Introduction

Recent studies have warned of an unexplained warming surge, with new global temperature records set in both 2023 and 20241,2,3,4,5,6,7,8. These temperature extremes were preceded by a multi-decadal decline in the proportion of solar radiation reflected back into space by low-level clouds, leading to an increasing amount of shortwave radiation reaching the ocean, with this trend accelerating in recent years7,9,10,11,12,13. The reductions in low-level cloud reflectivity reached a new extreme in the extratropical Northern Hemisphere in 2023 and 202414.

Regionally, North Atlantic sea surface temperatures (SSTs) have been rising faster than elsewhere12,14. In the Northeast Pacific, a period of persistent marine heatwaves beginning in 2014 has had devastating impacts on marine ecosystems and fisheries15,16. The exceptional sea surface temperatures cannot be explained by the historically observed Pacific Decadal Oscillation, which had been the dominant mode of long-term SST variability prior to 201417,18.

It is currently unknown whether the multi-decadal decline in cloud reflectivity has increased the likelihood of the temperature records in 2023 and 2024. This seems probable given that the time for the climate system to respond to a perturbation in radiative fluxes is measured in years to decades19,20. Overall, the causes of the recent rapid warming, and the cloud reflectivity reduction possibly driving it, remain poorly understood. Volcanic eruptions, variations in solar activity, and the El Niño Southern Oscillation are unlikely to have notably contributed14,21,22. Simulations conducted with different Earth System Model (ESM) ensembles have largely failed to replicate the full magnitude of the observed changes5,7,11,21. The daunting possibility exists that the current predictive capabilities are underestimating the warming of the Earth system21,22,23. In addition to a possible under-estimation in shortwave cloud feedbacks24, various studies have emphasized the need for better understanding of the role of aerosol changes on cloud reflectivity trends. Globally, emissions of aerosols and precursors have been declining for more than a decade, primarily due to the introduction of stricter standards for sulfur dioxide (SO2) in China25,26,27 and other countries in the Northern Hemisphere, consistent with earlier analyses of satellite observations28,29. Given that aerosols can form cloud droplets by serving as cloud condensation nuclei, reductions in aerosols have resulted in decreased cloud droplet concentrations in regions affected by declining emissions29. This could have reduced the cloud reflectivity30,31, but the extent of the climate impacts of these changes remains uncertain21,23. Complicating the matter, the chemical composition of the aerosols and therefore their impacts on climate can vary, depending on the source of the emissions32,33. Most significantly, assessments of the role of aerosols in the observed trends are affected by known biases and large uncertainties in simulated aerosol-cloud interactions in the current generation of ESMs34.

Here we first establish whether the rapidly declining cloud reflectivity in the North Atlantic and Northeast Pacific correlates with aerosol reductions in satellite observations and Earth Systen Model simulations. We then used a newly developed version of the global Canadian Atmospheric Model (CanAM5.1-PAM) to determine the contributions of aerosol reductions and SST changes to the observed aerosol and cloud trends. Finally, we analyzed the model simulations to determine the contributions of key aerosol-cloud interactions to the observed trends.

Results

Observed cloud trends

Our analysis of two decades of satellite observations from the Clouds and the Earth’s Radiant Energy System (CERES) Energy Balanced and Filled (EBAF) Ed4.2 dataset35,36,37 show that clouds have become considerably less reflective over the combined North Atlantic and Northeast Pacific regions over the recent 20-year long period from January 1, 2003 to December 31, 2022, consistent with earlier studies7,13,14. In these regions, the shortwave cloud radiative effect (CRE), which is the difference between the solar radiation that would be reflected back to space without clouds and the actual reflected radiation, weakened more rapidly than anywhere else in the Northern Hemisphere (Figs. 1, S1). Over these 2 decades, the annual mean CRE over the North Atlantic and Northeast Pacific steadily increased to less negative values by 0.91 ± 0.64 W m−2 decade−1 (95% confidence) and 1.18 ± 0.86 W m−2 decade−1, respectively. The mean trend for the combined North Atlantic and Northeast Pacific regions was 1.04 ± 0.45 W m−2 decade−1, a relative change of 2.8 ± 1.2% decade−1. This implies a contribution to the Earth’s global energy budget of 0.15 ± 0.06 W m−2 decade−1, given that these marine regions cover 14% of the Earth’s surface area.

