This study is focused mostly on the application and analysis of results obtained from the CPTEC-BAM coupled dynamic vegetation–atmosphere model. The CESM model is employed as a parallel model to specifically test the effects of deforestation and compare its results to those of other studies that employed this model to evaluate the physiological effects of eCO2 on Amazon rainfall (e.g. Kooperman et al., 2018).

CPTEC-BAM is a global atmospheric model created by the Centre for Weather Forecast and Climatic Studies (CPTEC) from Brazil's National Institute for Space Research (INPE), with a horizontal spectral grid T62 (∼ 1.875∘ lat × 1.875∘ long) and 28 vertical levels (hybrid sigma-pressure coordinates, with sigma close to the surface and pressure at the top of the atmosphere). Previous studies (e.g. Cavalcanti et al., 2002; Marengo et al., 2003) showed that this model was able to simulate the main climatic features of South America, although some systematic errors still remain, such as wet biases over the Andes. The land surface component of CPTEC-BAM is the Integrated Biosphere Simulator (IBIS) (Foley et al., 1996; Kucharik et al., 2000). The model simulates the coexistence of both grass and tree plant functional types (PFTs) in grid cells, and disturbances such as fires are represented by a fixed percentage of the biomass of all PFTs that is reduced each year. The estimation of stomatal conductance (gs) in IBIS is based on the model by Ball and Berry (1982) with an equation (Eq. 1) that describes the response of gs to the carbon assimilation rate (An), relative humidity (hs) and atmospheric CO2 concentration (cs) (Collatz et al., 1991): 1gs=mAnhscs+b, where m and b are the slope and intercept coefficients, respectively, and are obtained by analysing the linear regression of leaf gas exchange data in an environment with controlled ventilation and temperature (Ball et al., 1987). The coefficient m has values of 11 and 4 for tropical evergreen forest and tropical (C4) grasslands, respectively. The coefficient b has a value of 0.01 for tropical evergreen forest and a value of 0.04 for C4 grass. Hydraulic stress control over stomatal conductance is considered through the incorporation of a multiplying parameter based on soil water moisture, ranging from 0 to 1.

CESM is an Earth system model developed by the USA's National Center for Atmospheric Research (NCAR) that provides simulations of the Earth's climate (Hurrell et al., 2013). CESM is composed of five separate models representing the Earth's atmosphere (Community Atmosphere Model version 5-CAM5), ocean (Parallel Ocean Program-POP version 2), land (Community Land Model 4.5-CLM4.5), land-ice (Glimmer ice sheet model-G-CISM) and sea-ice (Community Ice CodE-CICE4) systems. These components communicate with each other through a central coupler component. The CESM system allows several resolution configurations and combinations of components and includes the potential for making simulations with only the surface component or with the surface component coupled with the atmospheric model, among many other combinations. The spatial resolution used is 0.9∘ lat × 1.25∘ long or approximately 100 km.

The numerical experiments employed here include simulations that consider the increase in the concentration of atmospheric CO2 affecting plant physiology as well as experiments that consider deforestation in the Amazon, as follows (Table 1):

Control. These are control runs with an atmospheric CO2 concentration of 388 ppmv, one with a dynamic and another with a static geographical distribution of vegetation types (for comparison with the Physiology and Deforestation scenarios, respectively).

For all model runs, sea surface temperature was considered the climatological mean annual cycle for the 1981–2010 period. In the experiments with increasing CO2, the dynamic vegetation scheme was turned on, meaning that the geographical distribution of vegetation types could vary throughout each model run according to the variations in the climatic variables (given that our analysis is focused on precipitation patterns over the Amazon region, the dynamic vegetation changes are not analysed here, especially because there are no significant changes from broadleaf forest to other vegetation types in the eCO2 runs). On the other hand, dynamic vegetation is disabled and C4 grass vegetation was prescribed and held constant until the end of the integration in the experiments representing the deforestation of the Amazon rainforest. The numerical experiments with dynamic vegetation were integrated for a period of 100 years, with constant CO2 concentrations as prescribed in Table 1. Both the control and sensitivity runs for the Deforestation scenario were run for a period of 30 years, given that these runs employed static vegetation. All control and experimental simulations were carried out using three different initial conditions derived from three distinct days (1 January 2003; 10 October 2007; 17 December 2017) of the USA's National Centers for Environmental Prediction (NCEP) reanalysis. The analysis of all scenarios relied on averaged results over the last 10 years of each simulation.

