WRF-CMAQ two-way coupled system with aerosol feedback

Abstract. Air quality models such as the EPA Community Multiscale Air Quality (CMAQ) require meteorological data as part of the input to drive the chemistry and transport simulation. The Meteorology-Chemistry Interface Processor (MCIP) is used to convert meteorological data into CMAQ-ready input. Key shortcoming of such one-way coupling include: excessive temporal interpolation of coarsely saved meteorological input and lack of feedback of atmospheric pollutant loading on simulated dynamics. We have developed a two-way coupled system to address these issues. A single source code principle was used to construct this two-way coupling system so that CMAQ can be consistently executed as a stand-alone model or part of the coupled system without any code changes; this approach eliminates maintenance of separate code versions for the coupled and uncoupled systems. The design also provides the flexibility to permit users: (1) to adjust the call frequency of WRF and CMAQ to balance the accuracy of the simulation versus computational intensity of the system, and (2) to execute the two-way coupling system with feedbacks to study the effect of gases and aerosols on short wave radiation and subsequent simulated dynamics. Details on the development and implementation of this two-way coupled system are provided. When the coupled system is executed without radiative feedback, computational time is virtually identical when using the Community Atmospheric Model (CAM) radiation option and a slightly increased (~8.5%) when using the Rapid Radiative Transfer Model for GCMs (RRTMG) radiation option in the coupled system compared to the offline WRF-CMAQ system. Once the feedback mechanism is turned on, the execution time increases only slightly with CAM but increases about 60% with RRTMG due to the use of a more detailed Mie calculation in this implementation of feedback mechanism. This two-way model with radiative feedback shows noticeably reduced bias in simulated surface shortwave radiation and 2-m temperatures as well improved correlation of simulated ambient ozone and PM2.5 relative to observed values for a test case with significant tropospheric aerosol loading from California wildfires.

changes; this approach eliminates maintenance of separate code versions for the coupled and uncoupled systems. The design also provides the flexibility to permit users: (1) to adjust the call frequency of WRF and CMAQ to balance the accuracy of the simulation versus computational intensity of the system, and (2) to execute the two-way coupling system with feedbacks to study the effect of gases and aerosols on short wave 15 radiation and subsequent simulated dynamics. Details on the development and implementation of this two-way coupled system are provided. When the coupled system is executed without radiative feedback, computational time is virtually identical when using the Community Atmospheric Model (CAM) radiation option and a slightly increased (∼8.5 %) when using the Rapid Radiative Transfer Model for GCMs (RRTMG) radiation 20 option in the coupled system compared to the offline WRF-CMAQ system. Once the feedback mechanism is turned on, the execution time increases only slightly with CAM but increases about 60 % with RRTMG due to the use of a more detailed Mie calculation in this implementation of feedback mechanism. This two-way model with radiative feedback shows noticeably reduced bias in simulated surface shortwave radiation and 25 2 m temperatures as well improved correlation of simulated ambient ozone and PM 2.5 relative to observed values for a test case with significant tropospheric aerosol loading from California wildfires.

