Elemental Mixing State of Aerosol Particles Collected in Central Amazonia during GoAmazon2014/15

Two complementary techniques, Scanning Transmission X-ray Microscopy/Near Edge Fine Structure spectroscopy (STXM/NEXAFS) and Scanning Electron Microscopy/Energy Dispersive X-ray spectroscopy (SEM/EDX), have been quantitatively combined to characterize individual atmospheric particles. This pair of techniques was applied to particle samples at three sampling sites (ATTO, ZF2, and T3) in the Amazon basin as part of the Observations and Modeling of the Green Ocean Amazon (GoAmazon2014/5) field campaign during the dry season of 2014. The combined data was subjected to k-means clustering using mass fractions of the following elements: C, N, O, Na, Mg, P, S, Cl, K, Ca, Mn, Fe, Ni, and Zn. Cluster analysis identified 12 particle types across different sampling sites and particle sizes. Samples from the remote Amazon Tall Tower Observatory (ATTO, also T0a) exhibited less cluster variety and fewer anthropogenic clusters than samples collected at the sites nearer to the Manaus metropolitan region, ZF2 (also T0t) or T3. Samples from the ZF2 site contained aged/anthropogenic clusters not readily explained by transport from ATTO or Manaus, possibly suggesting the effects of long range atmospheric transport or other local aerosol sources present during sampling. In addition, this data set allowed for recently established diversity parameters to be calculated. All sample periods had high mixing state indices (χ) that were >0.8. Two individual particle diversity (Di) populations were observed, with particles 0.5 µm particles having a Di of ~3.6, which likely correspond to fresh and aged aerosols, respectively. The diversity parameters determined by the quantitative method presented here will serve to aid in the accurate representation of aerosol mixing state, source apportionment, and aging in both less polluted and more developed environments in the Amazon Basin.


147
At the three sampling sites, atmospheric particle samples were collected on silicon nitride (Si3N4) 148 membranes overlaid on a 5 x 5 mm silicon chip frame with a central 0.5 x 0.5 mm window (100 nm thick 149 membrane, Silson Inc.). Samples were collected using a Micro-Orifice Uniform Deposit Impactor 150 (MOUDI, MSP MOUDI-110) on the dates and times shown in Table 1. HYSPLIT back trajectories were 151 examined for each sampling period to confirm the wind patterns seen in Figure 1. These samples were 152 then analyzed sequentially with the two spectromicroscopy techniques discussed in the following 153 sections.

159
The energy range of this STXM (200-600 eV) end station enables the quantitative study of carbon, 160 nitrogen, and oxygen. Energy selected soft X-rays were focused down to a ~30 nm spot size and directed 161 onto the sample surface. After a suitable 15x15 µm region was located, the sample stage was then raster 162 scanned, with 40 nm steps, using piezo-electric stages to capture an image at a specific energy. This 163 process was then repeated at multiple photon energies to produce a stack of images with an absorption 164 spectrum associated with each 40x40 nm pixel. For each element, photon energies were chosen before 165 and after the k-shell absorption edge: 278 and 320 eV for carbon, 400 and 430 eV for nitrogen, and 525 166 and 550 eV for oxygen [45]. Additional images were also taken near the carbon edge at 285.4 and 288.5 167 eV, for the RC=CR and RCOOR C1s→π* transitions respectively, in order to partly characterize the 168 molecular speciation of carbon [46].

169
Any displacement between images within a stack is corrected by a routine based on Guizar-Sicarios' 170 image registration algorithm [47]. Regions within a given stack were then identified as particles or 171 substrate using Otsu's method on that stack's average intensity image over all 8 energies [48].

172
Background subtraction of a given element's pre-edge intensity image from its post-edge image is then 173 performed to account for any absorbing species not attributed to that element.

174
The recorded intensity at each pixel determined to be a particle was converted to optical density 175 using: where OD is optical density, I is intensity of the pixel, and Io is the background intensity. This is followed 177 by a conversion to mass with the following formula: where m is the mass of a specific element at that pixel, A is the area of that pixel, and µpre and µpost refer to 179 the mass absorption coefficients for that specific element before and after the absorption edge, 180 respectively. Mass absorption coefficients have been both empirically and theoretically determined for a 181 variety of elements as tabulated in Henke et al., 1993 [45].

