Multi-Model Analyses of Dominant Factors Influencing Elemental Carbon in Tokyo Metropolitan Area of Japan

The first phase of the Urban air quality Model InterComparison Study in Japan (UMICS) has been conducted to find ways to improve model performance with regard to elemental carbon (EC). Common meteorology and emission datasets are used with eight different models. All the models underestimate the EC concentrations observed in Tokyo Metropolitan Area in the summer of 2007. Sensitivity analyses are conducted using these models to investigate the causes of this underestimation. The results of the analyses reveal that emissions and vertical diffusion are dominant factors that affect the simulated EC concentrations. A significant improvement in the accuracy of EC concentrations could be realized by applying appropriate scaling factors to EC emissions and boundary concentrations. Observation data from Lidar and radiosonde suggest the possible overestimation of planetary boundary layer height, which is a vital parameter representing vertical diffusion. The findings of this work can help to improve air quality models to that they can be used to develop more effective strategies for reducing PM2.5 concentrations.


INTRODUCTION
Fine particulate matter adversely affects human health (Pope and Dockery, 2006).The Japanese government has set 15 (annual mean) and 35 (daily mean) micrograms per cubic meter as the Environmental Quality Standard (EQS) for fine particulate matter smaller than 2.5 micrometers (PM 2.5 ) since 2009.The current PM 2.5 concentration is likely above the EQS in most parts of Japan (Ministry of the Environment of Japan, 2012).PM 2.5 over Japanese urban areas mainly consists of elemental carbon (EC), organic carbon (OC), sulfate, ammonium, and nitrate.EC concentration exhibits a decreasing trend, and the contribution of the remaining components, which are mainly produced from gaseous precursors via complex photochemical reactions in the atmosphere, is becoming important (Minoura et al., 2006).To seek effective strategies for reducing PM 2.5 concentration including secondary components, it is essential to use air quality models that represent physical and chemical processes in the atmosphere, such as emission, advection, diffusion, photochemical reactions, and deposition.However, a single model may bring inappropriate results owing to possible errors made by users and/or problems intrinsic to models and input data.Model intercomparison is a promising way for evaluating the performance of multiple models and sorting problems that are inevitable among the state-of-the-art models or are confined to a single model.The results obtained could also be useful for further improvement of models.Intercomparisons of regional air quality models have been carried out for a broad range of spatial scales.The models in the United States and Europe participated in Air Quality Model Evaluation International Initiative II (AQME II), and their results were evaluated in both continents (Solazzo et al., 2012).In the Model Intercomparison Study for Asia Phase II (MICS-Asia II), several models were applied to the domain covering East and Southeast Asian countries, and their results were compared from various aspects (Carmichael et al., 2008).Although CityDelta focused on evaluating the effects of emission reduction strategies on air quality in selected European cities (Cuvelier et al., 2007), the results of the participating models were also compared (Vautard et al., 2007).All of these works revealed the performance characteristics and limitations of the participating models in respective scales and regions.
As the first trial of the model intercomparison in Japan, the performance characteristics of four models on ambient concentrations of ozone and PM 2.5 components were evaluated and mutually compared (Morino et al., 2010a).Morino et al. (2010b) also indicated that the ensemble average of the four models was effective for evaluating ozone and inorganic aerosol components.One of the issues found in their studies was that the four models used different domains, meteorological fields, boundary concentrations, and emission datasets, which made it difficult to identify the causes of the differences observed among the modeling results.
The Urban air quality Model InterComparison Study in Japan (UMICS) is a model intercomparison project designed using the experiences described above.The target of UMICS is to make models suitable for considering effective strategies for reducing PM 2.5 .Most of major modeling groups in Japan have participated in this project.Common domains are specified, and datasets of common meteorological fields, emissions, and boundary concentrations are provided to them.The participants are requested to conduct simulations in their usual model configurations to evaluate a variety of configurations which may be applied for considering strategies in Japan.The performance and consistency of their simulation results are evaluated.In addition, UMICS serves as an efficient comprehensive sensitivity analysis which is difficult to carry out for a single user.Participants conduct sensitivity runs in their fields of expertise to examine the causes of the differences observed among the models or deviations from observations.The results gathered in the project could provide information valuable for improving air quality models.
The first phase of UMICS focuses on EC in Tokyo Metropolitan Area.EC is not directly affected by photochemical reactions in the atmosphere.Therefore, it is suitable to evaluate only the physical processes represented in the models as the first step of the project.Inorganic and organic aerosol components including secondary products will be considered in the forthcoming phases of this project (Shimadera et al., 2013a).

