No large kinetic limitations to partitioning are observed in the southeast, and partitioning according to vapor pressure is active on short timescales (Lopez-Hilfiker et al., 2016). The higher relative humidity (RH) in this region, which results in fast diffusion in isoprene-SOA containing particles (Song et al., 2015), may be at least partially responsible for this behavior. In some instances (e.g., for key IEPOX-SOA species), observations indicate that detected OA species are significantly less volatile than their structure indicates, likely due to thermal decomposition of their accretion products or inorganic–organic adducts in instruments (Lopez-Hilfiker et al., 2016; Hu et al., 2016; Isaacman-VanWertz et al., 2016; Stark et al., 2017).

Further research is needed regarding the role of organic partitioning into OA versus water and this can be evaluated using field data. If both processes occur in parallel in the atmosphere, vapor-pressure-dependent partitioning to OA may occur along with aqueous processing without significant double counting or duplication in models. However, due to the high relative humidity (average RH is 74 %, see Weber et al., 2016) and degree of oxygenation of organic compounds (OM / OC is 1.9–2.25, see below) in the southeastern US atmosphere, inorganic-rich and organic-rich phases may not be distinct (You et al., 2013) and more advanced partitioning algorithms accounting for a mixed inorganic–organic water phase may be needed (Pye et al., 2017, 2018).

Phase separation can be predicted based on the determination of a separation relative humidity (SRH), which is a function of the degree of oxygenation and inorganic constituent identity (You et al., 2013), and a comparison to the ambient relative humidity. For RH < SRH, phase separation occurs. Pye et al. (2017) predicted phase separation into organic-rich and electrolyte-rich phases occurs 70 % of the time during SOAS at CTR with a higher frequency during the day due to lower RH.

Primary organic aerosol (POA) concentrations are expected to be small in the southeast outside urban areas and we make no major recommendation for how to model them. Modelers should be aware that a fraction of primary organic aerosol based on the EPA National Emissions Inventory (NEI) is semivolatile (Robinson et al., 2007). However, not all POA is thought to be semivolatile – for example, OAs from sources such as soil are included in the NEI. Modeled POA may already include some oxidized POA (OPOAs) if the models include heterogeneous oxidation (as in CMAQ; Simon and Bhave, 2012) or hydrophilic conversion (as in GEOS-Chem; Park et al., 2003). Thus, care should be exercised in evaluating model species such as POA with aerosol mass spectrometer (AMS) positive matrix factorization (PMF) factors such as hydrocarbon-like OA (HOA). For semivolatile POA treatments, mismatches between POA inventories and semivolatile / intermediate volatility organic compounds (S / IVOCs) need to be carefully considered. Comparisons of model inventory versus ambient ratios of POA / ΔCO, POA / black carbon (BC) or POA / NOx can be used to indicate whether or not POA emissions are excessive (De Gouw and Jimenez, 2009). As these ratios can be affected by errors in the denominator species, it is important to also evaluate those carefully against observations. For models with limited POA information, the ratio of organic matter to organic carbon (OM / OC) should be adjusted to reflect the highly oxidized nature of ambient OA (as mass is transferred from hydrophobic/hydrophilic concentrations for example). The OM / OC ratio of bulk ambient OA in the southeastern US is 1.9–2.25 as measured during summer 2013 (Kim et al., 2015; Pye et al., 2017).

A biomass burning PMF factor (BBOA) was observed during SOAS and likely has a higher impact on brown carbon (BrC) than its contribution to OA mass would suggest, although overall BrC concentrations were very small (Washenfelder et al., 2015). Net SOA mass added via photochemical processing of biomass burning emissions is thought to be modest, relative to the high POA emissions (Cubison et al., 2011; Jolleys et al., 2012; Shrivastava et al., 2017).

Organic nitrates, primarily from monoterpene reactions with the nitrate radical, have been recognized as an important source of OA in the southeast, contributing from 5 to 12 % in the southeastern US in summer (Xu et al., 2015a, b; Ayres et al., 2015; Pye et al., 2015; B. H. Lee et al., 2016). In fact, this number could be an underestimate if some of these organic nitrates are susceptible to hydrolysis or photodegradation and thus are not detected as nitrates. We have high confidence that models should include SOA formation from nitrate radical oxidation of monoterpenes. Sesquiterpenes and isoprene may also contribute OA through nitrate radical oxidation, but the contribution is expected to be smaller (Pye et al., 2015; Fisher et al., 2016). A number of options exist for representing this type of aerosol including fixed yields, Odum 2-product parameterizations, volatility basis set (VBS) representations (Boyd et al., 2015) and explicit partitioning and/or uptake of organic nitrates (Pye et al., 2015; Fisher et al., 2016).

