In this study, we use the Weather Research and Forecasting model coupled with chemistry (WRF-Chem version 3.5.1), which is an online mesoscale model capable of simulating meteorological and chemical processes simultaneously (Grell et al., 2005; Fast et al., 2006). The model domain (Fig. 1) is defined on a Mercator projection and is centred at 22∘ N, 83∘ E with 274 and 352 grid points in the east–west and north–south directions, respectively, at the horizontal resolution of 12km×12km. The land use dataset is incorporated from the US Geological Survey (USGS) based on 24 land use categories. The ERA-Interim reanalysis dataset from ECMWF (http://www.ecmwf.int/en/research/climate-reanalysis/browse-reanalysis-datasets), archived at the horizontal resolution of about 0.7∘ and temporal resolution of 6 h, is used to provide the initial and lateral boundary conditions for the meteorological calculations. All simulations in the study have been conducted for the period of 26 February–31 May 2013 at a time step of 72 s. The model output is stored every hour for analysis. The first 3 days of model output have been discarded as model spin-up.

Radiative transfer in the model has been represented using the Rapid Radiative Transfer Model (RRTM) longwave scheme (Mlawer, 1997) and the Goddard shortwave scheme (Chou and Suarez, 1994). Surface physics is parameterized using the unified Noah land surface model (Tewari et al., 2004) along with ETA similarity option (Monin and Obukhov, 1954; Janjic, 1994, 1996), and the planetary boundary layer (PBL) is based on the Mellor–Yamada–Janjic (MYJ) scheme (Mellor and Yamada, 1982; Janjic, 2002). The cloud microphysics is represented by the Lin et al. scheme (Lin et. al., 1983), and cumulus convection is parameterized using the Grell 3-D ensemble scheme (Grell, 1993; Grell and Devenyi, 2002). Four-dimensional data assimilation (FDDA) is incorporated for nudging to limit the drift in the model-simulated meteorology from the ERA-Interim reanalysis (Stauffer and Seaman, 1990; Liu et al., 2008). Horizontal winds are nudged at all vertical levels, whereas temperature and water vapour mixing ratios are nudged above the PBL (Stauffer et al., 1990, 1991). The nudging coefficients for temperature and horizontal winds are set as 3×10-4 s-1 and as 10-5 s-1 for water vapour mixing ratio (Otte, 2008).

This study utilizes two different chemical mechanisms: RADM2 (Stockwell et al., 1990) and MOZART-4 (Emmons et al., 2010). RADM2 chemistry includes 63 chemical species participating in 136 gas-phase and 21 photolysis reactions. MOZART chemistry includes 81 chemical species participating in 159 gas-phase and 38 photolysis reactions. Aerosols are represented using the Modal Aerosol Dynamics Model for Europe/Secondary Organic Aerosol Model (MADE/SORGAM) (Ackermann et al., 1998; Schell et al., 2001) with RADM2 and Global Ozone Chemistry Aerosol Radiation and Transport (GOCART) (Chin et al., 2000) with MOZART. The photolysis rates are calculated using the Fast-J photolysis scheme (Wild et al., 2000) in the RADM2 simulations and the Madronich Fast Tropospheric Ultraviolet-Visible (FTUV) scheme in the MOZART simulation. In WRF-Chem, the Madronich FTUV photolysis scheme uses climatological O3 and O2 overhead columns. The treatment of dry deposition process also differs between RADM2 and MOZART due to differences in Henry's law coefficients and diffusion coefficients. The chemical initial and lateral boundary conditions are provided from 6-hourly fields from MOZART-4/GEOS5 (http://www.acom.ucar.edu/wrf-chem/mozart.shtml).