Participating Models
Eight models of seven groups participated in the first phase of the project.Their configurations are shown in Table 1.The models are labeled M1-M8 in this paper.The models except for M8 are different versions of the Community Multiscale Air Quality (CMAQ) Modeling System (Byun and Schere, 2006) with different adopted modules.They reflect the situation that CMAQ is widely applied in Japan.CMAQ is a community model and has multiple choices of modules for physical and chemical processes which are periodically updated.Therefore, different versions of CMAQ with different modules could cause large differences.M8 is Regional Air Quality Model 2 (RAQM2) developed by Kajino et al. (2012), and its treatment of physical and chemical processes is distinct from that of CMAQ.Participating models were requested to submit the simulated results of concentrations and dry deposition rates of EC.

Observation Data
The observation data obtained during the field monitoring campaign called Fine Aerosol Measurement and Modeling in Kanto Area (FAMIKA) (Hasegawa et al., 2008;Fushimi Zhang et al. (2001Zhang et al. ( , 2003) ) 1 cloud_acm 1 Modified by Kajino et al. (2012Kajino et al. ( ). et al., 2011) ) was compared with modeling results.Ambient PM 2.5 was collected on quartz fiber filters for six hours during three periods (July 31 st to August 3 rd , August 6 th to 10 th , and August 13 th to 16 th , 2007) at four monitoring stations (Komae, Kisai, Maebashi, and Tsukuba) shown in Fig. 1 in Tokyo Metropolitan Area.The samples obtained were analyzed with the DRI Model 2001 Thermal/Optical Carbon Analyzer (Atmoslytic Inc.) using the Interagency Monitoring of Protected Visual Environments (IMPROVE) protocol to quantify EC concentration.Its relative error is 5.3% (Fushimi et al., 2011).
The participating models showed similar performances for three observation periods.The simulation results from August 6 th to 10 th , 2007 among them are discussed in this paper.Tokyo Metropolitan Area was covered by a Pacific high-pressure system throughout the period after a typhoon passed through the west part of Japan on August 3rd.Wind speed was high in daytime and low at nighttime.Wind direction was southerly throughout the period at Komae and Tsukuba, while it turned to northerly at nighttime at Kisai and Maebashi owing to the circulation of land-sea and mountain-valley winds.No precipitation was observed within Tokyo Metropolitan Area.

Meteorological Field
The common meteorological field was generated with Weather Research and Forecasting (WRF) -Advanced Research WRF (ARW) model (Skamarock et al., 2008) version 3.1.1for nineteen days from July 29 th to August 16 th , 2007 covering the entire period of FAMIKA as well as a few spin-up days.The target domains are shown in Fig. 1.The two domains (Domains 1 and 2) are nested over the middle of Japan, and Domain 2 covers Tokyo Metropolitan Area.Grids in Domains 1 and 2 are 15 × 15 km and 5 × 5 km in size, respectively.Thirty vertical layers are set following the sigma-P coordinates from the surface to 100 hPa, and the height of the bottom layer is approximately 57 m.The vertical layers of M8, which follow the sigma-Z coordinates, were adjusted to minimize differences in both coordinates.The United States Geological Survey (USGS) land use data (30-seconds resolution) were used to prepare geological inputs.The National Centers for Environmental Prediction (NCEP) Final Analysis data (1degree resolution, every 6 hours) were fed as the initial and boundary conditions.No grid nudging was applied because nudging to analysis data which is too coarser than grids of the target domains does not always result in better results.The Monin-Obukhov (Janjic) scheme, Noah-MP landsurface model and Mellor-Yamada-Janjic TKE scheme were adopted for the options of the surface layer, land surface, and planetary boundary layer, respectively.These choices realized the best and reasonable performance among various configurations without nudging (Mori et al., 2011).

Emission and Boundary Concentration
Dataset of common emissions and boundary concentrations were provided to the participating models.The emissions are based on the database described in detail by Chatani et al. (2011).Emissions from large point sources and vessels were derived from the East Asian Air Pollutant Emissions Grid Database (EAGrid2000-Japan, Kannari et al., 2007), and those from the remaining anthropogenic sources other than vehicles were estimated with the Georeference-Based Emission Activity Modeling System (G-BEAMS, Nansai et al., 2004).Vehicle emissions were estimated using the Japan Auto-Oil Program (JATOP) vehicle emissions estimate model.The Model of Emissions of Gases and Aerosols from Nature (MEGAN, Guenther et al., 2006) version 2.04 with the common meteorological fields described in the previous section was utilized to estimate hourly biogenic emissions.The distribution of the EC emission rates in Domain 2 is shown in Fig. 1.EC emission rates are high at the heart of Tokyo, in the industrial areas along Tokyo Bay, along major highways, and from isolated large point sources.Boundary concentrations were set by the default values of CMAQ, but the values of sulfate were determined using the observation data obtained in FAMIKA.Emissions and boundary concentrations were divided into SAPRC99 (Carter, 2000) species groups and aerosol components (sulfate, nitrate, EC, OC, and others in fine fraction and nonspeciated coarse fraction), and were provided to the participating models.

