Occurrence and growth of sub-50 nm aerosol particles in the Amazonian boundary layer

New particle formation (NPF), referring to the nucleation of molecular clusters and their subsequent growth into the cloud condensation nuclei (CCN) size range, is a globally significant and climate-relevant source of atmospheric aerosols. Classical NPF exhibiting continuous growth from a few nanometers to the Aitken mode around 60-70 nm is widely observed in the planetary boundary layer (PBL) around the world, but not in central Amazonia. Here, classical NPF events are rarely observed in ::::: within : the PBL, but instead, NPF begins in the upper troposphere (UT), followed by downdraft injection of sub5 50 nm (CN<50) particles into the PBL and their subsequent growth. Central aspects of our understanding of these processes in the Amazon have remained enigmatic, however. Based on more than six years of aerosol and meteorological data from the Amazon Tall Tower Observatory (ATTO, Feb 2014 to Sep 2020), we analyzed the diurnal and seasonal patterns as well as meteorological conditions during 254 of such Amazonian growth events on 217 event days, which show a sudden occurrence of particles between 10 and 50 nm in the PBL, followed by their growth to CCN sizes. The occurrence of events was significantly 10 higher during the wet season, with 88 % of all events from January to June, than during the dry season, with 12 % from July to December, probably due to differences in the condensation sink (CS), atmospheric aerosol load, and meteorological conditions. Across all events, a median growth rate (GR) of 5.2 nm h-1 and a median CS of 0.0011 :::::::: 1.1×10−3 s-1 were observed. The growth events were more frequent during the daytime (74 %) and showed higher GR (5.9 nm h-1) compared to nighttime


Introduction
New particle formation (NPF) refers to the nucleation of nanometer-sized molecular clusters from gaseous precursors and their subsequent condensational growth (e.g., Kulmala et al., 2004;Dal Maso, 2005;Kirkby et al., 2011;Kulmala et al., 2012;Kerminen et al., 2018). Under favorable atmospheric conditions, the newly formed particles grow through condensation of 10 semi-and low-volatile gases as well as coagulation into the cloud-and, thus, climate-relevant size range with diameters, D, larger than ∼80 nm (see definitions in Kulmala et al., 2012;Kerminen et al., 2018). NPF has been observed worldwide in the course of ground-based observations in different environments, such as rural and remote continental areas, urban environments, the Arctic and Antarctica, marine areas, and mountain sites (Kerminen et al., 2018, and references therein). A brief overview of the current knowledge on the occurrence of NPF worldwide and the chemical mechanisms involved can be found in Andreae 15 et al. (2021). Its wide and frequent occurrence makes NPF a major and possible even dominant source of aerosol particle number concentrations and cloud condensation nuclei (CCN) on global scales (e.g., Merikanto et al., 2009;Spracklen et al., 2008;Nieminen et al., 2018;Yli-Juuti et al., 2020).
In the long list of locations where the 'classical NPF' has been detected in the planetary boundary layer (PBL) , the Amazon rain forest is a remarkable exception (e.g., Andreae, 2013;Rizzo et al., 2018;Wimmer et al., 2018). 20 Here, events have been observed that indeed resemble the classical 'banana plots' of NPF but differ clearly in the initial diameter of the growth curve. While the smallest diameters in, for instance, boreal forest is typically in the few nm range, the 'Amazonian bananas' rather start between about 20 to 40 nm (Kulmala et al., 2012;Kerminen et al., 2018). Rizzo et al. (2018) discussed the occurrence of such sub-50 nm particle growth events in the Amazon and found them only in 3 % of the 749 days examined, associated mainly with convective downdrafts. Accordingly, the 'Amazonian bananas' start at larger diameters and 25 are comparatively rare relative to the classical events e.g., in boreal forests (Nieminen et al., 2018;Dada et al., 2018).
This striking contrast to other environments has inspired researchers to investigate the underlying mechanisms that could explain the absence of NPF as well as alternative particle sources that sustain the Amazonian aerosol population. Reasons for the absence of NPF within the PBL could be: 1. Suppression by isoprene (e.g., Kiendler-Scharr et al., 2009;Kanawade et al., 2011;McFiggans et al., 2019;Yli-Juuti 30 et al., 2020), which is the most abundant volatile organic compound (VOC) in the Amazonian atmosphere (e.g., Yáñez-Serrano et al., 2020).
2. The very low concentrations of inorganic precursor gases such as sulfur dioxide (SO 2 , being converted into sulfuric acid H 2 SO 4 ) as well as the bases ammonia (NH 3 ) and amines (NR 3 ) Trebs et al., 2004), which play key roles in the binary H 2 SO 4 -H 2 O and ternary NH 3 -H 2 SO 4 -H 2 O nucleation mechanisms (Kirkby et al., 2011;Andreae et al., 2021).
3. The high levels of relative humidity (RH), which have been associated with a low occurrence of NPF (e.g., Bonn and 5 Moortgat, 2003;Hamed et al., 2011;Hyvönen et al., 2005). Figure 1. Conceptual scheme of sources, redistribution, processing, and removal of volatile organic compounds (VOC), aerosol particles and cloud condensation nuclei (CCN) over the Amazon (B) in combination with previously measured vertical profiles of particle concentrations and sizes (A). The scheme emphasizes the aerosol cycling in the course of deep convection, with an upward transport of volatile organic compounds (VOC) and aerosol particles, new particle formation in the free troposphere with initial particle growth , followed by the downward transport of sub-50 nm particles into the planetary boundary layer (Krejci et al., 2003;Wang et al., 2016). The three red arrows represent the main aerosol source categories in the Amazon, which are (i) primary and secondary particle formation from local and regional biogenic sources, (ii) long-range transport of dust, smoke and other aerosols, as well as (iii) the vertical transport of sub-50 nm particles from aloft. The figure integrates concepts and data from various previous studies (i.e., Krejci et al., 2003;Baars et al., 2012;Wang et al., 2016;Williamson et al., 2019;Lauer et al., 2022).
The occurrence of NPF is dependent on the local conditions at individual sites, including meteorology, biogenic emissions, and air pollution levels, but regional and synoptic scales are also very important for this process. Particle growth events lasting on the order of hours are particularly influenced by larger geographic scales. Nieminen et al. (2018) emphasized that the NPF occurrence and growth rates (GR) show a geographically inhomogeneous distribution, indicating that the underlying mechanisms are as manifold as complex. Typical atmospheric GR ranges from 1 to 12 nm h -1 (Yli-Juuti et al., 2020). Further, different meteorological conditions have been associated with the occurrence of NPF and particle growth. Specifically, photochemical reactions under daytime conditions have been regarded as a driving force for both, nucleation and condensational growth (e.g., Nieminen et al., 2018;Kerminen et al., 2018;Hamed et al., 2011;Ma and Birmili, 2015). This is in line with a 5 significantly higher occurrence of NPF under clear sky conditions, as observed for instance in Hyytiälä, Finland, and British Columbia (Dada et al., 2017;Andreae et al., 2021). In addition, an association between the occurrence of NPF and convective clouds has been observed at different marine to continental sites (e.g., Perry and Hobbs, 1994;Clarke, 1992;Waddicor et al., 2012;De Reus et al., 2001;Wehner et al., 2015). Evidence of this phenomenon in the Amazon has also been reported by . While different potential explanations have emerged, the exact mechanisms, precursors, and the spatial 10 distribution in the context of clouds have remained unknown .
NPF has likely been altered as a result of industrialization, when anthropogenic emissions started to influence the atmospheric concentrations of trace species . Relative to remote sites, rural and urban locations tend to show higher NPF frequencies (typically 10 − 30 % event days) and higher particle GR (4 − 12 nm h -1 ) Nieminen et al., 2018). This relates to the fundamental question whether atmospheric concentrations of certain, mainly an- 15 thropogenically derived, species such as H 2 SO 4 , have to exceed certain thresholds for NPF to occur. Recent evidence of pure biogenic ion-induced nucleation under controlled laboratory (Kirkby et al., 2016) and under real atmospheric conditions Zhao et al., 2020) highlights possible mechanisms for NPF pathways in a clean atmosphere.
Accordingly, environments with low anthropogenic influence are of particular interest to investigate processes under conditions that approximate a preindustrial state of the atmosphere. Amazonia is an ideal outdoor laboratory for such investigations 20 under pristine conditions (Hamilton et al., 2014;. Of particular relevance is the wet season with its episodic occurrence of pristine periods, which allows studying atmospheric processes -such as the occurrence of sub-50 nm particles (CN <50 ) -under conditions that approximate a preindustrial state of the rain forest atmosphere (Andreae et al., 2015). At the Amazon Tall Tower Observatory (ATTO) -which is located in a mostly untouched rain forest region and has become a landmark site for atmospheric research (Andreae et al., 2015;Pöhlker et al., 2019) -March to May represent the cleanest months 25 of the year with about 10 % of the time being considered as pristine periods . Figure 1 illustrates the main sources of aerosol particles and CCN in Amazonia, which can be broadly grouped into the following three categories: 1. A spectrum of biogenic emissions from the rain forest ecosystem, including the release of primary biological aerosol particles (i.e., pollen, spores, bacteria, fragments) (e.g., Huffman et al., 2012;Löbs et al., 2020;30 Prass et al., 2021), as well as the emission of biogenic VOCs followed by their atmospheric oxidation and conversion into secondary organic aerosols (SOA) (e.g., Chen et al., 2015;Liu et al., 2016;Saturno et al., 2018a).
3. Driven by deep convective clouds, biogenic VOCs are transported into the upper troposphere (∼10 km), where VOC oxidation, nucleation of new aerosol particles, and initial particle growth occurs, fostered by low temperatures and a low 5 preexisting aerosol surface area (Krejci et al., 2003;. Subsequently, the freshly formed particles are mixed downward into the PBL, where they continue to grow to CCN-relevant sizes (e.g., Krejci et al., 2003;Wang et al., 2016).
Several studies provide experimental and modeling support for the broad atmospheric relevance of UT particle production and the subsequent vertical mixing of the CN <50 (e.g., Krejci, 2003;Krejci et al., 2005;Wang et al., 2016;Andreae et al., 10 2018; Williamson et al., 2019;Leino et al., 2019;Zhao et al., 2020;Rizzo et al., 2018;Machado et al., 2021). Figure 1 shows the presence of a CN <50 pool in the Amazonian UT during the wet and dry seasons as well as an increase in D with decreasing altitude due to condensational particle growth. The downward motion of the CN <50 can be driven by strong convective downdrafts or weaker downward motions in stratiform cloud regions (Wang et al., 2016). Zhao et al. (2020) recently suggested that pure organic NPF based on biogenic VOCs dominates above 13 km, whereas ternary NPF involving organics and 15 H 2 SO 4 dominates between 8 and 13 km. In addition, an increase of sub-50 nm particles was observed, in particular in the early morning hours, suggesting a connection between these increased concentrations with vertical transport and deep convective clouds, as well as with lightning density. Major mechanistic questions regarding the vertical transport of the CN <50 remain open (e.g., Machado et al., 2021), however. This mechanism corroborates the fact that most of the observed sub-50 nm-Aitken mode particles in the PBL have relatively larger diameters (> 20 nm) due to the aging process while transported from the free 20 troposphere into the PBL. It means that the ion-induced biogenic nucleation in the uppermost troposphere potentially plays an important role, and, therefore substantially contributes to the particle population in the free troposphere and in the PBL Wang et al. (2016); ; Glicker et al. (2019b).
This study aims to identify and characterize the occurrence of particle growth events in the size range from 10 to 50 nm within the PBL of Central Amazonia. While previous studies have documented the occurrence and properties of freshly nucleated 25 particles at high altitudes (Krejci et al., 2003;Williamson et al., 2019), their growth in the course of downward transport (Wang et al., 2016), as well as the appearance of sub-50 nm particles in the PBL Wimmer et al., 2018), major questions remain open. With this study, we take a step beyond the existing knowledge, based on more than 6 years of aerosol measurements and complementary meteorological and satellite observations. In particular, we focus on a statistically broad characterization of Amazonian particle growth events ('Amazonian bananas') by means of GR, 30 CS, seasonality, diurnal cycle, as well as their relationship to meteorological variables and deep convection. We also document growth events under clear sky conditions and thus in the absence of deep convective mixing. Therefore, the knowledge obtained here about the sub-50 nm particle growth events addresses an important gap in our understanding of the Amazonian aerosol life cycle and will help to constrain the CCN sources and properties in this globally important ecosystem.
2 Measurements and data analysis

