The response of the Amazon ecosystem to the photosynthetically active radiation fields: integrating impacts of biomass burning aerosol and clouds in the NASA GEOS Earth system model

The Amazon experiences fires every year, and the resulting biomass burning aerosols, together with cloud particles, influence the penetration of sunlight through the atmosphere, increasing the ratio of diffuse to direct photosynthetically active radiation (PAR) reaching the vegetation canopy and thereby potentially increasing ecosystem productivity. In this study, we use the NASA Goddard Earth Observing System (GEOS) model with coupled aerosol, cloud, radiation, and ecosystem modules to investigate the impact of Amazon biomass burning aerosols on ecosystem productivity, as well as the role of the Amazon’s clouds in tempering this impact. The study focuses on a 7-year period (2010–2016) during which the Amazon experienced a variety of dynamic environments (e.g., La Niña, normal years, and El Niño). The direct radiative impact of biomass burning aerosols on ecosystem productivity – called here the aerosol diffuse radiation fertilization effect – is found to increase Amazonian gross primary production (GPP) by 2.6 % via a 3.8 % increase in diffuse PAR (DFPAR) despite a 5.4 % decrease in direct PAR (DRPAR) on multiyear average during burning seasons. On a monthly basis, this increase in GPP can be as large as 9.9 % (occurring in August 2010). Consequently, the net primary production (NPP) in the Amazon is increased by 1.5 %, or ∼ 92 TgCyr−1 – equivalent to ∼ 37 % of the average carbon lost due to Amazon fires over the 7 years considered. Clouds, however, strongly regulate the effectiveness of the aerosol diffuse radiation fertilization effect. The efficiency of this fertilization effect is the highest in cloud-free conditions and linearly decreases with increasing cloud amount until the cloud fraction reaches ∼ 0.8, at which point the aerosolinfluenced light changes from being a stimulator to an inhibitor of plant growth. Nevertheless, interannual changes in the overall strength of the aerosol diffuse radiation fertilization effect are primarily controlled by the large interannual changes in biomass burning aerosols rather than by changes in cloudiness during the studied period.

The Amazon experiences fires every year, and the resulting biomass burning aerosols, together 23 with cloud particles, influence the penetration of sunlight through the atmosphere, increasing the 24 ratio of diffuse to direct photosynthetically active radiation (PAR) reaching the vegetation 25 canopy and thereby potentially increasing ecosystem productivity. In this study, we use the 26 NASA Goddard Earth Observing System (GEOS) model with coupled aerosol, cloud, radiation, 27 and ecosystem modules to investigate the impact of Amazon biomass burning aerosols on 28 ecosystem productivity, as well as the role of the Amazon's clouds in tempering this impact. The 29 study focuses on a seven-year period (2010-2016) during which the Amazon experienced a 30 variety of dynamic environments (e.g., La Niña, normal years, and El Niño). The direct radiative 31 impact of biomass burning aerosols on ecosystem productivity-called here the aerosol diffuse 32 radiation fertilization effect -is found to increase Amazonian Gross Primary Production (GPP) 33 by 2.6% via a 3.8% increase in diffuse PAR (DFPAR) despite a 5.4% decrease in direct PAR 34 (DRPAR) on multiyear average during burning seasons. On a monthly basis, this increase in 35 GPP can be as large as 9.9% (occurring in August 2010). Consequently, the net primary 36 production (NPP) in Amazon is increased by 1.5%, or ~92 TgCyr -1 -equivalent to ~37% of the 37 average carbon lost due to Amazon fires over the seven years considered. Clouds, however, 38 strongly regulate the effectiveness of the aerosol diffuse radiation fertilization effect. The 39 efficiency of this fertilization effect is the highest in cloud-free conditions and linearly decreases 40 with increasing cloud amount until the cloud fraction reaches ~0.8, at which point the aerosol-41 influenced light changes from being a stimulator to an inhibitor of plant growth. Nevertheless, 42 interannual changes in the overall strength of the aerosol diffuse radiation fertilization effect are 43 primarily controlled by the large interannual changes in biomass burning aerosols rather than by 44 changes in cloudiness during the studied period. 45 46 1. Introduction 48 The Amazon is home to more than 34 million people and hosts a large variety of plants and 49 animals. The rainforest plays a vital role in the global climate, regulating temperatures and 50 storing vast quantities of carbon (Laurance 1999;Nepstad et al., 2008). It is matter of intense 51 research whether light or water is the limiting factor that controls plant growth over Amazonia. 