RESULTS
Fig. 2 shows the time series of the observed surface EC concentration for every six hours and the hourly surface EC concentration simulated by the eight participating models at Komae, Kisai, Maebashi, and Tsukuba for August 6 th to 10 th , 2007.The simulated values are lower than the observed values at all stations for most of the time.The observed values increase in the latter days, although the increases in simulated values for the same period are found only at Komae.The simulated values tend to increase in the morning and evening, and to decrease in the daytime and at night particularly at Komae and Maebashi whereas the decreases in observed values during the daytime are not clearly found.The values simulated by M8 are the highest among the eight models.The differences in the values simulated by the seven CMAQ are small.The versions and adopted modules of CMAQ which are widely used in Japan shown in Table 1 cause negligible differences.
Fig. 3 shows the vertical profiles of the EC concentrations simulated by M7 and M8 at Komae and Kisai at 6 AM on August 10 th , 2007.The EC concentration simulated by M8 is higher than those simulated by M7 within the bottom layer.M8 uses the 1.5-order TKE module for vertical diffusion, which is different from the modules applied in CMAQ.The different treatment of vertical diffusion could be one of the reasons for the higher surface EC concentrations simulated by M8.Fig. 4 shows the time series of the hourly EC dry deposition rates simulated by M7 and M8 at Komae, Kisai, Maebashi, and Tsukuba from August 6 th to 10 th , 2007.All of the simulated values are high in daytime and low at nighttime.However, the amplitude of temporal variation is significantly smaller in the values simulated by M8.M8 represents aerosol size distributions with multiple modes (Kajino, 2011) similarly to CMAQ (Binkowski and Roselle, 2003).However, M8 explicitly treats various mixing states of aerosols (Kajino and Kondo, 2011), while only internal mixing is assumed in CMAQ.Differences in the treatment of aerosol dynamics and resulting diameters could cause significant differences in dry deposition rates.

SENSITIVITY ANALYSES
The deviation of the simulated values from the observed ones cannot be explained only by uncertainties embedded in the models.The reasons for this deviation and any possible ways to improve models were further explored via sensitivity analyses conducted by the participants of the project who applied CMAQ.

Process Analyses
A capability of the integrated process rate (IPR) analysis is incorporated into CMAQ.IPR is widely used to quantify the effects of all the physical and chemical processes on model predictions (e.g., Khiem et al., 2010).To understand the effects of the physical and chemical processes on the surface EC concentrations simulated by the participating models, the simulated change rates of the EC concentrations caused by advection, vertical diffusion, emission, horizontal diffusion, dry deposition, and aqueous processes were derived by the IPR analysis.
Fig. 5 shows the simulated change rates of the surface EC concentrations caused by advection, vertical diffusion, emission, and all the other processes at Komae, Kisai, Maebashi, and Tsukuba from August 6 th to 10 th , 2007.Emission and vertical diffusion are dominant processes that increase and decrease the simulated values, respectively.
Their effects are high in daytime and low at nighttime.The absolute effect of vertical diffusion is comparable or slightly larger than that of emission in daytime.They result in the decreases of the simulated surface EC concentrations in daytime shown in Fig. 2. The effects of advection follow those of emission and vertical diffusion.The negative change rates caused by advection that are comparable to those of vertical diffusion are observed in the evening at Maebashi, which is located in the north of Tokyo Metropolitan Area, when the direction of the land sea breeze is northerly.
The simulated change rates of the surface EC concentrations caused by horizontal diffusion, dry deposition, and aqueous processes were small.Although the effect of aqueous process is negligible because of the absence of precipitation within Tokyo Metropolitan Area during this period, it could be a dominant removal process for EC under rainy conditions.The small effect of dry deposition means that the difference in dry deposition rates shown in Fig. 4 is not a major cause of differences in the EC concentrations simulated by CMAQ and RAQM2.As shown in Figs. 2 and 4, surface EC concentration is on the order of a few micrograms per cubic meter, and EC dry deposition rate is on the order of a few micrograms per square meter per hour.Moreover, the EC emission rates at the four stations are on the order of several tens of micrograms per square meter per hour, as shown in Fig. 1.Therefore, only a tiny fraction of EC in the bottom layer of 57 meters depth is removed by dry deposition.Therefore, the effects of the differences in dry deposition rates shown in Fig. 4 are almost negligible against the surface EC concentration and emission rate.Further investigation on the differences in dry deposition rates of CMAQ and RAQM2 is outside the scope of the current project.
The sensitivities of emission and vertical diffusion to the simulated EC concentration were further examined in the following subsections.