Introduction
3-D chemical transport models (CTMs) that are used in air quality research and regulatory applications are driven by 3-D meteorological fields provided by a priori runs of a meteorology model. Historically, the CTMs and meteorological models were developed over several decades along independent tracks with little regard for computational, nu-5 merical, or even scientific consistency between the two modeling systems. In recent years, however, there have been several efforts to combine meteorological and chemical transport models into single interactive systems (Grell and Baklanov, 2011). A primary driver for this trend has been the need to include the direct and indirect feedback effects of gases and aerosols on radiative forcing. While these feedback effects 10 are mainly important for climate applications, it is becoming evident that they have substantial effects on local meteorology and air quality in polluted regions (Jacobson et al., 1996;Mathur et al., 1998;Xiu et al., 1999). Zhang (2008) has provided an overview of several coupled meteorology-chemistry models including the WRF/chem (Grell et al., 2005) model in which chemistry has been added into the Weather Research and Fore- 15 casting model (Skamarock et al., 2008) at the science process level. Another approach is to couple historically independent meteorology and chemical transport models into a single executable. Advantages of this approach include maintaining consistency with existing separate loose coupled meteorology-chemistry systems that are being continuously and extensively applied and evaluated. Furthermore, the numerical and com- 20 putational techniques employed in meteorology models and CTMs differ considerably because of the greater need for strict mass conservation and positive-definiteness of transported scalars in the CTM. Also, CTMs generally use fractional integration of various processes while meteorology models use time split integration of all process rates. 2000 users from 90 different countries. CMAQ has been and continues to be extensively used to provide guidance in rulemaking such as CAIR (Clean Air Interstate Rule, http://www.epa.gov/cair/), by state and local agencies for air quality management analyses such as SIP (State Implementation Plan), by academia and industry for studying relevant atmospheric processes and model applications. CMAQ has also been adapted 5 into the real-time US National Air Quality Forecasting system (AQF) (Otte et al., 2005) and has been running operationally at National Weather Service since 2003 and was recently deployed for forecasting air quality for the 2010 Shanghai World Expo (Wang et al., 2010). In general, meteorological models are not built for air quality simulation purposes. 10 Hence, the meteorological model might not have the same map projection, coordinate system and grid format, and layer structure as CMAQ. The CMAQ model uses the Meteorology-Chemistry Interface Processor (MCIP)  to bridge this gap by providing transformed CMAQ-ready meteorological data. The transformation includes unit conversion, format conversion, vertical grid resolution related interpo- 15 lation, as well as calculations to create additional diagnostic variables that are required in CMAQ but not available in the meteorology model output. Typically, MCIP produces hourly meteorological data, based on storage requirement considerations, as input to CMAQ. The flow of information in this one-way coupled system is ( Fig. 1a): run a meteoro-20 logical model, like the Fifth-Generation Pennsylvania State University-National Center for Atmospheric Research Mesoscale Model (MM5) (Grell et al., 1994) or the Weather Research and Forecasting (WRF) model (Michalakes et al., 2005), process the meteorological model's output using MCIP, then run the CMAQ air quality model using the MCIP output. This whole coupling approach has been widely used in the research 25 community as well as in the real-time National Air Quality Forecasting system (Otte et al., 2005), it has several potential shortcomings. First, the integration time step of CMAQ is much finer than the typical hourly available meteorological data. Interpolation is used to handle this issue; however, interpolation accuracy is a problem for Introduction To address these potential shortcomings, we coupled the WRF meteorology model 15 and the CMAQ model to create a two-way coupled modeling system to facilitate feedbacks between chemistry and meteorology. Section 2 provides an overview of the scientific components in the model. Section 3 describes the software considerations for developing a flexible and efficient coupled modeling system, including domain decomposition and design issues. Preliminary results are presented in Sect. 4 while Sect. 5 20 summarizes the main results and presents a brief discussion of future work.

Overview of scientific components of the coupled model
The coupled modeling system consists of three components: the WRF meteorology model, the CMAQ model and the coupler. A high level view of the system is depicted in Fig. 1b

WRF
The Advanced Research WRF version 3 (WRF-ARW) is a state-of-the science mesoscale meteorology model (Skamarock et al., 2008) that is typically configured with horizontal grid resolutions ranging from 1-30 km. The dynamical equations numerically solved by the WRF-ARW model are fully compressible, Euler nonhydrostatic, 5 and are conservative for all scalar variables. The prognostic variables are the three velocity components, perturbation potential temperature, perturbation geopotential, and perturbation dry air surface pressure. Additional prognostic variables depend on the model physics options and include turbulent kinetic energy, water vapor mixing ratio, and several cloud microphysical scalars such as cloud water/ice mixing ratio, rain/snow 10 mixing ratio, and graupel mixing ratio. Both the WRF-ARW and the CMAQ model can be configured to use the exact same grid configurations and coordinate systems. Thus, no spatial interpolation of either meteorological or chemical data is required.

CMAQ
CMAQ version 4.7.1 (Foley et al., 2010) is a comprehensive atmospheric chemistry 15 and transport model that numerically integrates a set of independent chemical conservation of mass equations on a series of 3-D nested Eulerian grid meshes. The CMAQ model employs operator splitting to modularize the various physical and chemical processes including: subgrid turbulent vertical transport, horizontal and vertical advection, horizontal diffusion, cloud processes (i.e., aqueous chemistry, subgrid convective trans-Introduction

Coupler
The coupler is used to link these two models together and serves as an inter-model translator. The design and functionality will be described in the next section. The coupler also includes software (aqprep) to transfer meteorological fields from WRF to CMAQ and to transfer aerosol predictions from CMAQ back to WRF (feedback).