This study utilizes three different inventories for the anthropogenic emissions: the Emissions Database for Global Atmospheric Research – Hemispheric Transport of Air Pollution (EDGAR-HTAP), the Intercontinental Chemical Transport Experiment phase B (INTEX-B) and the Southeast Asia Composition, Cloud, Climate Coupling Regional Study (SEAC4RS), which are briefly described here. The HTAP inventory (Janssens-Maenhout et al., 2015) for anthropogenic emissions (http://edgar.jrc.ec.europa.eu/htap_v2/index.php?SECURE=_123) available for the year 2010 has been used. The HTAP inventory has been developed by complementing various regional emissions with EDGAR data, in which Asian region including India is represented by the Model Intercomparison Study for Asia (MICS-Asia) inventory, which is at a horizontal resolution of 0.25∘×0.25∘ (Carmichael et al., 2008). The resultant global inventory is regridded at the spatial resolution of 0.1∘×0.1∘ and temporal resolution of 1 month. HTAP includes emissions of CO, NOx, SO2, non-methane volatile organic compounds (NMVOCs), particulate matter (PM), black carbon (BC) and organic carbon (OC) from power, industry, residential, agriculture, ground transport and shipping sectors. The INTEX-B inventory (Zhang et al., 2009), developed to support the INTEX-B field campaign by the National Aeronautics and Space Administration (NASA) in spring 2006, is the second inventory used in this study. It provides total emissions for the year 2006 at a horizontal resolution of 0.5∘×0.5∘. The emission sectors include power generation, industry, residential and transportation. The SEAC4RS inventory (Lu and Streets, 2012), prepared for the NASA SEAC4RS field campaign, is the third inventory used in this study. It provides total emissions for the year 2012 at a spatial resolution of 0.1∘×0.1∘. SEAC4RS and INTEX-B did not cover regions in the north-western part of the domain, and therefore we complemented this region (longitude <75∘ E and latitude >25∘ N) by HTAP emission data. The emissions of CO, NMVOCs and NOx emissions among the three emission inventories, as included in the simulations, are shown in Fig. 2. Table 2 provides estimates of total emissions over different regions (as defined in Fig. 1) from the three inventories. The total emissions over all regions show that HTAP has about 43 % higher and SEAC4RS about 46 % higher NOx emissions compared to the INTEX-B inventory. Also, HTAP has about 37 % higher VOC emissions compared to SEAC4RS and about 49 % higher compared to the INTEX-B inventory. Hence, SEAC4RS, the most recent inventory of the three, has similar total NOx emissions to those in HTAP but the total VOC source is closer to INTEX-B, which is the oldest of the three inventories. Considering the non-linear dependence of O3 formation on precursors, numerical experiments are necessary to assess the influence of such large differences among the inventories. The emissions from biomass burning are included using the Fire Inventory from NCAR (FINN) version 1.0 (Wiedinmyer et al., 2011). The Model of Emissions of Gases and Aerosols from Nature (MEGAN) is used to include the biogenic emissions (Guenther et al., 2006) in the model.

The HTAP inventory is available at monthly temporal resolution, while INTEX-B and SEAC4RS are available as annual averages; however, seasonal variability in anthropogenic emissions may not have a major effect in this study as we focus here on spring (pre-monsoon), for which monthly emissions are similar to the annual mean (seasonal factor close to unity) (Fig. S1 in the Supplement; also see Fig. 2b in Kumar et al., 2012b). Nevertheless, seasonal influence during spring is strongest for biomass-burning emissions, which have been accounted for. The emissions from all inventories were injected in the lowest model layer. The diurnal profiles of the anthropogenic emissions of ozone precursors, specific to south Asia, are not available. A sensitivity simulation implementing the diurnal emission profile available for Europe (Mar et al., 2016 and references therein) showed a little impact on predicted noontime ozone over south Asia (Fig. S2).

We have conducted four different numerical simulations, as summarized in Table 3 and briefly described here. Three simulations correspond to three different emission inventories (HTAP, INTEX-B and SEAC4RS) for the anthropogenic emissions of ozone precursors, employing the RADM2 chemical mechanism. These simulations are named HTAP-RADM2, INTEX-RADM2 and S4RS-RADM2, respectively. The emissions of aerosols have been kept the same (HTAP) among these three simulations and aerosol–radiation feedback has been switched off to specifically identify the effects of emissions of O3 precursors on modelled ozone. An additional simulation (HTAP-MOZ) has been conducted to investigate the sensitivity of ozone to the employed chemical mechanism (MOZART vs. RADM2) by keeping the emissions fixed to HTAP.

Previous studies have shown that WRF-Chem accurately reproduces the synoptic-scale meteorology over the Indian region, justifying its use for atmospheric chemical simulations (e.g. Kumar et al., 2012a). Further, nudging towards reanalysis data limits deviations in simulated meteorology (e.g. Kumar et al., 2012a; Ojha et al., 2016; Girach et al., 2017). Nevertheless, we include an evaluation of model-simulated water vapour, temperature and wind speed against radiosonde observations (Fig. S3). Vertical profiles of the monthly average (April) water vapour mixing ratio (gKg-1), temperature (∘C) and horizontal wind speed (ms-1) have been obtained from radiosonde data (available at http://weather.uwyo.edu/upperair/sounding.html) for evaluation of modelled meteorology over Delhi (in north India), Bhubaneswar (in east India) and Ahmedabad (in west India). We find that model-simulated meteorology is in good agreement (within 1 standard deviation (SD) variability) with the observations.

