\gamma _{100yr} = - m\left( {1 + \beta } \right)\eta _{100yr}$$
(21)

where the η100yr is estimated on the 100-year timescale from temperature reconstructions and CO2 ice-core records over 1000–1850.

### Estimating β and γ and f(β, γ) from C4MIP and CMIP5 simulations

We followed the FEA approach based on a Taylor series expansion as defined in Friedlingstein et al.7, Arora et al.11, and Schwinger et al.14 of estimating β and γ and over a time period of N years from the biogeochemically-coupled (BGC) and radiatively coupled (RAD) CMIP5 simulations, respectively, using the COU-BGC, COU-RAD experiment pairs for Δt = N. In this study, we used the COU-BGC approach:

$$\beta ^{{\mathrm{BGC}}} = \frac{{\Delta C_B^{{\mathrm{BGC}}}\Delta T_A^{{\mathrm{COU}}} - \Delta C_B^{{\mathrm{COU}}}\Delta T_A^{{\mathrm{BGC}}}}}{{\Delta C_A^{{\mathrm{BGC}}}\left( {\Delta T_A^{{\mathrm{COU}}} - \Delta T_A^{{\mathrm{BGC}}}} \right)}} \approx \frac{{\Delta C_B^{{\mathrm{BGC}}}}}{{\Delta C_A^{{\mathrm{BGC}}}}}$$
(22)
$$\gamma ^{{\mathrm{COU}} - {\mathrm{BGC}}} = \frac{{\Delta C_B^{{\mathrm{COU}}} - \Delta C_B^{{\mathrm{BGC}}}}}{{\Delta T_A^{{\mathrm{COU}}} - \Delta T_A^{{\mathrm{BGC}}}}} \approx \frac{{\Delta C_B^{{\mathrm{COU}}} - \Delta C_B^{{\mathrm{BGC}}}}}{{\Delta T_A^{{\mathrm{COU}}}}}$$
(23)

$$\beta ^{{\mathrm{COU}} - {\mathrm{RAD}}} = \frac{{\Delta C_B^{{\mathrm{RAD}}}\Delta T_A^{{\mathrm{COU}}} - \Delta C_B^{{\mathrm{COU}}}\Delta T_A^{{\mathrm{RAD}}}}}{{\Delta C_A^{{\mathrm{COU}}}\Delta T_A^{{\mathrm{RAD}}}}} \approx \frac{{\Delta C_B^{{\mathrm{COU}}} - \Delta C_B^{{\mathrm{RAD}}}}}{{\Delta C_A^{{\mathrm{COU}}}}}$$
(24)
$$\gamma ^{{\mathrm{RAD}}} = \frac{{\Delta C_B^{{\mathrm{RAD}}}}}{{\Delta T_A^{{\mathrm{RAD}}}}}$$
(25)

where $$\Delta C_A^{{\mathrm{COU}}} = \Delta C_A^{{\mathrm{BGC}}}$$ for CMIP5 or C4MIP simulations. To calculate the nonlinear feedback parameter f(β, γ), we also defined:

$$f\left( {\beta ,\gamma } \right) = \frac{{\Delta C_B^{{\mathrm{COU}}} - (\Delta C_B^{{\mathrm{BGC}}} + \Delta C_B^{{\mathrm{RAD}}})}}{{\Delta C_A^{{\mathrm{COU}}}(\Delta T_A^{{\mathrm{COU}}} - \Delta T_A^{{\mathrm{BGC}}})}} \approx \frac{{\Delta C_B^{{\mathrm{COU}}} - (\Delta C_B^{{\mathrm{BGC}}} + \Delta C_B^{{\mathrm{RAD}}})}}{{\Delta C_A^{{\mathrm{COU}}}\Delta T_A^{{\mathrm{COU}}}}}$$
(26)

We set the reference time at 1880 and then calculated the βBGC, $$\gamma ^{{\mathrm{RAD}}}$$, γCOU-BGC, and f(β, γ) for the observation-overlapped period of 1880–2017 and the future emission scenario of 1880–2100 for the C4MIP models, and the 1pctCO2 140-year period for the CMIP5 models, respectively.