Fig. 1: Observed changes in clouds, sea surface temperature, and aerosols.
Fig. 1: Observed changes in clouds, sea surface temperature, and aerosols.
Full size image

The linear trends in annual mean shortwave cloud radiative effect (CRE, A), sea surface temperature (SST, B), cloud droplet number concentration at cloud top (Nd, C), and aerosol optical depth at 550 nm (D), from January 1, 2003 to December 31, 2022. Stippling indicates that the trends are statistically significant (95% confidence interval). The corresponding annual averages for the Northeast Pacific and North Atlantic regions, bounded by 180∘E to 5∘W and 15∘N to 55∘N (black rectangle), are shown in Fig. S1. Here and elsewhere, averages refer to the marine regions within these boundaries. See “Methods” for details.

Consistent with ref. 13 and ref. 14 we find that the CRE increase was by far the leading cause of the observed reduction in total reflected solar radiation (Figs. S2 and S3). Given the surface warming, a moderate increase in outgoing longwave radiation was observed, which acted to dampen the warming trend (ref. 14, Figs. S2 and S3). Furthermore, the CRE increase was associated with a reduction in total cloud fraction and increase in cloud droplet size (Figs. S2 and S3). Our simulations show that changes in cloud fraction and droplets both contributed to the observed CRE trend (Drivers of cloud trends).

Over the same time period, the observed SST in the HadCRUT5 dataset38,39 increased significantly over large regions of the ocean, with a particularly rapid warming of 0.19 ± 0.08 K decade−1 in the North Atlantic and Northeast Pacific, in close proximity to the observed extrema in the CRE and cloud fraction trends (Figs. 1, S1, and S2). These changes can be partly explained by the marine low cloud feedback, which can be expected to amplify the observed warming in these regions24. The fraction and reflectivity of clouds are known to be impacted by various meteorological controlling factors. ref. 40 concluded that a decrease in the strength of the temperature inversion and an SST increase caused the low cloud fraction to decrease in the Northeast Pacific region between 2001 and 2020. Similar, ref. 41 attributed reductions in cloud fraction and reflected shortwave radiation over the North Atlantic between 1970 and 2014 to greenhouse gas-induced warming. However, the extent to which cloud-controlling factors influenced the regional CRE trends remains uncertain.

Similar to ref. 42, we find that the observed CRE and SST trends cannot be solely attributed to the unforced internal climate variability in the trends in these regions, which we estimate to be within the ranges of ±0.38 W m−2 decade−1 for the CRE and ±0.11 K decade−1 for the SST (Methods). Therefore, a key question is whether the observed clouds trends were ultimately forced by changing fossil fuel and biofuel sources of aerosols. Indeed, we find evidence for rapidly changing cloud processes that are indicative of aerosol impacts on cloud reflectivity. Many cloud processes depend on the cloud droplet number concentration (Nd), including precipitation efficiency and cloud optical properties. Satellite observations of Nd using data from the Moderate Resolution Imaging Spectroradiometer (MODIS) Terra instruments (Methods) provide evidence for notable reductions in Nd over the North Atlantic (-6.7 ± 1.5 cm−3 decade−1) and Northeast Pacific (-2.9 ± 1.7 cm−3 decade−1). The co-location and magnitudes of the Nd and CRE trends raises the question whether aerosol reductions may have contributed to the CRE increase. Cloud droplets form as water vapor condenses on aerosol particles that are sufficiently large and water-soluble, widely referred to as cloud condensation nuclei. Both observations and models show that marine boundary layer cloud reflectivity is increased as the concentration of aerosols from emissions of SO2 increases43. Further, the reflectivity of the persistent region of boundary-layer clouds over the Northeast Pacific has been shown to be particularly susceptible to changes in aerosol concentrations44.

