Meteorological Modes of Variability for Fine Particulate Matter (PM2.5) Air Quality in the United States: Implications for PM2.5 Sensitivity to Climate Change

. We applied a multiple linear regression model to understand the relationships of PM 2 . 5 with meteorological variables in the contiguous US and from there to infer the sensitivity of PM 2 . 5 to climate change. We used 2004–2008 PM 2 . 5 observations from ∼ 1000 sites ( ∼ 200 sites for PM 2 . 5 components) and compared to results from the GEOS-Chem chemical transport model (CTM). All data were deseasonalized to focus on synoptic-scale correlations. We ﬁnd strong positive correlations of PM 2 . 5 components with temperature in most of the US, except for nitrate in the Southeast where the correlation is negative. Relative humidity (RH) is generally positively correlated with sulfate and nitrate but negatively correlated with organic carbon. GEOS-Chem results indicate that most of the correlations of PM 2 . 5 with temperature and RH do not arise from direct dependence but from covariation with synoptic transport. We applied principal component analysis and regression to identify the dominant meteorological modes controlling PM 2 . 5 variability, and show that 20–40 % of the observed PM 2 . 5 day-to-day variability can be explained by a single dominant meteorological mode: cold frontal passages in the eastern US and maritime inﬂow in the West. These and other synoptic transport modes drive most of the overall correlations of PM 2 . 5 with temperature and RH except in the Southeast. We show that interannual variability of PM 2 . 5 in the US Midwest is strongly correlated with cyclone frequency as diagnosed from a spectral-autoregressive analysis of the dominant meteorological mode. An ensemble of ﬁve realizations of 1996–2050 climate change with the GISS general circulation model (GCM) using the same climate forcings shows inconsistent trends in cyclone frequency over the Midwest (including in sign), with a likely decrease in cyclone frequency implying an increase in PM 2 . 5 . Our results demonstrate the need for multiple GCM realizations (because of climate chaos) when diagnosing the effect of climate change on PM 2 . 5 , and suggest that analysis of meteorological modes of variability provides a computation-ally more affordable approach for this purpose than coupled GCM-CTM studies.


Introduction
Air pollution is highly dependent on weather, and it follows that climate change could significantly impact air quality. The pollutants of most public health concern are ozone and fine particulate matter with diameter less than 2.5 µm (PM 2.5 ). Studies using chemical transport models (CTMs) driven by general circulation models (GCMs) consistently project a worsening of ozone air quality in a warming climate (Weaver et al., 2009). This finding is buttressed by observed correlations of ozone with temperature that are well reproduced by models Sillman and Samson, 1995;Rasmussen et al., 2012). By contrast, GCM-CTM studies of the effect of climate change on PM 2.5 show no consistency even in the sign of effect ). In previous work (Tai et al., 2010), we examined 3132 A. P. K. Tai et al.: Meteorological modes of PM 2.5 variability the observed correlations of PM 2.5 and its components in the US with meteorological variables as a means to understand PM 2.5 response to climate change. Here we develop this approach further to define meteorological modes of variability for PM 2.5 and interpret the observed correlations and modes using the GEOS-Chem CTM. We apply the Goddard Institute for Space Studies (GISS) GCM to illustrate how the modes enable effective diagnosis of the effect of climate change on PM 2.5 .
The uncertainty in assessing climatic effects on PM 2.5 reflects the complex dependence of different PM 2.5 components on meteorological variables. Higher temperatures can lead to higher sulfate concentrations due to faster SO 2 oxidation, but to lower nitrate and organic components due to volatility (Sheehan and Bowman, 2001;Aw and Kleeman, 2003;Dawson et al., 2007;Kleeman, 2008). Biogenic emissions of PM 2.5 precursors including agricultural ammonia, soil NO x , and volatile organic compounds (VOCs) increase with temperature and further complicate the PM 2.5temperature relationship (Pinder et al., 2004;Bertram et al., 2005;Guenther et al., 2006). Higher relative humidity (RH) promotes aqueous-phase sulfate production and ammonium nitrate formation (Koch et al., 2003;Liao et al., 2006;Dawson et al., 2007), but inhibits fires, which are important contributors to organic aerosols in many regions (Park et al., 2007;Spracklen et al., 2009). Changes in precipitation and in planetary boundary layer (PBL) depth have a consistent effect on PM 2.5 components but their projections in GCMs are highly uncertain .