182
Previously developed algorithms for determining the speciation of carbon using 278, 285.4, 288.5, 183 and 320 eV were applied to each Field Of View (FOV) as well. This mapping technique uses a series of 184 thresholds to identify inorganics, soot, and organic carbon. Total carbon is taken to be OD320 -OD278, 185 pixels with an OD278/OD320 ratio 0.5 or greater are rich in inorganics, and pixels with an elevated (0.35) 186 ratio of sp 2 bonding compared to total carbon (OD288.5 -OD278)/(OD320 -OD278) are indicative of soot [46].

188
The same sample windows previously imaged with STXM were imaged again with a computer 189 controlled scanning electron microscope (FEI, Quanta 3D FEG) coupled with energy dispersive X-ray spectroscopy (CCSEM/EDX). The SEM utilized a field emission tip to produce an electron beam which 191 was directed and focused onto a sample with an accelerating voltage of 20 kV which can cause core shell 192 atomic electrons to be ejected from the sample. Higher shell electrons then relax into the newly created 193 orbital hole, releasing an elementally characteristic photon recorded by an energy dispersive X-ray 194 detector (EDAX PV7761/54 ME with Si(Li) detector). As the electron beam was scanned over the sample, 195 the transmitted electron image was used to identify the exact same FOVs from the previous STXM 196 images. Once a FOV previously analyzed with STXM is located, a 10,000x image (30 nm/pixel resolution)

197
was captured. This image combines both transmitted and backscattered electron images to improve 198 particle detection [24]. A threshold contrast level was then set to identify which areas of the collected 199 image counted as particles using the "Genesis" software from EDAX, Inc. A software filter was then 200 applied which discounts particles that are too small (e.g. noise spikes) or too large (e.g. multiple nearby 201 particles counted as a single large particle). The electron beam was then directed towards each identified 202 particle in sequence and an EDX spectrum was collected. Afterwards software was used to fit the peaks 203 of eleven relevant elements selected for this study: Na, Mg, P, S, Cl, K, Ca, Mn, Fe, Ni, and Zn. Some 204 elements of interest have been included in the spectral fit, but omitted from quantitation, including Al, Si,

205
and Cu due to background sources of these elements: 1) the STXM sample holder where the Si3N4 206 windows sat was made of Al and was inserted into the SEM as well, 2) the mounting stage that holds 207 samples inside the microscope was fabricated from beryllium-copper alloy, 3) the EDX data was collected 208 using a Si(Li) detector with a 10 mm 2 active area. Each of these circumstances could contribute 209 background signal for the elements in question.

210
After data has been collected from both SEM and STXM, individual particle mass information is

216
Using the aforementioned methods, STXM yields quantitative, absolute mass information on a sub-217 particle basis. SEM/EDX is more limited in this aspect, being quantitative for elements with Z>11 (Na) 218 but only semi-quantitative for C, N, and O [24]. Due to the EDAX software used for EDX data collection 219 and analysis, there is an additional caveat to the quantitation of Z>11 elements: the software reports only 220 the relative mass percentages compared to the elements chosen during data processing. In order to 221 properly quantify the mixing state, the absolute mass of each element in each particle is necessary. To 222 determine these absolute masses, a system of equations was set up using the following equation types: For each pixel, ODi is the optical density taken at energy i, is the density and t is the thickness of the

243
The absolute mass of a given component , within a given particle , is labeled as ! ! where = 1, ..., A

244
(and A is the total number of components) and = 1, ..., N (the total number of particles). From this, the 245 following relationships are established: Mass fractions are then established from these relationships with: Where ! is the mass fraction of a particle within a sample, ! is the mass fraction of component a within a 248 sample, and ! ! is the mass fraction of component a within particle i.