Detailed modeling studies can provide additional insight into the interactions between monoterpene nitrate SOA and gas-phase chemistry, as well as the fates of specific organic nitrates. Explicit formation and treatment of organic nitrates, yields of which are parent hydrocarbon specific, can take into account hydrolysis of particle-phase organic nitrate. The hydrolysis should depend on the relative amounts of primary, secondary and tertiary nitrates which are produced in different abundances in photooxidation vs. nitrate radical oxidation (Boyd et al., 2015, 2017). Hydrolysis may also depend on the level of acidity and presence of double bonds in the organic nitrate (Jacobs et al., 2014; Rindelaub et al., 2016). In addition to hydrolysis, particle organic nitrates could photolyze and release NOx or serve as a NOx sink through deposition (Nah et al., 2016b).

Formation of organic nitrates should also be considered in the context of emerging evidence for the role of autoxidation, especially in the monoterpene system (Ehn et al., 2014). Autoxidation has been shown to occur in both photooxidation and ozonolysis of monoterpenes (Jokinen et al., 2015) and leads to highly oxidized species including organic nitrates (B. H. Lee et al., 2016; Nah et al., 2016b), many of which are low volatility. While some empirical representations (e.g., VBS or Odum 2-product) of monoterpene SOA may capture these species, autoxidation products may be very susceptible to chamber wall loss (Zhang et al., 2014; Krechmer et al., 2016) and missing from SOA parameterizations. The role of autoxidation in forming SOA in the southeastern US atmosphere remains to be determined. In this regard, future laboratory studies should carefully constrain the peroxy radical reaction channels (e.g., Schwantes et al., 2015; Boyd et al., 2015) and be conducted under regimes that are representative of ambient environments where the peroxy radical lifetimes can vary.

Due to the abundance of observations in the southeastern atmosphere (Budisulistiorini et al., 2016; W. W. Hu et al., 2015; Hu et al., 2016; Xu et al., 2015a, b, 2016), similarity between laboratory and field IEPOX-SOA determined by PMF analysis and availability of model parameterizations to predict IEPOX-SOA (Pye et al., 2013; Woo and McNeill, 2015; Marais et al., 2016; Budisulistiorini et al., 2017; Sareen et al., 2017), we have high confidence that IEPOX-SOA should be included in models. D'Ambro et al. (2017) predicts IEPOX will be the major precursor to SOA under low-NOx conditions when peroxy radical lifetimes are atmospherically relevant, which has not always been the case in older experiments. However, a number of parameters needed to predict IEPOX-SOA are uncertain and different modeling approaches, as well as the use of all available experimental constraints, could be beneficial. The mechanism of IEPOX-SOA formation involves gas-phase reactions followed by aqueous processing which can occur either in aerosols or cloud droplets, although the acid-catalyzed initiation step of the epoxide ring opening favors SE US aerosol conditions and makes this process less efficient in cloud water. This mechanism could be represented as heterogeneous reaction with a reactive uptake coefficient or more explicit partitioning and particle reaction (Table 1).

The correlation of IEPOX-SOA with sulfate (Xu et al., 2015a, 2016; W. W. Hu et al., 2015) can serve as a useful model evaluation technique as underestimates in sulfate could lead to underestimates in IEPOX-SOA in models (Fig. 2). Current pathways for IEPOX-SOA formation (Eddingsaas et al., 2010) involve acidity in aqueous solutions (Kuwata et al., 2015), but several studies suggest that IEPOX-SOA is not correlated well with aerosol acidity or aerosol water (Budisulistiorini et al., 2017; Xu et al., 2015a). Ion balances or other simple measures of aerosol acidity are likely inadequate to characterize particle acidity and thermodynamic models such as ISORROPIA II or AIM are more appropriate for modeling IEPOX-SOA (Guo et al., 2015; Weber et al., 2016). Currently, different observational data sets indicate different nominal ratios of ammonium to sulfate (Pye et al., 2018), so it needs to be kept in mind that some measurements report only inorganic sulfate (e.g., ion chromatography) while others report total (inorganic + organic) sulfate (e.g., AMS). A modeling study suggested that ammonia uptake might be limited by organics, thus affecting acidity (Kim et al., 2015; Silvern et al., 2017).

SAS observations also provide estimates of some components of IEPOX-SOA including 2-methyltetrols and IEPOX–organosulfates (Budisulistiorini et al., 2015; W. W. Hu et al., 2015). For modeling applications focusing on IEPOX-SOA, additional speciation of IEPOX-SOA (into tetrols, organosulfates, etc.) and oligomerization and volatility can be treated. Treating the monomers (e.g., 2-methyltetrols) explicitly with their molecular properties will likely lead to excessive volatility of the IEPOX-SOA (Lopez-Hilfiker et al., 2016; Hu et al., 2016; Isaacman-VanWertz et al., 2016; Stark et al., 2017).

