Fire air pollution reduces global terrestrial productivity

Fire emissions generate air pollutants ozone (O3) and aerosols that influence the land carbon cycle. Surface O3 damages vegetation photosynthesis through stomatal uptake, while aerosols influence photosynthesis by increasing diffuse radiation. Here we combine several state-of-the-art models and multiple measurement datasets to assess the net impacts of fire-induced O3 damage and the aerosol diffuse fertilization effect on gross primary productivity (GPP) for the 2002–2011 period. With all emissions except fires, O3 decreases global GPP by 4.0 ± 1.9 Pg C yr−1 while aerosols increase GPP by 1.0 ± 0.2 Pg C yr−1 with contrasting spatial impacts. Inclusion of fire pollution causes a further GPP reduction of 0.86 ± 0.74 Pg C yr−1 during 2002–2011, resulting from a reduction of 0.91 ± 0.44 Pg C yr−1 by O3 and an increase of 0.05 ± 0.30 Pg C yr−1 by aerosols. The net negative impact of fire pollution poses an increasing threat to ecosystem productivity in a warming future world.

F ire is an important disturbance to the terrestrial carbon budget. Every year, global fires directly emit 2-3 Pg (=10 15 g) carbon into the atmosphere 1 . This immediate carbon loss is partially compensated by an enhancement in net ecosystem productivity driven by changes in canopy composition and soil respiration 2 . In addition to the carbon emissions, fire plumes also generate short-lived climate pollutants, including ozone (O 3 ) and fine mode aerosols (e.g., PM 2.5 , particulate matter less than 2.5 μm in diameter), which may worsen air quality in the local and downwind regions 3 . Globally, fires contribute to 3.5% of the total tropospheric O 3 4 and 16% of total aerosol optical depth (AOD) 5 . Regionally, fires can enhance surface O 3 by 10-30 ppbv 6,7 during pollution episodes and dominate PM 2.5 with average contributions of 30-55% 8,9 during fire seasons.
While the impacts of fire O 3 and aerosols to public health have been widely recognized and studied 10,11 , the ecological impacts of this air pollution on ecosystem health and the land carbon cycle lack systematic quantification. Increases in O 3 and aerosols have strongly contrasting impacts on plant productivity 12 . O 3 is phytotoxic and reduces plant photosynthesis [13][14][15] . Aerosol pollution may promote photosynthesis by enhancing diffuse radiation 16,17 , and exert varied impacts on land carbon uptake through concomitant perturbations in meteorology [18][19][20] . Limited studies have accounted for the fire air pollution impacts on regional land carbon assimilation 21,22 , and shown that fire O 3 and aerosols result in comparable but opposite perturbations in vegetation productivity over the Amazon Forest. However, the combined effects of O 3 and aerosols from the same fires remain unquantified at both the regional and global scales.
Here, we examine the net impacts of fire O 3 and aerosols on ecosystem productivity using a suite of validated models in combination with multiple measurements (Methods). The chemical transport model GEOS-Chem 23 is used to predict fireinduced changes in aerosol optical depth and surface O 3 . The Column Radiation Model (CRM) is applied to quantify perturbations in diffuse and direct solar radiation caused by fire aerosols. The Yale Interactive terrestrial Biosphere (YIBs) model 24 , a dynamic global vegetation model, is used to quantify changes in ecosystem GPP due to O 3 inhibition and aerosol diffuse fertilization effects (DFE) originating from fire pollution. For this study, we do not consider ecosystem responses to aerosol-induced perturbations in meteorology and cloud due to the large uncertainties 19 . Joint simulations reveal an indirect terrestrial carbon loss caused by fire air pollutants, because fire O 3 strongly dampens ecosystem productivity of unburned forests and masks the benefit of increased diffuse radiation from fire aerosols.