### A box model for diagnosing the climate–carbon cycle feedback parameters

The coupled climate–carbon cycle system can be simplified as the combination of variations in temperature (TA = T0 + ΔTA) and CO2 (CA = C0 + ΔCA) over time intervals (Δt). The change in TA is assumed to increase logarithmically with CO217,66,

$$\Delta T_A = \frac{s}{{{\mathrm{ln}}(2)}} * {\mathrm{ln}}\left(\frac{{C_A}}{{C_0}}\right) + \varepsilon$$
(27)

where s is the impact of CO2 on the temperature that is suggested to be 1.5–4.5 K by the IPCC reports67, and ε is the residual term from climate internal variability. From Eq. (6),

$$\Delta C_A = \frac{{\Delta C_E - \gamma \ast \Delta T_A}}{{1 + \beta }}$$
(28)

We then have the three parameters box model with input from CO2 emissions,

$$C_A = C_0 + \frac{{\Delta C_E - \gamma \ast (T_A - T_0)}}{{1 + \beta }}$$
(29)
$$T_A = T_0 + \frac{s}{{{\mathrm{ln}}(2)}}{\mathrm{ln}}\left( {\frac{{C_A}}{{C_0}}} \right) + \varepsilon$$
(30)

where C0 and T0 are initial values at the first year, here referenced to 1850. To validate the estimated β and γ* for the industrial period, we applied this box model to predict temperature and CO2 over 1850–2017 using annual cumulated CO2 emissions, by setting β = 3.22 GtC ppm−1 = 1.52 GtC GtC−1, γ* = −10.9 GtC K−1, s = 3 ± 1.5 K and ε being the detrended TA anomaly time series from HadCRUT4 (Supplementary Fig. 2). Result shows that the increasing trend in predicted annual CO2 is very close to the observation ($$R^2 = 0.99,\,{\mathrm{RMSE}} = 3.5\,{\mathrm{ppm}}$$), while predicted temperature has a larger trend since 1980s ($$R^2 = 0.96,\,{\mathrm{RMSE}} = 0.17\,{\mathrm{K}}$$), compared to observed records (Supplementary Fig. 2).

### Uncertainty

We quantified the uncertainty by a ±1σ (standard deviation) that represents a 68% confidence interval. The uncertainty in CE (Fig. 1) resulted from the uncertainty in FE ($${\sigma}_{F_{\mathrm{E}}}$$) that was calculated by $${\sigma}_{F_{\mathrm{E}}} = \sqrt {\sigma _{F_{{\mathrm{FF}}}}^2 + \sigma _{F_{{\mathrm{LUC}}}}^2}$$. The uncertainty in FFF ($${\sigma}_{F_{{\mathrm{FF}}}}$$) was ±5% of FFF22. This is consistent with a more detailed recent analysis of uncertainty for FFF68. The uncertainty in FLUC ($${\sigma}_{F_{{\mathrm{LUC}}}}$$) was ±0.7 GtC year−1 at each year1. To estimate the β and its uncertainty (σβ), we first estimated the βi and $${\sigma}_{\beta ,i}$$ ($$i = 1, \cdots ,4$$) for each combination of 1 CO2 × 1 CO2 emission × 4 temperature datasets (HadCRUT4, GISTEMP, Berkeley Earth, and NOAA GlobalTemp) that included uncertainty on CE and the regression. We then estimated the mean value of β by $$\overline \beta = \frac{1}{4}{\sum} {\beta _i}$$ and the σβ by $${\sigma}_\beta = \frac{1}{4}\sqrt {{\sum} {\sigma _{\beta _i}^2} }$$ (Supplementary Table 1). Similarly, we estimated uncertainty in γ (σγ) for the industrial period at a given timescale by $${\sigma}_\gamma = \frac{1}{4}\sqrt {{\sum} {\sigma _{\gamma _i}^2} }$$. As the uncertainty in γ (σγ) for the preindustrial period arises from uncertainties in ice-core CO2 records and temperature reconstructions, we estimated the σγ over 1000–1850 or 1400–1700, or 1000–1300 periods at a given timescale by calculating the standard deviation of the ensemble of γ from 3 ice-core CO2 records × 5 temperature reconstructions (or the EnOBS): $${\sigma}_\gamma = \sqrt {\frac{1}{{n - 1}}\mathop {\sum}\nolimits_{i = 1}^{n - 1} {(\gamma _i - \overline \gamma )^2} }$$ (Supplementary Table 2, Fig. 3c, d). The estimated uncertainties in β and γ for C4MIP or CMIP5 models were calculated as the standard deviations of their ensembles of β and γ for each model (Supplementary Tables 3, 4). As $$G = \frac{1}{{1 - g}} \approx 1 + g$$, uncertainty in G (σG) was σg, and uncertainty in relative change of G from C4MIP models compared to OBS ((GMODGOBS)/GOBS) was $$\frac{{G_{{\mathrm{MOD}}}}}{{G_{{\mathrm{OBS}}}}} * \sqrt {\left( {\frac{{\sigma _{{\mathrm{GMOD}}}}}{{G_{{\mathrm{MOD}}}}}} \right)^2 + \left( {\frac{{\sigma _{{\mathrm{GOBS}}}}}{{G_{{\mathrm{OBS}}}}}} \right)^2}$$.

