The data used in this study were collected as part of the ATTO project, a bilateral initiative between Brazil and Germany. Since 2012, ATTO has carried out continuous measurements, as described by , located in an area of pristine tropical forests in the central Amazon (Fig. ), which contains the Instant Tower of 81 m (-2.1441° S, -58.9999° W).

The Instant tower is located 150 km from the city of Manaus in the state of Amazonas, Brazil, at an altitude of 120 m above sea level on a plateau covered by terra firme forests with an average crown height of 40 m . In this landscape, wind speeds are relatively low, around 1 m s⁻¹ immediately above the forest canopy, and above the canopy, the wind speed increases logarithmically with height . The main wind direction at the site is from the NE–E. It passes through areas of minimal anthropogenic influence in the northeast, a clean fetch region covered by tropical forests .

The climate is tropical humid and characterized by two seasons (wet and dry), driven by seasonal shifts of the Intertropical Convergence Zone over the Amazon Basin . The wet season is characterized by more than 200 mm of rainfall per month and an average temperature of around 25 °C at the forest-atmosphere interface. In contrast, the dry season sees less than 100 mm of rainfall per month and an average temperature of around 27.7 °C .

The dataset used in this study was measured at the ATTO site from 2016 to 2022 (see Table ). Wind speed, sensible heat flux (H), latent heat flux (LE), and carbon dioxide flux (FCO₂) data were calculated as 30 min averages using EddyPro® software (LI-COR), as derived from fast-response sonic anemometers, according to . The other variables (radiation, thermodynamics and aerosols) were obtained as 30 min averages, including net radiation (Rn) and its radiative components: incoming and outgoing shortwave radiation (SWin and SWout), and atmospheric and terrestrial longwave radiation (LWatm and LWterr), respectively. Additionally, diffuse shortwave radiation (SWd) was measured using a SPN1 Pyranometer (Delta-T Devices) installed at 75 m above ground level. However, SWd data were available only for 2021, prior to this year, SWd was not measured at the ATTO site, and data from 2022 were excluded due to technical issues with the sensor.

Based on and , these data were organized by seasonality into four periods: (i) the wet season (February to May), which has a cleaner atmosphere, (ii) the wet-dry transition (June to July), (iii) the dry season (August to November), which has higher levels of pollution, and (iv) the dry-wet transition (December to January).

To eliminate cloud interference and investigate the role of aerosols in surface energy fluxes, the central objective of this study, we used data from the Aerosol Robotic Network (AERONET) at the ATTO site, specifically AOD (version 3, level 2). These data are free of cloud contamination due to pre and post-field calibration . Based on this, 30 min averages were calculated between 2016 and 2022 for which AOD data from AERONET were available, the initial combined dataset comprised 10 890 observations, including all variables listed in Table 1. This matched dataset served as the starting point for the subsequent quality control and filtering procedures. First, the turbulent fluxes underwent quality control following , who defined that only data with flags “0” (best quality) and “1” (acceptable for general analysis) should be used; data with flag “2” (poor quality) were discarded. Second, this study only considered the daytime period (from 07:00 to 17:00 LT) because the highest Rn values occur during this time. After filtering, the resulting dataset is summarized in Table S1 and S2 in the Supplement.

Using the values for humidity and temperature (variables shown in Table ), it was possible to calculate the vapor pressure deficit (VPD) using Eqs. (1) to (3) according to . 1VPD=es-ea The water vapor saturation pressure (es) as a function of temperature (T) was calculated according to the equation . 2es(T)=6.112exp⁡17.67⋅TT+243.5 The actual vapor pressure (ea) was obtained by relating it to the relative humidity (RH). 3ea=RH⋅es

Daily averages of AOD values were obtained to investigate seasonal variability. Our analysis distinguishes two contrasting atmospheric conditions at the ATTO site, defined as “Clean” and “Polluted” using AOD thresholds derived from the dry-season distribution of AOD. The Clean and Polluted regimes correspond to the 10th (AOD ≤0.13) and 90th (AOD ≤0.40) percentiles, respectively. Further details on the seasonal aerosol analysis are provided in Sect. 3.1 and Table S3. Subsequently, 30 min AOD averages between 07:00 and 17:00 LT were computed to ensure temporal consistency with the surface flux data and enable direct comparisons. To improve the visualization of the mean diurnal patterns, a 4th-order polynomial curve was applied exclusively as a smoothing technique to the observational data. This curve fitting was used solely for graphical purposes and does not represent a physical or predictive model. All analyses were based on the measured data. For comparisons between Clean and Polluted regimes, only the interval from 10:00 to 14:00 local time was considered, as this period corresponds to the maximum net radiation at the study site and minimizes the influence of low solar elevation angles.