Results
Evaluation of GPP responses to O 3 and diffuse radiation. The YIBs model calculates O 3 damages to GPP as a function of stomatal O 3 flux 25 , which is dependent on both O 3 concentrations ([O 3 ]) and leaf stomatal conductance (Methods). Compared with an observation-based meta-analysis (Supplementary Data 1), the model reasonably captures GPP sensitivity to enhanced [O 3 ] for most plant functional types (PFTs), especially for the temperate deciduous trees, C4 herbs, and crops (slopes of lines, Fig. 1). For these PFTs, GPP is found to decrease by 0. 24 GPP increases with photosynthetically active radiation (PAR) but at different rates for diffuse and direct light. Analyses at 24 FLUXNET sites ( Supplementary Fig. 1) show that GPP grows faster with light when light is more diffusive ( Supplementary  Fig. 2). The YIBs model generally reproduces GPP responses to both diffuse and direct light. Observed GPP increases by 0.3-0.7% for 1 W m −2 enhancement in diffuse PAR (Fig. 2). The modeled GPP sensitivities are close to observations for tree species but show large uncertainties for grass and crops. These sensitivities are derived assuming no changes in direct light or other meteorological variables. Appearance of aerosols (or cloud) will increase diffuse radiation but decrease direct radiation. The reduction of direct light can offset the benefit of diffuse fertilization, leading to positive GPP responses at low-medium diffuse fraction but negative responses at high diffuse fraction 16,17 .
Thresholds for cloud diffuse effects on GPP. Site-level studies reveal that plant photosynthesis reaches the maximum when diffuse fraction is between 0.4 and 0.6, no matter for cloud or aerosols 16,17,26 . Theoretically, each location on the Earth has a specific cloud amount leading to optimized GPP. It is important to estimate such threshold so as to quantify potentials for aerosol DFE. Here, we perform 20 sensitivity experiments with cloud scaling factors from 0 to 10 without aerosol radiative effects (Methods). Cloud feedbacks to climate (e.g., cooling and rainfall) are not considered due to the large uncertainties. Simulation results are compared with identify the maximum GPP and derive the corresponding cloud scaling factor (defined as cloud scaling threshold). The model predicts maximum global GPP at cloud scaling of 0.9 ( Supplementary Fig. 3), suggesting that ecosystem productivity is in general almost optimized at the current cloud amount. Regionally, cloud scaling thresholds are usually <1 for boreal and tropical forest (Fig. 3), where cloud cover is too dense (Supplementary Fig. 4) and limits plant photosynthesis. In contrast, the cloud scaling threshold is >1 over arid and semi-arid areas (Fig. 3), suggesting that light availability is not saturated with present-day cloud amount over those regions and increasing atmospheric aerosol loading has the potential to further increase plant photosynthesis.  Fig. 5a), suggesting that most fire activity occurs in wildlands or less developed regions where anthropogenic emissions are usually low. In addition to O 3 , fires contribute to regional AOD by 39.2% in southern Africa, 33.0% in tropical Amazon, and 31.8% in Indonesia on an annual mean basis ( Supplementary Fig. 5b 25 . We calculate the ozone damage as the central value of the low and high sensitivity impact results, and evaluate uncertainties as the range from low to high sensitivity. With all emissions except from fires, global GPP is reduced by 4.0 ± 1.9 Pg C yr −1 (2.9 ± 1.4%, mean ± uncertainties) due to O 3 damage losses compared with O 3 -free GPP (Fig. 4a).  (Fig. 4b).
Large  GPP changes by aerosol DFE. In contrast to the O 3 impacts, high AOD does not always result in large responses in GPP (Fig. 4c).