52 Considerable evidence demonstrates that sunlight indeed drives Amazon forest growth (Doughty 53 et November) (Myneni et al., 2007). Vegetation index maps also show that a majority of Amazonia 58 is greener in the dry season than in the wet season (~mid-December -mid-May) (Huete et al., 59 2006). It is in the dry season, when more light reaches the canopy level, that the Amazon forest 60 thrives. 61 62 Plant photosynthesis requires sunlight to reach the leaves of the canopy. While aerosols and 63 clouds in the atmosphere decrease the total amount of light that reaches the canopy, they also 64 increase scattering, thereby increasing the ratio of diffuse radiation to direct radiation. This is 65 important The situation is more profound during the Amazon dry season when intensive seasonal fires 83 release large amounts of primary aerosol particles as well as gas precursors that form secondary 84 organic and inorganic aerosols. Using stand-alone radiation and vegetation models, Rap et al. 85 (2015) concluded that fires over the Amazon dry season increase Amazon net primary 86 production (NPP) by 1.4-2.8% by increasing diffuse radiation. This enhancement of Amazon 87 basin NPP (78-156 Tg C a -1 ) is equivalent to 33-65% of the annual regional carbon emissions 88 from biomass burning and accounts for 8-16% of the observed carbon sink across mature 89 Amazonian forests. Moreira  transpiration depend non-linearly on solar radiation. The canopy is assumed to consist of sunlit 237 leaves and shaded leaves, and the DRPAR and DFPAR absorbed by the vegetation is 238 apportioned to the sunlit and shaded leaves as described by Thornton and Zimmermann (2007  Robotic Network (AERONET) sun photometer network (http://aeronet.gsfc.nasa.gov). We also 276 use MODIS collection 6.1 level-3 AOD product 277 (http://modis.gsfc.nasa.gov/data/dataprod/index.php), which is characterized by observations 278 with large spatial coverage. 279 280 MODIS cloud products (https://modis-atmosphere.gsfc.nasa.gov/data/dataprod/), specifically 281 total cloud fraction and cloud optical depth in liquid and ice particles, are used to evaluate the 282 model cloud simulation. We use the cloud data from MODIS collection 6.1 MYD08_D3, a level-283 3 1°×1° global gridded monthly joint product derived from the MODIS level-2 pixel level 284 products. MODIS level 2 cloud fraction is produced by the infrared retrieval methods during 285 both day and night at a 5×5 1-km-pixel resolution. Level 2 cloud optical thickness used in this 286 study is derived using the MODIS visible and near-infrared channel radiances from the Aqua 287 platform. 288 289 The

Experiment setup 304
All experiments were run with the coupled atmosphere and land components of the NASA 305 GEOS ESM system discussed above. The sea surface temperature (SST) for the atmospheric 306 dynamic circulation is provided by the GEOS Atmospheric Data Assimilation System (ADAS) 307 that incorporates satellite and in situ SST observations and assimilates Advanced Very High 308 Resolution Radiometer (AVHRR) brightness temperatures. The experiments were run in replay 309 mode, which means that the model dynamical variables (winds, pressure, temperature, and 310 humidity) were set, every 6 hours, to the values archived by the Modern-Era Retrospective 311 Analysis for Research and Applications version 2 (MERRA-2) meteorological reanalysis (Gelaro 312 et al. 2017); a 6-hourly forecast provided the dynamical and physical fields between the 6-hour 313 resets. In effect, the replay approach forces the atmospheric "weather" simulated in the model to 314 agree with the reanalysis. This nudging of the GEOS dynamic fields toward the MERRA2 315 reanalysis ensures that the atmospheric conditions of our four simulations (see below) remain 316 close to each other, allowing a more focused study of radiative impact on ecosystem. All  Index (ONI,  319 https://origin.cpc.ncep.noaa.gov/) ( Figure S1). Information regarding long-term BB OA 320 emissions (i.e., 1997-2016) and long-term MERRA2 cloud fraction anomalies (i.e., 1995-2018) 321 is shown in Figure S2. The selected period of 2010-2016 represents well the long-term period in 322 terms of the variation of BB emissions and cloud coverage. 323 324 Our experimental design makes extensive use of GEOS's highly flexible configuration. First, the 325 GEOS GOCART module includes a tagged aerosol mechanism. Each specific aerosol 326 component in GOCART is simulated independently from the others, and the contribution of each 327 emission type to the total aerosol mass is also not interfered by that of other emission types. 328 Thus, additional aerosol tracers can easily be "tagged" according to emission source types. This 329 makes it possible for GOCART to calculate and transfer two sets of aerosol fields (e.g., one with 330 and one without a biomass burning source) to the radiation module.