The Amazon Tall Tower Observatory (ATTO) site
The Amazon Tall Tower Observatory (ATTO) is located 150 km northeast of Manaus, Brazil, in a forest reserve. Detailed descriptions of the site, its location, instrumentation, and scientific missions can be found elsewhere (Andreae et al., 2015;Pöhlker et al., 2019). At the ATTO site, the first aerosol measurements were initiated in 2011 (e.g., 5 Saturno et al., 2018b). Since 2014, multiple continuous measurements of physical and chemical particle properties have been established and gradually extended (e.g., Pöhlker et al., 2016Holanda et al., 2020;Saturno et al., 2018a;Schrod et al., 2020).

Terminology
According to Pöhlker et al. (2016), we define the Amazonian seasons as follows: the wet season spans from February to 10 May, followed by the wet to dry transition period (WtoD) including June and July. The dry season extends from August to November, followed by the dry to wet transition period (DtoW) including December and January. For the Amazonian submicron particle population, which is characterized by a multi-modal size distribution, we use the widely established terms Aitken mode (50 − 100 nm) and accumulation mode (100 − 1000 nm) Machado et al., 2021). In addition, we introduce the term sub-50 nm mode, defined as particles between 10 and 50 nm. We avoid using the term nucleation mode 15 for this particle population as this term is typically defined as particles < 25 nm and refers to an aerosol population relatively soon after nucleation (Kulmala et al., 2012;Nieminen et al., 2018). The sub-50 nm particles analyzed here, however, have experienced initial aging and growth to diameters between 10 and 50 nm already. As an abbreviation, we use CN <50 to refer to the particle fraction in the sub-50 nm mode. For the particle number concentrations in the individual modes, we use the symbols N <50 , N Ait , and N acc . N CN is defined as the total particle number concentration. 20