New information on glyoxal SOA is emerging in this area but its importance in the southeast remains unclear. Glyoxal has been suspected to be the dominant aqueous SOA source under high-NOx (RO2+ NO) oxidation conditions (McNeill et al., 2012) and the southeast has a mix of high-NOx and low-NOx (RO2+ HO2) conditions (Travis et al., 2016). In addition, abundant isoprene emissions can lead to substantial glyoxal concentrations. Modeling for the southeastern US indicates significant SOA can form from glyoxal (Marais et al., 2016; Pye et al., 2015; Knote et al., 2014; Li et al., 2016; Chan Miller et al., 2017). Implementation in models may require modifications to the gas-phase chemistry to specifically track glyoxal which may be lumped with other aldehydes (e.g., in CB05). Recent model studies do not find that a large SOA source from glyoxal is required to match observations, but more field measurements and laboratory studies, especially of the yield from isoprene oxidation and the aerosol uptake coefficient, are required to constrain the process.

Results from SOAS and SEAC4RS indicate only a modest enhancement of OA due to cloud processing over the SE US, which was not statistically significant (Wagner et al., 2015). In addition, epoxide reactions in cloud droplets are predicted to lead to minor amounts of SOA due to the pH dependence of IEPOX hydrolysis (Fahey et al., 2017; McNeill, 2015).

While the rural southeast is assumed to be dominated by SOA from biogenic precursors (which may be influenced by anthropogenic pollution) as a result of high modern carbon (Hidy et al., 2014), SOA from anthropogenic VOCs is known to play a role from fossil carbon measurements (∼ 18 % at Centerville; Kim et al., 2015), but it is not directly apportioned otherwise. We note that since ∼ 50 % of urban POA and 30 % of urban SOA is non-fossil (Zotter et al., 2014; Hayes et al., 2015); an urban fraction of ∼ 28 % for the SOAS site is consistent with observations (Kim et al., 2015). This source is as large as most of the other individual sources discussed in this section and should not be neglected in modeling studies. A simple parameterization based on CO emissions (Hayes et al., 2015) may be adequate for incorporating this source in modeling studies and has shown good results for the southeastern US (Kim et al., 2015), but care should be taken to evaluate the CO emissions when using it.

We list several caveats for the process of comparing model results to surface network observations. OC measurements from IMPROVE surface sites may be biased low in the summer due to evaporation of organic aerosols during the sample collection and handling (Kim et al., 2015). On the other hand, SEARCH measurements agree well with research community instruments in the Centerville site, such as AMS. Therefore the SEARCH data should be considered as the reference.

Decreases in mass concentrations of particulate sulfate and nitrate over the past decades are consistent with environmental policy targeting their gas-phase precursors, namely SOx and NOx emissions. Reductions in particulate organic carbon in the southeastern US over the past decade (Blanchard et al., 2016, 2013) are more difficult to reconcile because in the summertime it is predominantly modern and there is no control policy aimed at reducing biogenic VOCs. Decreased SOx (Kim et al., 2015; Xu et al., 2015b; Blanchard et al., 2013) and NOx emissions modulate the amount of organic aerosol formation through the gas-phase impacts described above and impacts on the absorbing medium amount (T. K. V. Nguyen et al., 2015; Attwood et al., 2014) and chemical composition.

In addition to sources and sinks of OA, attention should also be paid to the role of dry deposition of gases in determining mass loadings, as this process can have a large impact on model predictions and is very poorly constrained (Glasius and Goldstein, 2016; Knote et al., 2015).

A motivating goal of the southeast studies was to examine PM mass measurements at the surface and satellite-measured AOD (aerosol optical depth) to facilitate improved prediction of the total aerosol loading. Aerosol mass aloft contributes to AOD (Wagner et al., 2015), and this complicates the relationship to surface concentrations. Relative humidity, vertical structure of the daytime PBL and aerosol liquid water (not measured by surface networks) influences remotely sensed AOD (Brock et al., 2016a, b; Kim et al., 2015; Nguyen et al., 2016). AOD is also complicated by aerosol composition. Attwood et al. (2014) finds that the steeper decrease in sulfate aerosol relative to organic from 2001 to 2013 has changed the hygroscopicity of SE US aerosol, leading to lower aerosol liquid water and thus lower optical extinction and AOD.