Check atmospheric BB aerosol impact on plants via radiation fields during 2010-2016 Pair 2 callaer Standard all, w/ AERbb emission fixed at 2010 The NASA GEOS ESM model, including its aerosol, cloud, radiation, and ecosystem modules as 378 used in the baseline simulation (i.e., experiment allaer), has been evaluated extensively and 379 utilized in a number of scientific studies. However, very few of the past studies with GEOS was 380 concentrated on detailed model evaluation over South America. We provide such an evaluation 381 here. 382 383 The simulated tracer fields are compared with measurements over the Amazon in Figures 1 and  384 2. Figure 1 shows results for surface OA concentration, surface CO concentration, and the OA 385 concentration vertical profile. We focus primarily on the OA evaluation since it is the major 386 component of biomass burning aerosols. Figure 1a shows the comparison of surface daily OA 387 concentration between the model simulation and the GoAmazon measurements at Manaus, 388 Brazil, in 2014 (The location is indicated in Figure 2c with an open-diamond). The simulated OA 389 broadly captures the seasonal trend in OA concentrations measured at Manaus, but it is lower 390 than observed OA values by ~24% during Sept-Oct and ~ 30% annually. For the period of 391 interest, the model simulates a large fire signal in August that is not seen in the measurements. 392 However, this strong August biomass burning signal does show up in the CO measurements 393 (Figure 1b), which should also be from biomass burning. The reasons for such discrepancy from 394 observations are not clear. 395 396

398
When compared with aircraft G-1 measurements over a ~2°×2° region around the center of 399 Manaus during the biomass burning season (Sept. 6 -Oct. 4, 2014) (Figure 1c), the simulated 400 vertical OA concentrations underestimate the measurements above 1 km altitude but 401 overestimate them under it, although they overlap within their standard deviations for all 402 altitudes. Here the model data have been sampled spatially and temporally along the G-1 flight 403 paths. This surface OA overestimation by the model seems to contradict the model's 404 underestimation seen in Figure 1a, indicating that capturing aerosols at the right times and 405 locations is a challenge. 406 407 Figure 2 shows the AOD (550nm) and SSA (440nm) comparison at the AERONET station of 408 Alta-Floresta, which is located close to the area of the most intensive Amazon fires (location is 409 marked in Figure 2c as a filled-in circle). The model-simulated, AERONET-measured, and 410 MODIS-retrieved AOD at this site agree within 20% (Figure 2a), all showing a peak of AOD 411 during the biomass burning season. SSA during the burning season generally ranges between 412 0.85 -0.95 (Figure 2b). The model agrees with the measurements with accurate better than 5% 413 except during the first half of August, when the model aerosols are too scattering. However, it is 414 puzzling to observe the extremely low measured SSA in the beginning of August given that the 415 AOD is still low then, as shown in Figure 2a. It could be the quality of AERONET SSA is not 416 "reliable" at low AOD (Chin et al., 2009). Because of the low sensitivity to the absorption when 417 aerosol loading is low, SSA is retrieved with sufficiently high accuracy only when the 418 total AOD at 440 nm is equal or higher than 0.4 and solar zenith angle is 50 degree or higher 419 (Dubovik et al., 2000(Dubovik et al., , 2002. Regionally over the Amazon region, defined throughout the study 420 as the land area within 80°W-30°W, 25°S-5°N (shaded land area in Figure 2d), the model-421 simulated AOD (0.22 in Figure 2d) during the biomass burning season generally agrees with 422 MODIS satellite retrievals (0.21 in Figure 2c). A simulated high bias is seen over the east 423 Amazon; however, though this region is in our area of interest, the bias should have only a minor 424 impact on our study given that the area is relatively bare, with little vegetation coverage .  