Moderate AOD promotes plant photosynthesis by increasing diffuse radiation, while dense aerosol loading plays the opposite role due to the strong attenuation of direct light 16 . In addition, aerosol DFE is highly sensitive to cloud amount which may have saturated the supply of diffuse radiation at the cost of total light availability 17,27 . Sensitivity tests show that aerosols are beneficial to plant photosynthesis with zero or low cloud coverage, but cause limited or even negative GPP responses when cloud fraction is high 28,29 . Our simulations with zero cloud coverage show widespread GPP enhancement of 6-30% over subtropical continents ( Supplementary Fig. 7a), especially in China where dense aerosols significantly increase regional diffuse fraction (Supplementary Fig. 8a). However, with the inclusion of cloud effects, regions with ΔGPP > 2% are limited to western Asia, North Africa, and central US (Fig. 4c), most of which are arid or semiarid regions with average cloud fraction of 39.4% ( Supplementary  Fig. 4). For these regions, the cloud scaling threshold is higher than 0.8 (Fig. 3) so that local cloud is not too abundant to inhibit aerosol DFE. For other land regions where the cloud scaling threshold is lower than 0.8 (Fig. 3), the dense cloud masks aerosol DFE and leaves very limited ΔGPP (Fig. 4c). The average GPP is 1.2 g C m −2 day −1 over the arid/semi-arid regions with ΔGPP > 2%, much lower than the average of 3.5 g C m −2 day −1 in the wet regions with ΔGPP < 2%, leading to a limited potential for aerosol DFE on the continental scale. Globally, aerosols from non-fire c d e f Fire aerosols alone cause limited impacts on plant photosynthesis except over central and southern Africa, central South America, and eastern Siberia (|ΔGPP| > 0.2%, Fig. 4d), where fire aerosols enhance diffuse fraction by 1.4%, 0.7%, and 0.5% ( Supplementary Fig. 8d). We find contrasting responses of GPP to fire aerosols in Africa, with negative changes along the western coast while positive responses occur in inner areas. Such discrepancy is related to the seasonality of biomass burning and cloud fraction. African fire pollution is confined between 0-15 o S in boreal summer (June-August) and shifts northward to 0-8 o N in winter (December-February). Meanwhile, dense cloud overlaps the fire domain (especially over the western coast) in winter but shifts northward to an offset location with fires in the summer ( Supplementary Fig. 4). On average, the summertime cloud fraction is 72.0% over the coastal regions with negative aerosol effects, but is much lower at 48.5% over the inner areas where ΔGPP > 0.2%. The dense clouds limit DFE of fire aerosols along the coast due to the attenuation of total available sunlight. The largest GPP percentage change of 1% is found in eastern Siberia, where diffuse fraction is increased only 0.5% by fire aerosols (Supplementary Fig. 8d). Sensitivity tests without cloud effects also show large responses in GPP at high latitudes (Supplementary Fig. 7b), where the changes in diffuse fraction are smaller than those in tropical regions ( Supplementary Fig. 8b). Plants at high latitudes have much larger shaded portion than that at low latitudes because of the smaller solar zenith in high latitudes (Methods). As a result, the boreal forest is more sensitive than tropical trees to the same fractional change in diffuse radiation. On the global scale, fire aerosols enhance GPP by 0.05 Pg C yr −1 , lower than the value of 0.7 Pg C yr −1 under clear-sky conditions ( Supplementary Fig. 7b).
Net effects of O 3 and aerosols to GPP. Combined aerosol and O 3 impacts on GPP show offset spatial patterns globally (Fig. 4e). Aerosol radiation fertilization effects are more dominant over arid and semi-arid areas, where human activity is usually limited. In contrast, O 3 damage effects are prominent over regions with high population density, where both anthropogenic emissions and vegetation coverage are high. The net impact of fire pollution on GPP is dominated by the O 3 effect (Fig. 4f), though both O 3 and aerosols are generated by the same fires. On the global scale, we quantify that fire emissions indirectly reduce GPP by 0.86 Pg C yr −1 (0.6%) during 2002-2011, resulting from a reduction of 0.91 Pg C yr −1 by O 3 and an increase of 0.05 Pg C yr −1 by aerosols (Fig. 5a). The fertilization effect of fire aerosols is very limited, likely because fire emissions predominantly occur in tropical forests where dense cloud masks the aerosol effects. The overlap of tropical forests and cloud typifies the intimacy of biosphere-atmosphere coupling: the vegetation grows in regions with abundant precipitation while strong evapotranspiration generates more cloud and consequent rainfall 30 .