249
These mass fractions are used to calculate the Shannon entropy (also called information entropy) for 250 each particle, each component, and for the bulk using Equations 10, 11, and 12 respectively.
Each type of mass fraction can be thought of as a probability, and thus the collection of mass The diversity values contain the same type of information, but represent it in another way. Each 257 diversity value represents the effective number of species (weighted by mass) within a given population 258 (i.e. Di represents the number of species within a specific particle, Dα is the average number of species 259 within any given particle, and Dγ represents the number of species within the entire sample). From these 260 diversity values the mixing state index is defined as This definition compares how many species exist, on average, within individual particles, with the 262 total number of species identified in the sample. χ is at a minimum of 0 when Dα is 1, corresponding to 263 each particle being comprised of exactly one species. A mixing state index of 1 occurs when Dα and Dγ 264 are equal, meaning that each particle has the same composition as the bulk sample.
where ! is the mean mass of the a th component and is the mean mass of particles within the sample.

275
From this, the standard error (for a 95% confidence level) can be determined for ! and which is then 276 propagated through Equations 9-13.

277
The statistical uncertainty in Dα was found by first rearranging and combining Equations 11 and 13:

301
Most values listed for the 12/Sept and 13/Sept sample at T3 are consistent with their sample-period 302 monthly averages, even considering time of day each sample was collected. The data from 14/Sept, 303 however, shows a marked increase in particle concentration, nitrate, organic, and CO concentration,

304
along with a small increase in BC. This is indicative of a heavy pollution plume which, in this case, had 305 recently passed over the T3 sampling site (see Figure S2). AMS and particle concentration data for the T3 306 sites show a reasonable agreement with either background or polluted conditions previously reported, as 307 do ozone measurements [53,54]. The monthly average values for 13/Sept and 14/Sept are often similar 308 due to the similar (though not identical) sampling time from 8:00 to 12:00 and from 9:00 to 12:00 309 respectively. The similarity between monthly average particle concentrations for 12/Sept and 13/Sept are 310 purely coincidental.

311
Particle concentration and AMS/ACSM data were not available for the ZF2 site during this study.

312
Concentrations of CO, ozone, and BC values agreed well with their sample-period monthly averages with 313 the lone exception of ozone levels for the 3-6/Oct sample period. This increase is also reflected, albeit to a 314 much lesser degree, in an increase of CO and BC levels. From Figure S3 it appears that sample collection 315 began in the middle of a period of higher than average pollution levels.

322
Unsurprisingly, the ATTO site shows the lowest levels of almost all presented aerosol and gas 323 components. The sample collected on 15/Oct also appears to be fairly average with respect to the sample- The supporting data tabulated here has been collected from different instruments at the three sites 328 and so a direct comparison could be suspect, especially in the case of BC measurements [56]. However,

329
considering the agreement with published literature and qualitative use of Table 2 in the current work,

330
we believe any associated error is acceptable.

367
The SEM grayscale image shows the slightly different views presented by the two techniques, with 368 particle shapes appearing different between them along with a higher spatial resolution image (10 nm vs.

369
40 nm with STXM). Soot inclusions identified in the C speciation map are also seen as bright spots in the 370 SEM grayscale image in addition to many of the inorganic inclusions [24]. From the EDX data collection, 371 mass fraction maps for each element (on a per-particle basis) were used to calculate individual particle 372 diversity (Di) values for each particle. Another aspect of the maps is the varying background level 373 between SEM images, seen especially in the high background of the ZF2 image. This is a consequence of 374 the brightness and contrast levels being set before EDX acquisition and was performed to ensure that the 375 maximum number of particles were detected by the CCSEM particle detection software.   Table 3 outlines the assigned colors and labels, as well as 391 some relevant descriptive statistics for each cluster. As can be seen in the average particle diversity column in Table 3, most clusters have a Dα value near either 2.4 or 3.6 (with a single exception). These 393 two values define the "low" or "high" diversity referred to in the cluster names and are discussed in 394 more detail in section 3.5. A similar source apportionment was discussed in a previous SEM based study, 395 however, it was conducted during the wet season when biogenic aerosols dominate [38]. During the dry 396 season, these biogenic particles are still present but are overwhelmed by aerosols derived from biomass 397 burning.

398
One notable aspect of Figure 4 is the ratio of elemental Cl to S in each of the clusters shown. From 399 the EDX spectroscopy data presented here, the mass fraction of Cl is often greater or at least similar to 400 that of S. This is apparently contradicted by    Table 3.