The importance of aerosol-cloud interactions for the CRE trends is further supported by the high correlation between the Nd trends and reductions in aerosol optical depth (AOD) over the North Atlantic (-0.0063 ± 0.0035 decade−1) and Northeast Pacific (-0.0069 ± 0.0035 decade−1), as derived from the Multiangle Imaging SpectroRadiometer (MISR) Level 3 Global Joint Aerosol monthly product V002 (refs. 45,46,47, Figs. 1 and S1). The pronounced reductions in AOD along the eastern boundaries of Eurasia and North America indicate that these were caused by declining anthropogenic emissions of SO2 and other aerosol precursors over large regions of these continents. Annual anthropogenic emissions of SO2 from China and the United States of America declined rapidly over the 20 year period of the observational data (-15.9 and -7.1 Mt decade−1, respectively, Figs. S4 and S5), largely owing to the introduction of stringent standards for emissions from power plants aimed at improving air quality in these countries25,48,49,50,51.

Earth System Model Simulations

To analyze the causes of the observed CRE increase we first turned to results from 24 ESMs that participated in the Coupled Model Intercomparison Project Phase 6 (CMIP652, Supplementary Table 1). These models provide the radiative fluxes required for the calculation of the CRE, using the future SSP2-4.5 scenario11 between January 1, 2015 and December 31, 2022 (Methods).

We find that the both the spatial extent and magnitude of the CMIP6 multi-model mean simulated CRE increase are weaker than observed (Fig. 2A). According to the CERES satellite observations, the CRE experienced an increase in 80% of the combined North Atlantic and Northeast Pacific regions, whereas the simulated CRE increased in 64% of these regions within the multi-model ensemble (ranging from 45 to 86%). The CRE trend surpassed 2 W m−2 decade−1 in 21% (ranging from 5 to 46%) of the regions within the multi-model ensemble, which is considerably less than the observed 34%.

Fig. 2: The spatial extents and magnitudes of the trends.
Fig. 2: The spatial extents and magnitudes of the trends.
Full size image

Cumulative area within the North Atlantic plus Northeast Pacific regions (Fig. 1) for which the decadal trend of the cloud radiative effect (CRE) falls between nil and the value indicated on the abscissa (A). The trends of the sea surface temperature (SST, B) and cloud droplet number concentration (Nd, C) in these locations are also shown. The black line corresponds to the observational data sets as described in “Observed cloud trends”. The gray lines refer to results from simulations with 24 Coupled Model Intercomparison Project Phase 6 (CMIP6) Earth System Models, which includes a subset of models that provided Nd. Results from simulations with the model CanAM5.1-PAM are shown in red, which incorporates the Twomey and Albrecht effects (Earth System Model simulations). Thick lines refer to ensemble mean results and dashed lines to the individual ensemble members. See “Observed cloud trends”, “Earth System Model simulations”, and the Supplementary Notes for details.

Similarly, the multi-model mean underestimates the extensive spatial extent and the intensity of the observed SST warming, exhibiting a comparable pattern to the CRE trends (Fig. S6). Consequently, the CMIP6 multi-model mean CRE and surface temperature trends in these regions are weaker than observed, with simulated CRE and temperature trends of 0.42 W m−2 decade−1 (range from -0.2 to 1.15 W m−2 decade−1) and 0.15 K decade−1 (range from 0 to 0.42 K decade−1), respectively. In comparison, the observed trends are 1.04 ± 0.45 W m−2 decade−1 and 0.19  ± 0.08 K decade−1 (Fig. 3), respectively.

Fig. 3: Regionally averaged trends in Earth System Models and observations.
Fig. 3: Regionally averaged trends in Earth System Models and observations.
Full size image

The mean trends in cloud radiative effect (CRE), sea surface temperature (SST), and cloud droplet number concentration (Nd) in the North Atlantic and Northeast Pacific regions in the observations (white circle) and in the simulations with the Coupled Model Intercomparison Project Phase 6 (CMIP6) Earth System Models. Where available, the magnitude of the trends in Nd is indicated by the size of the circles. Otherwise, where Nd is not available, squares are used. The CMIP6 models are grouped into “best” (red and purple) and “other” (black) models, as described in the text. The purple square refers to results from the CanESM5 model. 95% confidence intervals are shown for the observed trends (dark gray shading), the ranges associated with internal unforced climate variability (lighter gray shading), and the model ensemble member trends (gray, red, and purple whiskers).