Isoprene emissions measured by the NOAA P3, using the mixed boundary layer budget method, and NCAR/NSF C-130 and NASA DC-8 aircraft using direct eddy covariance flux measurements were within the wide range of observations reported by previous studies. The two methods of estimating isoprene emissions agreed within their uncertainties (Yu et al., 2017). Solar radiation and temperature measured by the aircraft along the flight tracks and available from regional model and assimilations (e.g., WRF, NLDAS-2) enabled estimation of emissions using models including BEIS3.12, BEIS3.13, MEGAN2.0, MEGAN2.1 with default land cover, MEGAN2.1 with revised land cover and MEGAN3. Isoprene emissions are highly sensitive to solar radiation and temperature, and biases in the values used to drive emission models can result in errors exceeding 40 %, complicating efforts to evaluate biogenic emission models. As has previously been noted in the southeastern US, MEGAN2.1 predicted isoprene emissions in the southeastern US were about twice as high as BEIS3.13. The measurements fall between the two models and are within the model and measurement uncertainties (Warneke et al., 2010). Isoprene mixing ratios were modeled with (a) WRF-Chem using BEIS and with (b) CAMx using MEGAN, and the results were consistent with the measurement–inventory comparison: WRF-Chem was biased low and CAMx biased high (Warneke et al., in preparation).

Land cover characteristics including leaf area index (LAI) and tree species composition data are also critical driving variables for BEIS and MEGAN isoprene and monoterpene emission estimates. Airborne flux measurements agreed well with MEGAN2.1 for landscapes dominated by southeastern oaks, which are high-isoprene-emitting tree species, but landscapes that had an overstory of non-emitters, with the high-isoprene emitters in the understory, showed emissions lower than expected by the model. The isoprene emission factor (EF) was linearly correlated with the high-isoprene-emitter plant species fraction in the land cover data set. This may indicate a need for models to include canopy vertical heterogeneity of the isoprene emitting fraction (Yu et al., 2017).

A simplification used in current biogenic emission models including BEIS3.13, BEIS3.6 and MEGAN2.1 is that all high-isoprene-emitting species are assigned the same isoprene emission factor. For example, all North American species of Quercus (oak), Liquidambar (sweetgum), Nyssa (tupelo), Platanus (sycamore), Salix (willow), Robinia (locust) and Populus (poplar and aspen) are assigned a single value based on the average of an extensive set of enclosure measurements conducted in North Carolina, California and Oregon in the 1990s (Geron et al., 2001). Earlier studies had reported isoprene emission factors for these tree species that ranged over more than an order of magnitude (Benjamin et al., 1996). Geron et al. (2001) showed that by following specific measurement protocols, including leaf cuvettes with environmental controls and ancillary physiological measurements such as photosynthesis, the variability dropped from over an order of magnitude to about a factor of 3. They concluded that this remaining variability was due at least as much to growth conditions as to species differences and so recommended that a single isoprene emission factor be used for all of these species. Recent aircraft flux measurements (Misztal et al., 2016; Yu et al., 2017) indicate that there is at least a factor of 2 difference in the isoprene emission factors of these species. This could be due to a genetic difference in emission capacity and/or differences in canopy structure. The aircraft measurements indicate that sweetgum and tupelo emission factors are similar to the value used in BESI3.13 and BEIS3.6, while the California oak emission factor is similar to that used in MEGAN2.1. The aircraft-based estimate of southeastern oak emission factors falls between the BEIS3.6 and MEGAN2.1 values. As a result, aircraft flux measurements in the southeastern US are higher than BEIS3.13 and BEIS3.6 and lower than MEGAN2.1. The MEGAN3 emission factor processor provides an approach for synthesizing available emission factor data and can be used to account for the emission rate variability observed by these aircraft flux studies (Guenther et al., 2018).

Modeling monoterpene emissions is even more challenging than isoprene emissions for reasons that include multiple emission processes (e.g., both light-dependent and light-independent emissions), stress-induced emission capability present in many plant species but not always expressed and the potential for enclosure measurements to dramatically overestimate emissions due to release of monoterpenes from damaged storage pools. The eddy covariance flux measurements on the NCAR/NSF C-130 are similar to the values estimated by MEGAN2.1 for needle leaf forests, considered to be high-emission regions, but are higher than the modeled monoterpene emissions from other landscapes (Yu et al., 2017). They conclude that unaccounted processes, such as floral and stress emissions, or sources such as non-tree vegetation may be responsible for the unexpectedly high monoterpene emissions observed by the aircraft.