442
In addition to LDS1 lacking the P, K, and Ca that HDS has, LDS1 also has a smaller amount of Na,

443
Mg, S, and Cl which results in the lower average particle diversity.

458
HDOrg is comprised of small particles with their carbon being entirely organic dominant. This 459 cluster has a substantial amount of the heavier elements (Z=11 (Na) and above) driving the diversity up.

460
The presence of P specifically is important as these elements, coupled with the carbon speciation, suggest

465
LDOrg is similar in carbon speciation, morphology, and size but lacks the heavier elements

472
As discussed further on in section 3.2, both organic clusters are unique in that they make up a 473 sizeable fraction of particles at all sampling sites.

476
Carbon speciation, however, shows a clear inorganic core with an organic coating. This leaves only a few 477 options for the identity of the inorganic cores seen here. One possibility is that the inorganic core is 478 composed of elements not analyzed here, such as Si or Al. Due to the Al mounting plate and the Si (Li) 479 detector used, we are not able to quantitatively detect Al and Si. However, a more likely possibility is 480 that, as mentioned above, the inorganic cores that were initially detected with STXM were particularly 481 sensitive to electron beam damage leading to these sensitive inorganics (possibly ammonium sulfate)  coating here is substantial and is likely due to aging as aerosols are transported inland to the ATTO site.

488
The transport of particles inland over a large distance is also reflected in the O/C ratio, where the particles 489 (specifically the organic coatings) may have oxidized more than other clusters.

490
HDI2 is characterized by many small inorganic inclusions speckled throughout the particles which 491 are not as localized as with the HDI1 cluster. There are small soot inclusions and an increased presence of P, K, and S as compared to HDI1. These particles are mainly seen at the ZF2 site with a smaller portion 493 present at ATTO. Thus it is possible that this cluster is associated with spore rupturing but further 494 investigation is needed to apportion this cluster [75].

495
The VHDI cluster is unique in that it possesses the highest Dα value of any of the clusters at 4.83, 496 well above both the nominal "high diversity" value of 3.6 and the second highest Dα value of 3.83. This 497 cluster also has a large statistical error of 1.92 (at a 95% confidence level), which could indicate multiple 498 disparate groups are present in this cluster. This cluster is comprised mostly of particles from ZF2, but 499 ATTO and T3 particles contribute substantially as well. The VHDI cluster's elemental composition is 500 similar to that of HDI2, but with a decreased C and O mass fraction and an enhancement of the other 501 elements, especially K (often seen in inorganic salt grains from biomass burning) [31,76]. Inorganics are 502 seen both as large localized inclusions, and as many small inclusions speckled throughout the particle.

503
This cluster's high diversity and larger statistical spread may also be indicative of the varied biomass 504 burning fuels and burning conditions present.

513
The species of carbon found in the HDM cluster's particles are well mixed with soot, non-514 carbonaceous inorganic, and organic carbon found in varying ratios. The large soot inclusions, high 515 diversity, and substantial presence of higher Z elements may point to an industrial or automotive origin.

516
Although the sizeable representation of the HDM cluster at T3 supports this, a slightly larger 517 representation is seen at site ZF2. This raises the possibility that some emissions from service vehicles

527
This last particle type may contain adjacent particles being erroneously deemed a single particle by 528 our detection algorithm because of overlap of the organic coating upon impaction. This grouping of 529 multiple individual particles into agglomerations much larger than expected for the given MOUDI stage 530 could have caused them to be placed in the Misc. cluster.

531
One notable particle type seen in this cluster is a collection of particles with a rectangular inorganic 532 core with a small patch of organic carbon in the center. Some of these inorganic cores wrap around the 533 carbon center while some others have a side missing but they all retain the same basic shape. The  547 Figure 5. Contribution of the twelve particle-type clusters identified in the samples from stage 7 (nominal 548 aerodynamic size range: 560-320 nm) and stage 8 (320-180 nm) at each sampling site.

549
As the ATTO sampling site is less polluted and representative of biogenic aerosols, the presence of 550 both organic clusters as well as two inorganic clusters with possible biogenic origins is expected.