Despite the differences in simulated SST and CRE trends, observations and models consistently indicate that the warming was more pronounced in areas with rapidly increasing CRE (Fig. 2B). The simulated and observed relationships between CRE and SST trends exhibit surprisingly similar patterns across many models, despite significant differences in the trends. The reasons for the similarity in these patterns remain unclear. ref. 53 found SST variations in phase with solar activity across the Indian, Pacific, and Atlantic oceans. By analyzing the relationship between these variations they were able to derive an ocean climate sensitivity to changing solar irradiance at the top of the atmosphere of 0.08 K W−1 m2 on decadal and 0.14 K W−1 m2 on interdecadal time scales, the same order as the simulated changes in Fig. 2B.

To understand the differences in the CMIP6 multi-model ensemble simulations, we separated the models into two groups. Ten of the CMIP6 models simulate CRE trends that are not significantly different from the observed CRE trend (95% confidence), with a mean simulated CRE trend of 0.84 W m−2 decade−1 (range from 0.15 to 1.15 W m−2 decade−1). The mean SST trend simulated by these models (0.20 K decade−1, range from 0.03 to 0.34 K decade−1) agrees well with the observed SST trend. Given the good agreement of the simulated and observed CRE and SST trends we refer to these as the “best” models. Cloud droplet number concentrations are available for 4 of these 10 models, which we used to calculate a multi-model mean trend of -4.45 cm−3 decade−1 (range from −9.53 to -0.73 cm−3 decade−1), which also agrees well with the observations (-4.86 ± 1.32 cm−3 decade−1).

In contrast, the remaining 14 models simulate CRE trends that are significantly weaker than the observed CRE trend, with a mean simulated trend of 0.12 W m−2 decade−1 (range from -0.20 to 0.38 W m−2 decade−1). None of these “other” models simulate trends exceeding the observed trend. Furthermore, the mean temperature trend tends to be lower than for the “best” models (0.12 K decade−1, range from 0 to 0.42 K decade−1). The mean Nd trend in the three models that provide Nd data also tends to be considerably weaker than observed (-2.02 cm−3 decade−1, range from -5.13 to -0.31 cm−3 decade−1).

While the results from the “best” and “other” models appear to be consistent with our expectation that the CRE and SST should increase more rapidly in models that simulate strong reductions in Nd, we were unable to robustly determine the extent to which biases in simulated Nd and aerosol-cloud interactions contribute to the weak CRE and surface temperature trends in the CMIP6 multi-model ensemble because there are not sufficient data available for Nd.

Drivers of cloud trends

To assess the impacts of aerosols and cloud microphysical processes on CRE trends we conducted additional simulations with CanAM5.1-PAM, which combines the established climate modeling capabilities for the global atmosphere in CanAM5.154,55 with new and improved modeling capabilities for aerosol-cloud interactions. In particular, it includes an advanced numerical representation of the aerosol size distribution56 and of the formation of cloud droplets through aerosol activation57. Only minor differences exist between CanAM5.1 and the earlier model version CanAM554 in the Earth System Model CanESM5 (ref. 58, Methods).

Results from CanAM5.1-PAM for aerosols agree well with in-situ observations from field campaigns and long-term monitoring56,57,59, with a simulated aerosol radiative forcing within previously assessed ranges (Methods). By specifying the SSTs in CanAM5.1-PAM, following the CMIP6 protocol, it is possible to robustly distinguish and compare the impacts of SSTs and aerosols on the regional cloud trends.

Different configurations of CanAM5.1-PAM are available that allowed us to attribute trends in simulated CRE trends to two key processes. First, decreasing anthropogenic aerosol emissions can lead to lower cloud reflectivity via the Twomey effect. In particular, a decrease in Nd results in larger cloud droplets if the cloud liquid water content does not change, which reduces the cloud albedo30. Second, if the cloud is precipitating, the cloud liquid water content may decrease as aerosol concentrations decrease. This is because larger droplets take less time to coalesce into raindrop size, enhancing precipitation efficiency. Consequently, the lifetime of the cloud is reduced, which reduces the reflectivity further, beyond the Twomey effect. This is commonly referred to as the Albrecht effect31.