During the experiment direct observations of fluxes for a variety of species from large aircraft were conducted, enabling a first direct estimate of fluxes over a regional domain (Wolfe et al., 2015; Yuan et al., 2015; Kaser et al., 2015). These data have the potential for enabling analyses of strengths and weaknesses of current emission and deposition schemes and their implementation within chemical transport models. Vertical flux profiles also contain information on the chemical production and loss rates, providing a new observational constraint on the processes controlling reactive gas budgets. An LES model was used to simulate isoprene, NOx and their variability in the boundary layer. The results showed good agreement between the measurements and the model. The atmospheric variability of isoprene, the altitude profile in the boundary layer of isoprene, and NOx mixing ratios and fluxes were well reproduced in the model, which was used to validate the eddy covariance and mixed boundary layer methods of estimating isoprene fluxes (Kim et al., 2016a; Wolfe et al., 2015).

Travis et al. (2016) utilizing the GEOS-Chem model report that NOx emissions are significantly overestimated by the NEI 2011 and suggest that mobile source and industrial emissions of NOx need to be lowered by 30–60 % to be consistent with aircraft measurements collected over the southeastern US during the SEAC4RS study. These results are consistent with modeling studies performed during the DISCOVER-AQ field campaign, which also found that the NEI 2011 overestimated NOx emissions (Anderson et al., 2014; Souri et al., 2016). However, a later study by Li et al. (2018) utilizing the AM3 model during the SENEX study suggests that overestimates in NEI 2011 NOx emissions may be smaller than reported in the Travis et al. study (∼ 14 % vs. 30–60 %). McDonald et al. (2018) using WRF-Chem found mobile source emissions in the NEI 2011 to be overestimated by ∼ 50 % and a factor of 2.2 for NOx and CO, respectively, when evaluated with SENEX aircraft measurements. Due to rapidly declining trends in vehicle emissions (McDonald et al., 2013, 2012), some of the emissions overestimate was attributed to utilizing a 2011 inventory in 2013 model simulations. However, roadside measurements of vehicular exhaust also suggest systematic overestimates in emission factors used by the EPA's vehicle emissions model (MOVES), likely contributing to the consistent reporting to date of overestimated mobile source NOx emissions (Anderson et al., 2014; Souri et al., 2016; Travis et al., 2016). When NOx emissions were reduced from mobile sources by this amount, model predictions of O3 over the southeastern US were improved both for mean concentrations and O3 extreme days (McDonald et al., 2018), consistent with modeling by Li et al. (2018) demonstrating the sensitivity of O3 to NOx emissions in the southeastern US over the 2004–2013 timespan.

Along with other aircraft field campaigns and tall tower measurements in the Upper Midwest, data from the SENEX study was used to assess anthropogenic emissions of VOCs in the NEI and a global inventory (RETRO). L. Hu et al. (2015) found that RETRO consistently overestimates US emissions of C6–C8 aromatic compounds by factors of 2–4.5; the NEI 2008 overestimates toluene by a factor of 3 but is consistent with top-down emission estimates for benzene and C8 aromatics. The study also suggests that East Asian emissions are an increasingly important source of benzene concentrations over the US, highlighting the importance of long-range transport on US air quality as domestic sources of emissions decline (Warneke et al., 2012).

Two studies have quantified top-down emissions of oil and gas operations, derived from aircraft measurements for VOCs and methane from SENEX P-3 data (Peischl et al., 2015; Yuan et al., 2015). The oil and gas regions measured during SENEX account for half of the US shale gas production, and loss rates of methane to the atmosphere relative to production were typically lower than prior assessments (Peischl et al., 2015). Yuan et al. (2015) explored the utility of eddy-covariance flux measurements on SENEX and NOMADSS aircraft campaigns and showed that methane emissions were disproportionately from a subset of higher emitting oil and gas facilities. Strong correlations were also found between methane and benzene, indicating that VOCs are also emitted in oil and gas extraction. High wintertime O3 has been found in the Uintah Basin, UT (Ahmadov et al., 2015; Edwards et al., 2014), though it is unclear at this time how significant oil and gas emissions of VOCs could be in an isoprene-rich source region on tropospheric O3 formation. Future atmospheric modeling efforts of oil and gas emissions are needed.

During the SENEX and SEAC4RS studies, research aircraft measured agricultural fires over the southeast. Liu et al. (2016) reported emission factors of trace gases, which were consistent with prior literature. In general, the authors found emissions of SO2, NOx and CO from agricultural fires to be small relative to mobile sources (< 10 %). However, within fire plumes, rapid O3 formation was observed, indicating potential air quality impacts on downwind communities. To represent the impact of biomass burning, air quality models need improved treatments of initial VOC and NOx emissions and near-source chemistry. Sub-grid parameterizations, based on detailed models like the Aerosol Simulation Program (ASP; Alvarado and Prinn, 2009) and which incorporate gas-phase chemistry, inorganic and organic aerosol thermodynamics, and evolution of aerosol size distribution and optical properties, could improve coarse model representations of chemistry near biomass burning plumes. Zarzana et al. (2017) investigated enhancements of glyoxal and methylglyoxal relative to CO from agricultural fires and report that global models may overestimate biomass burning emissions of glyoxal by a factor of 4. This highlights large uncertainties and variability in fire emissions and a need for additional observational constraints on inventories and models.