Statistical differences in meteorological variables and surface fluxes between the Clean and Polluted regimes were assessed using the Mann-Whitney U test. This non-parametric approach was selected because the observational data violated the assumption of normality, as confirmed by preliminary Shapiro-Wilk tests. The Mann-Whitney U test was used to determine whether the median values of the two independent regimes differed significantly (p<0.05), offering a robust framework for analyzing non-normally distributed atmospheric data .

The distribution of atmospheric aerosols, expressed as AOD, exhibits a clear seasonal cycle at the ATTO site (Fig. ). The lowest values occur during the wet season, with an average of 0.07 in April, while the dry season is marked by higher AOD values, reaching an average of 0.28 in September. Furthermore, this seasonal variation in AOD values has previously been observed at other sites in the Amazon region . , for example, used data measured at the ZF2 site, located 60 km northwest of Manaus in central Amazonia, to show that AOD values were close to 0.4 (with peaks above 0.5) in the dry season and less than 0.2 in the wet season. Attention is drawn to the AOD values observed in the southern region of the Amazon basin, which is influenced by the arc of deforestation, an agricultural frontier zone with intense burning activity during the dry season . Several studies in this region have shown that AOD values often exceed 4 in the dry season, whereas in the wet season they rarely exceed 0.2 .

The main distinction between the AOD values measured at the ATTO site and those measured in the southern Amazon is the magnitude of these values. In particular, the AOD values at the ATTO site are approximately 15 times lower than those in the region close to the arc of deforestation during the dry season . and for example, investigated the seasonal contrast of aerosols at the ATTO site, highlighting that parts of the wet season resemble preindustrial conditions with minimal human impact.

Figure  shows the average daily AOD values for the dry and wet seasons, from 2016 to 2022. It is clear to see that the highest average AOD values were obtained during the dry season, with values reaching 1.5, while in the wet season these values did not exceed 0.5, a result similar to that already reported in Fig. . It should also be noted that during the dry season, the 90th and 10th percentiles of the AOD values are 0.40 and 0.13, respectively. During the wet season, these percentiles were 0.13 and 0.04, respectively. In other words, the AOD values above the 90th percentile in the wet season are slightly higher than the values observed for the 10th percentile in the dry season. This reinforces what was already mentioned in Fig. , that the wet season in the ATTO region is quite “Clean” compared to the dry season. As the main goal of this work is to investigate the impact of aerosols on surface turbulent fluxes, the analysis focuses on data from the dry season. In addition, during the dry season there is more aerosol data since the cloud interference is much less pronounced than during the wet season. Two aerosol regimes were defined based on percentile thresholds of the dry-season AOD distribution. Several percentile combinations were tested to assess the robustness of the regime separation. Based on this analysis, the 10th and 90th percentiles were selected to define the Clean (AOD ≤0.13) and Polluted (AOD ≥0.40) regimes, respectively, as they preserve physically meaningful differences between aerosol regimes (See Table S1).

As described in Sect. 2.3, the comparisons between Clean and Polluted regimes were restricted to the 10:00–14:00 LT period, corresponding to the maximum net radiation. The full diurnal cycles of shortwave, longwave, and net radiation during the dry season (2016–2022) show that the maximum values occur between 10:00 and 14:00 LT (Fig. ), supporting the choice of this time window for the subsequent analyses. The average values of the radiation balance components during this period are summarized in Table . The negative sign in the difference between the Polluted and Clean regimes indicates that the radiative components decrease during this period. The Rn fell the most in relative terms, by around -4 %. Outgoing shortwave radiation (SW_out) showed a non-significant increase of 3.3 % (p=0.07). As is well known, the longwave balance is always negative during the daytime in the Amazon region since LW_terr is greater than LW_atm. However, pollution reduced the difference between LW_atm and LW_terr by around 2 Wm-2 compared to the Clean regime, indicating a slightly less radiative surface and a slightly warmer atmosphere.

Quantifying the impact of aerosols on radiative flux remains a significant challenge in climate system studies, with persistent uncertainties . However, the relationship between aerosols and radiative flux has been investigated for decades in the Amazon region . There is a consensus in the literature that an increase in AOD reduces SW_in, which consequently also causes a reduction in Rn. However, the magnitude of these reductions varies considerably. Studies carried out during the dry season in the Amazon rainforest using different methods to estimate direct aerosol radiative forcing (ARF) illustrate this variability. For example, reported an average daily ARF of -20±7Wm-2 per unit of AOD at 550 nm in the Amazon rainforest. Consistent with these findings, estimated an average ARF of -20.77±5.04 Wm-2 for the dry season in the central Amazon. found daily ARF values ranging from -21 to -74Wm-2 in the deforestation arc, an area with higher levels of pollution than the central Amazon. investigated this central region and reported a daily ARF value of -32 Wm-2.