Synoptic-scale transport should be an important factor driving the effect of climate change on PM 2.5 . Previous studies have used principal component analysis (PCA) to identify important meteorological modes of variability for PM 2.5 air quality (Cheng et al., 2007;Thishan Dharshana et al., 2010). Thishan Dharshana et al. (2010) found that as much as 30 % of PM 2.5 daily variability in the US Midwest is associated with passages of synoptic weather systems. Cold fronts associated with mid-latitude cyclone passages provide the dominant ventilation pathway for the eastern US (Cooper et al., 2001;Li et al., 2005). A general reduction in the frequency of these cyclones is expected as a result of greenhouse warming (Lambert and Fyfe, 2006;Christensen et al., 2007;Pinto et al., 2007), potentially leading to more frequent and prolonged stagnation episodes (Mickley et al., 2004;Murazaki and Hess, 2006). Leibensperger et al. (2008) found a strong anticorrelation between summer cyclone frequency and ozone pollution in the eastern US for 1980-2006, and further showed evidence of a long-term decline in cyclone frequency over that period that significantly hindered attainment of ozone air quality standards. Tai et al. (2010) projected a PM 2.5 enhancement of up to 1 µg m −3 in the Midwest from 2000-2050 climate change due to more frequent stagnation.
In this study, we first apply the GEOS-Chem global CTM to interpret the observed correlations between PM 2.5 com-ponents and meteorological variables in the contiguous US. As we will see, interpretation is complicated by the covariation of meteorological variables with synoptic transport. To address this issue, we use PCA and regression to determine the dominant meteorological modes of observed daily PM 2.5 variability in different US regions, and show how spectral analysis of these modes enables a robust estimate of the effect of climate change on PM 2.5 air quality.  Tai et al. (2010) show maps of the annual mean data for total PM 2.5 (1998-2008) and individual components (2000-2008). We do not discuss ammonium and elemental carbon (EC) here because ammonium is mainly the counter-ion for sulfate and nitrate, and the correlation patterns of EC with meteorological variables generally follow those of OC (Tai et al., 2010).

GEOS-Chem simulations
We used the GEOS-Chem global CTM to conduct fullyear simulations of coupled gas-phase and aerosol chemistry. GEOS-Chem (http://geos-chem.org) uses assimilated meteorological data from the NASA Global Earth Observing System (GEOS-5) with 6-h temporal resolution (3-h for surface variables and PBL depth), 0.5 • latitude by 0.667 • longitude (0.5 • × 0.667 • ) horizontal resolution, and 47 hybrid pressure-sigma vertical levels. We conducted GEOS-Chem simulations at three different horizontal resolutions: native 0.5 • × 0.667 • , 2 • × 2.5 • , and 4 • × 5 • . The coarser resolutions have been used previously with meteorological fields from the GISS GCM to investigate effects of climate change on air quality (Wu et al., 2008;Pye et al., 2009;Leibensperger et al., 2011a). For the native resolution simulation we used a nested continental version of GEOS-Chem over North America (140-40 • W, 10-70 • N) with 2 • × 2.5 • resolution for the rest of the world Zhang et al., 2011). The native simulation was conducted for one year (2006)  by Holtslag and Boville (1993) and implemented in GEOS-Chem by Lin and McElroy (2010). GEOS-Chem includes a fully coupled treatment of tropospheric ozone-NO x -VOCaerosol chemistry Liao et al., 2007). Gasaerosol phase partitioning of the sulfate-nitrate-ammoniumwater system is calculated using the ISORROPIA II thermodynamic equilibrium model (Fountoukis and Nenes, 2007). In-cloud SO 2 oxidation uses liquid water content information from the GEOS-5 archive (Fisher et al., 2011). Secondary organic aerosol (SOA) formation is computed with a standard mechanism based on reversible gas-aerosol partitioning of semi-volatile VOC oxidation products (Chung and Seinfeld, 2002). SOA precursors include isoprene, terpenes, and aromatic hydrocarbons .