Using satellite data, Goldstein et al. (2009) diagnosed summertime enhancements in AOD of 0.18 over the southeast, relative to winter, and hypothesized that secondary organic aerosol from biogenic emissions accounts for this enhancement. Goldstein et al. (2009) further estimated a regional surface cooling of -0.4 W m-2 in response to annual mean AOD over the southeast. These findings seemed at first at odds with surface PM2.5 measurements, which reveal little seasonal enhancement in summer. Using SEAC4RS measurements and GEOS-Chem, Kim et al. (2015) determined that the relatively flat seasonality in surface PM2.5 can be traced to the deeper boundary layer in summer, which dilutes surface concentrations.

In response to emission controls, aerosol loading over the southeast has declined in recent decades. For example, wet deposition fluxes of sulfate decreased by as much as ∼ 50 % from the 1980s to 2010 (Leibensperger et al., 2012a). Over the 2003–2013 time period, surface concentrations of sulfate PM2.5 declined by 60 %. Organic aerosol (OA) also declined by 60 % even though most OA appears to be biogenic and there is no indication of a decrease in anthropogenic sources (Kim et al., 2015). Model results suggest that the observed decline in OA may be tied to the decrease in sulfate, since OA formation from biogenic isoprene depends on aerosol water content and acidity (Marais et al., 2016, 2017). Consistent with these surface trends, 550 nm AOD at AERONET (Aerosol Robotic Network) sites across the southeast has also decreased, with trends of -4.1 % a-1 from 2001 to 2013 (Attwood et al., 2014). Xing et al. (2015a) reported a roughly -4 % decrease in remotely sensed AOD across the eastern United States, as measured by the Moderate Resolution Imaging and Spectroradiometer (MODIS) on board Terra and Aqua. These large declines could potentially have had a substantial impact on regional climate, both through aerosol–radiation interactions and aerosol–cloud interactions.

Even as global mean temperatures rose over the 20th century in response to increasing greenhouse gases, significant cooling occurred over the central and southeastern United States. This cooling, referred to as the US warming hole (Pan et al., 2004), has been quantified in several ways. For example, Fig. 3 shows that annual mean temperatures across the southeast decreased by ∼ 1 ∘C during the 1930–1990 timeframe (Capparelli et al., 2013). A different temperature metric, the 20-year annual return value for the hot tail of daily maximum temperatures, decreased by 2∘ from 1950 to 2007 (Grotjahn et al., 2016). Over a similar time frame, Portmann et al. (2009) diagnosed declines in maximum daily temperatures in the southeast of 2–4∘ per decade, with peak declines in May–June, and linked these temperature trends with regions of high climatological precipitation. Since the early 2000s, the cooling trend has appeared to reverse (Meehl et al., 2015).

The causes of the US warming hole are not clear. Most freely running climate models participating in the Coupled Model Intercomparison Project (CMIP5) cannot capture the observed 20th century temperature trends over the southeast (Knutson et al., 2013; Kumar et al., 2013; Sheffield et al., 2013); this failure likely arises from either model deficiency or natural variability not included in the simulations. Indeed, several studies have argued that naturally occurring oscillations in sea surface temperatures (SSTs) influenced the large-scale cooling in the southeast (Robinson et al., 2002; Kunkel et al., 2006; Meehl et al., 2012; Weaver, 2013; Mascioli et al., 2017). Kumar et al. (2013), for example, linked the June–July–August indices of the Atlantic Multidecadal Oscillation (AMO) to annual mean temperatures across the eastern US for the 1901–2004 period. Mauget and Cordero (2014), however, pointed out inconsistencies in these two time series, with the AMO index sometimes lagging temperature changes. A recent study has argued that the transition of the Interdecadal Pacific Oscillation (IPO) phase from positive to negative in the late 1990s may have triggered a reversal of the warming hole trend (Meehl et al., 2015).

The cool period in the southeast coincided with heavy aerosol loading over the region, and several studies have suggested that trends in aerosol forcing may have also played a role in driving the US warming hole. For example, Leibensperger et al. (2012a, b) found that the regional radiative forcing from anthropogenic aerosols led to a strong regional climate response, cooling the central and eastern US by 0.5–1.0∘ from 1970 to 1990 (Fig. 3), with the strongest effects on maximum daytime temperatures in summer and autumn. In that study, the spatial mismatch between maximum aerosol loading and maximum cooling could be partly explained by aerosol outflow cooling the North Atlantic, which strengthened the Bermuda High and increased the flow of moist air into the south-central United States. Another model study diagnosed positive feedbacks between aerosol loading, soil moisture and low cloud cover that may amplify the local response to aerosol trends (Mickley et al., 2012). The strength of such positive feedbacks may vary regionally, yielding different sensitivities in surface temperature to aerosol forcing.