551
Conversely the relative absence of soot clusters or the mixed clusters further highlights the ATTO site's 552 remoteness from regional anthropogenic (urban) influences. However, even this site is far from being 553 pristine, as shown by the presence of significant amounts of BC, presumably from long-range transport.

554
While the ZF2 site contains many of the same clusters present at the ATTO site, there are some 555 notable differences. The presence of the HDI1 cluster is diminished (~1% as compared to ATTO's 26%),

556
and both mixed clusters are seen in substantial amounts. The largest difference between the two stages is 557 the enhancement of the LDM cluster in stage 7 data and the minor increase in all three soot clusters in 558 stage 8 particles.

559
Site T3 shows the presence of many clusters, with all three soot clusters present in substantial 560 amounts. This is expected, as automotive exhaust or energy production through fuel oil burning will 561 produce soot particles that travel to site T3 [77]. Both organic clusters are present with a slight 562 enhancement in stage 8 particles. Because both organic clusters are seen in reasonable amounts at each 563 sampling site, these particles may be part of the aerosol background inherent to sampling in a heavily forested region. Stage 8 particles are also devoid of LDI, HDI1, and Misc. clusters, but few of these were 565 seen in stage 7 and so this absence may be due to insufficient sampling.

566
Another aspect of Figure 5 is how many clusters make up most of each site's aerosol population, for 567 which we use the following metric. Each site's cluster contribution is sorted in descending order and an 568 effective number of clusters is found using ( ) = ! , where r is the rank of each cluster's 569 contribution to that site's population (with 1 assigned to the cluster with the largest contribution), fr is the 570 fraction of that site's population that cluster r accounts for, and E(r) is the effective number of clusters.

571
This metric will vary, in this case from 1 to 12, where the lower the effective number of clusters, the better 572 a given site is characterized by fewer clusters. The values calculated from this metric are listed in Table 4.

573
This metric highlights the increased diversity of sites T3 and ZF2 with respect to the ATTO site. Site 574 ZF2's cluster composition is more varied. This is possibly due to specific events occurring during

582
(placing many of them in the same cluster).

586
Although relatively few supermicron particles were collected, most clusters included some fraction 587 of both sub-and supermicron particles. Only 1 cluster (Misc) was exclusively supermicron in size 588 whereas three clusters (LDS1, LDI, LDOrg) included exclusively submicron particles. Referencing Figure   589 6, the only clusters observed in the supermicron size range were those labeled: Misc (located around 2 590 µm with a very small percentage), LDM, HDM, HDI2, LDS2, and HDI1. Supermicron particles in the 591 clusters HDS, HDOrg, and VHDI particles were also observed, but in very small numbers. Many clusters 592 that make up the supermicron range represent more aged species.

593
The submicron range is composed of many more clusters relative to the supermicron range. Many 594 clusters in the submicron range were often labeled as less aged than the ones found in the supermicron 595 range. This qualitative observation is supported by Figure 7 where there is an increasing trend in 596 individual particle diversity (Dα) with increasing particle size and the notion that Dα is correlated with the 597 extent of particle aging.  appear to be the least diverse. However, the error analysis described below, renders this merely 604 suggestive rather than conclusive. As particle size increases, two things are observed: 1) average particle 605 diversity increases slightly and 2) the fraction of inorganics increases. Because of the ubiquity of C, N,

606
and O in aerosol particles, the average particle diversity will almost always be slightly above 3. Given the  that only 32 particles with diameters >2 µm were analyzed which is why this region is fairly noisy. Error 618 bars are not shown when only a single particle of that size was measured. Only 9 elements are labeled 619 (with P and K seen as small slivers) whereas the others are too small to be seen in the figure.

620
After clustering, most clusters were assigned so that their average particle diversity (Dα) was close to 621 one of the two modes present in Figure 8. This clear distinction between the two diversity modes is what 622 the high and low diversity cluster names are referring to.

623
The bimodality seen here may represent the separation between fresh and more aged aerosol

632
The two dimensional histograms between Di and CED in Figure 8 serve to reinforce the idea that 633 smaller particles tend to be less diverse. These smaller, less diverse particles are also less spread out 634 whereas the more diverse particles show a wider spread in both diversity and size.