Anthropogenic emissions of sulfur, ammonia and NO x emissions in the US are from the EPA 2005 National Emissions Inventory (http://www.epa.gov/ttn/chief/ net/2005inventory.html), and primary anthropogenic OC and EC emissions are from Cooke et al. (1999). Non-US anthropogenic emissions are described by Park et al. (2006). Biomass burning emissions of OC and EC are from the Global Fire Emissions Database (GFED v2) (Giglio et al., 2006). These emissions are included in the model as monthly averages and do not contribute to day-to-day variability of PM 2.5 . In contrast, soil NO x emissions (Yienger and Levy, 1995) and biogenic emissions of isoprene, terpenes, and methylbutenol (Guenther et al., 2006) are updated locally every three hours as a function of temperature, solar radiation, and precipitation. Scavenging of PM 2.5 by precipitation follows the scheme of Liu et al. (2001). Dry deposition follows a standard resistance-in-series scheme (Wesely, 1989) as implemented by Wang et al. (1998).
Maps of annual mean PM 2.5 concentrations from our simulation are included in the Supplement. Total PM 2.5 in GEOS-Chem is taken to be the sum of sulfate, nitrate, ammo-nium, OC and EC. Detailed evaluations of the GEOS-Chem simulation of PM 2.5 and its components over the US have been presented in a number of publications using observations from surface sites, aircraft, and satellites (Heald et al., 2006Park et al., 2006;van Donkelaar et al., 2006van Donkelaar et al., , 2008Fu et al., 2009;Drury et al., 2010;Leibensperger et al., 2011a;Zhang et al., 2012). These evaluations mainly focused on seasonal concentrations and showed no prominent biases. Here we will focus on the ability of the model to reproduce observed correlations of PM 2.5 with meteorological variables.

Multiple linear regression
We examined the correlations of PM 2.5 and its components with meteorological variables for 2004-2008 (EPA-AQS) and 2005-2007 (GEOS-Chem) by applying a standardized multiple linear regression (MLR) model: where y represents the deseasonalized daily PM 2.5 concentration (total PM 2.5 or individual component), x k represents the eight deseasonalized meteorological variables from GEOS-5 listed in Table 1,x k andȳ are the temporal means of x k and y, s k and s y are their standard deviations (see Supplement), β k is the dimensionless, normalized regression coefficient, and t is time. To compare observed with simulated correlations, we interpolate the EPA-AQS data onto the GEOS-Chem grid (Tai et al., 2010) and use the interpolated PM 2.5 fields for regression. The MLR model is applied to each individual grid cell for both the observed and simulated PM 2.5 fields. All data (x k and y) are deseasonalized and detrended by subtracting the 30-day moving averages from the original data so that x k =ȳ = 0. This allows us to focus on synoptic-scale variability and avoid aliasing from common seasonal or interannual variations. The standardized regression coefficients β k allow direct comparisons between the correlations of different PM 2.5 components with different meteorological variables (Kutner et al., 2004). The original regression coefficients β k * in units of µg m −3 D −1 , where D is the dimension of meteorological variable x k in Table 1, can be recovered by The observed coefficients of determination (R 2 ) for the MLR model have values ranging from 0.1 (in the west-central US where data are sparse) to 0.5 (in the Midwest and Northeast), agreeing with previous studies (Wise and Comrie, 2005;Tai et al., 2010). In addition to the standardized MLR analysis, we also conducted a stepwise MLR analysis with interaction terms as described by Tai et al. (2010). The interaction terms were generally found to be insignificant. We conducted the MLR analysis for the model at all three resolutions (0.5 • × 0.667 • , 2 • × 2.5 • , 4 • × 5 • ) and found the patterns of correlations to be similar. Figure 2 shows as an example (to be discussed later) the simulated and observed relationships of nitrate with temperature as measured by the recovered regression coefficient β 1 * in Eq. (2). In general, 2 • × 2.5 • and 4 • × 5 • regression results agree well with each other for all meteorological variables and all components. The native-resolution regression does not show as Table 1. Meteorological variables used for PM 2.5 correlation analysis a .