Similar “Control” and “Deforestation” experiments were carried out using the CESM model for a comparison with the “Physiology” runs conducted using this model in other studies (Cao et al., 2010; Kooperman et al., 2018). These CESM simulations were configured with only the atmospheric and land surface components enabled to produce simulations that could be comparable with those of CPTEC-BAM.

The similarity of the changes in average precipitation, evapotranspiration and moisture convergence between the Physiology and Deforestation scenarios reveals the strength of the forest's ecophysiological (i.e. stomatal) control on the regional climate (Fig. 4). The effect that a higher CO2 concentration has on reducing gs (Eq. 1) overcomes the positive effect of increased gross primary productivity (GPP) (Physiology: +7.0 µmol CO2 m-2 s-1 (+58 %); Deforestation: -1.0 µmol CO2 m-2 s-1 (-16 %)) on gs, resulting in a net reduction in stomatal conductance in the Physiology run of -0.10 mol m-2 s-1 (-26 %), related to a decrease of -0.35 mm d-1 (-18 %) in canopy transpiration (Table 2). On the other hand, the decreases in precipitation and evapotranspiration obtained in the Deforestation run (Fig. 4) do not result in the considerably lower gs that is generally maintained by C4 grasses (-0.02 mol m-2 s-1; -4 %). However, the considerable reduction in the leaf area index (-72 %) and a slightly decreased GPP are associated with an average decrease in transpiration (-0.42 mm d-1; -22 %) in the Deforestation scenario. Nevertheless, a counterintuitive increase in specific moisture along the vertical atmospheric profile above the planetary boundary layer is found in the Physiology model run with CPTEC-BAM (+0.32 g kg-1), whereas the same model shows a decrease in specific humidity in the Deforestation run (-0.07 g kg-1) (Fig. 5b and d).

As previously modelled in the study by Kooperman et al. (2018) using CESM, eCO2 is related to convective heating over Central Africa that drives anomalous eastward flows across the tropical Atlantic Ocean, ultimately affecting the flow of humidity into the Amazon basin (Fig. 5). In fact, there is also a strengthening of the Walker cell observed in CPTEC-BAM over the Amazon region, with increased moisture convergence in northern South America (also helped by stronger westerlies from the Pacific in this region) that is not as strong as that observed in CESM but is sufficient to result in a precipitation increase in the north Andes and an atmospheric stabilization with precipitation decreases across most of the Amazon.

The atmospheric circulation changes are completely different in the Deforestation scenario (Fig. 5c), in which there is a pronounced increase in easterlies across the entire Amazon region as a result of decreased roughness length of surface vegetation (2.65 m in tropical evergreen forest and 0.08 m in C4 grass; Sampaio et al., 2007) and the reduced pumping of deep soil moisture to the atmosphere, especially in the dry season (June to October) (Fig. 6d). Figure S5 shows the meridional mean planetary boundary layer height at the Equator over the Amazon under the different scenarios. In the Deforestation scenario, there is an average decrease of 10 % in the boundary layer height, attributable to the considerably lower surface roughness length of pastures compared to that of tropical forests. On the other hand, there is an average increase of 21 % in the boundary layer height in the Physiology run, associated with the increased heating of the surface. As a result, eCO2 causes a higher, drier and warmer boundary layer over the Amazon that acts as a barrier to a humidity-enriched, though shallower, simulated free troposphere with less deep convection (see Langenbrunner et al., 2019). On the other hand, the strong increase in westward moisture advection, aligned with the increased albedo and decreased vertical mixing (Fig. S5) seems to best explain the nearly unchanged surface temperature seen in the Deforestation scenario. The superposition of the spatial pattern of changes in moisture convergence over the 850 hPa atmospheric circulation anomalies shows that different circulation patterns produce similar changes in the region's atmospheric moisture budget (Fig. 5a and c).

The reduction in latent heat flux in our simulations (Fig. 3c and Table 2) also helps reduce convection over the Amazon region, tending to cool the upper atmosphere and reinforce atmospheric stabilization.