Aerosol measurements
This study focuses on particle number size distributions (PNSDs) obtained from a Scanning Mobility Particle Sizer (SMPS) with an inlet located at 60 m above ground. The inlet used to sample the aerosols is installed on an 80 m high tower (02°0 8.602'S, 59°00.033'W; 130 m a.s.l.) at the ATTO site. The SMPS is manufactured by TSI Inc., and we used as classifiers: model 3080 and, later, model 3082, coupled to a condensation particle counter (CPC) 3772. The inlet height was chosen to 25 be approximately 30 m above the average canopy height, which enables measurements close to the canopy without direct contact with the largest trees. The SMPS is located in an air-conditioned laboratory container at the foot of the mast. Sample air is transported through a 25 mm diameter stainless steel tube (finetron tubes, Dockweiler AG, Neustadt-Glewe, Germany) and dried to a relative humidity (RH) below 40 %. An automatic regenerating silica gel adsorption aerosol dryer, as described in Tuch et al. (2009) was installed upstream of the instruments in 2014 and was replaced by a custom-built and automated 30 6 condensation aerosol dryer in March 2020. For more detailed information on the aerosol measurements setup, see Andreae et al. (2015).
The SMPS measurements cover the particle size range from 10 to 400 nm and yield a temporal resolution of 5 min. The PNSD data covers more than six years, from February 2014 to September 2020, covering 1 596 measurement days, and comprising 426 272 sample runs in total. The data coverage of ∼67 % over the entire time frame (i.e., Feb 2014 to Sep 2020) can 5 be considered as a robust data foundation and statistical basis for the observations and conclusion presented here.
The sizing accuracy of the SMPS was frequently checked with monodisperse polystyrene latex particles. Additionally, the data quality was continuously verified by complementary measurements with a Condensation Particle Counter (CPC, model 5412, Grimm Aerosol Technik, Ainring, Germany) measuring the total particle number concentration (N CN ) >4 nm.
All particle data were visually inspected for malfunction and contamination, further corrected for standard temperature and CN <50 are particularly prone to diffusion losses at surfaces (e.g., the tube surfaces of the inlet lines) (von der Weiden et al., 2009). Accordingly, the generally sparse occurrence of CN <50 in the Amazon frequently raises questions about whether these results are (systematically) biased by unaccounted diffusion losses (e.g., in the 60 m long inlets). The observations outlined 15 below suggest the absence of large and unaccounted-for particle losses in the size range that is particularly relevant here (i.e., 10 to 100 nm) and further indicate that the observed PNSDs correctly reflect the actual atmospheric aerosol distribution: -The inlet and particle transport is optimized for high particle transmission efficiency and short residence time of the sample air. According to the particle loss calculator provided by von der Weiden et al. (2009) and corresponding sensitivity tests, the 50 % transmission efficiency of the inlet at the lower end of the PNSD is reached at D 50% ≈ 8 nm. All PNSDs 20 in this study have been corrected for diffusional, sedimentation, and inertial losses according to von der Weiden et al. -Experiments with the SMPS running at the 60 m inlet line and a separate and mobile CPC running without inlet lines at the height of 60 m directly on the tower agreed well, which underlines that no significant fractions of CN <50 were lost in the inlet lines.

Multi-modal log-normal fitting of PNSDs
Each measured PNSD was fitted by a multi-modal log-normal distribution function, according to Heintzenberg (1994): 30 where D p is the particle diameter and n is the number of aerosol size modes to be fitted (with n ≤ 3, see Sec. 2.2). Each mode is characterized by 3 main parameters: the mode number concentration, N i , the mode geometric median diameter D i , and the mode geometric standard deviation σ i . A script was developed -similarly to the procedure in Hussein et al. (2005) -to provide an automatic user-free decision algorithm to obtain the size modes according to the following steps: 1. In the first step, the maximum particle number concentration and the corresponding particle diameter, D dom , are determined within the particle number size distribution. Within the size range of −30 % to +20 % of D dom , a one-modal log-normal distribution is fitted. 5 2. The first one-modal fit is assigned as accumulation (D Acc ∈ [100, 300]), Aitken (D Ait ∈ [50, 100)) or sub-50 nm mode (D <50 ∈ [9, 50)), and two additional one-modal log-normal distributions are added for the remaining modes. The parameters of the three log-normal distributions are then varied within the mentioned diameter range, the standard deviation, and for concentrations less than the maximum of the particle number size distribution.
3. The geometric standard (σ i ) deviation of all modes was constrained within the range of 1.1 to 1.55, which was optimized 10 for the ATTO conditions.
4. Subsequently, a joint optimization of the previously obtained fit parameters (D i , σ i , and N i ) for the modes was conducted.
The procedure is developed by fixing two of the modes and leaving the third free so that its parameters are again optimized by minimizing the least-squares. The optimization order in this process was to optimize the sub-50 nm mode, then the accumulation mode, and, finally, the Aitken mode. In this case, all the free diameters of the modes could vary 15 between 0.5D i and 1.5D i . As a measure of fitting quality, for each particle number size distribution, the algorithm compares the particle number concentrations of each bin of the measured and the fitted curve and obtains the R 2 value.
We considered only fits in which the agreement returned R 2 > 0.8, which means that about 97 % of the data are covered by the developed mode fitting. Examples of fits can be seen in Figure S1.
5. Comparisons between the integrated particle number concentration from the SMPS measurements (N conc, SMPS ) and log-20 normal fitted size distributions (N conc, n-modes ) were made to further assure the quality of the fits. Within this data set, on average, fits with R 2 = 0.97 were obtained, which yielded a linear fit of N conc,SMPS and N conc, n-modes with R 2 = 0.99 ( Figure S2).

Identification of particle growth events
We analyzed the occurrence and properties of particle growth events ('Amazonian bananas plots') in the sub-50 nm and Aitken 25 mode size range. Characteristic examples of such growth events are shown and discussed in Sect. 3. The growth event identification is based on the following main steps: 1. All data were smoothed to eliminate single exceptionally high or low values to avoid possible bias due to short intense particle peaks or dips. The moving window has two dimensions: one in time and the other in size. The SMPS measurements last 5 minutes to get a full-size distribution. In order to reduce noise, we average the time window at 25 minutes. 30 The second parameter is the particle size window, which accounts for 5 SMPS-bins. This is also made to reduce noise in terms of particle size. These two choices were shown as ideal to get reliable data, following suggestions by Kulmala et al. (2012). We also performed several tests to verify if sudden events were missing, and found that, in general, the method could cover them without significant losses.
2. All PNSD data were divided into 24 h subsets.
3. Particle growth event days were then automatically flagged based on the guidelines in Kulmala et al. (2012). These 5 guidelines were slightly modified by increasing the size threshold for the initial growth event identification from 20 nm to 40 nm to account for the characteristics of the Amazonian banana plots and PNSDs.
4. Further following Kulmala et al. (2012), the total particle number concentration of particle diameters > 40 nm was then subtracted from the total particle number concentration of particle diameters 10 ≤ D p ≤ 40 nm. Positive values in the PNSDs are marked as regions of interest for the occurrence of CN <50 that could result in particle growth events. Days 10 fulfilling these criteria are flagged as particle growth event days.
5. This method is sensitive to the integral particle number concentration in the Aitken and accumulation modes, and their seasonal variation, which might result in false positive or false negative event flagging. To account for that, the results from the automated identification routine were visually inspected and potentially misinterpreted events were excluded from the analysis. The inspection followed the procedure described in Dal Maso (2005), in which a particle growth event 15 is characterized by i) the appearance of a distinct new mode of particles in the PNSD, ii) the particle size is inside the sub-50 nm mode, iii) the mode prevails for more than one hour, and iv) it shows signs of growth in time.