Variable Meteorological parameter Geopotential height at 850 hPa (km) East-west wind direction indicator cos θ (dimensionless) d x 8 North-south wind direction indicator sin θ (dimensionless) d a Assimilated meteorological data with 0.5 × 0.667 • horizontal resolution from the NASA Goddard Earth Observing System (GEOS-5). All data used are 24-h averages, and are deseasonalized and detrended as described in the text. b At 6 m above the surface (0.994 sigma level).
c Calculated from the horizontal wind vectors (u, v). (1). Values are for deseasonalized and detrended variables and are only shown when significant with 95 % confidence (p-value < 0.05). The right panels show the direct effects of temperature on sulfate, nitrate and OC as determined by applying a global +1 K temperature perturbation in the GEOS-Chem simulation, and normalizing the results to the standard deviations of deseasonalized concentrations and temperatures to allow direct comparison to β 1 .

Correlations with temperature
Figure 3 (left and middle panels) shows the observed and simulated relationships of sulfate, nitrate, and OC with temperature as measured by the standardized regression coefficient β 1 in Eq. (1). The relationships may reflect both a direct dependence of PM 2.5 on temperature and a covariation of temperature with other meteorological variables affecting PM 2.5 . To separate the two effects, we conducted a direct sensitivity analysis with GEOS-Chem by increasing temperatures by 1 K throughout the troposphere while keeping all other meteorological variables constant. The resulting sensitivities are shown in the right panels of Fig. 3, normalized to the standard deviations of deseasonalized concentrations and temperature to make them directly comparable to the standardized regression coefficients β 1 in the left and middle panels.
Sulfate in the observations shows a positive relationship with temperature over most of the US. The model is generally consistent with the observations but does not capture the Southwest maximum. Results from the direct sensitivity analysis, however, show a generally negative dependence of sulfate on temperature particularly in the West. This contrasts with a previous CTM sensitivity analysis by Dawson et al. (2007) that found a positive dependence of sulfate on temperature, though much weaker than the observed relationship (Tai et al., 2010). Dawson et al. (2007) attributed their result to faster SO 2 oxidation kinetics at higher temperature, but we find in GEOS-Chem that this is more than offset by the increased volatility of H 2 O 2 and SO 2 , slowing down the incloud aqueous-phase production of sulfate. In any case, it is clear from the model that the observed positive relationship of sulfate with temperature must reflect covariation of temperature with meteorological variables rather than a direct dependence. We elaborate on this in Sect. 4. Nitrate in the observations shows a negative relationship with temperature in the Southeast but a positive relationship in the North and the Southwest. The model reproduces these results except for the positive relationship in the Southwest. The negative relationship in the model is too strong in the South but the higher-resolution 0.5 • × 0.667 • simulation does not show such a bias (Fig. 2). The direct sensitivity of nitrate to temperature in the model is negative everywhere, with magnitude comparable to that found by Dawson et al. (2007), and reflecting the volatility of ammonium nitrate (Stelson and Seinfeld, 1982). We see from Fig. 3 that this direct dependence could account for most of the observed negative relationship of nitrate with temperature in the Southeast, but it is more than offset in the North by the positive association of temperature with southerly flow importing polluted air. The observed positive relationship of nitrate with temperature in the Southwest may reflect the temperature dependence of ammonia and fire emissions; in the model these emissions are specified as monthly means.