5
A subroutine called aqprep prepares meteorological fields in forms compatible for use in CMAQ's generalized coordinate formulation. The preparation includes extracting data such as pressure and wind field directly from WRF and calculating additional variables that are used in CMAQ such as the vertical coordinate Jacobian and the fractional area of each land use category in each grid cell. In essence, aqprep includes 10 the functionality currently embodied by the MCIP  preprocesses in the offline WRF CMAQ system.
An important benefit of two-way coupling between meteorology and air quality models is the ability to use aerosol fields simulated by the air quality model to affect processes in the meteorology model. The first feedback implemented in the WRF-CMAQ system is the direct effects by which chemical species calculated in CMAQ are transferred to WRF for calculating their influence on radiation computed in WRF. In addition to the data coupling described in the next section, implementation of direct feedback requires new subroutine for the calculation of the aerosol optical properties: extinction optical depth, single scattering albedo, asymmetry parameter, and forward scat-20 tering fraction, for short-wave spectral bands (19 bands in Community Atmosphere Model (CAM) and 14 bands in Rapid Radiative Transfer Model for GCMs (RRTMG)). The aerosol chemical species calculated by CMAQ are combined into five groups: water-soluble, insoluble, sea-salt, black carbon, and water. The refractive indices for these species are taken from the OPAC (Optical Properties of Aerosols and Clouds) 25 (Hess et al., 1998) database using linear interpolation to the central wavelength of the RRTMG wavelength intervals. These direct feedbacks tend to reduce SW radiation reaching the ground in areas of high aerosol loading, thereby reducing daytime surface temperatures, as shown in Sect. 4. In addition, absorbing aerosols, such as black carbon, tend to warm the air in layers with high concentrations. There are also secondary effects on PBL heights and cloud properties. An efficient numerical quadrature method calculates the extinction and scattering coefficients along with the asymmetry factor by integrating the Mie codes over the log-5 normal size distributions representing the Aitken, accumulation, and coarse modes produced by CMAQ.

Software considerations for the coupler
An air quality model may utilize a different map projection, time integration, grid orientation, grid cell size, and/or vertical coordinate that is different from its meteorological 10 driver. In order to facilitate communication between models to exchange relevant information that is usable by each individual model, a coupler is devised.

Modeling domain
Both the WRF-ARW and CMAQ use the Arakawa C horizontal grid staggering, and in the coupler no spatial interpolation of meteorological or chemical data is required. In 15 addition, both WRF and CMAQ use the same map projection, so the coupler inherits the map projection from WRF. The vertical coordinate in WRF is a hydrostatic sigmapressure, and CMAQ uses a modified, generalized form of that coordinate. Unlike in the offline WRF-CMAQ system, the coupler must use the same number and configuration of vertical layers (i.e., no layer collapsing, Otte and Pleim, 2010). Figure 2 illustrates the 20 typical domain configurations of the WRF-CMAQ coupled system; in this the chemistrytransport calculations are performed for a sub-domain of the larger WRF domain to avoid numerical effects associated in the vicinity of the WRF domain boundaries. In general, users can choose how many grid cells to trim off at run time, but five grid cells is the recommended minimum. Users can define any CMAQ domain as long as  Fig. 2. In addition, the user is required to provide the value of delta x and delta y which defines the lower left corner of the CMAQ domain relative to the WRF domain. Because the WRF model is the driver and CMAQ is called as a subroutine within WRF, a global timer that is based on the WRF advection time step is used to synchronize WRF and CMAQ in the coupled 5 system.