The cool period in the southeast coincided with heavy aerosol loading over the region, and several studies have suggested that trends in aerosol forcing may have also played a role in driving the US warming hole. For example, Leibensperger et al. (2012a, b) found that the regional radiative forcing from anthropogenic aerosols led to a strong regional climate response, cooling the central and eastern US by 0.5–1.0∘ from 1970 to 1990 (Fig. 3), with the strongest effects on maximum daytime temperatures in summer and autumn. In that study, the spatial mismatch between maximum aerosol loading and maximum cooling could be partly explained by aerosol outflow cooling the North Atlantic, which strengthened the Bermuda High and increased the flow of moist air into the south-central United States. Another model study diagnosed positive feedbacks between aerosol loading, soil moisture and low cloud cover that may amplify the local response to aerosol trends in the eastern US, including the southeast (Mickley et al., 2012). The strength of such positive feedbacks may vary regionally, yielding different sensitivities in surface temperature to aerosol forcing.  More recent modeling studies, however, have generated conflicting results regarding the role of aerosols in driving the warming hole. For example, the model study of Mascioli et al. (2016) reported little sensitivity in southeast surface temperatures to external forcings such as anthropogenic aerosols or even greenhouse gases. In contrast, Banerjee et al. (2017) found that as much of 50 % of the observed 1950–1975 summertime cooling trend in the southeast could be explained by increasing aerosols. Examining multi-model output, Mascioli et al. (2017) concluded that aerosols accounted for just 17 % of this cooling trend in summer. These contrasting model results point to the challenges in modeling climate feedbacks, such as those involving cloud cover or soil moisture.

These early model studies have been accompanied by more observationally based efforts to link trends in surface temperature to aerosol loading. A key first step is to determine whether changes in surface solar radiation are related to changes in aerosol loading. Measurements from the Surface Radiation network (SURFRAD) reveal increases of +0.4 Wm-2 a-1 in total surface solar radiation across the east during 1995–2010 (Gan et al., 2014). An attempt to reproduce the trend in total surface radiation with a regional chemistry–climate model found a reasonable match with observations over the east when aerosol–radiation interactions were included (Xing et al., 2015a). Most of the observed increase in surface solar radiation, however, appears due to increasing diffuse radiation, at odds with the decline in AOD, which should instead increase direct radiation (Gan et al., 2015, 2014). Using satellite data and assimilated meteorology, Yu et al. (2014) showed that trends in spatially averaged AOD and cloud optical depth declined over the 2000–2011 time period over the eastern US, while daily maximum temperatures and shortwave cloud forcing increased. These opposing trends suggest that aerosol–cloud interactions may have influenced the observed ∼ 1∘ warming trend in the southeast over this 10-year time period, with the decline in anthropogenic aerosols driving a decrease in cloud cover and a rise in surface temperatures. Yu et al. (2014) confirmed this hypothesis using a chemistry–climate model. In contrast, the observational study of Tosca et al. (2017), which also relied on satellite AOD, pointed to aerosol–radiation interactions as the driver of surface temperature trends in the southeast. Analysis of ground-based observations in Mississippi, however, found little covariability between AOD and clear-sky solar radiation at the surface, casting doubt on the importance of aerosol–radiation interactions in driving the observed cooling in this region (Cusworth et al., 2017).

Continued improvements of PM2.5 air quality in the southeast may further influence regional climate. Y. Lee et al. (2016) projected a warming of about +0.5 Wm-2 over the eastern US, including the southeast, over the 2000–2030 timeframe due to anticipated improvements in air quality and the associated reduction in AOD. Xing et al. (2015b) have pointed out that an overlooked beneficial effect of aerosol reduction is increased ventilation of surface air, a positive feedback that leads to further decline in surface PM2.5 concentrations. The feedback arises from changes in the temperature profile, with warmer temperatures at the surface and cooler temperatures aloft, which together enhance atmospheric instability and ventilation as aerosol-induced cooling is reduced. The feedback may lead to unexpected health benefits of clearing PM2.5 pollution (Xing et al., 2016).

Pollution episodes in the southeastern United States are correlated with high temperatures, low wind speeds, clear skies and stagnant weather (Camalier et al., 2007; Jacob and Winner, 2009). The spatial extent of the Bermuda High also plays a role in modulating air quality in the southeast (Zhu and Liang, 2013).