Although these studies provide estimates of the reduction in surface radiation from aerosols in the Amazon, they do not converge on a single consensus value. This is because, in addition to the different methodologies used to obtain ARF values, , and point out that uncertainties lie mainly in the complex interactions between types and concentrations of aerosols, surface characteristics, atmospheric conditions, and solar angle.

SW_out is directly related to surface albedo and the fact that it did not change significantly in our data between regimes (maintaining albedo at ∼0.11) indicates that pollution has a secondary effect compared to the characteristics of the surface itself. There is a wide range of surface characteristics in the Amazon that directly influence albedo, as observed by and : (i) degree of vegetation cover; (ii) soil and vegetation water conditions; (iii) solar elevation; (iv) cloud cover and; (v) wind speed and direction.

However, the behavior of longwave radiation was quite interesting. It shows that because of their interaction with the incident shortwaves, aerosols increase the emission of thermal energy toward the surface. At the same time, they act as a barrier to the total energy reaching the surface, thus impacting the amount of thermal energy emitted by the surface itself. The increase in LW_atm and the decrease in LW_terr in the Polluted regime result in a smaller longwave balance in this regime. also observed this effect in their experiments involving biomass burning aerosols in South America: a subtle variation in longwave intensity attributed to the presence of aerosols.

With reduced solar energy input on the surface during the Polluted regime, cooling occurs at the forest-atmosphere interface, accompanied by a decrease in VPD compared to the Clean regime, as illustrated in Fig. . The cooling between the 10:00 and 14:00 LT regimes implies an average reduction in canopy surface temperature of 0.9 °C, based on infrared surface temperature measurements, and a corresponding reduction in air temperature of 0.3 °C, resulting in a -2 hPa (13%) decrease in VPD.

As the curve for the Clean regime is consistently above that for the Polluted regime at all shown temperatures, it is suggested that the Clean regime will first achieve a reduction in evapotranspiration, given the approximately linear relationship between temperature and VPD.

These cooling values are consistent with the effects documented in other studies. For example, found a reduction in 1.2 °C above the Amazon region, while identified a 1.8 °C and a decrease in 35% in VPD in the central Amazon. In the deforestation arc, found an average cooling effect of between 3 and 4 °C, as well as reductions of between -2 and -3 hPa in VPD.

investigated temperature variations in the Amazon using a radiative transfer model. By simulating a scenario without aerosols (AOD =0) and comparing it with real conditions, they observed an increase in temperature in the scenario without aerosols. They identified a correlation between relative irradiance, air temperature, and VPD. Meanwhile, and reinforce the idea that AOD significantly influences temperature variations, particularly on a regional scale. For instance, Palácios et al. (2024) observed positive linear correlations between AOD and air temperature across distinct climatic phases, attributed to the absorption of solar radiation by biomass burning emissions resulting in atmospheric heating. Similarly, Herbert and Stier (2023) utilized reanalysis data to demonstrate that 2 m air temperature increases as a function of AOD, consistent with localized heating of the smoke layer due to strong absorption of solar radiation.

and also highlight that the physical characteristics of the aerosols present in the atmosphere, such as size, mixing state and presence of coatings, as well as the chemical characteristics, such as the ability to absorb or scatter light and hygroscopicity, determine their direct impact on temperature and VPD through radiative interaction, as well as their indirect impact by influencing cloud properties and evapotranspiration rates. These are essential components of the atmosphere's energy balance.

The interaction between aerosols, radiation, and evapotranspiration affects not only temperature and VPD, but also the fluxes of energy and matter on the surface. This has a direct impact on atmospheric and ecosystem processes. Figure  illustrates the impact of aerosols on these fluxes. It shows that for the Polluted regime, the values were lower than those observed during the Clean regime, especially during periods of high solar radiation, i.e. between 10:00 and 14:00 LT. The most significant reductions in the energy available to the surface occur during this period, with Rn falling by -4%, as reflected in the energy partitions. The surface energy balance closure was 0.89 for the Clean regime and 0.88 for the Polluted regime, comparable to values reported in the literature . The corresponding residuals were of similar magnitude (70 Wm-2 for Clean and 75 Wm-2 for Polluted), indicating that the observed differences in energy fluxes are not related to differences in energy balance closure.