Domain decomposition
The main task of the coupler is to transfer needed data between these two models correctly. Both models were designed to run in a parallel computing environment. WRF supports MPI and OpenMP but the current version of CMAQ only supports MPI. 10 They both use domain decomposition as the basic parallelization approach. However, the details of the decomposition are quite different in both models. Runtime System Library (RSL) (Michalakes, 1994) and RSL-lite (Michalakes, 1998), which both handle high-level stencil and inter-domain communication, irregular domain decomposition, automatic local/global index translation, distributed I/O, and dynamic load balancing, 15 are used in WRF to parallelize the code. RSL-lite is a bit faster than RSL. Both RSL and RSL-lite were used in WRF version 2 implementation, but RSL has been removed since version 3. Besides performance, the main difference between RSL and RSL-lite is the domain decomposition algorithm. Since RSL has been removed, the description of the decom-20 position algorithm is focused on RSL-lite. The mapping between processor and sub domain is in row-wise fashion. The starting point and order of assigning the remainder row elements is at the bottom and then top, and then bottom until all the remaining elements are distributed (Fig. 3a). Similarly, for the column dimension, it starts at the left and then right and moving towards the center. 25 In CMAQ, the domain decomposition uses the same processor and sub-domain mapping as in WRF (Fig. 3b), but the starting point and order of assigning the remainder column or row elements is different. The remainder elements starts from the 2425 Introduction bottom and moves toward the top for row dimension and starts from left and moves toward the right for column dimension, rather than alternating inward from the top/bottom and left/right. When CMAQ is executed, users can choose a particular processor configuration based upon the number of processors allocated. For instance, if the number of avail-5 able processors is 16, the user can choose between a 4×4, 8×2, 2×8, 16×1 or 1×16 processor configuration. In WRF, the processor configuration with a "square" orientation, is the default but it is user-definable. In the WRF-CMAQ coupled system, CMAQ's processor configuration is inherited from WRF. The coupler's main task is to compute the mapping between the WRF and CMAQ domains with respect to each sub domain 10 with consideration of the position of the CMAQ domain relative to the WRF domain. This mapping information will be used for data transfer between these two models in the forward and feedback steps.

Data exchange
We have considered different tools, such as ESMF (Earth System Modeling Frame-15 work) (http://www.earthsystemmodeling.org/), Cpl6 (Craig et al., 2005) and MCT (Larson et al., 2005), for data exchange between the two models. CMAQ uses the IOAPI3 (Input/Output Applications Programming Interface version 3, http://www.baronams. com/products/ioapi/AVAIL.html) to handle physical file I/O. With the consideration of minimal code change, we chose to use IOAPI3 for the coupling. The actual data trans-20 fer is performed in memory through IOAPI3 buffered files. IOAPI3 is third party software written by Baron Advanced Meteorological Systems (BAMS). It is written to handle various types of files: volatile real files which are used in CMAQ to deal with I/O, using the netCDF format; buffered virtual files which facilitate data exchange within the same program through memory; coupling-mode virtual files which use the PVM3.4 mailbox 25 mechanism to exchange data among models executed concurrently; and native-binary real files which are the same as the volatile real files except the file is stored in native binary format instead of netCDF. The type of file, whether it is a volatile real file or a buffered virtual file, used in the application is determined at run time. The capability of handling various file types within IOAPI3 is transparent to the user application code; hence, code modification is not needed. The coupler will create/open the same number of files as in the offline run that uses physical files. In the stand-alone CMAQ model, each processor is able to 5 access the entire file and only extract relevant data for the sub-domain portion from the file. In addition, hourly meteorological input is interpolated to the current time step in various science processes within CMAQ throughout the execution. In order to make the same code work for buffered files while reducing memory consumption, each buffered file is exactly the same size as a sub-domain and corresponds to that sub-domain only. 10 Two time steps of data are stored in each circular buffered file.

System structure
The coupler consists of two major components: aqprep and feedback. The prepared data is placed in the corresponding buffered files which have similar attributes as the physical files used in the uncoupled stand-alone CMAQ. In CMAQ, the same IOAPI3 15 calling interface as in the stand alone model is used to access these buffered files. This design provides flexibility to read and write either buffered or disk files, enabling consistent coupled and uncoupled modeling paradigm. Thus an inherent advantage of using the IOAPI3 to handle the file format in the coupler is that minimal code changes are needed in CMAQ. The feedback part is called within the aerosol module in CMAQ 20 and computes several variables that are needed for direct aerosol feedback to the WRF radiation module. Various information such as coarse mode diameter and Aitken mode natural log of standard deviation, is used to compute soluble mass, elementary carbon mass and other parameters.
WRF integrates at a very fine time step, e.g., one minute for 12 m horizontal grid cell 25 size. In CMAQ each physical process, (e.g., transport and chemistry), has a different time step requirement that is based on individual process characteristic time scales and numerical stability criteria. As a stand-alone model, CMAQ determines the minimum 2427 Introduction is 30 s, the CMAQ time step will be two minutes. Consequently, the computational burden for the coupled system increases substantially as the CMAQ calling frequency increases. The non-linear increase in computational intensity is related to inherent nonlinearity in atmospheric processes and numerical solution of the governing equations. Figure 4 depicts the calling sequence for the coupled system. In general, CMAQ is 10 called after an aqprep step except the very first time. This implementation ensures two steps of WRF data are always available in case temporal interpolation of meteorological information is needed in CMAQ. The feedback step takes place in CMAQ within the aerosol module. Since CMAQ is a subroutine of WRF it occupies a portion of a WRF step. There are two versions of the radiation calculation within the two-way coupling 15 model: one comes with the WRF code (Ra) and the other was modified to include feedback capability (Ra ). Ra is not called until feedback information is available. That is why Ra is used for the very first step.