Global fire emissions show interannual variations even with regional offsets 1 . The largest increase of fire O 3 of 2.1 ppbv occurs in the year 2003, leading to the highest GPP reduction of 0.99 Pg C yr −1 globally (Fig. 5c). In contrast, the smallest GPP reduction of 0.85 Pg C yr −1 is found in the year 2011, when the average fireinduced [O 3 ] is 1.9 ppbv globally. On average, the interannual variation (0.05 Pg C yr −1 , one standard deviation for 2002-2011) accounts for 5% of the mean ΔGPP by fire pollution, much smaller than the uncertainties of 43% driven by O 3 damage sensitivity (0.37 Pg C yr −1 , half of the range from low to high sensitivities). Regionally, fire perturbations cause large year-to-year variability (Fig. 5d). For example, ΔGPP ranges from −0.9%

Discussion
Compared with simulations by Pacifico et al. 21 , which estimated that fire O 3 reduces GPP by 230 Tg C yr −1 in Amazon forest, this study predicts a lower O 3 damage of 137 Tg C yr −1 over the same region (Fig. 4b). Our predicted GPP reduction is smaller, likely because Pacifico et al. 21 (Fig. 4d). The GEOS-Chem simulation with satellite-based emission inventories may underestimate fire AOD compared with MODIS 32 . However, sensitivity simulations with doubled or tripled fire aerosols does not increase Amazon GPP substantially ( Supplementary Fig. 9). The average cloud amount in this region is 69% on an annual mean basis and remains as high as 50% during the dry season ( Supplementary Fig. 4). A previous study reported decreases in plant photosynthesis when cloud cover is higher than 48% 17 . The dense cloud over Amazonia implies a low potential for increasing diffuse radiation, and as a result limits DFE of aerosols from fires. On the other hand, we predict relatively high responses (0.2-0.4%) in GPP south of Amazonia ( Supplementary Fig. 9), where cloud fraction is 44% during the dry season.
Derived fire pollution effects on ecosystem productivity are influenced by uncertainties in O 3 damage sensitivity, cloud amount, and aerosol properties. For O 3 damage, applying low to high sensitivity parameters results in varied ΔGPP from −1.29 to −0.55 Pg C yr −1 (Fig. 5b). Predicted GPP reduction by fire O 3 is high over tropical Africa (Fig. 4b), where the dominant PFT is evergreen broadleaf forest (EBF). The O 3 damage scheme proposed by Sitch et al. 25 was developed based on temperate plants. Compared with limited measurements, the derived GPP sensitivity is larger by 37.5% for EBF (Fig. 1). A recent meta-analysis 33 shows that the plant O 3 sensitivity is closely related to leaf mass per area (LMA), suggesting that tropical trees may have higher resistance to ozone damage due to their larger LMA (average 107-121 g m −2 ) 34 than temperate trees (usually < 100 g m −2 ) 33 . As a check, we reduce the O 3 sensitivity coefficients of EBF in the YIBs model by 33% while holding unchanged the coefficients for other PFTs (Methods). The new simulations show that global GPP is reduced 3.7 ± 1.8 Pg C yr −1 by fire-free O 3 and 0.81 ± 0.40 Pg C yr −1 by fire O 3 . These values are slightly smaller than the results of 4.0 ± 1.9 Pg C yr −1 and 0.92 ± 0.44 Pg C yr −1 calculated with original sensitivity coefficients, but do still mean that fire O 3 damage dominates over fire aerosol effects at the global scale.