OC in the observations shows a positive relationship with temperature throughout the US, and the same is found in the model although the relationship is steeper. The direct sensitivity study in the model also shows a positive dependence of OC on temperature. Dawson et al. (2007) previously found a negative dependence due to OC volatility but did not consider the temperature dependence of biogenic VOC emissions, which is included in our analysis and more than offsets the volatility effect. Day and Pandis (2011) similarly found an increase in OC at higher temperatures mainly due to increased VOC emissions. We see from Fig. 3 that the direct temperature dependence may be a significant contributor the positive relationship between OC and temperature in the Southeast, where biogenic emissions are particularly high, but it has little effect elsewhere. Figure 4 shows the observed and simulated correlations of sulfate, nitrate, and OC with RH, expressed as the standardized regression coefficient β 2 in Eq. (1). The relationships are generally positive for sulfate and nitrate both in the observations and the model. The OC-RH relationship is generally negative with some model biases in the Great Plains and Midwest. Results from a model perturbation simulation similar to that for temperature are also shown in Fig. 4, indicating negligible direct dependence of sulfate and OC on RH, but a significant positive relationship for nitrate due to more favorable ammonium nitrate formation at higher RH (Stelson and Seinfeld, 1982). The direct positive sensitivity of nitrate in the southeastern coast is offset by the negative influence from the association of high RH with clean marine air, leading to the weak overall correlation there. Figure 5 shows the observed and simulated relationships of total PM 2.5 with precipitation and wind speed as measured by β 3 and β 6 in Eq. (1). Similar effects are found for all individual PM 2.5 components (Tai et al., 2010). The observations show strong negative relationships reflecting aerosol scavenging and ventilation. These are generally well captured by the model. The precipitation effect appears to be primarily driven by large-scale rather than convective precipitation in the US. Fang et al. (2011) similarly illustrated the dominance of large-scale precipitation in wet scavenging of soluble pollutants.

Major meteorological modes controlling PM 2.5 variability
Results from the previous section show that much of the correlation of PM 2.5 with individual meteorological variables is driven by covariance between meteorological variables, with an apparent major contribution from synoptic transport. To resolve this covariance we turn to principal component analysis (PCA) of the meteorological variables to identify the meteorological modes controlling PM 2.5 variability.

Principal component analysis and regression
We conducted a PCA for the 2004-2008 GEOS-5 data by averaging spatially over each region of Fig. 1 where X k represents the regionally averaged GEOS-5 variable,X k and s k the temporal mean and standard deviation of X k ,, and α kj the elements of the orthogonal transformation matrix. Each PC represents a distinct meteorological regime or mode. We identified the nature of meteorological mode by examining the values of α kj in Eq. (3). PCs with high |α kj | values (e.g., greater than 0.3 and topping the other |α kj | values) for geopotential height, pressure tendency, and wind direction are presumably associated with synopticscale weather systems, and can be referred to as synoptic transport modes. We then followed U j (t) day by day and visually examined the corresponding weather maps for multiple months during 2004-2008. From this we assigned a generalized meteorological feature for a given PC when the same feature could be associated with the majority of peaks and troughs of U j (t). The PCs are ranked by their variances, usually with the leading three or four PCs capturing most of the meteorological variability. For instance, in the eastern US, a single mode representing cyclone and cold frontal passages (discussed further in Sect. 4.2) typically accounts for ∼20 % of total meteorological variability.   We then applied a principal component regression (PCR) model to correlate observed and simulated PM 2.5 concentrations with the eight PCs for each region

Atmos