We used CanAM5.1-PAM with specified SSTs to conduct two distinct sets of simulations. One set incorporates only the Twomey effect, while the other set encompasses both the Twomey and Albrecht effects (Methods). In the CanAM5.1-PAM simulations in which aerosol-cloud interactions are governed only by the Twomey effect, the CRE increases in 73% (range 70−74%) of the combined North Atlantic and Northeast Pacific regions (Figs. S7–S10), which is less than observed. With the addition of the Albrecht effect in CanAM5.1-PAM, the CRE increases in 82% of the regions (range 78−86%, Figs. 2 and S11–S14), in good agreement with the observations (Observed cloud trends). The mean simulated CRE trend of 1.21 W m−2 decade−1 (range from 1.07 to 1.35 W m−2 decade−1) is statistically indistinguishable from the observed trend (Fig. 4B) and agrees well with the mean trend simulated by the 10 “best” CMIP6 models. Similar to the observations and results from the “best” CMIP6 models, the CRE increase is associated with a notable decrease in Nd of -4.37 cm−3 decade−1 (range from -4.54 to -4.11 cm−3 decade−1). Wide-spread reductions of 5−10% are simulated, with the most rapid losses occurring for CRE trends greater than ~2 W m−2, similar to the observations.

Fig. 4: The factors influencing cloud trends in the CanAM5.1-PAM model.
Fig. 4: The factors influencing cloud trends in the CanAM5.1-PAM model.
Full size image

The mean trends in cloud radiative effect (CRE) and cloud droplet number concentration (Nd) in the North Atlantic and Northeast Pacific regions in the observations (white circle) and in three atmospheric global climate model ensembles with specified sea surface temperature (SST): Simulations of the Twomey effect in the models CanAM5.1 (A), CanAM5.1-PAM (B), and the combined Twomey and Albrecht effects in CanAM5.1-PAM (C). For (B, C) simulations using specified SSTs from 2003 to 2022 (yellow) are compared to simulations with specified perpetual SSTs from year 2000 (blue) and with specified perpetual emissions from year 2000 (orange). 95% confidence intervals are shown for the observed trends (dark gray shading), the ranges associated with internal unforced climate variability (lighter gray shading), and the model ensemble member trends (whiskers). The magnitude of the trend in Nd is indicated by the size of the circles.

In the simulation with the Albrecht effect the loss of cloud fraction nearly doubles to  −3.2% decade−1 (range -4.2 to  − 2.0 decade−1) from  −1.7 decade−1 (range -2.9 to  −0.5 decade−1) in the simulation with the Twomey effect. The former is similar to the observed trend of  −4.2 ± 0.3 decade−1.

To determine the contribution of the SST changes to the CRE trends, we compared the CRE trends to results from simulations with perpetual monthly anthropogenic emissions from the year 2000 instead of the time-varying emissions. This produced aerosol and cloud droplet concentrations that remained nearly constant (Fig. 4). Although the CRE also increases in these simulations, the probability that the simulated CRE increase falls within the standard error of the observations (0.23 W m−2 decade−1) is only 5% (Supplementary Notes), which indicates that the SST increase is unlikely to be the sole cause of the observed CRE increase.

Based on our simulations with the different configurations of CanAM5.1-PAM we find that 0.80 W m−2 decade−1 (range 0.5 to 1.0 W m−2 decade−1), or 69% (range 55−85%) of the total simulated CRE increase can be explained by the impact of the declining emissions on the combined Twomey and Albrecht effects. Individually, the Twomey and Albrecht effects contributed 0.29W m−2 decade−1 (range 0.1 to 0.5 W m−2 decade−1) and 0.51 W m−2 decade−1 (range 0.15 to 0.85 W m−2 decade−1), respectively (Fig. 5). In comparison, the SST changes since 2000 contributed 0.38 W m−2 decade−1 (range 0.12 to 0.65 W m−2 decade−1), or 31% (range 15−45%). The latter may be overestimated, given that cloud feedbacks in this model are generally stronger than in many other ESMs60,61,62.