For aerosol effects, simulated aerosol DFE is sensitive to cloud properties. In the current runs, daily cloud fraction and liquid water path is applied though cloud property may have distinct spatiotemporal variations. As a check, we use 3-hour cloud variables from CERES to update fire aerosol DFE (Methods). Simulated ΔGPP by fire aerosols with 3-hour cloud is 0.01 Pg C yr −1 , slightly lower than the value of 0.05 Pg C yr −1 obtained using daily cloud (Fig. 5b). Current simulations employ monthly fire emissions, which may lead to inconsistency between fire events and cloud status. Fires tend to occur in dry periods when the cloud amount is relatively low, and fire aerosols can inhibit cloud and precipitation through fire-meteorology interactions 35 . To place bounds on the DFE uncertainty due to clouds, we perform two sensitivity experiments that alter the cloud fraction and liquid water path. First, we reduce the observed cloud fraction and liquid water path by 50% (Methods). In this case, fire aerosols cause a stronger fertilization effect of 0.40 Pg C yr −1 to global GPP. At the same time, fire aerosols themselves can increase cloud condensation nuclei and consequently increase cloud depth and amount 36 . Therefore, a second sensitivity experiment is performed with doubled observed cloud fraction and liquid water path. In this case, fire aerosols result in a negative GPP change of −0.13 Pg C yr −1 (Fig. 5b).
In addition to cloud properties, uncertainties in aerosol radiative effects can affect GPP responses. Since monthly fire pollution may underestimate high density of particles during episodes, we perform two simulations with doubled and tripled fire aerosols (Methods). Predicted ΔGPP in these two simulations are respectively, 0.06 Pg C yr −1 and 0.04 Pg C yr −1 , very similar to the value of 0.05 Pg C yr −1 with the original fire emissions. Fire aerosols are dominated by organic (OC) and black carbon (BC), which have distinct optical properties. Here, two sensitivity experiments are performed that alter the BC loadings and BC/OC ratios. Simulation BC05 assumes that the fraction of BC in the fire pollution is reduced by 50% while OC increases by 5%, because the emission factor of BC is~1/10 of that for OC 37 in fires. Simulation BC20 assumes the loading of BC is doubled while OC decreases by 10%. Predicted diffuse fraction is similar between BC05 and BC20, but the total PAR is smaller in BC20 because BC is more absorptive than OC. Consequently, ΔGPP by fire aerosols in BC05 is 0.10 Pg C yr −1 , higher than the value of −0.06 Pg C yr −1 in BC20. In summary, uncertainties in composition and optics of fire aerosols cause varied GPP responses from −0.06 to 0.10 Pg C yr −1 , generally smaller than those of −0.13 to 0.40 Pg C yr −1 due to changed cloud properties (Fig. 5b). Finally, aerosols alter meteorology (e.g., temperature and precipitation) and cloud (both semi-direct and indirect effects), which in turn affect plant photosynthesis and carbon assimilation 19,20,38 . We do not include these feedbacks due to the large uncertainties. Furthermore, deposition of fire aerosols may cause physiological (e.g., altering stomatal functions 39 ) and biogeochemical (e.g., nitrogen deposition effects 40 ) responses but these potential impacts are ignored in the current study.
This study reveals an indirect carbon loss due to the combined effects of O 3 stomatal uptake and aerosol diffuse fertilization by fire emissions. The damage to ecosystem productivity not only occurs in the fire regimes, but also over the downwind areas through long-range transport of air pollution. The net impact is estimated to be a 0.6% reduction per year in global GPP by fire pollution. In comparison, drought is estimated to reduce net primary production by 0. 55   Uncertainties for ozone effects mainly originate from damaging sensitivity and relatively lower responses of evergreen broadleaf forest, while uncertainties for aerosol effects are mainly from cloud amount and perturbations in fire aerosols (e.g., content of black carbon, or amount of total fire particles, see Supplementary Table 8). c Global GPP changes due to fire ozone (blue) and aerosols (red) show moderate interannual variations. Blue shadings represent uncertainties due to ozone damaging sensitivity. Red shadings indicate uncertainties of aerosol effects caused by cloud scalings from 0.5 to 2.0. d Fire pollution causes distinct responses to regional GPP (see Supplementary  Fig. 5 for region names), with the maximum perturbations in southern Africa (SAF), the minimum in boreal North America (BNA), and the largest interannual variation in Indonesia (IND). Error bars indicate ranges of GPP changes from low to high years. The years leading to maximum regional GPP changes are denoted fire pollution-induced productivity inhibition (0.6% per year) is considerably stronger than the drought effect (0.1% per year) for the same period. Regionally, fire pollution causes GPP losses of up to 3.6% during some large fire years. Over tropical Africa, fire pollution induces stable GPP reductions of 1.5-2% every year. Such perturbations may result in the loss of land carbon storage and exacerbate the global warming trend due to increasing the atmospheric CO 2 burden. Although some studies 42 suggest that increased temperature is beneficial to the biosphere by enhancing GPP at the rate of 0.8-0.9 Pg C yr −1°C−1 , the associated increase in fire activity 43,44 as well as background O 3 45 may weaken the global and regional GPP response to a warming world.