These changes in horizontal circulation imply, in the Physiology scenario, that less moisture enters the Amazon region from the Atlantic (-4.9 kg m-1 s-1) and less moisture leaves the regions towards the Andes (-2.6 kg m-1 s-1) (this latter is somewhat compensated by a stronger moisture convergence from the Pacific to the Andes, as shown in Fig. 5b). In the Deforestation scenario, there is an increase in the input of humidity to the Andes at the surface level (on the order of 3.0 kg m-1 s-1), which is also perceptible in the western part of the vertical humidity profile near the surface levels (Fig. 5d). The lower evapotranspiration capacity, aligned with the lower vertical mixing due to pasture's lower roughness length (than that of forests), results in an atmospheric volume that is depleted of moisture and shows a decreased uplifting of air masses. In the Physiology scenario, despite the decreased evapotranspiration capacity, the increased surface heating increases vertical mixing at low levels (up to 700 hPa), associated with a deeper boundary layer and higher mixing layer, which is, in turn, connected to the increase in humidity throughout the free tropospheric volume (above the boundary layer) over the region. However, after such atmospheric heights, there are strong subsidence anomalies seen in the Physiology run (Fig. 5b), which decrease deep convection that is ultimately associated with lower rainfall rates. The same vertical circulation patterns have been demonstrated well in previous (separate) studies that modelled the large-scale deforestation of the Amazon and, more recently, the isolated physiological effects of eCO2 on the region's climate (see Langenbrunner et al., 2019).

A decrease in the surface sensible heat (-1.34 W m-2) in the Deforestation run (Fig. 3c), alongside a decrease in the latent heat, results in a negative net surface radiation balance in the Deforestation run, associated with a small decrease in the average 2 m air temperature (-0.2 ∘C) (Table 2) (but also with an increase of +0.4 ∘C in surface temperature). On the other hand, in the Physiology scenario, an increase in sensible heat (+3.96 W m-2) is observed, associated with an average increase in the 2 m air temperature of +2.1 ∘C. While the decrease in latent heat is also directly connected to a lower evapotranspiration capacity, the opposite results shown in each scenario regarding sensible heat are also associated with opposite changes in near-surface atmospheric circulation patterns: in the Deforestation run, there is an increase in near-surface atmospheric advection, whereas in the Physiology scenario, this advection is considerably decreased (as explained in Sect. 3.2, Atmospheric circulation). Shortwave radiation is increased due to decreased nebulosity in both model scenarios (Physiology: -1.4 %; Deforestation: -2.2 %), but such an increase in the shortwave radiation balance is stronger in the Deforestation scenario due to the albedo change. The same pattern is also obtained for the surface balance of longwave radiation, which increases in both scenarios but increases more strongly in the Deforestation run (Physiology: 2.7 W m-2; Deforestation: 6.9 W m-2), which is probably a combination of the lower evapotranspiration capacity and increased horizontal advection in the latter scenario.

Precipitation is consistently below the control values year-round in the Physiology and Deforestation experimental model runs (Figs. 6a and S4a). However, differences regarding monthly precipitation between the Physiology and Deforestation scenarios are evident at the end of the dry season and at the onset of the rainy season (August to December). In this regard, precipitation seasonality is stronger in the Deforestation scenario than in the Physiology model run. This is closely linked to changes in evapotranspiration given that the permanence of the forest in the Physiology scenario supports a higher evapotranspiration flux during the dry season compared to that in the Deforestation run (Fig. 6c). On the other hand, the evaporation values in the Deforestation run are, for most of the year, above the control values, which explains the higher evapotranspiration observed during the rainy season in comparison to that seen in the Physiology scenario (although evapotranspiration is reduced in comparison to the control run, following the reduction in precipitation).

As shown, for example, by Kooperman et al. (2018), the physiological effects of eCO2 on the region's climate take place, namely, in the wet season, when GPP is higher and transpiration is lower (see Fig. 6d and h), even though our results also show a considerable rainfall reduction during the dry season. Conversely, it has been demonstrated (e.g. by Lawrence and Vandecar, 2015) that large-scale deforestation causes climatic changes specifically during the dry season, when transpiration is particularly reduced, as was also shown in our results (Fig. 6a and d).

These seasonal variations in evapotranspiration are at least partly explained by the opposing seasonal patterns of canopy transpiration in the Physiology and Deforestation scenarios (Fig. 6d). On the one hand, the highest values of this variable in the Physiology run occur during the dry season, when a high vapour pressure deficit increases the evapotranspiration demand that trees can fulfil (at least partially) even under the given eCO2. On the other hand, the lowest canopy transpiration values in the Deforestation run occur during the dry season as a result of seasonal decreases in the pasture leaf area index and root depth in this scenario.