425  426  427  428  429  430  431  432  433  434  435  436  437  438  439  440  441  442  443  444  445  446  447  448  449  450  451  452  453  454  455  456  457  458  459  460 The accurate simulation of cloud fields is also important for our study. In Figure 3   products. Here the GEOS data have been sampled with MODIS overpass time and location. 463 GEOS generally captures the magnitude and main features of the cloud fields observed in 464 MODIS, though with some differences; the model overestimates the cloud quantities over the 465 central Amazon and underestimates them in northwest South America. The overall difference 466 over the Amazon region between simulated and MODIS-based estimates is less than 7% for 467 cloud cover fraction, 10% for liquid water cloud optical depth, and 15% for ice cloud optical 468 depth. The seasonality of these cloud quantities is shown in Figure S5a     based products and the GEOS simulation. The overall spatial distributions of GEOS GPP ( Figure  496 7c) over South America show similar spatial pattern to both of the observations-based datasets 497 (Figures 7a and 7b) with higher values over the eastern part of the domain but lying between the 498 two datasets in other areas. Over the studied period and the Amazon region, the GEOS GPP is 499 comparable to the FluxSat GPP and is about 35% higher than the FluxCom GPP. 500 The seasonality of GPP over the Amazon region from FluxCOM, FluxSat and GEOS during 501 2010-2015 is shown in Figure S7, and the corresponding time series of monthly means is shown 502 in Figure S8. During all four seasons, regional FluxCom GPP is the lowest and FluxSat GPP is 503 the highest. All datasets show higher GPP during Nov-Apr than during May-Oct. GEOS 504 multiyear annual average GPP is close to that of FluxSat but is higher than that of FluxCom. 505 Although

511
Although the evaluations of global and regional multiyear average GPP conducted above 512 (Figures 5-7) are needed for the examination of the model's fundamental mechanisms including 513 photosynthesis, a more direct evaluation to address the model's accuracy in simulating observed 514 GPP response to changes in diffuse and direct surface radiation is shown in Figure 8. of BBAOD even if the incident diffuse radiation decreases below its peak value, though for some 547 value of BBAOD, the reduction in total radiation will be large enough to overwhelm the impact 548 of increased diffuse radiation, and plant photosynthesis will be lower than that for clean sky 549 conditions. 550 551 Figure 9. The ratio of Rdir@srf to Rtot@toa (blue), which presents the clearness index for the direct radiation portion (CIdir), the ratio of Rdiff@srf to Rtot@toa (red) for the diffuse radiation portion (CIdiff), and the ratio of Rtot@srf to Rtot@toa (green). Here, Rtot@toa is incoming total solar flux at the top of atmosphere (TOA), Rdir@srf is surface downward direct solar flux, Rdiff@srf is surface downward diffuse solar flux, and Rtot@srf is sum of Rdir@srf and Rdif@srf. All Rs are over 400-700 nm. 9a) the change of the radiative flux ratios in BBAOD = 0-3 under clear sky condition. 9b) same as left panel but under cloudy conditions (cloud fraction =1) with COD=1. 9c) same as middle panel but for COD=10. Calculations use fast-JX radiation model column version adopting a standard atmospheric condition of typical tropics at ozone column = 260 Dobson Units, SZA = 15°, and surface albedo = 0.1.