Emission and Boundary Concentration
The differences between the observed and simulated values may be partly caused by the underestimation of the EC emissions and boundary concentrations.A simple approach similar to that of Hu et al. (2009) was applied to investigate how much EC emissions and boundary concentrations need to be increased to match the simulated values to the observed ones.
First, the hourly sensitivities of the EC emissions of three sources (vehicles, other anthropogenic sources, and open agricultural burning) and the EC boundary concentrations to the simulated hourly EC concentration were calculated by the so-called "brute force method", in which gaps of simulated EC concentrations in two runs with and without the EC emission of each source or the EC boundary concentrations are considered as their sensitivities.The sensitivities of vehicles and other anthropogenic sources were dominant, and those of open agricultural burning and boundary concentrations were only a few percents.
Then, it was assumed that the differences between the observed and simulated values are explained by the sensitivities and multiplying factors as expressed in where EC obs and EC sim are the observed and simulated surface EC concentrations, S veh , S otr , S bng , and S bdy are the sensitivities of EC emissions (vehicles, other anthropogenic sources, and open agricultural burning) and EC boundary concentrations obtained by the brute force method, and f veh , f otr , f bng , and f bdy are the multiplying factors of the EC emissions of the three sources and EC boundary concentrations, respectively.The subscript i represents each hour and station.Any spatial or temporal variation in multiplying factors was ignored.Suitable multiplying factors were estimated by multiple regression analysis utilizing pairs of observed and simulated surface EC concentrations for all periods at the four stations.The estimated multiplying factors were 4.0, -0.8, 41.7, and 22.4 for the EC emissions of vehicles, other anthropogenic sources, and open agricultural burning, and for EC boundary concentrations, respectively, and the multiple correlation coefficient was 0.90.
An additional run (named as posteriori) was executed in which the EC emissions of vehicles and open agricultural burning, and EC boundary concentrations were multiplied by 4, 40, and 20, respectively, based on the results of multiple regression analysis.The time series of the observed and simulated surface EC concentrations of the two runs at Komae, Kisai, Maebashi, and Tsukuba from August 6 th to 10 th , 2007 are shown in Fig. 6.The model performance was significantly improved in terms of not only the absolute value of EC concentration but also the temporal variation.Although no evidences for such high multiplying factors are currently available, the results imply that the EC emissions not originating from urban areas such as those from the open agricultural burning and the transport of EC from the outside of Tokyo Metropolitan Area, should be paid more attention.Shimadera et al. (2013b) conducted a simulation including the same period as in this study.Their simulation well reproduced the surface EC concentration observed at Chichijima, which is located at about 1000 km south of the center of Tokyo in the Pacific Ocean.They also indicated that the contribution of the emissions outside of Japan to the simulated EC concentration in Tokyo Metropolitan Area was negligible during the same period.On the basis of their results, the effect of EC transported from regions far away from Tokyo Metropolitan Area may be small.Effects of emission sources outside of Tokyo Metropolitan Area but within Japanese islands and surrounding oceans as well as within Tokyo Metropolitan Area should be examined further.Fushimi et al. (2011) indicated that a few tens percent of