where Y represents the regionally averaged PM 2.5 concentration, γ j the PC regression coefficients, andȲ and s Y the temporal mean and standard deviation of Y . The ratio of regression to total sum of squares (SSR j /SST) for each PC is calculated by where the summation is over the entire time series Y (t) and U j (t). This ratio quantifies the fraction of variance of PM 2.5 that can be explained by a single PC. From Eqs. (3) and (4), the fraction (f k ) of the overall correlation of PM 2.5 with a given meteorological variable X k (e.g., in Figs. 4 through 6) that is associated with a particular PC can be estimated by where the summation is over the m PCs that have a significant effect on PM 2.5 (p-value < 0.01). Here the denominator represents the total effect of X k on PM 2.5 that is equivalent to a regionally averaged version of β k in Eq. (1). The PCR model was applied to both the full-year data and to seasonal subsets. Figure 6 shows as an example the dominant meteorological mode contributing to total PM 2.5 variability in the Midwest as determined by the highest SSR j /SST ratio in Eq. (3), consisting of low temperature, high precipitation, low and rising pressure, and strong northwesterly winds. From weather maps we can verify that high positive values of this PC represent the center of an eastward-propagating mid-latitude cyclone with a precipitating cold front at the southwest tail end. High negative values indicate the "opposite" regime -warm and dry stagnant condition at the tail end of an anticyclone. Figure 6 (top and bottom right) shows, for instance, that as U j (t) rose from a minimum to maximum between 28 and 30 January 2006 in the Midwest, a mid-latitude cyclone was approaching and the associated cold front swept over the region bringing down total PM 2.5 by 9 µg m −3 . Figure 7 shows as another example the dominant meteorological mode of PM 2.5 variability in California, demonstrating again a strong anticorrelation between the time series of this mode and PM 2.5 concentrations (r = −0.80). This mode has similar meteorological composition to that in Fig. 6 except for wind direction. Positive phases of this mode represent ventilation by cold maritime inflows associated with synoptic disturbances, whereas negative phases represent warm, stagnant conditions associated with highpressure systems. The bottom panel shows, for instance, that between 6 and 8 January 2005, a precipitating maritime inflow reduced PM 2.5 by 16 µg m −3 .  Table 2 summarizes the characteristics of the dominant PC controlling PM 2.5 variability for five selected regions. In the eastern US (Northeast, Midwest and Southeast), the observed dominant modes resemble that for the Midwest described above (Fig. 6). In the Northeast, another mode representing southwesterlies associated with high pressure over the western North Atlantic is equally important. In the Pacific Northwest, the dominant mode resembles that for California (Fig. 7). In general, the PCR results illustrate the importance of synoptic-scale transport in controlling the observed daily variability of PM 2.5 . As shown in Table 2, this control appears to be well represented in GEOS-Chem, supporting the ability of the model to describe the variability in PM 2.5 associated with this transport.

Dominant meteorological modes of PM 2.5 variability
Using Eq. (6), we find overall that the synoptic transport modes account for more than 70 % of the observed correlations of PM 2.5 components with temperature in the Northeast and Midwest. This reflects the association of elevated temperature with southerly flow and stagnation. In the Southeast, however, we find that more than 60 % of the observed correlations of nitrate and OC with temperature and RH arise from a single non-transport mode consisting of low temperature and high RH. Nitrate has a positive dependence on that mode because of ammonium nitrate thermodynamics, while OC has a negative dependence reflecting biogenic VOC emissions and the occurrence of fires. The weaker importance 3140 A. P. K. Tai et al.: Meteorological modes of PM 2.5 variability of transport in driving the nitrate-temperature relationship in the Southeast likely reflects the lower frequency of cold fronts. In California, the transport and non-transport modes are comparably important in shaping the observed correlations of PM 2.5 components with temperature and RH.