Growth rate and condensation sink
The GR and the condensation sink (CS) -both important physical parameters in the characterization of growth events -were calculated following the procedures of Dal Maso (2005) and Kulmala et al. (2012). The GR is defined as the rate at which the 20 mean geometric diameter D p of the CN <50 population changes linearly with time: where D p1 is the geometric diameter of the sub-50 nm mode obtained by the multi-modal fit at the beginning of the growth event at time t 1 and D p2 is the geometric diameter at the end of the growth event at time t 2 . Thereby, the beginning of a growth event is defined as the moment at which D p starts to increase. The end of a growth event is reached when either (i) D p 25 (10 − 50 nm) stops to grow, or (ii) the growth is interrupted due to sudden changes in air masses, or (iii) D p reaches the Aitken mode -in this case, we selected D 2 as the last observed growth D p inside the sub-50 nm mode. There were a few events in which the growth stopped for a while and, afterward, restarted again. In these cases, we considered the second growth as a new growth event. The growth events considered in this study have a duration of at least one hour.
A moving average smoothing filter was applied at the mean geometric diameter interval D p1 ≤ D p ≤ D p2 and the fit was 30 obtained by applying a linear model fit at the referred diameter interval. The model returned the following parameters: R 2 , p-value, and GR. To assure the data quality during the analyses, fits were statistically tested and only fits with R 2 > 0.6 and p-value < 0.05 were accepted. Additionally, we performed visual inspections of the quality of each of the fits. It is worth mentioning that fits statistically tested with R 2 > 0.6 were able to represent the widest possible variability of growth events, without compromising the analyses.
The CS was calculated from the particle number concentration as (Dal Maso et al., 2002): where N i is the particle concentration at the diameter D p,i of the i-esm size bin; D is the diffusion coefficient of the precursor condensable vapor, and β m is the transition-regime correction (Fuchs and Sutugin (1971)), defined as: which depends on the dimensionless Knudsen number, Kn = 2λ/D p . The Kn parameter represents the ratio of two length 10 scales, where λ is the effective mean free path of the vapor molecules in the gas (Dal Maso et al., 2002).
Physically, CS is a parameter that quantifies the ability of particles to remove condensable vapors from the atmosphere, incorporating them into the particle population and directly influencing the particle growth process. In this study, CS was calculated assuming D = 0.117 cm −2 s −1 , i.e., the value for sulfuric acid (H 2 SO 4 ) (Gong et al., 2008), which is commonly used in the literature, allowing comparisons to other studies. We used the term CS growth as the average CS during the particle 15 growth event.

Meteorological parameter measurements
The meteorological parameters, air temperature (T ), incoming shortwave radiation (SW ), rainfall (P ATTO ), air pressure (p), and relative humidity (RH) were measured at an 80 m high tower (02°08.647'S, 59°59.992'W; 130 m a.s.l.) located approximately 100 m from the ATTO aerosol mast. The measurements performed at the 80 m tower ranged from 2013 to 2018. 20 Specifically, SW and P ATTO were measured at the top of the tower, whereas T , p, and RH were measured at 55 m, 55 m, and 81 m, respectively. From January 2019 to September 2020, the meteorological parameters air temperature (T ), rainfall (P ATTO ), air pressure (p), and relative humidity (RH) were measured at the 321 m ATTO Tall Tower with a compact weather station (Lufft, WS600-LMB, G. Lufft Mess-und Regeltechnik GmbH, Fellbach, Germany). Overall, meteorological parameters span the time frame from May 2013 to September 2020. Furthermore, an optical fog sensor (OFS, Eigenbrodt GmbH, 25 Königsmoor, Germany) measured the near-field visibility since October 2015 at the height of 50 m. Fog occurrence is defined as visibility below 5000 m, which represents a threshold for light fog. Detailed information on the meteorological instruments can be found in Andreae et al. (2015).

Equivalent potential temperature
Variations of the equivalent potential temperature, θ e , have been used as a proxy to indicate downdraft occurrences (Machado et al., 2002;Betts et al., 2002;Rizzo et al., 2018;Wang et al., 2016;Gerken et al., 2016). θ e quantifies the temperature of an air parcel, when lifted to a certain height where it condenses (characterized by its lift temperature, T L ), releasing the latent heat, and lowered adiabatically to 1000 hPa. In this study, θ e was calculated from meteorological parameters measured in-situ and 5 was analyzed similarly to Wang et al. (2016) and Rizzo et al. (2018), using the definition described in Bolton (1980) as: where T k is the ambient temperature in Kelvin, p and r are the ambient pressure ( hPa) and the water mixing ratio (g/kg), 10 respectively, and T L is the lifting condensation level temperature in Kelvin. To obtain the variations in θ e , it was necessary to subtract seasonality and diurnal variations.
The steps of this process are illustrated in Figure S3 and are described as follows: θ e was obtained by subtracting the mean seasonal trend values, in which we considered both wet and dry seasons, for each year of the time series. Then, the calculated mean diurnal cycle of θ e was subtracted from θ e , at the same time of the day, resulting in a new time series: ∆θ e . The quantity 15 ∆θ e is the anomaly in θ e , and represents the deviation of θ e from its expected value for that time of the day and season. Values of ∆θ e < 0 are a proxy to the occurrence of downdrafts and indicate a decrease in θ e due to air masses from the free troposphere that enter the PBL, typically related to the occurrence of rain (Wang et al., 2016;Rizzo et al., 2018). Other processes may also be related to a decrease in θ e , such as evaporation of rainfall, river breeze, and advection mechanisms.
It should be mentioned that two time series of meteorological data were used to calculate ∆θ e : the first one, with measure-20 ments conducted close to the canopy (2013 -2018), and the second one, with measurements conducted at 325 m elevation (2019 -Sep 2020). This was necessary because meteorological data are not available at the 80 m tower for the final period of analysis. The consistency of the ∆θ e calculation was verified by comparing ∆θ e for a 2-month period with overlapping measurements at the two height levels (January and February 2019), as shown in Figure S4. Figure S5 shows the correlation between ∆θ e obtained at the two levels, with the statistical results of the comparison. Although there are very small differences 25 for single pairs of measurements, the overall agreement is reasonably good, and therefore, does not impact the conclusions of the analysis. These results encouraged us to use the meteorological data measured at 325 m height, which enabled us to extend the data analysis to the years 2019 and 2020. We can not rule out a very fine scale stratification phenomenon close to the canopy, as observed by Zha et al. (2018), which could have some influence on ∆θ e , but to a minor extent, as observed by the comparison analysis. Further studies are required to examine this aspect in detail for the ATTO site. This study uses infra-red brightness temperature (T ir) data obtained by the Geostationary Operational Environmental Satellite (GOES), GOES-16, from November 2017 to April 2020. This data set comprises a total of 914 days with measurements every 10 min, as an indication of the troposphere's meteorological conditions. The Advance Baseline Imager (ABI) -a stateof-the-art 16-band radiometer on board GOES-16 -was employed in this study, specifically, Band 13, the infrared window 5 at 10.3µm. These measurements are less sensitive than other infrared bands to gas absorption, which allows estimating the cloud-top brightness temperature. An area of 3x3 pixels centered at ATTO was selected for obtaining the time series of T ir, representing around 6.0×6.0 km 2 . Meteorological conditions representing shallow clouds/clear sky are described by a warm T ir and deep convection conditions by a cold T ir. Here, we considered T ir > 280K as a nearly clear sky condition, 245 ≤ T ir < 280K corresponding to shallow clouds and cumuliform clouds, and T ir < 245K to all convective clouds associated with deep 10 convection (Machado and Rossow, 1993;Machado et al., 2002). T ir under nearly clear sky conditions corresponds roughly to the temperature in the PBL.
3 Results and discussion