Stomatal closure driven by eCO2 is related to higher water use efficiency (the amount of water used (in transpiration) per unit of carbon assimilated through photosynthesis), but even so, the net effect is a small decrease (∼ 2 %) in the available soil water in the Physiology scenario due to the simulated decrease in precipitation. This decrease is more pronounced in the Deforestation run (reaching a reduction of 30 % at the peak of the dry season in September) because the GPP is considerably lower at this time of year in pasture grasslands, which, together with the lower evapotranspiration and the decreased input of rainwater, acts to decrease the soil water in the dry season in the Deforestation scenario (Fig. 6f).

There is a long-known and overwhelming agreement among models that the whole-basin deforestation of the Amazon is associated with a warmer (average of 1.9 ∘C (±1.8 ∘C) vs. -0.2 ∘C in 2 m air temperature (+0.4 ∘C in surface temperature) in the current simulation with CPTEC-BAM) and drier (average -15 % vs. -9 % from CPTEC-BAM) climate over the region, driven namely by an increase in trade winds due to the considerably smaller roughness length of pastures than that of forests (Lawrence and Vandecar, 2015; Sampaio et al., 2007; Spracklen and Garcia-Carreras, 2015; Sud et al., 1996). Fully interactive coupling between the atmosphere and oceans results in twice the rainfall reduction in comparison to that output by non-coupled simulation such as those conducted in the present study (Nobre et al., 2009). Although previous modelling and observational studies (e.g. Saad et al., 2010; Silva Dias et al., 2002) have shown that small-scale deforestation is combined with a localized increase in rainfall, there is now modelling and observational evidence that widespread and large-scale deforestation in the Amazon drives rainfall reductions (Lawrence and Vandecar, 2015; Nobre et al., 2016; Sampaio et al., 2007) and/or the lengthening of the dry season (Dubreuil et al., 2012; Fu et al., 2013). This latter effect is also in line with our results (Fig. 6a).

While the conceptual model proposed and reviewed by Lawrence and Vandecar (2015) suggests that whole-basin deforestation should lead to rainfall reductions of >30 %, we argue that the longitudinal gradient in rainfall recycling should be considered in these estimates: the rainfall reductions observed with CPTEC-BAM in both the Deforestation and Physiology scenarios are within the estimated range of precipitation recycling in the east Amazon (10 %–30 %, Zemp et al., 2017), which is the region where the subsidence branch of the Walker cell acts most strongly in these simulations.

In contrast to the effect of deforestation on Amazon rainfall, observational or experimental evidence on the effects of eCO2 on water fluxes in tropical forests is scarce. Most of the knowledge on the ecosystem-scale effects of eCO2 comes from low-diversity temperate forests (Ainsworth and Long, 2005; Ainsworth and Rogers, 2007; De Kauwe et al., 2013), laboratory studies with seedlings or saplings (e.g. Aidar et al., 2002), or growth rings obtained from trees at the fringes of tropical forests (van der Sleen et al., 2014). For example, the +150 ppm Oak Ridge free-air CO2 enrichment (FACE) experiment conducted in a broadleaf temperate forest resulted in an average reduction in transpiration of 17 % (De Kauwe et al., 2013). A reduction of 20 % in gs was found in the +150 ppm, single-species, eucalyptus FACE (EucFACE) experiment conducted in woodlands in New South Wales, Australia (Gimeno et al., 2016). Both results are comparable to the 18 % reduction in gs and the 20 % reduction in transpiration found in the Physiology scenario. However, water-use efficiency (calculated here as the ratio between GPP and transpiration) increased by 35 % in the 11-year-long Oak Ridge FACE experiment and by 30 %–35 % in the 1850–2000 period, as assessed from growth rings from trees at the fringes of tropical forests (van der Sleen et al., 2014). Our simulation yielded a much higher value of 94 % in the Physiology scenario, owing to a stronger increase in GPP in CPTEC-BAM (+13 % in Oak Ridge FACE; +58 % in CPTEC-BAM). Although the temperature dependence of Rubisco kinetics implies that the effects of eCO2 on GPP and NPP in the tropics should, in principle, be stronger than those in temperate regions (Hickler et al., 2008), the GPP in CPTEC-BAM seems to be oversensitive to eCO2, as is the case for other vegetation models that do not consider nutrient cycling (De Kauwe et al., 2013). Phosphorus, for example, is a highly limiting nutrient in Amazon soils, and the consideration of such a limitation would decrease the expected eCO2-induced gains in the GPP and NPP simulated by models without nutrient constraints by 42 % and 50 %, respectively, after 10 years (Fleischer et al., 2019). Observations from the strongly P-limited EucFACE site even showed a 12 % increase in the GPP of mature Eucalyptus tereticornis stands after 4 years of CO2 fertilization (Jiang et al., 2020). Should our simulations consider the combined effect of P limitation, gs and therefore canopy transpiration would be even lower, and Amazon rainfall reduction could be even stronger in the Physiology scenario compared to that in the Deforestation scenario.