Cyclone frequency as a metric for climate change effect on PM 2.5
Mid-latitude cyclones and their associated cold fronts are known to provide the dominant year-round mechanism for ventilating the US Midwest and Northeast (Cooper et al., 2001;Li et al., 2005), and they emerge in our analysis of Sect. 4 as the dominant meteorological mode of PM 2.5 variability. Previous studies diagnosing cyclone frequency have relied on identifying local pressure minima (Mickley et al., 2004;Lambert and Fyfe, 2006;Lang and Waugh, 2011) or used storm tracking algorithms (Geng and Sugi, 2001;Bauer and Del Genio, 2006;Bengtsson et al., 2006). Here we diagnose cyclone frequency by applying a fast Fourier transform (FFT) to the time series of the dominant Midwest PC representing cyclone and frontal passages as shown in Fig. 6. We use 1999-2010 meteorological data from the NCEP/NCAR Reanalysis 1 (Kalnay et al., 1996;Kistler et al., 2001), which provides a longer record than GEOS-5. PCA of the NCEP/NCAR data yields essentially the same meteorological modes as GEOS-5. Figure 8 (gray thin line) shows the FFT spectrum for the dominant cyclone mode in the Midwest for 1999-2010. The low-frequency structure (with periods > 20 d) is an artifact of the 30-day moving average applied to the meteorological data to remove seasonality. We smooth the time series with a second-order autoregressive (AR2) filter (Wilks, 2006), indicating a median spectral frequency of 52 a −1 (cyclone period of about 7 days). We applied the spectral-autoregressive method above to find the median cyclone frequencies and periods for individual years of the 1999-2010 record. Figure 9 shows the time series of annual mean anomalies in total PM 2.5 concentrations and cyclone periods for the Midwest, where the correlation is strongest (r = 0.76) corresponding to a PM 2.5 -tocyclone period sensitivity of 0.94 ± 0.43 µg m −3 d −1 (95 % confidence interval). Leibensperger et al. (2008) previously found a strong interannual correlation of summer ozone with cyclone frequency in the Northeast using the 1980-2006 record of NCEP/NCAR data. Our analysis does not show the same for PM 2.5 in this region, possibly because of the short record (12 years) available for PM 2.5 . Cyclone frequencies found by Leibensperger et al. (2008) are generally lower, possibly because their storm-tracking algorithm may neglect weaker cyclones and fronts.
The strong interannual correlation of PM 2.5 with cyclone frequency, at least in the Midwest, encourages the use of cyclone frequency as a metric to diagnose the effect of climate change on PM 2.5 . We used for this purpose an ensemble of five realizations of 1950-2050 climate change generated by Leibensperger et al. (2011b) with the GISS GCM III (Rind et al., 2007) applied to the IPCC A1B scenario (Nakicenovic and Swart, 2000) and including time-dependent aerosol radiative forcings. For each realization we examined the change in median cyclone frequency between the presentday (1996)(1997)(1998)(1999)(2000)(2001)(2002)(2003)(2004)(2005)(2006)(2007)(2008)(2009)(2010) and the future (2036)(2037)(2038)(2039)(2040)(2041)(2042)(2043)(2044)(2045)(2046)(2047)(2048)(2049)(2050), by applying the spectral-autoregressive method to the dominant cyclone PC for each 15-year time series, and using a Monte Carlo method to diagnose the probability distribution and significance of the change based on variability of the AR2 parameters. Three out of the five realizations indicated statistically significant decreases in cyclone frequencies between 1996-2010 and 2036-2050 of −3.2, −3.4 and −1.5 a −1 (pvalue < 0.05). One realization showed a significant increase of 2.7 a −1 and another showed no significant change. Figure 10 shows the combined probability distribution of cyclone frequency change in the Midwest from all five realizations and the corresponding responses of annual mean PM 2.5 based on the PM 2.5 -to-cyclone period sensitivity reported above, indicating a roughly 70 % probability of reduced cyclone frequency and elevated PM 2.5 in the Midwest by 2050. This corresponds to a mean decrease in cyclone frequency of −1.1 ± 4.8 a −1 and a resulting increase in annual mean PM 2.5 of 0.13 ± 0.60 µg m −3 .
Previous GISS-GEOS-Chem GCM-CTM studies of the effects of 2000-2050 climate change on PM 2.5 air quality projected a mean increase of 0.1-0.5 µg m −3 in the Midwest in the 2050 climate based on one GCM realization (Pye et al., 2009;Lam et al., 2011). Their estimates are within the range of our projection from the cyclone frequency trend alone. However, the large variability of the cyclone trends (including in sign) across five realizations of the same GCM underscores the imperative need for multiple realizations in diagnosing the effect of climate change on PM 2.5 air quality. All GCM-CTM studies in the literature reviewed by Jacob and Winner (2009) have used single climate realizations and this may partly explain the inconsistency in their results.