Particle number size distributions for wet and dry season
In agreement with previous studies, our long-term PNSD measurements showed the distinct characteristics of the Amazonian 15 wet and dry season aerosol populations (e.g., Roberts et al., 2001;Gunthe et al., 2009;Artaxo et al., 2013;Pöhlker et al., 2016;Rizzo et al., 2018). We chose a different representation of the typical PNSD shapes in Figure 2 by showing them as frequency distributions (FDs). The PNSDs differ significantly between both seasons: during the wet season, clear Aitken and accumulation modes stand out, separated by a distinct Hoppel minimum (Hoppel et al., 1986). On average, the Aitken mode is centered at 71 nm, the Hoppel minimum is centered at 102 nm, and the accumulation mode is centered at 153 nm. In contrast, 20 Figure 2b shows the typical dry season PNSDs characterized by a strong mono-modal shape with a dominating accumulation mode, reflecting the prevalence of biomass burning pollution (e.g., Rissler et al., 2006;Brito et al., 2014). On average, the accumulation mode is centered at 146 nm. In addition, the distribution in Figure 2b reveals a contribution of the Aitken mode, centered at 68 nm (mean) and visible as a small shoulder on the dominant accumulation mode.
Regarding the abundance of CN <50 , the FDs in Figure 2 reveal that aerosols in this size range are rather sparse -though  from Manaus (e.g., Cirino et al., 2018;Fan et al., 2018;Wimmer et al., 2018;Glicker et al., 2019a). These urban-influenced PNSDs are characterized by strongly enhanced particle concentrations below about 20 nm. As the mean and median PNSDs do not sufficiently reflect the abundance and properties of CN <50 in Figure 2, the following paragraphs summarize the in-depth analysis that allowed us to extract their event characteristics, seasonal and diurnal variability, and their estimated significance.
3.2 Particle growth event characterization 5 The abundance of CN <50 , which show up as a weak, though noticeable, mode during the wet season (see the overall outline in Figure 2a), results from the episodic occurrence of CN <50 events and their subsequent growth. Figure 3 shows a typical example of an 'Amazonian banana plot' representing two subsequent growth events, as frequently observed at ATTO. The first example in Figure 3 starts in the morning hours around 08:30 local time (LT) with an average initial diameter slightly larger than 30 nm. The particles grow for about four hours reaching the Aitken mode size range up to ∼ 60 nm. On the same day, a 10 second growth event starts around noon with an average initial growth diameter slightly larger than 20 nm, growing during the afternoon hours. The initial diameters at the onset of the growth events in Figure 3 are well above the lower size limit of the SMPS (i.e., 10 nm), which implies that the event characterization is not distorted or limited by the effectively measured size range.
The growth events shown here resemble the events reported by Wang et al. (2016). Note that in all previous studies in 15 Amazonia, the growth events were observed during the wet season, suggesting that this event type is a typical wet season The GR frequency distribution in Figure 4a shows a clear peak centered around the median of 5.2 nm h -1 . The median GR of this study agrees well with the median GR of 5.5 nm h -1 obtained by Rizzo et al. (2018). Figure S6a contrasts the median GR obtained at different sites in Amazonia and worldwide and shows that the median GR from this study is within the GR ranges obtained at remote boreal (0.5 to 5.3 nm h -1 ) or polar sites (0.2 to 5.5 nm h -1 ). The CS growth frequency distribution in Figure 4b 10 shows a clear peak centered around the median of 1.1×10 −3 s -1 . In contrast, the median CS calculated for all observation days is 3.2×10 −3 s -1 , which corroborates that particle growth events at remote sites are expected when CS values are low ( Figure 6d). The Amazonian CS also agrees with what is observed in other remote regions (Figure S6 b). For example, boreal sites have average CS ranging from 9.8 ×10 −4 to 3.9×10 −3 s -1 . Figure 5a shows the FD of PNSDs exclusively for the periods of growth events, starting 3 hours before the event's onset 15 and lasting until the time when growth stopped being observed (according to Sec. 2.6). Figure 5b shows the separated CN <50 , Aitken and accumulation modes for the median SD of Figure 5a. The characteristic multi-modal shape of the wet season PNSDs stands out. The FD for the growth events further underlines the sparse particle abundance below 20 nm. Figure 5b shows that the Aitken and accumulation modes are the dominant modes of the median PNSD. The CN <50 mode, although small and not readily perceptible from the median PNSD, has a significant contribution, however, and is centered at D CN<50 = 37 nm (mean), with σ CN<50 = 1.5. For comparison, the diameters selected as the initial values for the growth events, D p,i , have a median value shows that the initial diameter of growth events in the Amazon is typically larger than reported in other regions (Nieminen et al. (2018)).  Figure 6a shows the typical seasonality of precipitation at ATTO, P ATTO . The highest rain rate occurs during the wet season, with P ATTO peaking in March and April, while the minimum in P ATTO occurs between July to September. A similarly pronounced seasonality can be found in various aerosol properties. The N CN in the size range between 10 to 400 nm had its 15 minimum in the wet season months March and April with a median of ∼280 cm -3 and its maximum in the dry season months August to November with a median of ∼1400 cm -3 (Fig. 6c). The same pattern can be found in the monthly median CS in Figure 6d. The strong seasonal differences in the physical aerosol properties -here manifested in a wide range of N CN and different PNSD shapes -have a substantial influence on the concentration of available cloud condensation nuclei (CCN) and, thus, cloud micro-physical processes in the Amazon Basin Lauer et al., 2022). These results agree well with long-term measurements at another central Amazonian site (i.e., the ZF2 site) presented by Rizzo et al. (2018) and allow to put earlier campaign-wise measurements into 5 a broader context (e.g., Roberts et al., 2001;Roberts, 2003;Zhou, 2002;Rissler et al., 2004Rissler et al., , 2006Martin et al., 2010). Figure 6e shows the seasonal pattern in the CN <50 growth event frequency. We found the highest frequencies during the wet season, peaking in April with about 26 %, and dropping down during the transition period (WtoD, Jun and Jul) to a minimum with almost zero events in August. The growth event occurrence stayed remarkably low during the dry season months with frequencies mostly below 3 % from July to November. Frequency levels increase again during the transition period (DtoW, Dec   10 and Jan). The seasonality in growth event occurrence corresponds well with the seasonality in monthly rainfall and appears inversely related to the seasonality in N CN and CS. This agrees with previous studies in the Amazon, suggesting a close link between generally low particle concentrations and the appearance of CN <50 and Aitken mode particles in the PBL (e.g., Krejci, 2003;Wang et al., 2016). The low CS might further favor the characteristic growth patterns of these events through the condensation of semi-and low-volatile gaseous compounds on the CN <50 particle fraction as available surfaces (see example in The N CN , CS, and growth event occurrence in Figure 6 are all based on the same multi-year SMPS data set. The underlying data availability is documented in Figure 6b as the number of valid measurement days. To provide a seasonal overview, Tab. 1 shows statistical information on GR and CS growth during the observed growth events, separated by the month of the year. In the wet season, when ∼88 % of all growth events occurred, the GR values fluctuate around the median GR of 5.2 nm h -1 .