635
The increased spread seen in the aged aerosol group may be due to the variety of ways that aerosols 636 can age and differences in distances traveled from the aerosols origin. Because the same variability isn't 637 seen in the smaller, less diverse, fresh aerosol group we suspect these particles have sources closer to 638 where they were sampled. By sampling particles with nearby sources, the elemental composition and, by 639 extension, particle diversity will be determined by the method of production and therefore be much less 640 variable.

646
Entropy metrics were used to quantify mixing state for each sample analyzed here. Figure 9 shows 647 the mixing state index (χ) corresponding to particles in each sample. In this case, the variation in mixing 648 state index is small, with all samples having a χ bounded between 0.8 and 0.9. This is a result of Dα and

673
The spread in Dα values among samples within a sampling site is much wider than that of Dγ. A 674 large spread in Dα is expected when a singular diversity value is calculated from samples containing the 675 variety of distinct particle types seen in Figure 3. This value is more susceptible to change from one sampling date to another compared to Dγ and depends on how much of each aerosol type is collected 677 during a given sampling time.

678
The increase in the average particle diversity (Dα) with respect to increasing particle size is hinted at 679 here, albeit in less certain terms. Focusing on samples collected where both stage 7 and 8 data were 680 analyzed, average Dα values appear to be larger for stage 7 particles.

681
Samples collected at site T3 were expected to have a lower mixing state than either the ATTO or the 682 ZF2 site. This hypothesis was borne from the quantity of fresh emissions in Manaus, specifically soot 683 production from combustion, which would serve to drive the mixing state downwards towards total 684 external mixing. However, the end result of the error calculations in section 2.7 is that the values of χ for 685 each point in Figure 8a become statistically unresolvable as can be seen from the large error bars. This is 686 also not an issue that would be solved with any reasonable amount of extra data collection but is instead 687 mainly the consequence of the intrinsic spread in Dα.

689
Presented here is a quantitative combination between two complementary per-particle 690 spectromicroscopy techniques, STXM/NEXAFS and SEM/EDX, on the exact same data set.

691
STXM/NEXAFS data was collected at the C, N, and O K-edges on a sub-particle level. This allowed not 692 only the quantitative determination of C, N, and O absolute masses, but also carbon speciation and 693 morphology. SEM/EDX allowed the approximate composition of the inorganic fraction to be determined 694 and then quantified along with the STXM data. The combination of these two techniques enables almost 695 all atmospherically relevant elements to be quantitatively probed on a per-particle basis. The potential 696 issue with S detection discussed above could be mitigated entirely in future measurements by conducting 697 STXM measurements at the S L-edge to obtain S mass fractions. This combined technique could be 698 especially useful for identifying aerosol sources using elemental tracers or unique elemental 699 compositions.

700
Using particle-specific elemental composition, size, carbon speciation, and individual particle 701 diversity (Di), k-means clustering was used to separate particles into 12 clusters. The cluster average of 702 these same parameters allowed for potential sources to be assigned. It was found that the stage 7 of the 703 T3 site had a more varied population of particles (as defined by the effective cluster number) and 704 contained more soot-containing clusters than either the ATTO or ZF2 site. Clusters also exhibited size 705 dependence, with a large portion of supermicron particles assigned to high diversity clusters which have 706 been hypothesized to represent more aged particles. This approach could be used for even larger data 707 sets, especially those located at long-standing measurement facilities. From this, diurnal, seasonal, or 708 yearly changes in the aerosol population could be monitored directly. 727 and size-resolved particle composition to be used rather than assumed to improve model performance 728 [79][80][81]. Even though this type of individual particle microscopy study is time consuming, regions that 729 are important to global climate models (such as the Amazon) may benefit from the improved accuracy of 730 an experimentally determined mixing state.

731
The quantitative mixing state index presented is a useful tool, but its utility can be readily expanded.

732
Two of the advantages that this combined spectromicroscopy technique has are the ability to identify 733 morphology both of the particles as a whole and of the constituents within the particle. Due to the 734 general nature of the mixing state parameterization, the mixing state index and its interpretation is   From right to left: 3/Oct 11:00 -6/Oct 11:00 and 6/Oct 14:00 -8/Oct 12:00.