Methods
Models. We apply a suite of models in offline mode for this study ( Supplementary  Fig. 10) Wavelength-and species-dependent optical parameters, including extinction coefficient, single-scattering albedo, and asymmetry parameter, are adopted from multiple sources with reasonable calibrations 28 . The CRM is driven with hourly 3-D meteorology (e.g., temperature and humidity) from MERRA reanalyses and daily cloud profile (e.g., cloud fraction and liquid water path) from CERES SYN1deg assimilation data product (http://ceres.larc.nasa.gov), which is developed based on satellite retrievals from the Moderate Resolution Imaging Spectroradiometer (MODIS) and the Visible and Infrared Sounder (VIRS).
YIBs is a process-based vegetation model that simulates the global carbon cycle with dynamical prediction of leaf area index (LAI) and tree height 24 . The model considers eight PFTs, including evergreen needleleaf forest, deciduous broadleaf forest, EBF, shrubland, tundra, C3/C4 grass, and C3 crops. The satellite-based land types and cover fraction 49 are aggregated into these eight PFTs and used as input. The model is driven with hourly 2-D meteorology (e.g., surface air temperature, specific humidity) and 3-D soil variables (e.g., soil temperature and moisture) from WFDEI 50 (WATCH Forcing Data methodology applied to ERA-Interim data). Leaf-level photosynthesis is simulated with the well-established Farquhar et al. 51 scheme and is coupled to stomatal conductance with the Ball-Berry scheme 52 . The canopy is divided into multiple layers with separation of diffuse and direct light components on sunlit and shaded leaves 53 . Light intensity decreases exponentially with LAI and becomes more diffusive when penetrating into the deep canopy. At each layer, plant photosynthesis is calculated as: where A sl and A sh are the photosynthetic rates of sunlit and shaded leaves, respectively. The sunlit leaves receive both direct and diffuse light while the shaded ones receive only diffuse radiation. The areal fraction of sunlit leaf f sl is calculated as: where L is the LAI at a certain canopy layer. The extinction coefficient σ = 0.5/cosα is dependent on solar zenith α, and will be larger at higher latitudes, leading to smaller (larger) fraction of sunlit (shading) leaves. The YIBs model calculates O 3 damage to photosynthesis following the Sitch et al. 25 flux-based scheme, which assumes GPP reduction rate F as a function of stomatal O 3 flux F O3 : Here, a is the damaging sensitivity coefficient and T O3 is the flux threshold. The coefficient a varies among PFTs and shows low to high sensitivities for species within the same PFT 25 . The coefficient a can be calibrated using observed GPP-O 3 responses (Fig. 1)  Simulations and validation. We perform two GEOS-Chem runs, three CRM runs, and eleven YIBs runs to isolate ozone and aerosol impacts on GPP from fire and non-fire sources (Supplementary Table 1 Fig. 9). The YIBs model simulates reasonable GPP spatial pattern compared with a benchmark product upscaled from FLUXNET site data ( Supplementary Fig. 11). The simulation matches benchmark well with high correlation coefficient of 0.9 (p < 0.01) and low relative bias of 3.6%, except that simulated GPP is higher by 37.6% in Amazon and 44.1% in tropical Africa. However, site-level simulations at eight EBF sites from FLUXNET network ( Supplementary Fig. 12) show that the model actually overestimates GPP by 9-22% in Amazon (BR-Sa3 and GF-Guy) and 19% in Africa (GH-Ank). The YIBs also predicts