Seasonality
The CS growth shows a similar behavior, fluctuating around 1.1×10 −3 s -1 . The lowest median GR of the wet season is found in March, with 3.7 nm h -1 , while the lowest CS growth of the wet season is found in April, with 8.0×10 −4 s -1 . The number of particle growth occurrences drops in the dry season, comprising only ∼12 % of the total growth events. During December, in the transition from dry to wet season, the percentage of growth events starts to increase to ∼6 %. The average CS growth (during growth events, Tab. 1) in the dry season is lower compared to the average monthly CS (cf. Fig. 6) indicating that growth events 5 occur on days with a cleaner atmosphere compared average dry season conditions. The only exceptions are the events from October, whose median CS growth was 4.4×10 −3 s -1 . As outlined above, our results indicate that the occurrence of particle growth events is linked to the pronounced seasonality of the aerosol properties and the related underlying meteorological conditions and prevailing air mass history. In the dry season, regional and long-range transported biomass burning aerosol from the Southern Hemisphere dominates the aerosol population 10 in the lower troposphere (e.g., Moran-Zuloaga et al., 2018;Saturno et al., 2018b;Holanda et al., 2020). The resulting strong accumulation mode and, hence, high CS likely suppresses the growth of the smaller particles in the sub-50 nm mode. With the increase in precipitation (thus, in wet deposition) and the change in the average air mass history towards northern hemispheric origins, less polluted air masses arrive at the ATTO site (e.g., Moran-Zuloaga et al., 2018;. The number of particles of the accumulation mode and CS decreases significantly coinciding with an increase in the occurrence of growth 15 events.

Diurnal trends
The diurnal patterns of the growth event occurrence are shown in Fig. 7 in relation to meteorological parameters, such as air temperature (T ), SW , P ATTO , RH, and near-field visibility representing fog. Note that we contrasted the diurnal cycles for the entire observation period (i.e., Feb 2014 to Sep 2020, shown as solid lines) and the wet season months (i.e., Feb to May, shown as dashed lines), since the majority of the growth events are observed during the wet season. The meteorological variables, T , 5 SW , RH, and visibility were not discriminated with respect to days with and without events, because no significant differences between event and non-event days were observed. Possible effects of deep convection, associated rainfall, and cloudiness are investigated in the following Sec. 3.5. In the diurnal cycles considering the whole period of observations, T and SW show the typical tropical rain forest conditions with about 12 h daylight and a remarkably low amplitude in T , spanning on average only 5 • C. Rainfall is most intense in the afternoon hours, with the highest precipitation intensity at 15:00 LT. The data also show a 10 secondary maximum in the early morning, which has been associated with nocturnal long-lived mesoscale systems (Machado et al., 2021). The RH levels reach on average 100 % during the early morning and decrease during the day to around 75 %.
Fog typically occurs in the second half of the night and often in the early morning before sunrise (i.e., between 03:00 and 07:00 LT), when T is lowest. Sporadically, fog also occurs shortly after rain showers, which is not reflected in the average conditions. Figure 7e shows the diurnal cycle of the median CN <50 number concentration, N <50 , during the particle growth event days 15 (blue line) and for comparison, the median diurnal cycle of N <50 comprising all measured PNSDs (black line). The particle concentration on growth event days is somewhat higher than that including all analyzed PNSDs, with median daily values and interquartile range of 64 (38 -108) cm −3 , compared to 49 (29 -81) cm −3 for all days. The diurnal cycles of both, the particle concentration during growth events and for the entire measurement period show also some similarities. Around 9:00 LT in the morning, there is a remarkable decrease until noon, which is followed by an increase towards midnight. The decrease starting 20 in the late morning is likely due to increased mixing after the breakup of the nocturnal boundary layer and the development of a well-mixed boundary layer. Towards the late afternoon, the well mixed layer is then shrinking due to decreasing solar radiation. The N <50 diurnal cycle for growth event days shows further a strong increase from midnight to 09:00 LT, peaking at N <50 = 88 cm −3 , compared to N <50 = 56 cm −3 at the same time for the total data. At this time, the difference between the growth event particle concentration and the average is largest and coincides with the highest occurrence of growth events. 25 The diurnal cycle of the growth event onsets has a rather broad maximum in the early morning hours from 06:00 to 10:00 LT.
It peaks at about 07:00 LT and then gradually decreases towards noon (see Fig. 7f), which is in agreement with what is observed in Figure 7e. In addition to the PBL development, particles are subject to atmospheric aging and likely condensation of semiand low-volatile compounds resulting in particle growth and a decrease in particle number concentration. It is interesting to note that the morning growth event maximum coincides with a maximum in RH and the occurrence of fog (cf. Fig. 7c,d). A second 30 local and less pronounced maximum is visible from 13:00 to 15:00 LT. The growth events reported during daytime likely correspond to rainfall events as reported by Machado et al. (2021) and probably the vertical transport of CN <50 and Aitken size particles due to strong downdrafts in the course of convective rainfall, and the injection of these particle populations into the PBL as reported in Wang et al. (2016) and . incoming shortwave radiation (SW , orange) and air temperature (T , red) at 26 m (canopy level) height. b) average local rain rate (PATTO, blue). c) relative humidity (RH) at 26 m. d) visibility (blue) measured by a fog monitor and fog occurrence (black) with lines representing the first quartile, e) median diurnal cycle of N<50 for all data (black) and only for the days in which particle growth events were observed (blue). And f) the diurnal cycle of the particle growth event onsets. All error bars denote the standard error. The data shown represent all observations (full lines) and wet season subsets (dashed lines).
P ATTO shows two maxima: A pronounced and rather defined maximum in the early morning at around 07:00 LT, which follows a gradual increase in precipitation during the second half of the night, and a broader maximum during the afternoon hours between 13:00 and 17:00 LT. Although about 74 % of the particle growth events occur during the day, there are still ∼26 % that take place during night conditions, between 19:00 -05:00 LT. In particular, the occurrence of growth events from 01:00 -05:00, which represents about 16 % of the total observed events, is evidence for complexity in the causes and mechanisms of particle injection and growth.
The evolution of the PBL also has a strong influence on the diurnal pattern. At night, the nocturnal PBL close to the forest 5 canopy is decoupled from the residual layer above (Fisch et al., 2004). In the morning hours -as soon as convection becomes effective -air masses transported into and within the residual layer are mixed into lower levels and measured at the canopy level. Consequently, CN <50 and Aitken mode particles advected with the residual layer will be mixed downwards and appear at the 60 m inlet in the morning hours, typically around 8:00 LT. This behavior is in agreement with that observed in Figure 7e, with the increase of CN <50 throughout the night and in the early morning. Machado et al. (2021) discuss this daily mechanism 10 of particle growth in more detail. Section 3.5 further discusses the meteorological conditions regarding convective downdrafts and the atmospheric conditions under which the growth events are observed.
A contrast in GR and CS is observed when day and night events are compared, as shown in Figure 8. Daytime events, which correspond to ∼74 % of the events, have significantly higher GR and CS growth , at 5.9 nm h -1 and 1.2×10 −3 s -1 , respectively.
The nighttime events, which account for ∼26 %, have GR and CS growth of 4.0 nm h -1 and 9×10 −4 s -1 , respectively. To verify the 15 statistical significance of the difference between day and nighttime values, the Wilcoxon rank-sum test was applied. The p-value obtained for GR is 3.6× 10 −6 , while the p-value obtained for CS growth is 2×10 −2 , indicating that the data groups regarding day and night are statistically different considering a significance level of 5×10 −2 . The observed differences are likely due to the different atmospheric mechanisms during daytime and nighttime. Figures 7a and 7f clearly indicate an increase of growth events with the increase in the intensity of solar radiation after sunrise. With increasing solar radiation, atmospheric aging 20 and oxidation of possible precursor intensifies. Sunlight also drives the dynamics of the PBL. Other phenomena also play an important role in daytime events, such as the peaks of precipitation that coincide with the peaks of growth events, which is discussed in more detail in Sec. 3.5. In contrast, the occurrence of nocturnal events may be related to different mechanisms such as local meteorological phenomena, horizontal advection, entrainment of air and particles from the free atmosphere into the PBL, and perhaps the contribution of biogenic sources from the surface.