552
The Amazon dry season is characterized by high biomass burning aerosol loading combined with 553 low cloud cover, a good match to obtain more diffuse radiation without the loss of too much total 554 radiation. However, as we have pointed out, cloud impacts on radiation typically dominate those 555 of aerosols. To examine this, we repeated the radiation model calculations after adding, at the top 556 of the aerosol layer (~3.5km), a cloud layer with a cloud fraction of 1.0 and a cloud optical depth 557 (COD) of 1 ( Figure 9b) and 10 ( Figure 9c). The latter COD is close to the mean liquid cloud 558 COD over the Amazon dry season (Figure 3). The impact on Rdir@srf and Rdiff@srf is quite 559 large even with a very thin overhead cloud (Figure 9b). Without BBaer, the clouds already 560 produce abundant diffuse light that can reach the surface (i.e., CIdiff > 50%, as seen in both 561 Figure 9b-c), while almost shutting down the direct light (i.e., CIdir < 1% in Figure 9c). 562 Accordingly, for full cloud coverage, a clean sky (i.e., no aerosols) would provide the best 563 conditions for plant growth. When fires start, the diffuse light declines rapidly, reducing the 564 potential for plant growth. At BBAOD ~ 3 the ratios among Figure 9a-c look similar, that is, 565 essentially very little radiation reaches the surface. 566 The simple examples in Figure 9 illustrate the complicated responses of direct and diffuse light 567 to the presence of aerosol and cloud. Measurements indicate that plant growth peaks for a 568 clearness index (CI, defined as CIdir+CIdiff) of about 0.4-0.7 for some forest ecosystems (Butt 569 et al., 2010, Letts and Lafleur, 2005). This CI range translates, based on Figure 9, to a BBAOD 570 range of about 0.3~1.5 in clear sky and 0~0.5 in cloudy-sky conditions. 571 572 3.3 How the ecosystem responds to the BBaer diffuse radiation fertilization effect 573 574 Figure 10. GEOS simulated daily values of total cloud fraction (CLDFRC, %), biomass burning AOD (BBAOD), direct PAR (DRPAR, Wm -2 ), diffuse PAR (DFPAR, Wm -2 ), and gross primary growth (GPP, µg/m 2 /s) from the two experiments of pair1 at a selected site (54°W, 15°S; marked with a diamond in Figure 11) during Aug-Oct 2010. The grey dashed line in the bottom panel shows the absolute GPP difference (dGPP) between allaer and nobbaer.

575
We first examine the two experiments in pair1 by taking a close look at the time series of 576 aerosol, cloud, radiation, and ecosystem responses generated at a selected site (54°W, 15°S) 577 during Aug-Oct 2010 ( Figure 10) (site location marked in Figure 11), with the aim of extending 578 the general understanding gained in section 3.2 to a real case study at a single site in the 579 Amazon. This is an interesting site and period, showing a large DFPAR change (Figure 11f) and 580 providing a wide variety of conditions for study -the sky alternates between clear and cloudy 581 conditions in August, is relatively clear in September but relatively cloudy in October, and the 582 biomass burning aerosols increase in August, peak in September, and diminish greatly in early 583 October (Figure 10). During August-September, when the atmosphere experiences biomass 584 burning pollution, the allaer (with BBAOD light fertilizer) and nobbaer (without BBAOD light 585 fertilizer) results differ significantly: DRPAR for allaer (solid line) lies below that for nobbaer 586 (dotted-line), while DFPAR and GPP for allaer are generally higher than those for nobbaer. In 587 October, the sky is almost clean (i.e., low BBaer), leading to very similar results for DRPAR, 588 DFPAR, and GPP between the two experiments. Looking closer, we see that the changes of 589 DRPAR, DFPAR, and GPP between allaer and nobbaer are more prominent when the 590 atmosphere has low cloudiness and high aerosol (e.g., at the end of August), confirming both that 591 BBaer does transform some of the direct light at the surface into diffuse light and that plants are 592 more efficient in their use of diffuse light. When both cloudiness and aerosols are high (e.g., at 593 the end of September), the influence of aerosols is overwhelmed by clouds, and the impact of the 594 aerosols on radiation and the ecosystem becomes secondary. 595 596 We now evaluate BB aerosol impacts on radiation and ecosystem fields over the Amazon during 597 August 2010, when the aerosol has its largest impact. Figure 11 shows the simulated Amazon 598 DRPAR, DFPAR, and GPP fields from the two experiments comprising pair1 (nobbaer and 599 allaer). The distribution of DRPAR shows a clear spatial gradient, with low values in the 600 northwest and high values in the southeast, and the spatial pattern of DFPAR shows the reverse 601 pattern. These features are primarily controlled by the cloud distribution ( Figure 3). Comparing 602 the nobbaer and allaer results by calculating field relative change (i.e., (allaer-nobbaer)/allaer), 603 we find that BBaer decreases DRPAR by 16% and increases DFPAR by 10% over the Amazon 604 region, with maximum local changes of up to -50% for DRPAR and 25% for DFPAR. 