## Code availability

The NCL processing codes are available via GitHub at https://github.com/xuanzezhang/climate–carbon-cycle-feedback (https://doi.org/10.5281/zenodo.4575812) or from the corresponding author upon request.

## References

1. Le Quéré, C. et al. Global carbon budget 2017. Earth Syst. Sci. Data 10, 405–448 (2018).

2. Shevliakova, E. et al. Historical warming reduced due to enhanced land carbon uptake. Proc. Natl Acad. Sci. USA 110, 16730–16735 (2013).

3. Holden, P. B. et al. Climate–carbon cycle uncertainties and the Paris Agreement. Nat. Clim. Change 8, 609–613 (2018).

4. Friedlingstein, P. et al. Uncertainties in CMIP5 climate projections due to carbon cycle feedbacks. J. Clim. 27, 511–526 (2014).

5. Friedlingstein, P. Carbon cycle feedbacks and future climate change. Philos. Trans. A Math. Phys. Eng. Sci. 373, 20140421 (2015).

6. Cox, P. M. et al. Acceleration of global warming due to carbon-cycle feedbacks in a coupled climate model. Nature 408, 184–187 (2000).

7. Friedlingstein, P. et al. Climate–carbon cycle feedback analysis: results from the C4MIP model intercomparison. J. Clim. 19, 3337–3353 (2006).

8. Dufresne, J. L. et al. On the magnitude of positive feedback between future climate change and the carbon cycle. Geophys. Res. Lett. 29, 43-41–43-44 (2002).

9. Bolin, B. et al. Carbon Cycle Modelling (SCOPE 16) (John Wiley, New York, 1981).

10. Gregory, J. M., Jones, C. D., Cadule, P. & Friedlingstein, P. Quantifying carbon cycle feedbacks. J. Clim. 22, 5232–5250 (2009).

11. Arora, V. K. et al. Carbon–concentration and carbon–climate feedbacks in CMIP5 Earth system models. J. Clim. 26, 5289–5314 (2013).

12. Ciais, P. et al. Carbon and other biogeochemical cycles. in Climate Change 2013: The Physical Science Basis. Contribution of Working Group I to the Fifth Assessment Report of the Intergovernmental Panel on Climate Change (eds. Stocker, T. F. et al.) (Cambridge University Press, 2013).

13. Denman, K. L. et al. Couplings between changes in the climate system and biogeochemistry. in Climate Change 2007: The Physical Science Basis. Contribution of Working Group I to the Fourth Assessment Report of the Intergovernmental Panel on Climate Change (eds. Solomon, S. et al.) (Cambridge University Press, 2007).