Fig. 5: Summary of the key processes driving the cloud trends.
Fig. 5: Summary of the key processes driving the cloud trends.
Full size image

The linear cloud radiative effect (CRE) trends and the factors that were driving them in the combined North Atlantic and Northeast Pacific regions. The gray shading refers to the 95% confidence intervals of the observed trend (dark gray) and the internal unforced climate variability (light gray). In (A), the trends simulated by the Coupled Model Intercomparison Project Phase 6 (CMIP6) multi-model ensemble are displayed on the far left. The box extends from the first to the third quartile and whiskers from the minimum to the maximum simulated 20 year mean CRE trend. The median and mean CMIP6 trends are indicated by the horizontal line and the black bullet, respectively. The CMIP6 results are compared to our simulations of the combined Twomey and Albrecht effects in the CanAM5.1-PAM model. In (B), the latter is broken down into the separate contributions associated with the Twomey and Albrecht effects and the sea surface temperature trend. See “Earth System Model simulations” and “Methods” for details.

Discussion

Our analysis of the observed cloud trends in the combined North Atlantic and Northeast Pacific regions from 2003 to 2022 is consistent with the scientifically widely accepted view that atmospheric aerosols from fossil fuel and biofuel combustion sources have historically masked some of the warming effect that results from the emissions of greenhouse gases9,33,63,64. After sulfur dioxide emissions peaked sometime in the early 2000’s65, and with that aerosol impacts on clouds and climate, the extent to which the greenhouse gas-induced warming is being masked by aerosols has lessened. Specifically, the aerosol reductions resulted in a decrease in cloud droplet number concentration, consequently decreasing cloud reflectivity. Both the Twomey and Albrecht effects contributed to this cloud reflectivity decrease. This increased the absorption of solar radiation by the ocean and caused an acceleration in the warming of the surface ocean. These changes are significant over the North Atlantic and Northeast Pacific where aerosols have been rapidly declining.

In addition to increasing the absorption of solar radiation, the decrease in cloud reflectivity likely contributed to the warming of the sea surface through cloud feedbacks, which enhanced the loss of cloud reflectivity further as these regions were warming24. However, we found that the warming-induced loss of cloud reflectivity, from the aerosol and greenhouse gas changes, was less impactful than the decrease in cloud droplet number.

Our analysis shows that reductions in emissions of sulfur dioxide and other aerosol precursors were responsible for a significant fraction of the recent radiative forcing changes in these regions. For example, CO2 concentrations at the Mauna Loa observatory increased by 22.7 ppm decade−1 from January 1, 2003 to December 31, 2022. This corresponds to an increase in the global CO2 radiative forcing at the top of the atmosphere of 0.31 W m−2 decade−1 (Methods). In comparison, over the same period, the cloud radiative effect in the combined North Atlantic and Northeast Pacific regions increased by 1.04 ± 0.45 W m−2 decade−1. This likely resulted in a disproportionate warming impact, considering the spatial inhomogeneity of the radiative forcing changes66. However, it is unclear how the recent radiative forcing and temperature changes in these regions compare to those observed in other regions and how changes in different regions are related.

Furthermore, the cloud reflectivity reduction exceeded the reduction in clear-sky atmospheric reflectivity in these regions, which is consistent with our model simulations. In comparison, the analysis by ref. 67 indicates that the reduction in global reflection of shortwave radiation to space between 2000 and 2020 was caused almost equally by reductions in cloudy and clear-sky atmospheric scattering. It would be beneficial for future research to assess aerosol impacts on clouds and clear-sky scattering in other regions, and globally.

The shared socio-economic pathway scenarios used by the Intergovernmental Panel on Climate Change and a range of clean air policy scenarios project that the aerosol concentrations and the number of cloud droplets will continue to decline over the next few decades in the combined North Atlantic and Northeast Pacific regions23,33,65. Given the fundamentally non-linear nature of aerosol-cloud interactions, even relatively minor additional reductions in cloud droplet number could decrease cloud reflectivity notably as the air becomes increasingly clean. The expected continuation of emission reductions may therefore play a significant role in affecting near-term climate trends in these regions, adding to the warming produced by the aerosol emission reductions to date. However, the extent to which climate projections beyond 2022 could be affected by aerosol emission-driven cloud reflectivity changes is currently unknown.

The recent cloud changes are an unintended consequence of efforts to improve air quality and reduce health risks68. Our analysis based on different versions of the global Canadian Atmospheric Model (CanAM) draws attention to the need to better account for aerosol-cloud interactions in the global Earth System Models, many of which have been found to underestimate the observed cloud reflectivity reductions in simulations driven by historical and scenario emissions. This limits the capacity to accurately assess near-term changes in the sea surface temperature based on projections with these models. A comprehensive understanding of recent changes in aerosols and clouds, and modeling capabilities to simulate these, is therefore essential to inform climate mitigation and adaptation decisions.