605 Interestingly, these maxima are not co-located, though the spatial patterns of perturbations do 606 agree with each other. The mismatch in the locations of the maxima in the difference fields 607 implies a nonlinear response of direct and diffuse light to aerosol and cloud particles (see section 608 3.2). In response to the inclusion of BBaer, the Amazon GPP increases by 10%. That is, the 609 increase in GPP stemming from the increase in the diffuse light fraction overwhelms a potential 610 reduction in GPP from a reduction of total PAR. When we consider all burning seasons over the 611 7-year studied period, the biomass burning aerosol increases DFPAR by 3.8% and decreases 612 DRPAR by 5.4%, allowing it to increase Amazon GPP by 2.6%. However, the 7-year averaged 613 GPP increases by 0.99% (Table 2), which is much less than the value during burning seasons. 614 615 We also examine the multi-year (2010-2016) BBaer impacts on net primary production (NPP), 616 that is, the rate at which carbon is accumulated (GPP) in excess of autotrophic respiration. In 617 essence, NPP can be considered a proxy for the net plant sink of atmospheric carbon. Figure 12  618 shows monthly and long-term averaged NPP over the Amazon Basin from the two experiments 619 comprising pair1. The monthly change of NPP (i.e., dNPP = NPP(allaer) -NPP(nobbaer)) is 620 shown in the figure as a green line. Each year, during the August-September period when BBaer 621 is high and cloudiness is low over the Amazon, BBaer is seen to enhance NPP. The percentage 622 difference of annually-averaged NPP (dNPP/NPP(nobbaer)*100) in % is 4.2, 0.06, 1.9, 0.5, 1.   Figure 11. August 2010 Amazon DRPAR (W m-2 ) (a, b, c), DFPAR (W m-2 ) (d, e, f), and GPP (kg m -2 s -1 ) (g, h, i) from the nobbaer (a, d, g) and allaer (b, e, h) GEOS experiments. The (c, f, i) shows the relative change between allaer and nobbaer. All values are the Amazon regional average except the GPP values of (g, h) are regional total. Further analyses on the (c, f, i) diamond locations are given in Figure 10.

634
To assess how our simulated GPP/NPP response compares with other existing model estimates, 635 we summarize all relevant studies in

How clouds adjust the BBaer diffuse radiation fertilization effect 666
Our second objective in this study is to investigate how the presence of clouds modulates the 667 ability of BBaer to affect GPP. We highlight the cloud impact because even at the same biomass 668 burning aerosol optical depth (BBAOD), the surface downward DRPAR and DFPAR can be 669 very different between cloudy and cloud-free conditions (see section 3.2). As mentioned above, 670 the Amazon's so-called "dry season" still features a considerable amount of cloud, and the 671 cloudiness levels vary significantly from year to year. This raises some questions: How do 672 clouds affect the aerosol impact on radiation fields during the Amazon biomass burning season? 673 Could different levels of background clouds have different impacts on the efficacy of the BBaer 674 DRFE? There are two distinctive features in clouds and aerosols that require us to treat them 675 differently in their impact on the radiation flux to the ecosystem. First, like our distinction of 676 natural and anthropogenic aerosols in their impact on air quality and climate, the cloud is a more 677 natural phenomenon, while biomass burning aerosols (BBaer) can be, at least partially, 678 controlled by humans. Second, clouds are much more efficient in controlling both direct and 679 diffuse radiation fields than aerosol ( Figure 6). What is the potential range of the variation of 680 Amazon aerosols. The correlation presumably stems from the fact that biomass burning aerosols increase 694 the diffuse PAR reaching the canopy (dashed pink line) although they decrease the total PAR 695 (dotted purple line) via decreasing direct PAR (Table 3 and Table S1a). This aerosol-radiation-696 GPP relationship is seen to vary with cloud amount with clouds acting to reduce the aerosol 697 impact; both the diffuse radiation and the GPP show larger changes with BBAOD under clear 698 sky conditions. The overall (i.e., all-sky) aerosol impact on dGPP is similar to that for a cloud 699 coverage of 0.3-0.6, simply because the averaged cloud coverage over the Amazon during the 700 studied period is roughly in that range. 