Other climatic factors than cyclone and frontal frequency may also affect future PM 2.5 air quality in the US. Mean temperature increases may be particularly important for the Southeast as discussed previously. Changes in precipitation and PBL depth are obviously important. As scavenging within a precipitating column is highly efficient (Balkanski et al., 1993), precipitation frequency, often modulated by synoptic weather, may be more relevant as a predictor than climatological mean precipitation.

Conclusions
Projecting the effects of climate change on PM 2.5 air quality requires an understanding of the dependence of PM 2.5 on meteorological variables. We used here a multiple linear regression model to correlate both observed (EPA-AQS) and  simulated (GEOS-Chem) daily mean concentrations of total PM 2.5 and its major components with a suite of meteorological variables in the contiguous US for 2004-2008. All data were deseasonalized to focus on synoptic correlations. We applied principal component analysis (PCA) and regression to identify the dominant meteorological modes controlling PM 2.5 variability, and showed how trend analysis for these modes can be used to estimate the effects of climate change on PM 2.5 .
We observe strong positive correlations of all PM 2.5 components with temperature in most of the US, except for nitrate in the Southeast where the correlation is negative. A temperature perturbation simulation with GEOS-Chem reveals that most of the correlations of PM 2.5 with temperature do not arise from direct dependence on temperature but from covariation with synoptic transport. Exceptions are ni-34 1 Figure 9. Anomalies of annual mean PM2.5 concentrations and median cyclone periods for the 2 US Midwest (Fig. 1). 3 4 Fig. 9. Anomalies of annual mean PM 2.5 concentrations and median cyclone periods for the US Midwest (Fig. 1). trate and OC in the Southeast, where the direct dependence of ammonium nitrate thermodynamics and biogenic VOC emissions on temperature contributes significantly to the correlations. RH is generally positively correlated with sulfate and nitrate but negatively correlated with OC; the correlations also appear to be mainly driven by covariation of RH with synoptic transport. Total PM 2.5 is strongly negatively correlated everywhere with precipitation and wind speed.
We find from the PCA and regression that 20-40 % of the observed PM 2.5 day-to-day variability in different US regions can be explained by a single dominant synoptic meteorological mode: cold frontal passages in the eastern US and maritime inflow in the West. These and other transport modes are found to contribute to most of the overall correlations of different PM 2.5 components with temperature and RH except in the Southeast.
We show that the interannual variability of annual mean PM 2.5 in the Midwest for 1999-2010 is strongly correlated with cyclone frequency as diagnosed from a spectralautoregressive analysis of the dominant meteorological mode of variability, with a PM 2.5 -to-cyclone period sensitivity of 0.9 ± 0.4 µg m −3 d −1 . We conducted an ensemble of five realizations of 1996-2050 climate change using the GISS GCM III with A1B greenhouse and aerosol forcings. Three of these found a significant decrease in cyclone frequency over the US Midwest, one found no significant change and one found a significant increase. From this ensemble we derive a likely increase in annual mean PM 2.5 of 0.13 ± 0.60 µg m −3 in the Midwest in the 2050s climate. This is consistent with previous GCM-CTM studies using the same GCM and suggests that cyclone frequency may be a major driver of the effect of climate change on PM 2.5 air quality. However, the variability of cyclone trends (including in sign) across multiple realizations of the same GCM with identical forcings demonstrates the importance of multiple climate change realizations in GCM-CTM studies because of climate chaos. All GCM-CTM studies to date have used single realizations because of computational expense, and this may partly explain the wide inconsistencies in their projections of PM 2.5 response to climate change. The climate trend analysis in this study, using the Midwest as an illustration, is preliminary. A comprehensive analysis using outputs from various GCMs will be the topic of a future paper.