Methods

Observations

The CERES instruments on the Terra and Aqua platforms separately measure filtered radiances in the shortwave between 0.3 and 5 μm, total between 0.3 and 200 μm, and window region between 8 and 12 μm wavelengths. The filtered radiances are converted to unfiltered radiances, which in turn are converted to instantaneous radiative fluxes at the top of the atmosphere using empirical angular distribution models35. The impacts of calibration uncertainties on trends are small compared to the internal unforced climate variability in radiative fluxes. ref. 69 estimated the trend uncertainty in the CRE retrievals due to instrument drift to be <0.085 W m−2 decade−1.

Four different sampling strategies were used by ref. 70 to estimate the gridded, monthly-mean Nd from the MODIS collection 6.1 (MOD06_L2) cloud optical depth and effective radius retrievals71. For Figs. 1 and 2 we averaged the results of the different sampling strategies for the Terra platform available for 2.1 μm, after extending these data sets to December 2022. We also repeated the calculations using data from the MODIS Aqua platform, which produced similar trend patterns, with statistically significant reductions in the North Atlantic and Northeast Pacific regions that are of interest here (Figs. S15 and S16).

The regional patterns in AOD trends based on the MISR Level 3 Global Joint Aerosol monthly product V002 are consistent with various other satellites and surface-based remote sensing products even though the uncertainties in AOD retrievals are substantial28,29.

The carbon dioxide data on Mauna Loa was obtained from the NOAA Global Monitoring Laboratory72. The corresponding radiative forcing of CO2 at the top of the atmosphere was calculated using the method proposed by ref. 73.

Modeling

CanAM5.1-PAM is a spectral atmospheric general circulation model that employs 49 vertical levels that extend from the surface up to 1 hPa. The spectral representation of the horizontal fields is triangularly truncated at a wave number of 63 (T63). The simulations were performed following the Atmosphere Model Intercomparison Project (AMIP) configuration employing specified boundary conditions of sea-surface temperature and sea-ice concentration. Changes in greenhouse gas abundances, land-use, volcanic aerosols, and the solar constant are also specified52.

Aerosol chemical and microphysical processes in CanAM5.1-PAM are represented using the Piecewise-Lognormal Aerosol Model (PAM)56,57,74, which replaces the bulk aerosol scheme and parameterization of cloud droplet number used in CanAM5.155. CanAM5.1 differs from the earlier version CanAM554 in CanESM558 by using an improved representation of bulk dust aerosol concentrations. Consequently, these previous model versions differ from CanAM5.1-PAM in their representation of the Twomey effect and aerosol-radiation interactions, as distinct aerosol schemes are simulated in these models, while maintaining otherwise identical atmospheric process representations. In addition, the Albrecht effect is not included in simulations with CanAM5.1 and CanESM5. The simulated CRE trends in these models tends to be lower than observed, which aligns with the CanAM5.1-PAM simulations in which aerosol-cloud interactions are governed solely by the Twomey effect (Figs. 3, 4A).

Simulations of key climate variables and model biases are similar in CanAM5.1-PAM and CanAM5 (Figs. S18–23). Stratocumulus and cumulus regimes are well represented in the model. For instance, ref. 75 found that the simulated flux changes along the equator and Eastern Pacific during July 2000 to June 2017 in CanAM5 agree well with the observed changes, compared to other models. Simulated cloud fractions of shallow cumulus clouds in CanAM5 and CanAM4 also agree well with observations76,77. The structure of cloud biases in CanAM5 agrees better with observations than CanAM4 and the results are largely consistent with previous multi-model studies54.