Meteorological and cloud conditions during growth events
This section explores the weather and convective transport conditions linked to particle growth events. Figure 9a shows a histogram of the growth event frequency as a function of the associated anomaly of the equivalent potential temperature (∆θ e ) at the onset of the growth events (see Sec. 2.8). A negative ∆θ e is an indicator for air mass downdrafts from higher altitudes, as has been shown for the events analyzed by Wang et al. (2016). We found that ∼63 % of events were likely associated 30 with air mass downdrafts (∆θ e < 0K), whereas ∼37 % were associated with ∆θ e > 0K. Figure 9b shows a histogram of all growth events as a function of the associated cloud brightness temperature, which is an indicator for deep convective clouds (T ir < 245 K) vs. clear sky/shallow cloud conditions (T ir > 280 K). For all events after 2017 (when T ir data is available), we found that ∼36 % were likely associated with deep convective clouds (T ir < 245 K, red), ∼26 % with clear sky/shallow clouds (T ir > 280 K, blue), and ∼38 % with mixed sky conditions (245 K < T ir < 280 K, gray).
The most contrasting T ir conditions were observed between the lower quartile (T ir < 228 K), representing deep convective cloud conditions, and the upper quartile (T ir > 281 K), representing clear sky conditions, with 36 events contained in each group. As an example, four of these 'extreme' events were selected ( Figure 10) for daytime and nighttime conditions. The 5 growth events under clear sky conditions are characterized by a trimodal aerosol population, with accumulation and Aitken modes as well as a third mode below 50 nm. As an example, the event on 14 March 2019 (daytime, clear sky) was characterized by the occurrence of a significant amount of CN <50 . The particle concentration for particles smaller than 50 nm increased already during nighttime (starting after 02:00 LT) and resulted in a pronounced peak around 08:00 LT. While during the late night and early morning, the mean modal diameter for these small particles is rather constant, particle growth started around 10 08:00 LT and lasted for about 10 h. These events could be associated with advection processes, e.g., by a downdraft in the gust front (clear sky nighttime) or by nighttime rainfall. Afterwards, subsequent growth begins in the early morning coinciding with the evolution of the PBL.
The events under deep convective conditions -both, during daytime and nighttime -resemble the events reported by Wang et al. (2016). Here, downdrafts transport air masses into the lower atmosphere and, hence, inject CN <50 upper tropospheric 15 particles into the PBL, followed by particle growth into the Aitken mode. In both these cases, the atmosphere is very clean, with low concentrations in the accumulation mode. Before the growth event, most of the particle population is in the Aitken size mode, which is removed by the injection of upper or mid-tropospheric air during the downdraft event, so that only the CN <50 aerosol population remained. Note that about 4 hours after the start of the growth event, the accumulation mode particles (re)appeared, probably by mixing with surrounding airmasses. in lower CS growth values, and the presence of clouds reduces the availability of sunlight and thereby suppresses photochemical production of condensable species.
To further investigate typical conditions or processes related to the observed growth events, we here discriminate between different groups based on their daily frequency distribution. The growth events (see Figure 7f) were divided into four groups (G1 to G4), where G1 and G4 represent nighttime and G2 and G3 represent daytime events. The daytime events were divided 5 considering the occurrence of two frequency peaks: the first peak, representing 53 % of the growth events, is included in G2, covering the time from 06:00 to 11:59 LT. The afternoon increase, representing 21 % of the growth events, is included in G3, covering the time from 12:00 to 17:59 LT. The nocturnal events were divided according to the evolution stage of the PBL.
Events between 18:00 and 00:59 (G4), with 10 % of the events are still influenced by some PBL turbulence, but may also have had some influence from convective events in the late afternoon. The growth events between 01:00 and 05:59 (G1), with 16 % Figure 11. Boxplot of a) GR and b) CSgrowth for clear sky conditions (T ir > 281 K) vs deep convection conditions(T ir < 228 K). Boxes represent the quartiles, whiskers represent 90th and 10th percentiles, and the horizontal line represents the median. PBL by, e.g., intermittent turbulence (Dias-Júnior et al., 2017), nighttime rainfall events, or even an unknown biogenic source could play a role in the aerosol particle dynamics of in this time period.
The results indicate differences between the four groups (G1 to G4) regarding the T ir and ∆θ e conditions during the event onset. The nocturnal groups G1 and G4 have median ∆θ e of -0.8 and 0 K, respectively. Their median T ir indicates conditions closer to low clouds and clear skies, with median values equal to 269 and 274 K for G1 and G4, respectively. Figure 12 shows 5 the average (mean and standard deviation) T ir for the entire event days separated for the different groups of events (blue). The T ir diurnal cycle shows clear differences between the groups. Within the G1 period, T ir shows a minimum at 04:00 LT, when the average brightness temperature reaches 256 K, indicating convective activity and early precipitation compared to all days. The G4 group, with the smallest number of growth events, is not significantly different from the median diurnal cycle for all days, suggesting that this 10 % of growth cases appears not be related to specific meteorological events. There are signs of convection at 16 LT, which coincides with the precipitation peak in the afternoon. Afterward, T ir increases, going to clear sky conditions during the night. The median GR for G1 and G4 varies from 3.8 to 4.2 nm h -1 , respectively, and CS growth is approximately constant around 1.0×10 −3 s -1 , while the median D p,i is similar for both nocturnal groups. The daytime groups 5 G2 and G3 have the lowest median ∆θ e at the onset of the events, with -1.9 and -3.6 K, respectively, indicating that convective downdraft activity plays an essential role during these growth events. For these two groups, the diurnal cycle of T ir presents lower values during the whole day compared to the overall average, and the group G2 also shows minima in the early morning coinciding with the precipitation peaks, and also in the afternoon (Figure 7b). The G3 group is the one with the most convective characteristics. The T ir values indicate an occurrence of strong convective systems throughout the day, mainly in the afternoon. In particular, the pronounced decrease at 15:00 LT reaches deep convection conditions, with T ir = 249 K, approximately 1 hour earlier than expected considering the entire observation days. The median GR ranges from 6.3 to 5.6 nm h -1 in G2 and G3, respectively, and the median CS growth remains similar for both groups, at 1.2×10 −3 s -1 . The smaller value of GR in G3 compared to G2 (diurnal events) could be associated with deep convection 5 systems, which reduce the solar irradiance and thereby influence the photochemical processes. The presence of clouds has been associated with lowering GR and even lower occurrences of NPF and particle growth events (Dada et al., 2017;Kerminen et al., 2018). The median D p,i for G3 is the lowest of the four groups, at 24.6 nm, indicating that the strong convective downdrafts are more effective in transporting smaller particles from the free troposphere into the PBL, which agrees with what has been observed previously (Wang et al., 2016). Therefore, the growth events of groups G2 and G3 are probably influenced 10 by the strong convective systems during daytime. In particular, the G3 group has the most significant characteristics of deep convection and intense occurrences of downdraft throughout the day.
When discriminating the growth events by positive or negative ∆θ e at the event's onset (here defined as ∆t = 0) and looking 10 hours before and after this time, the different behavior of ∆θ e and T ir near the growth event is evident. Figure 13 shows the behavior of the mean ensembles of ∆θ e and T ir around ∆t = 0 for ∆θ e less than the 25th percentile (-5.3 K) and higher 15 than the 75th percentile (+1.5 K) at the event's onset. Considering the case where ∆θ e < -5.3 K at the event's onset, both ∆θ e and T ir strongly decrease from 10 h before the event and reach a minimum at ∆t = 0, where ∆θ e = −7.6 K and T ir = 238 K, which represents deep convection conditions with strong downdraft occurrence. Both parameters increase afterwards to cleaner sky conditions and out-of-downdraft conditions.
In contrast, for ∆θ e > +1.5K at the event's onset, the ensembles show an opposite behavior of the parameters. The ∆θ e presents an increasing tendency from 10 h before the event until ∆t = 0, reaching a value of +4 K. Afterward, ∆θ e decreases, but always retains positive values, indicating that there is a class of growth events that may not be driven by convective 5 downdrafts. The T ir values from 10 h before the events up to ∆t = 0 are equivalent to conditions close to clear skies (Tir ∼280 K), which agrees with the results obtained for ∆θ e , as an indication of sky conditions not dominated by convection systems around the event's onset. After ∆t = 0, T ir values decrease, but without presenting deep convection conditions. This shows that clear sky events can be associated either with advection or subsidence bringing particulates from another area (e.g., a nearby rain event) or from the upper troposphere.