Software modification in WRF and CMAQ to support the coupled system
In the coupled system, CMAQ is implemented as a subroutine in WRF. Since CMAQ 20 is a community model with a wide user community which has used the model in an offline mode, all the coupling related functions are encapsulated in Fortran 90 modules that will not be invoked when running CMAQ in a stand-alone offline mode. Thus, the same version of CMAQ can be consistently used both in an offline or coupled system mode. and eliminates software maintenance of separate model versions for on-line and offline configurations. The aerosol optical depth for nine selected wavelength bands estimated from the simulated aerosol distribution is added to the WRF output for examination and validation purposes. This requires nine new variables in the WRF Registry. Two routines 5 in WRF were also modified for the coupled system: solve em.F (to invoke CMAQ) and the radiation calculations (to add the nine variables which are passed into various parts of the radiation calculation and to process aerosol feedback). These modifications can be easily ported to newer versions of WRF as its science is updated.
Run time switches have been implemented to disable the aerosol feedback in the 10 coupled system as well as to run WRF in stand-alone mode (i.e., without calling CMAQ). These options provide flexibility to perform sensitivity studies on the effects of the feedback mechanism and the coupling

Preliminary performance of the coupled system
The performance of the WRF-CMAQ two-way coupled system is evaluated for both 15 computational impact and scientific advancement. In order to be of general use to the CMAQ community, the coupled system must not add such a computational burden that it would become prohibitive for the average user to run. In addition, it is desirable that the coupled system scales well computationally as processors are added to the configuration. Of equal importance is the need for the coupled system to demonstrate 20 a scientific advantage beyond what could be achieved by using the WRF and CMAQ models sequentially.

Computational performance
We conducted a series one-day simulation (1 August  encompassing Eastern US, discretized with 12 km horizontal grid spacing while the vertical extent ranging from the surface to 50 hPa was discretized using 34 layers of variable thickness. Here, the WRF domain size is 290×251 grid points, and the CMAQ domain size is 279 × 240 grid cells, which allows for the five-cell boundary along the perimeter of the WRF domain to be excluded from CMAQ (Fig. 2). WRF-ARW ver-5 sion 3.1 was built with CMAQ version 4.7.1 to form the WRF-CMAQ two-way coupled system. Initial and lateral boundary conditions for WRF were derived from a combination of North American Mesoscale (NAM) model analyses and forecasts at 3 h intervals that were developed by the National Center for Environmental Prediction and obtained from the National Climatic Data Center. In WRF, the model options included the WRF 10 single-moment 6-class (WSM6) microphysics scheme (Hong et al., 2004), version 2 of the Kain-Fritsch (KF2) cumulus cloud parameterization (Kain, 2004), the Asymmetric Convective Model version 2 (ACM2) for the planetary boundary layer (Pleim, 2007a, b), and the Pleim-Xiu land-surface model (Xiu and Pleim, 2001) with soil moisture and temperature nudging (Pleim and Xiu, 2003;Pleim and Gilliam, 2009 ) radiation schemes were tested and run in the coupled system to contrast the simulated impact of radiative feedback from CMAQ to WRF using multiple radiation schemes. Also in WRF, analysis nudging was included for temperature and humidity above the PBL and for winds at all model levels (Stauffer 20 et al., 1991). In CMAQ, the CB05 chemical mechanism was used. The same subgrid vertical transport of meteorological and chemical species was used in both WRF and CMAQ following ACM2. For these tests, a one-minute WRF time step was used, and CMAQ was called every five WRF time steps (ratio of 5:1). The simulations were run on 32 processors on a Linux cluster.
25 Table 1 presents the execution time of the offline WRF-CMAQ system, the WRF-CMAQ two-way coupled system with and without radiative feedback. When the coupled system is executed without radiative feedback (but with increasing the temporal frequency of the WRF meteorological fields available for CMAQ), computational time is virtually identical when using the CAM radiation option and a slightly increased (∼8.5 %) when using the RRTMG radiation option compared to the offline WRF-CMAQ system. Once the feedback mechanism is turned on, the execution time increases only slightly with CAM but increases about 60 % with RRTMG. The numerical techniques used to compute aerosol optical characteristics (extinction, scattering, and asymmetry 5 factor) with the Mie approach used in RRTMG are much more computationally intensive than the Mie approximation (Evans and Fournier, 1990) used in the CAM implementation. However, this new scheme used for RRTMG is more accurate and robust over a wider range of refractive indices. The individual models used in the coupled system, WRF and CMAQ, are fully par-10 allelized. As a result, the scalability of the coupled system is inherited from both components. By doubling the processors from 32 to 64 on this domain, a speedup of ∼1.6 was achieved for both CAM and RRTMG configurations, regardless of whether the radiative feedback was enabled (Table 2). Increasing the number of processors by a factor of four (32 to 128) resulted in a speedup of ∼2.3-2.7. The addition of radiative 15 feedback in the coupled system does not adversely affect the scalability using either CAM or RRTMG, but results in a greater relative speedup on more processors.