In addition to the Twomey and Albrecht effects simulated in CanAM5.1-PAM, larger droplets make the cloud less susceptible to evaporation with the entrainment of dry air from above the cloud. Larger droplets sediment more rapidly, so are less likely to be exposed to dry entrained air. This tends to decrease cloud evaporation and enhance cloud water content, which dampens the Twomey and Albrecht effects78. These and other similarly complex cloud adjustments act to modulate the Twomey effect and are not usually fully accounted for in the CMIP6 ESMs, including CanAM5.1-PAM79. Currently it is unknown whether the cloud adjustments could systematically enhance or reduce the CRE trends80. Even if cloud adjustments were fully accounted for, it is important to consider in studies of aerosol-cloud interactions that there are large differences in the magnitudes of the simulated Twomey and Albrecht effects across currently available ESMs, leading to a large range in radiative forcing through aerosol-cloud interactions, even though the importance of these processes for aerosol-cloud interactions has been well established43,81,82,83.

For the simulation of the combined Twomey and Albrecht effects, the predicted Nd57 is used to calculate the optical properties and autoconversion of cloud water to rain84, which results in a global aerosol effective radiative forcing of -1.72 W m2 for the combined Twomey and Albrecht effects in 2000–2014, relative to 1850, which is within the range of previous assessments34. The simplified version of the model that only accounts for the Twomey effect simulates a global aerosol effective radiative forcing of -1.36 W m2. In this model version the predicted Nd is only used for the simulation of cloud optical properties and a specified constant Nd is used in the calculation of autoconversion, similar to other studies80,85.

Emissions of SO2, black and organic carbon from anthropogenic fossil fuel and biofuel sources in CanAM5.1-PAM are taken from the Community Emissions Data System v_2021_04_21 Release Emission Data (CEDSv202151,86), with the exception of shipping emissions. The latter are from the Finnish Meteorological Institute (FMI) inventory87, which captures the International Maritime Organization (IMO) emission regulations implemented in 2020. Biomass burning emissions for 2002−2014 were drawn from the CMIP6 historical inventory, and those for 2015−2022 were taken from the Global Fire Emissions Database (GFED v4.1s88).

The CEDSv2021 inventory in CanAM5.1-PAM contains lower aerosol emissions than the combined CMIP6 historical and SSP2-4.5 inventories in the CMIP6 models89,90, due in large part to China’s clean air policies, which are underestimated in the CMIP6 historical inventory before 201450. However, regional linear trends in SO2 emissions are similar in the combined inventories over the 20 years of the simulations (Fig. S5). The difference in the emissions causes negligible impacts on the simulated radiative fluxes in CanAM5.1-PAM. These are well within the uncertainty of the model ensemble simulations, according to additional simulations with the combined CMIP6 emissions (Fig. S25), similar to the conclusions by ref. 11.

Nd at cloud top in CanAM5.1-PAM is diagnosed using the simulated liquid cloud droplet radius near cloud top and cloud optical thickness according to ref. 91. The calculations supplement the Cloud Feedback Model Intercomparison Project Observational Simulator Package (COSP), which provides a collection of observation proxies that translate model-simulated cloud properties to synthetic observations that are directly comparable to satellite retrievals from CERES and MODIS.

The contributions of the Twomey and Albrecht effects and the SST changes in Fig. 5B were determined by differencing the CanAM5.1-PAM simulation results from the additional simulations with specified monthly emissions in year 2000 and SSTs, respectively.

To diagnose the unforced internal variability in the regionally averaged CRE and SST trends we used the CMIP6 preindustrial control simulation with CanESM5, wherein forcing agents are fixed at 1850 values and the model uses freely evolving lower boundary conditions. Similar to ref. 42, we diagnosed the ensemble of trends that can occur in a 20 year period, arising from internal atmospheric and oceanic variability in the model. This provided 38 consecutive trend values for each variable. We subsequently determined the unforced internal variability in the trends by multiplying the ensemble standard deviations by two42. We note that the results are representative of CMIP6 model results, given that the interdecadal global mean surface air temperature (GMST) variability in CanESM5 matches the CMIP6 multi-model ensemble median value of 0.07 °C92. Further, the unforced internal variability in globally averaged CRE trends in CanESM5 (±0.17 W m−2 decade−1) agrees well with the estimate by ref. 42 for the CMIP6 multi-model ensemble (± 0.18 W m−2 decade−1).

Summaries describing the CMIP6 models used here, including CanESM5, are provided in the Supplementary Table 1. Linear trends in the annual mean shortwave cloud radiative effect and sea surface temperature in these models were calculated using the monthly mean gridded data sets52, as described in the Supplemental Notes.