Vertical Diffusion
Vertical diffusion coefficients control the strength of vertical diffusion in the simulation.They are calculated using the equations of Holstlag and Boville (1993) and Liu and Carroll (1996) below and above the planetary boundary layer (PBL) height, respectively, within the acm2 scheme of CMAQ.The former tends to produce higher values and also utilizes a value of the PBL height itself in the equation.Therefore, the vertical diffusion coefficients calculated in the simulation are highly affected by the PBL height.
Three sensitivity runs in which the values of the PBL height were changed to zero, half and twice the original value were performed.The PBL height is tightly linked to other meteorological parameters like winds in the real atmosphere, but only the values of the PBL height were changed here to evaluate its importance.The time series of the observed and simulated surface EC concentrations of the four runs at Komae, Kisai, Maebashi, and Tsukuba for August 6 th to 10 th , 2007 are shown in Fig. 7.The simulated values are significantly affected by the PBL height.They tend to increase in daytime when smaller values of the PBL height are used because lower vertical diffusion coefficients are applied to more vertical layers.
We attempted to roughly estimate the PBL height by using the two types of available observation data (lidar and radiosonde) at Tsukuba.The PBL height was determined as the altitude at which the gradient of the spherical particle extinction coefficients higher than 5.0 × 10 -5 for 532 nm measured by the lidar falls below -8% per 30 meters, and as the altitude at which the gradient of the potential temperature observed by the radiosonde exceeds 6 K per kilometer (Kuribayashi et al., 2011).Fig. 8 shows the PBL height determined from the two observation data and the values simulated in the common meteorological field at Tsukuba from August 5 th to 11 th , 2007.The values of the PBL height determined by the two observation data coincide; thus the two methodologies used to determine the PBL height appear to be reliable.Although the PBL is low at nighttime and high in daytime in both of the observed and simulated values, the simulated values in daytime are higher than the observed ones.These results imply a possibility that the simulated PBL height in daytime is overestimated.
It must be noted that the simulated EC concentrations do not reach the observed values even if the PBL height is set to zero.The deviation of the simulated values from the observed ones cannot be attributed only to vertical diffusion.

SUMMARY
A model intercomparison project in Japan called UMICS has been conducted.The first phase of UMICS focused on model performance characteristics and sensitivities to EC in Tokyo Metropolitan Area because it is not directly affected by photochemical reactions in the atmosphere and is suitable for evaluating only the physical processes represented in the models.Common meteorological field and emission datasets were provided to eight participating models.The simulated EC concentrations were compared with each other and with the observation data obtained in FAMIKA.All of the participating models underestimated the surface EC concentration.
Sensitivity analyses were performed to investigate the reasons for the deviation of the simulated values from the observation data.Emission and vertical diffusion were identified as the two dominant factors that induce the  variation in simulated EC concentration via the process analyses.The performance of the simulation was significantly improved when multiplying factors obtained by multiple regression analysis were applied to EC emissions as well as to EC boundary concentrations.They implied that the EC emissions not originating from urban areas should be paid more attention.The PBL height estimated from the observation data of lidar and radiosonde implied that the simulated PBL height was possibly underestimated.These outcomes are useful to improve air quality models further and to make models suitable for considering effective strategies for reducing PM 2.5 concentration.
The target period for the first phase of UMICS was focused on the summertime in which major winds are southerly and photochemical reactions are active.Large-scale transport is significant in spring and autumn, and stagnant air in winter causes heavy air pollution in winter.Therefore, model performance characteristics also vary with season.UMICS has tight linkages with the observation and emission research groups.The remaining issues described above will be reflected in their next observation campaign not only in summer and emission improvement, and then their results will be utilized in the second and following phases of UMICS.Such iteration should contribute to the improvement of air quality models.
The effectiveness of the regional air quality model intercomparison designed in UMICS has been strongly recognized by the participants.Further valuable results are expected in the forthcoming phases of UMICS on secondary aerosol components (Shimadera et al., 2013a) as well as the remaining issues for EC described in this paper.

Fig. 1 .
Fig. 1.Common target domains of simulation (left) and EC emission rates in Domain 2 (right).The locations of the four monitoring stations of FAMIKA are also shown in the right figure.

Fig. 2 .Fig. 3 .Fig. 4 .
Fig. 2. Time series of observed surface EC concentration for every six hours and hourly surface EC concentration simulated by eight participating models at Komae, Kisai, Maebashi, and Tsukuba from August 6 th to 10 th , 2007.

Fig. 7 .
Fig. 7. Time series of observed and simulated surface EC concentrations of three runs, in which PBL height was changed to zero, half and twice, as well as base case at Komae, Kisai, Maebashi, and Tsukuba from August 6 th to 10 th , 2007.

Fig. 8 .
Fig. 8. PBL heights determined from observation data of lider and radiosonde, and that simulated in common meteorological field (MCIP) at Tsukuba from August 5 th to 11 th , 2007.The values under clear sky (no-cld) and cloudy (cld) conditions are represented by different symbols.

Table 1 .
Configurations of participating models.
Time series of observed and simulated surface EC concentrations of two runs in which multiplying factors are applied (posteriori) and not applied (base) at Komae, Kisai, Maebashi, and Tsukuba from August 6 th to 10 th , 2007.modern carbon not originating in fossil fuels are included in the total carbon collected during FAMIKA which is much higher than the sensitivity of open agricultural burning obtained in the simulation.It implies possible underestimation of EC emissions from open agricultural burning.