Sensible heat decreased by an average of -21.7 Wm-2 (13.5%), reflecting reduced energy transfer to the atmospheric boundary layer. Similarly, LE decreased by -8.9 Wm-2 (2%), indicating limited evapotranspiration due to the reduced radiative energy available. The Bowen ratio, which relates H and LE, recorded 0.38 in the Clean regime and 0.33 in the Polluted regime, suggesting that a higher proportion of energy was allocated to latent processes, as expected in forest environments. The ground heat flux (G) also decreased by -1.0 Wm-2 (54.5%), demonstrating its greater sensitivity to variations in Rn compared to turbulent fluxes.

In addition to their effect on energy fluxes, aerosols were found to have a significant influence on CO2 flux, becoming more negative by an average of 4.9 µmolm-2s-1 (39.5%) in the Polluted regime compared to Clean conditions between 10:00 and 14:00 LT. This is when the difference between the Polluted and Clean regimes is most pronounced, indicating that the forest absorbs more CO2 in the Polluted regime (Fig. ). The reductions in H, LE, G, and FCO₂ shown in Figs.  and  were also observed across individual years (see Fig. S3 in the Supplement).

In the Polluted regime, CO2 fluxes were more negative (Fig. ), indicating increased CO2 uptake by vegetation related to photosynthetic activity. Such enhanced photosynthesis may be linked to changes in stomatal regulation that allow greater CO2 uptake without a proportional increase in transpiration, reflecting higher stomatal conductance efficiency . However, analysis of the LE, which represents the fraction of available energy converted into evapotranspiration, shows a consistent decrease in the Polluted regime compared to the Clean regime (Fig. ).

The apparent paradox of an increase in CO2 absorption alongside an equilibrium in LE can be explained by water use efficiency (WUE). According to and , WUE is defined as the ratio of carbon assimilated to water transpired by vegetation. In this study, WUE was estimated using FCO₂/LE as a proxy. WUE was significantly higher under Polluted compared to Clean regime (mean values of 0.042 and 0.029 µmol J⁻¹, respectively, p<0.05). This indicates that, under Polluted regimes, vegetation assimilates more carbon per unit of water lost, consistent with the observed equilibrium in latent heat flux (Fig. ) despite enhanced CO2 uptake (Fig. ).

In forests in the USA, conducted experiments to quantify the impact of aerosols on turbulent surface fluxes, observing reductions in H and LE ranging from 10% to 30%. Few studies have examined the relationship between H, LE and AOD in the Amazon region. , for example, used regional modeling with an AOD threshold of 0.3 to obtain a daily average reduction of -15 Wm-2 for H and -5 Wm-2 for LE. In the deforestation zone, observed a decrease of -67 Wm-2 (36%) for H and -4 Wm-2 (2%) for LE when simulating Clean conditions (AOD =0) and comparing them with real conditions involving the presence of aerosols. These results suggest that regional climate models may underestimate the reduction in LE, highlighting the importance of biological processes, such as transpiration, in compensating for these effects.

In contrast, numerous studies in the Amazon have demonstrated the significant impact of aerosols on CO2 assimilation by forests. This occurs by increasing the diffuse fraction of photosynthetically active radiation reaching forest shade zones, thereby intensifying photosynthesis. Simultaneously, it reduces the net direct solar radiation reaching the canopy surface, thereby generating photosynthetic enhancement in this region . This diffuse fraction, which falls within the wavelengths of interest for vegetation (0.4 to 0.7 µm), can increase from around 19% (the typical value of a Clean atmosphere) to 80% under biomass burning conditions .

We quantified the diffuse radiation fraction (Fd=SWd/SWin) for the available period (2021) and compared Fd between Clean and Polluted aerosol regimes. Our results indicate higher Fd values under Polluted regime compared to Clean regime (Fig. S1). Specifically for the 10:00 and 14:00 LT interval, the mean Fd values were 0.43 and 0.27 for Polluted and Clean regime, respectively, indicating an absolute difference of 0.16 between the two regimes (p<0.05). This is consistent with enhanced scattering of solar radiation associated with increased aerosol loading . Moreover, daytime CO2 fluxes showed a non-linear dependence on Fd, with net CO2 uptake increasing up to an Fd threshold (≈0.6) and decreasing at higher Fd values (Fig. S2). This behaviour was consistent with the response of net ecosystem exchange for diffuse radiation reported by for four forest sites in China and aligns with the global-scale mechanisms proposed by . These results provide observational support for the proposed mechanism linking aerosol loading, radiation partitioning, and ecosystem carbon exchange.