Scientific performance
To examine the scientific performance of the coupled system with radiative feedback, we conducted a ten-day simulation (20-29 June 2008) of a wildfire event in California. 20 Widespread wildfires (Fig. 5) resulted in significant particulate matter (PM) pollution during mid-late June 2008 in California and surrounding states. The coupled model using the options discussed in Sect. 4.1, was applied to a domain covers California and portion of the surroundings states (Fig. 6); the RRTMG radiation scheme was used and the vertical extent up to 100 mb was discretized using 22 layers. This simulation 25 uses the latest fire emission data ( Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | meteorology measurements from the Meteorological Assimilation Data Ingest System (MADIS), radiation measurements from the Integrated Surface Irradiance Study (ISIS) Network, and concentration measurements from the US Environmental Protection Agency's Air Quality System (AQS) network. Figure 6 presents an illustration of the direct aerosol feedbacks simulated by the 5 coupled WRF-CMAQ modeling system during the time period of these wild fires in California. In general relatively high aerosol optical depths are noted in regions of high surface and boundary-layer particulate matter pollution ( Fig. 6a and b). Shown in Fig. 6c is the difference in surface shortwave radiation between a run with aerosol feedback and one without, while Fig. 6d presents an illustration of a similar difference 10 in the modeled planetary boundary layer (PBL) height at the same time. As illustrated, aerosol direct radiative effects associated with scattering and absorption of incoming radiation, result in a reduction of short-wave radiation reaching the surface (Fig. 6c), which then translate to reduction in temperature at the surface as well as a reduction in PBL height (Fig. 6d). These effects are particularly pronounced in regions with high 15 aerosol loading with simulated reductions of over 250 W m −2 in instantaneous surface shortwave radiation and corresponding reductions of over 500 m in PBL heights. Figure 7 presents comparisons of the surface shortwave radiation simulated for the cases with and without direct aerosol feedback with measurements at an Integrated Surface Irradiance Study (ISIS) site in Hanford, California (see Fig. 5, red star). The 20 observations show that there was significant reduction of incident shortwave radiation at the surface (peak observed values of 900-1000 W m −2 ) on 24, 26 and 27 June due to the smoke plumes from the wildfires. Without aerosol feedback effects, the model overestimates shortwave radiation by ∼100-200 W m −2 . When the aerosol feedbacks were included, the model bias was significantly reduced, though a slight overestimation 25 still persists. Thus including the aerosol feedback in the coupled system is important for better simulating the shortwave radiation fields in WRF.
The presence of aerosols from the wildfires reduces the shortwave radiation at the surface, which also acts to reduce the maximum daytime temperatures near the surface. Figure 8 shows a comparison of 2 m temperatures averaged from four sites in the Sacramento Valley, (Oroville, Red Bluff, Redding, and Sacramento, see blue triangles in Fig. 5) with model simulations for the cases with and without radiation feedback, for the ten-day simulation period. Typically the 2 m temperature was overestimated in the simulation that did not consider any aerosol feedback effects. This overestimation was reduced in the simulation with the aerosol feedbacks. For example, on 25,26,27 and 29 June, when the wildfires were most actively affecting these sites, inclusion of the aerosol feedback reduced or eliminated the persistent over prediction evident in the simulation without feedback. Figure 9 presents comparisons of the day time (08:00 a.m. to 06:00 p.m. local time) 10 model and observed ambient levels of ozone and PM 2.5 for all sites and data pairs; model results for both simulations with and without the feedback effects are shown. Figure 10 presents similar comparisons but only for data pairs where the simulated AOD > 0.5. While somewhat arbitrary this criteria helps examine the model performance for cases of significant aerosol loading, and consequently wherein radiative 15 feedback effects on temperature and PBL heights could in turn influence the subsequent chemistry-transport simulation. Though modest, the simulation including the aerosol feedback effects exhibits slight higher correlation coefficients than the one without.