Summary and conclusions
This study reports the statistical characterization of aerosol particle growth events in the sub-50 nm size range (10 -50 nm), based on continuous measurements (February 2014 to September 2020) of PNSDs at a remote site in Central Amazonia. In total, 254 particle growth events were detected, comprising about 14 % of the analyzed days. Of all events, 88 % were found between January and June and 12 % between July and December. The diurnal cycle of the growth events shows that most of 15 them occur during the daytime, accounting for 74 %, which leaves still a significant fraction of 26 % occurring during the night.
During daytime, 53 % of the growth events start in the morning between 06:00 and 11:00 00 LT, with a pronounced peak at 07:00 LT, showing a relation to the photochemistry and links with the evolution of the PBL. These events also coincide with a precipitation peak in the morning. The nocturnal increase of N <50 is likely related to convective systems that result in the precipitation peak in the morning. The subsequent decrease of N <50 is likely due to condensation of semi-and less-volatile 20 organic species on the sub-50 nm particles, resulting in growth. A second, less pronounced but significant peak occurs around 15 h LT, coinciding with the strongest precipitation peak, also suggesting a relation to atmospheric convective systems.
The median GR, considering all the growth events, is 5.2 nm h -1 , which agrees with what was reported by Rizzo et al. (2018).
The median CS growth is 1.1×10 −3 s -1 . Monthly variations in GR and CS show that during the wet season the growth events occur under low CS values, although the average CS growth does not change much from month to month, oscillating around 25 1.1×10 −3 s -1 . A remarkable contrast is observed when comparing daytime (median GR of 5.9 nm h -1 and median CS growth of 1.2×10 −3 s -1 ) and nighttime growth events (median GR of 4.0 nm h -1 and median CS growth of 9×10 −4 s -1 ).
Daytime events are directly influenced by sunlight, which controls photochemistry and hence the oxidation of SOA precursors. In contrast, nocturnal events may have different causes and mechanisms. One particular mechanism for nighttime growth events could be supported both by continued deposition of condensables formed during the day and by production of condens-30 ables at night by ozonolysis reactions. However, the direct influence of meteorology, entrainment of air masses, and perhaps the contribution of biogenic sources can not be ruled out. Upcoming flight and in-situ campaigns are expected to provide new important insights on the main drivers of the sources for the different particle growth events. It is worth noting that, for primary biogenic emissions, there is still the need to identify the particle sources, although earlier studies have shown that the growth of secondary aerosol particles can be initiated by biogenically emitted potassium-salt-rich particles . Also, a primary source close in the canopy would be fundamental for the aerosol particle maintenance in the PBL ).
An analysis performed using ∆θ e and T ir revealed that diverse atmospheric dynamics play different roles during particle growth event days. Many event onsets coincide with downdraft occurrences, when N <50 appear and grow afterward. We also 5 observed that accumulation mode particles processed in clouds appear sporadically, causing a prominent Hoppel minimum.
The growth events occurring under clear skies present GR and CS higher than those related to deep convection: the median GR and CS for clear sky conditions are 7.0 nm h -1 and 1.6×10 −3 s -1 , whereas under deep convection conditions, the median GR and CS are 3.8 nm h -1 and 5×10 −4 s -1 .
The events were further classified according to their frequency of occurrence throughout the day, showing that they are 10 mostly driven by local convective activities (73 %). However, when analyzing the growth events by ∆θ e at the event onset, the occurrence of downdrafts does not explain all the cases. The contrast is more evident in the ensembles of ∆θ e and T ir when growth events and their respective occurrence days are selected by ∆θ e > 75th percentile (∆θ e =+1.5 K) at the event onset. For these events, representing about 27 % of the growth events, ∆θ e is maximum and positive at time ∆t = 0, and even ten hours before or after the growth event, it did not present negative values that could indicate the occurrence of convective downdrafts. 15 Also, T ir fluctuates at around 270 K over the observed period, which represents shallow clouds, instead of convective systems conditions. In contrast, events with ∆θ e < 25th percentile (∆θ e =-5.3 K) at the event onset are associated with downdraft occurrences: ∆θ e strongly decreases 10 hours after the growth event up to ∆t = 0, which is also followed by a strong decrease in T ir.
Sources that could explain growth events in the absence of deep convection are perhaps related to primary biogenic aerosols 20 emitted by the forest, smooth entrainment of air masses from the free troposphere into the PBL in the early morning, or even different meteorological mechanisms such as gravity waves and particle production by lightning in the free troposphere, as reported by Machado et al. (2021). Another possible explanation is related to nighttime downdrafts far upwind, which get trapped above the nocturnal boundary layer, and travel in the jet above the nocturnal inversion for potentially quite a large distance, being mixed down into the PBL after sunrise, as suggested by Krejci et al. (2005). Therefore future studies are required 25 to unveil the aerosol sources that could explain the diversity of particle growth events observed in the lower troposphere over Central Amazônia.
Data availability. The data of the key results presented here have been deposited in supplementary data files for use in follow-up studies. Particle number size distributions for the entire observation period are available via the ATTO data portal through https://www.attoproject.org/.
For data requests beyond the available data, please refer to the corresponding authors.

30
Author contributions. MAF and FD contributed equally to this work. MAF, FD, and CP designed the study. MAF and FD analyzed the data.
MAF, FD, LAK, BAH, FGM, JS, SC, JB, and SW collected and processed the ATTO aerosol data. LATM processed the satellte data. AA and MS collected the micrometeorological data at the INSTANT tower at ATTO. SW, FGM, PA, MOA and UP provided essential scientific support for the ATTO measurements. DW supported the data management. MAF, CP, and FD wrote the paper. LATM, MOA, LVR, HMJB, JPN, FGM, MP, SC, JB, and SC contributed with valuable ideas and comments to the analysis and the manuscript. All authors contributed to 5 the discussion of the results as well as the finalization of the manuscript. PA and CP supervised the study.
Competing interests. The authors declare that they have no known competing interests that could have influenced the work reported in this paper.
Disclaimer. This Rabelo Costa for technical, logistical, and scientific support within the ATTO project.