704 Figure 13 and  variation of biomass burning emissions and BBAOD,  711  table S1e). In addition to the detailed information given in Tables S1a-e and S2a-e, we  712  summarize in Table 3 the averaged GPP, DFPAR, DRPAR, CLDFRC, and BBAOD during Aug-713 Sept, 2011-2016 over the Amazon region in all-sky conditions. Also given in Table 3 is the  714 multi principle, similar to the method of aerosol radiative forcing (RF) estimation (i.e., estimating 724 aerosol radiative effect (RE) with and without aerosol for present-day (pair1) and pre-industrial 725 (pair2) conditions and then deriving RF as a difference between the two pair REs). Here we 726 study the sensitivity of the aerosol DRFE to a unit change of AOD. We call it susceptibility of 727 the DRFE to BB aerosols. That is, on a daily basis, the sensitivity of a variable X to a change in 728 the biomass burning AOD is calculated as: ddX/dAOD = ((dX)1-(dX)2)/(AOD1-AOD2). Here, the 729 X represents GPP, DRPAR, and DFPAR, and the subscripts 1 and 2 represent the pair1 or pair2 730 experiment, respectively. 731 732 ddX/dAOD is computed on a gridded daily basis over August-September of 2011-2016. The 733 calculations are then catalogued according to daily cloud cover fraction -we combine the results 734 within each of 10 cloud fraction bins (0-0.1, 0.1-0.2, …, 0.9-1.0). To examine the maximum 735 impact of interannual cloud change during our study period, the binned ddX/dAOD vs. CLDFRC 736 relationship is also computed separately from daily (August-September) values in 2013 and from 737 corresponding daily values in 2015, as these are the years for which monthly cloud cover is 738 around the maximum (0.44) and minimum (0.35), respectively ( Figure 13 and table S1e). 739 Figure 14 shows the results. An almost linear relationship is seen between the ddX/dAOD values 740 and cloud cover fraction. BB aerosols increase GPP in clear sky conditions (e.g., 29.6 kgm -2 s -1 ) 741 but decrease it under full cloudiness conditions (e.g., -5.8 kgm -2 s -1 ). The cloud fraction at which 742 BB aerosol switches from stimulating to inhibiting plant growth occurs at ~0.8. Cloud conditions 743 thus not only affect strongly the strength of the aerosol DRFE but can also change the 744 fundamental direction of the effect. The lines produced for the three different study periods are 745 fairly similar, indicating that the relationship of ddX/dAOD to CLDFRC is fairly stable within 746 the range of cloud cover seen over the Amazon during the period of interest. Figure 14 also  747 indicates that the dGPP can change from 18.5 to 15.5 (kgm -2 s -1 ) with a unit AOD of burning 748 particles released to the atmosphere under the range of Amazon interannual cloud variation in 749 dry season, which is 0.35 to 0.44 in our study period. In other words, there is ~20% dGPP 750 uncertainty adjusted by background Amazon cloud. Our work demonstrates quantitively the role 751 of clouds in tempering the aerosol diffuse radiation fertilization effect. 752 753 754

Conclusions 756
We use the NASA GEOS ESM system with coupled aerosol, cloud, radiation, and ecosystem 757 modules to investigate the impact of biomass burning aerosols on plant productivity across the 758 Amazon Basin under the natural background cloud fields experienced during 2010-2016 -a 759 period containing a broad range of cloudiness conditions. We find that the biomass burning 760 aerosol DRFE does stimulate plant growth and has a notable impact on Amazon ecosystem 761 productivity during the biomass burning season (August-September). In the long-term mean, the 762 aerosol light fertilizer increases DFPAR by 3.8% and decreases DRPAR by 5.4%, allowing it to 763 increase Amazon GPP by 2.6%. On a monthly basis, the DRFE can increase GPP by up to 9.9%. 764 Consequently, biomass burning aerosols increase Amazonia yearly NPP by 1.5% on average, 765 with yearly increases ranging from 0.06% to 4.2% over the seven years studied. This 1.5% NPP 766 enhancement (or ~92TgC yr -1 ) is equivalent to ~37% of the carbon loss due to Amazon fires. 767 768 The aerosol DRFE is strongly dependent on the presence of clouds, much stronger in clear sky 769 conditions and decreases with the increase of cloudiness. A fairly robust linear relationship is 770 found between cloud cover fraction and the sensitivity of radiation and GPP change to a change 771 in biomass burning AOD. BB aerosols stimulate plant growth under clear-sky conditions but 772 suppress it under full cloudiness conditions. Over the Amazon region within our study period, 773 the cloud fraction at which a unit AOD switches from stimulating to inhibiting plant growth 774 occurs at ~0.8. Note, however, that while our results show a clear sensitivity of the aerosol 775 DRFE to cloudiness, interannual variations in the aerosol light fertilizer's overall effectiveness 776