14. Schwinger, J. et al. Nonlinearity of ocean carbon cycle feedbacks in CMIP5 Earth system models. J. Clim. 27, 3869–3888 (2014).

15. Cox, P. M. et al. Sensitivity of tropical carbon to climate change constrained by carbon dioxide variability. Nature 494, 341–344 (2013).

16. Frank, D. C. et al. Ensemble reconstruction constraints on the global carbon cycle sensitivity to climate. Nature 463, 527–530 (2010).

17. Scheffer, M., Brovkin, V. & Cox, P. M. Positive feedback between global warming and atmospheric CO2 concentration inferred from past climate change. Geophys. Res. Lett. 33, L10702 (2006).

18. Cox, P. & Jones, C. Climate change. Illuminating the modern dance of climate and CO2. Science 321, 1642–1644 (2008).

19. Friedlingstein, P., Dufresne, J. L., Cox, P. M. & Rayner, P. How positive is the feedback between climate change and the carbon cycle? Tellus B 55, 692–700 (2003).

20. Matthews, H. D., Gillett, N. P., Stott, P. A. & Zickfeld, K. The proportionality of global warming to cumulative carbon emissions. Nature 459, 829–832 (2009).

21. Jones, C. D. & Friedlingstein, P. Quantifying process-level uncertainty contributions to TCRE and carbon budgets for meeting Paris Agreement climate targets. Environ. Res. Lett. 15, 074019 (2020).

22. Boden, T. A., Marland, G. & Andres, R. J. Global, Regional, and National Fossil-Fuel CO2 Emissions. Carbon Dioxide Information Analysis Center, Oak Ridge National Laboratory, U.S. Department of Energy, Oak Ridge, Tenn., USA (2017).

23. Houghton, R. A. & Nassikas, A. A. Global and regional fluxes of carbon from land use and land cover change 1850-2015. Glob. Biogeochem. Cycles 31, 456–472 (2017).

24. Hansis, E., Davis, S. J. & Pongratz, J. Relevance of methodological choices for accounting of land use change carbon fluxes. Glob. Biogeochem. Cycles 29, 1230–1246 (2015).

25. Myhre, G., Highwood, E. J., Shine, K. P. & Stordal, F. New estimates of radiative forcing due to well mixed greenhouse gases. Geophys. Res. Lett. 25, 2715–2718 (1998).

26. He, J. et al. Transient climate sensitivity depends on base climate ocean circulation. J. Clim. 30, 1493–1504 (2017).

27. Rubino, M. et al. Low atmospheric CO2 levels during the Little Ice Age due to cooling-induced terrestrial uptake. Nat. Geosci. 9, 691–694 (2016).

28. Bauska, T. K. et al. Links between atmospheric carbon dioxide, the land carbon reservoir and climate over the past millennium. Nat. Geosci. 8, 383–387 (2015).

29. Pongratz, J., Reick, C., Raddatz, T. & Claussen, M. A reconstruction of global agricultural areas and land cover for the last millennium. Glob. Biogeochem. Cycles 22, L02709 (2008).

30. Rubino, M. et al. A revised 1000 year atmospheric δ13C-CO2 record from Law Dome and South Pole, Antarctica. J. Geophys. Res. -Atmos. 118, 8482–8499 (2013).

31. Siegenthaler, U. et al. Supporting evidence from the EPICA Dronning Maud Land ice core for atmospheric CO2 changes during the past millennium. Tellus B 57, 51–57 (2005).

32. PAGES2k Consortium. Continental-scale temperature variability during the past two millennia. Nat. Geosci. 6, 339–346 (2013).

33. Mann, M. E. et al. Proxy-based reconstructions of hemispheric and global surface temperature variations over the past two millennia. Proc. Natl Acad. Sci. USA 105, 13252–13257 (2008).