Fu et al. (2015) used models and observations to examine the sensitivity of August surface ozone in the southeast to temperature variability during 1988–2011. This study finds that warmer temperatures enhance ozone by increasing biogenic emissions and accelerating photochemical reaction rates. However, variability in ozone advection into the region may also explain much of the variability of surface ozone, with possibly increased advection occurring during the positive phase of the Atlantic Multidecadal Oscillation. Applying empirical orthogonal functions (EOF) analysis to observed ozone, Shen et al. (2015) determined that the sensitivity of surface ozone in the southeast can be quantified by the behavior of the west edge of the Bermuda High. Specifically, for those summers when the average position of the west edge is located west of ∼ 85.4∘ W, a westward shift in the Bermuda High west edge increases ozone in the southeast by 1 ppbv deg-1 in longitude. For all summers, a northward shift in the Bermuda High west edge increases ozone over the entire eastern United States by 1–2 ppbv deg-1 in latitude.

The influence of meteorology on PM2.5 in the southeast is not well quantified. Tai et al. (2010) found that observed sulfate and OC concentrations increase with increasing temperature across the region due to faster oxidation rates and the association of warm temperatures with stagnation and biogenic and fire emissions. Nitrate PM2.5, however, becomes more volatile at higher temperatures and decreases with temperature. Using local meteorology, however, Tai et al. (2010) could explain only about 20–30 % of PM2.5 daily variability in the southeast. Both Thishan Dharshana et al. (2010) and Tai et al. (2012b) diagnosed a relatively weak effect of synoptic-scale weather systems on PM2.5 air quality in the southeast, especially in the deep south. Shen et al. (2017), however, extended the statistical studies of Tai et al. (2012a, b) by taking into account not just the local influences of meteorology on PM2.5 air quality but also the relationships between local PM2.5 and meteorological variables in the surrounding region. These authors developed a statistical model that explains 30–50 % of PM2.5 monthly variability in the southeast. Shen et al. (2017) further reported that many atmospheric chemistry models may underestimate or even fail to capture the strongly positive sensitivity of monthly mean PM2.5 to surface temperature in the eastern United States, including the southeast, in summer.  In GEOS-Chem, this underestimate can be traced to the overly strong tendency of modeled low cloud cover to decrease as temperatures rise (Shen et al., 2017).

Emissions of US pollution precursors are expected to decline in coming decades (Lamarque et al., 2013; Fiore et al., 2015), which may offset any potential climate penalty. Background ozone, however, may increase due to increasing methane (West et al., 2012). A major challenge in quantifying the future trends in surface air quality is our lack of knowledge in temperature-dependent isoprene emissions and photochemistry (Achakulwisut et al., 2015).

Using a regional chemistry–climate model, Gonzalez-Abraham et al. (2015) found that daily maximum 8 h average (MDA8) ozone concentrations in the southeast would likely increase by 3–6 ppbv by the 2050s due solely to climate change and land use change. Changes in anthropogenic emissions of ozone precursors such as methane could further enhance MDA8 ozone in the southeast by 1–2 ppbv. Rieder et al. (2015), however, determined that large areas of the southeast would experience little change in surface ozone by the 2050s, but that study neglected the influence of warming temperatures on biogenic emissions. Shen et al. (2016) developed a statistical model using extreme value theory to estimate the 2000–2050 changes in ozone episodes across the United States. Assuming constant anthropogenic emissions at the present level, they found an average annual increase in ozone episodes of 2.3 days (> 75 ppbv) across the United States by the 2050s, but relatively little change in the southeast. In fact, a key result of this work is the relative insensitivity of ozone episodes to temperature in the southeast. However, Zhang and Wang (2016) have suggested that warmer and drier conditions in the southeast future atmosphere could extend the ozone season, leading to ozone episodes in October.

Model studies differ on the effects of future climate change on PM2.5 in the southeast. Tai et al. (2012a, b) analyzed trends in meteorological modes from an ensemble of climate models and found only modest changes in annual mean PM2.5 (±0.4 µg m-3) by the 2050s in the southeast, relative to the present-day. Using a single chemistry–climate model, Day and Pandis (2015) calculated significant increases of ∼ 3.6 µg m-3 in July mean PM2.5 along the Gulf coast by the 2050s and attributed these increases to a combination of decreased rain-out, reduced ventilation and increased biogenic emissions. Building on the statistical model of Tai et al. (2012a,b), Shen et al. (2017) found that PM2.5 concentrations in the southeast could increase by 0.5–1.0 µg m-3 by 2050 on an annual basis and as much as 2.0–3.0 µg m-3 in summer, assuming anthropogenic emissions remained at present-day levels. These authors found that the driver for these increases was rising surface temperature, which influences both biogenic emissions and the rate of sulfate production.