reasonable GPP responses to O 3 and diffuse radiation. To test ozone damaging scheme, we perform 17 sensitivity tests (Supplementary Table 2 Table 3), because a recent meta-analysis shows that tropical trees may have lower sensitivity to ozone damage due to their larger LMA 33 . For diffuse effect, we perform simulations at FLUXNET sites providing long-term (>3 years) hourly diffuse radiation. In total, 24 sites are selected with wide geographic distribution ( Supplementary Fig. 1). For these sites, we separate hours dominated by diffuse (diffuse fraction > 0.8) and direct (diffuse fraction < 0.2) radiation, and then aggregate both observed and simulated GPP into discrete bins based on PAR intervals. In this way, we isolate GPP responses to diffuse/direct radiation by smoothing influences of other meteorological forcings ( Supplementary Fig. 2).
Sensitivity experiments are performed to estimate GPP uncertainties due to cloud and aerosol properties. To derive cloud thresholds for optimized GPP, we perform 20 sets of simulations with cloud scaling factors ranging from 0 to 10 (Supplementary Tables 4, 5). In this way, 3-D cloud profile is retained and cloud fraction in multiple layers can be 100%. Predicted total PAR in these runs always decrease with increased scaling factors (Supplementary Table 4). However, diffuse fraction increases at beginning but decreases with large cloud amount, because diffuse light reduces after the scaling of 1.5 ( Supplementary Fig. 3). To estimate the uncertainties of aerosol DFE, another 12 sets of runs are performed (Supplementary Tables 6, 7) by assuming doubled cloud amount (C20), half cloud amount (C05), 3-hour cloud variables (C3H), no cloud (C00), doubled fire BC (BC20), half fire BC (BC05), and doubled (FA2) and tripled (FA3) total fire aerosols. Results from these sensitivity experiments are shown in the main text with instructions listed in Supplementary Table 8.
Simulated AOD based on aerosols from GEOS-Chem model ( Supplementary  Fig. 13a) reproduces the observed (Supplementary Fig. 13c) spatial distribution retrieved by satellites. The model overestimates AOD by 25.1% in eastern China, but underestimates AOD by 34.7% in Amazonia, 44.6% in India and 58.0% in western US. The satellite data show no values over North Africa, where the model predicts high AOD due to dust, because the bright surface results in the failure of retrieval. On the global scale, simulated AOD correlates with observations at R = 0.63 (p < 0.01) with a relative bias of 5.1% (Supplementary Fig. 13e). Observations of surface [O 3 ] are very limited. Here, we combine long-term monitoring data from US and Europe 54 with records at 1580 sites from Chinese Ministry of Ecology and Environment (http://english.mep.gov.cn/, Supplementary Fig. 13d). The comparison shows that GEOS-Chem simulation ( Supplementary Fig. 13b) generally overestimates surface O 3 , leading to a positive bias of~20% on the global scale ( Supplementary Fig. 13f). With simulated aerosols from GEOS-Chem, the CRM reproduces surface solar insolation with biases of only 1.8% for all-sky and 2.6% for clear-sky conditions ( Supplementary Fig. 14). Simulated clear-sky aerosol direct radiative effect (DRE) shows reasonable seasonality compared with other models summarized in Yu et al. 55 (Supplementary Fig. 15). On the annual mean basis, the model predicts aerosol DRE of −6.3 W m −2 at surface and −3.1 W m −2 at top of atmosphere, close to the averages of −7.6 and −3.0 W m −2 from an ensemble of models.