34. Mann, M. E. et al. Global signatures and dynamical origins of the Little Ice Age and Medieval Climate Anomaly. Science 326, 1256–1260 (2009).

35. Moberg, A. et al. Highly variable Northern Hemisphere temperatures reconstructed from low- and high-resolution proxy data. Nature 433, 613–617 (2005).

36. Koch, A., Brierley, C., Maslin, M. M. & Lewis, S. L. Earth system impacts of the European arrival and Great Dying in the Americas after 1492. Quat. Sci. Rev. 207, 13–36 (2019).

37. Willeit, M. et al. Time-scale and state dependence of the carbon-cycle feedback to climate. Clim. Dyn. 42, 1699–1713 (2014).

38. Terrer, C. et al. Nitrogen and phosphorus constrain the CO2 fertilization of global plant biomass. Nat. Clim. Change 9, 684–689 (2019).

39. Liu, Y. et al. Field-experiment constraints on the enhancement of the terrestrial carbon sink by CO2 fertilization. Nat. Geosci. 12, 809–814 (2019).

40. Arora, V. K. et al. Carbon–concentration and carbon–climate feedbacks in CMIP6 models and their comparison to CMIP5 models. Biogeosciences 17, 4173–4222 (2020).

41. Zickfeld, K. et al. Nonlinearity of carbon cycle feedbacks. J. Clim. 24, 4255–4275 (2011).

42. Paris Climate Agreement (UNFCCC, 2015).

43. IPCC. in Global Warming of 1.5°C (eds Masson-Delmotte, V. et al.) (World Meteorological Organization, 2018).

44. Raupach, M. R. The exponential eigenmodes of the carbon-climate system, and their implications for ratios of responses to forcings. Earth Syst. Dyn. 4, 31–49 (2013).

45. Gloor, M., Sarmiento, J. L. & Gruber, N. What can be learned about carbon cycle climate feedbacks from the CO2 airborne fraction? Atmos. Chem. Phys. 10, 7739–7751 (2010).

46. Huntzinger, D. N. et al. Uncertainty in the response of terrestrial carbon sink to environmental drivers undermines carbon-climate feedback predictions. Sci. Rep. 7, 4765 (2017).

47. Cheng, L. et al. Warming enhances old organic carbon decomposition through altering functional microbial communities. ISME J. 11, 1825–1835 (2017).

48. Schuur, E. A. G. et al. Climate change and the permafrost carbon feedback. Nature 520, 171–179 (2015).

49. Piao, S. et al. Plant phenology and global climate change: current progresses and challenges. Glob. Change Biol. 25, 1922–1940 (2019).

50. Randerson, J. T. et al. Multicentury changes in ocean and land contributions to the climate-carbon feedback. Glob. Biogeochem. Cycles 29, 744–759 (2015).

51. Stocker, B. D., Yu, Z. C., Massa, C. & Joos, F. Holocene peatland and ice-core data constraints on the timing and magnitude of CO2 emissions from past land use. Proc. Natl Acad. Sci. USA 114, 1492–1497 (2017).

52. Ballantyne, A. P. et al. Increase in observed net carbon dioxide uptake by land and oceans during the past 50 years. Nature 488, 70–72 (2012).

53. Keeling, C. D. et al. Atmospheric carbon dioxide variations at Mauna Loa Observatory, Hawaii. Tellus 28, 538–551 (1976).

54. Dlugokencky, E. & Tans, P. Trends in atmospheric carbon dioxide. http://www.esrl.noaa.gov/gmd/ccgg/trends/global.html (2018).

55. MacFarling Meure, C. et al. Law Dome CO2, CH4 and N2O ice core records extended to 2000 years BP. Geophys. Res. Lett. 33, https://doi.org/10.1029/2006GL026152 (2006).

56. Morice, C. P., Kennedy, J. J., Rayner, N. A. & Jones, P. D. Quantifying uncertainties in global and regional temperature change using an ensemble of observational estimates: the HadCRUT4 data set. J. Geophys. Res. -Atmos. 117, D08101 (2012).