One must also consider that in a hyper-diverse ecosystem such as the Amazon forest, the response to eCO2 in terms of gs may vary considerably from one tree species to another or from one functional group/strategy of trees to another (Domingues et al., 2014). It is now known that different Amazon tree species can have rather different strategies regarding water usage and saving (Bonal et al., 2000). Such a variety of responses and more subtle implications of eCO2 on Amazon forest functioning have yet to be incorporated in vegetation models or surface schemes (Lapola, 2018).

Therefore, even if our results for the Physiology scenario are aligned with observational results from non-tropical forest ecosystems and modelling results (namely, from the studies by Cao et al., 2010; Kooperman et al., 2018), there is a considerable level of uncertainty in the Physiology scenario projection of CPTEC-BAM (and of CESM, Cao et al., 2010; Kooperman et al., 2018). This level of uncertainty will stay as such until there are direct field-based data on the ecosystem-level effects of eCO2 in the Amazon forest (Norby et al., 2016).

As such, we suggest that future research on this topic should focus on gathering such field-based experimental evidence on the ecosystem-level effects of eCO2 in the Amazon forest and that the basin-wide effects of eCO2 on Amazon rainfall should be projected with models that consider the potential limitations of soil phosphorus and interacting oceans. Additionally, multi-factorial ensemble simulations with gradual increases in CO2 concentrations and deforestation levels (sensu Sampaio et al., 2007) could be valuable for understanding when and how the effects of increasing CO2 and deforestation dominate the rainfall responses in the Amazon region. Last, the similarity of the results obtained for rainfall and evapotranspiration reduction with CPTEC-BAM under the 1.5×CO2 experiment and the results from CESM under the 2×CO2 (Cao et al., 2010) and 4×CO2 (Kooperman et al., 2018) scenarios might be a result, first, of the strong sensitivity of GPP and transpiration to eCO2 in CPTEC-BAM but could also be a consequence of the saturation of eCO2 effects on gs that takes place between 600 and 1000 ppmv, as shown for a variety of plant species with instantaneous measurements (e.g. Domingues et al., 2014; Zheng et al., 2019), although the long-term (beyond the execution time of the FACE experiments) acclimation changes of gs to eCO2 are still poorly known (Xu et al., 2016).

One should interpret the implications of the results presented here with care, keeping in mind the different responsibilities involved in the two anthropogenic disturbances considered in this modelling exercise: deforestation and elevated atmospheric CO2 concentration. Avoiding the significant rainfall reductions projected here involves halting deforestation in the Amazon and reducing global CO2 emissions or actively removing CO2 from the atmosphere. On the one hand, the curbing of deforestation in the Amazon is something that invariably has to be carried out by different actors within the nine Amazonian countries (France and French Guyana included), although international markets and institutions can play important roles as well (Nepstad et al., 2014; Rajão et al., 2020). On the other hand, the increase in atmospheric CO2 concentration is a global process, the mitigation of which demands a concerted effort by all countries, especially the historical and current top emitters (Peters et al., 2015). In this sense, even if Amazon deforestation is stopped in the near future, forest functioning and structure can still be jeopardized by eCO2 and consequent climatic changes. Therefore, while both anthropogenic disturbances analysed in this study–deforestation and elevated atmospheric CO2 concentrations–are associated with equivalent reductions in Amazon rainfall, this result should be interpreted as evidence that both regional and global responsibilities are at stake to mitigate potential future climatic change and its impacts in the region (Lapola et al., 2018).