Summary and future work 20
A two-way coupled meteorological and air quality modeling system has been developed by linking the WRF and CMAQ models. The system represents advancement over the traditional offline WRF-CMAQ system because the aerosols predicted by CMAQ are able to impact the clouds, radiation, and precipitation simulated by WRF in a consistent online coupled manner. In addition, because CMAQ is called directly from WRF, 25 the temporal interpolation of meteorological fields from WRF is eliminated thereby improving consistency in the use of meteorological information in the chemistry-transport calculations. A coupler is developed to efficiently link the two model systems. WRF fields to drive CMAQ, and provides aerosols feedback information from CMAQ to the WRF. The coupler is encapsulated in Fortran 90 modules so the details of the twoway coupling are transparent to the users. This software design also enables WRF and CMAQ to be detached and executed as stand-alone models as in the traditional offline paradigm. The single-source coding approach minimizes software maintenance so sci-5 entific updates to both WRF and CMAQ can be readily incorporated into the coupled WRF-CMAQ system. In addition to scientific and software maintenance issues, the coupled modeling system was designed to maximize user flexibility for research and applications by imposing minimal restrictions on domain specifications and physics options in both WRF and 10 CMAQ. Furthermore, the coupler allows users to choose the call frequency of CMAQ to balance the computational burden against the scientific accuracy, depending on the availability of computational resources. The coupled modeling system also includes a run-time switch to disable aerosol feedback and emulate the traditional offline paradigm albeit with greater frequency of communication of meteorological information from WRF 15 to the CMAQ model; this option can be used for further sensitivity tests examining the potential effects of temporal interpolation of meteorological data in the traditional offline paradigm.
When aerosol feedback is disabled, the computational time for of the coupled model is virtually identical to the offline WRF-CMAQ system. When the radiative feedback 20 is enabled, there is slight increase in execution time (compared to the case without feedback) using the CAM radiation scheme. However, adding radiative feedback with the RRTMG scheme results in an increase in run time of about 60 %, which is largely attributed to the more computationally intensive Mie calculation used in the implementation of the feedback effects with the RRTM scheme. Improving the computational 25 efficiency of the more accurate Mie scheme and its coupling with the RRTM is currently being investigated. In general, the coupled WRF-CMAQ modeling system scales well as the number of processors increase, regardless of the radiation model chosen in WRF or whether or not simulation of feedback effects is enabled. To demonstrate the improvements in simulated atmospheric dynamical and chemical features with the inclusion of aerosol radiative effects, we conducted a ten-day simulation of a wildfire event in California, a case characterized by significant tropospheric aerosol loading. Including radiative feedbacks in the model noticeably reduced the bias in simulated surface shortwave radiation and 2 m temperatures as well improved the 5 correlation of simulated ambient ozone and PM 2.5 relative to observed values. This preliminary analysis suggests that for cases with high aerosol loading (such as from wildfires, or in regions with significant anthropogenic pollution), including the radiative effects of aerosols improves the accuracy of both the meteorology and air quality simulations. 10 Further model evaluation studies are continuing, including efforts to examine the direct aerosol effects using a closed set of aerosol and radiation observations from the DOE/ARM Southern Great Plains site. Ozone has absorption bands in the long wave radiation bands and can thus absorb outgoing radiation. Efforts are underway to implement ozone feedback in the coupled WRF-CMAQ system and study the impact 15 of ozone on long wave radiation using the RRTMG long wave radiation scheme. An initial implementation of indirect aerosol forcing has also recently been completed and is under further testing and evaluation. These features will be made available in future versions of the 2-way coupled WRF-CMAQ modeling system.  Figure 7. Short wave radiation comparison between measurements at Hanford, CA (blue), and with 2 (yellow) and without (red) direct aerosol feedback 3