57. Rohde, R. et al. A new estimate of the average Earth surface land temperature spanning 1753 to 2011. Geoinfor. Geostat.: Overv. 01, 1–7 (2013).

58. Hansen, J., Ruedy, R., Sato, M. & Lo, K. Global surface temperature change. Rev. Geophys. 48, RG4004 (2010).

59. Vose, R. S. et al. NOAA’s merged land–ocean surface temperature. Anal. B. Am. Meteorol. Soc. 93, 1677–1685 (2012).

60. Lee, T. C. K., Zwiers, F. W. & Tsao, M. Evaluation of proxy-based millennial reconstruction methods. Clim. Dyn. 31, 263–281 (2007).

61. Woodwell, G. M. et al. Biotic feedbacks in the warming of the earth. Clim. Change 40, 495–518 (1998).

62. Torrence, C. & Compo, G. P. A practical guide to wavelet. Anal. B. Am. Meteorol. Soc. 79, 61–78 (1998).

63. Zhang, X. et al. Dominant regions and drivers of the variability of the global land carbon sink across timescales. Glob. Change Biol. 24, 3954–3968 (2018).

64. Athreya, K. B. & Lahiri, S. N. Measure Theory and Probability Theory (Springer, 2006).

65. Stoica, P. & Moses, R. Spectral Analysis of Signals (Prentice Hall, 2005).

66. Budyko, M. I. The Earth’s Climate, Past and Future, 1st edn (Elsevier, 1982).

67. IPCC. Climate change 2013: the physical science basis. Contribution of Working Group I to the Fifth Assessment Report of the Intergovernmental Panel on Climate Change. Cambridge, United Kingdom and New York, NY, USA (2013).

68. Andres, R. J., Boden, T. A. & Higdon, D. A new evaluation of the uncertainty associated with CDIAC estimates of fossil fuel carbon dioxide emission. Tellus B 66, 23616 (2014).

## Acknowledgements

This study was funded by the National Key Research and Development Program of China (2017YFA0604603), the CAS Pioneer Talent Program, and the National Key Research and Development Program of China (2016YFA0602501). X.Z.Z. was also sponsored by the National Natural Science Foundation of China (42001019) and the Shanghai Sailing Program (19YF1413100). Y.P.W. was supported by the National Environmental Science Program (climate change and earth system science). P.C. acknowledges support from the European Research Council Synergy grant ERC-2013-SyG-610028 IMBALANCE-P. We gratefully thank Dr. Pierre Friedlingstein for providing constructive comments. We acknowledge the following data providers and model developers: the Global Carbon Project, the University of East Anglia Climatic Research Unit (CRU), the NASA Goddard Institute for Space Studies, the NOAA, and Berkeley Earth. We also gratefully thank Drs Chris Jones and Vivek Arora for providing the outputs of 11 C4MIP and nine CMIP5 models.

## Author information

Authors

### Contributions

X.Z.Z. and Y.P.W. designed the study. X.Z.Z. performed the analysis. X.Z.Z., Y.P.W., Y.Q.Z., and P.C. drafted the paper. P.J.R. proposed the conception of box model. K.H., Y.Q.L., S.L.P., Z.L.W., J.Y.X., W.Z., X.G.Z., and J.T. contributed to the interpretation of the results and to the text.

### Corresponding authors

Correspondence to Xuanze Zhang, Ying-Ping Wang or Yongqiang Zhang.

## Ethics declarations

### Competing interests

The authors declare no competing interests.

Peer review information Nature Communications thanks the anonymous reviewers for their contributions to the peer review of this work. Peer review reports are available.

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

## Rights and permissions

Reprints and Permissions

Zhang, X., Wang, YP., Rayner, P.J. et al. A small climate-amplifying effect of climate-carbon cycle feedback. Nat Commun 12, 2952 (2021). https://doi.org/10.1038/s41467-021-22392-w

• Accepted:

• Published:

• DOI: https://doi.org/10.1038/s41467-021-22392-w