Surface ozone data are acquired from various studies and sources, as given in Table 4. In general, surface O3 measurements over these stations have been conducted using the well-known technique of UV light absorption by ozone molecules at about 254 nm, making use of the Beer–Lambert law. The accuracy of these measurements is reported to be about 5 % (Kleinmann et al., 1994). The response time of such instruments is about 20 s and instruments have a lower detection limit of 1 ppbv (Ojha et al., 2012). Here, we have used the hourly and monthly average data for the model evaluation. The details of the instrument and calibrations at individual stations can be found in the references given in Table 4. It is to be noted that most of the observations are conducted generally inside the campuses of universities/institutes, reasonably away from the direct roadside emissions/exhaust (see references provided in Table 4) and therefore not influenced by concentrated local pollution sources.

As simultaneous measurements at different stations are very sparse over south Asia, the model evaluation has often been conducted using observations of the same season/month of a different year (e.g. Kumar et al., 2012b, 2015; Ojha et al., 2016). However, to minimize the effect of temporal differences, we preferentially used measurements of recent years; i.e. the observations at the stations used in this study are for the period of 2009–2013. For four stations – Delhi (north India), Jabalpur (central India), Pune (west India) and Thumba (south India) – the observations and simulations are for the same year (2013). Finally, we investigated the effects of temporal differences on the results and model biases presented here by conducting another simulation for a different year (2010) (Fig. S4).

There is also a need to evaluate precursor mixing ratios over the region to further reduce uncertainties in modelled ozone over south Asia. However, very limited data are available for ozone precursors in India and adjacent countries (especially for NMVOCs). We include an evaluation of modelled NOx, ethane and ethene mixing ratios against several recent observations in the Supplement (see Table S1). More sensitive techniques (e.g. blue light converter for NO2) in future would provide better insights into model performance in reproducing NOx over India.

The spatial distribution of WRF-Chem simulated 24 h monthly average ozone during April is shown in Fig. 3a for the three different emission inventories (HTAP, INTEX and SEAC4RS). Generally, the months of March and May are marked with seasonal transition from winter to summer and summer to monsoon, respectively. Hence, the month of April is chosen to represent the pre-monsoon season, as it is not influenced by these seasonal transitions, and the observational data are available for a maximum number of stations during this month for the comparison. The 24 h average ozone mixing ratios are found to be 40–55 ppbv over most of the Indian subcontinent for all the three inventories. Model-simulated ozone levels over the coastal regions are also similar (30–40 ppbv) among the three inventories. The highest ozone mixing ratios (55 ppbv and higher) predicted in the south Asian region are found over northern India and the Tibetan Plateau. The WRF-Chem simulated spatial distributions of average ozone shown here are in agreement with a previous evaluation study over south Asia (Kumar et al., 2012b). Further, it is found that qualitatively as well as quantitatively the HTAP, INTEX-B and SEAC4RS can lead to very similar distributions of 24 h average ozone over most of the south Asian region. The 24 h monthly average ozone from observations is superimposed on the model results in Fig. 3a for comparison. WRF-Chem simulated distributions of average O3 are in general agreement with the observational data (Fig. 3a), except at a few stations near coasts (e.g. Kannur and Thumba) and in complex terrain (Pantnagar and Dibrugarh). In contrast to the distribution of 24 h average O3, the noontime (11:30–16:30 IST) O3 mixing ratios over continental south Asia exhibit significant differences among the three emission inventories (Fig. 3b). HTAP clearly leads to higher noontime O3 mixing ratios, the difference being up to 10 ppbv over the Indo-Gangetic Plain (IGP), 20 ppbv over central India and 30 ppbv over southern India, compared to INTEX-B and SEAC4RS. The mean bias (MB) (model–observation) for 24 h and noontime average ozone at individual stations is provided in the Supplement (Tables S2 and S3). A sensitivity simulation is conducted to reveal the influence of a different cumulus parameterization (Kain–Fritsch scheme) on our conclusions. The differences in the modelled surface ozone mixing ratios over most of the Indian domain are found to be within ±5 % (Fig. S5). The relatively large differences over some of the Indian region indicate that the Kain–Fritsch scheme tends to predict higher surface ozone mixing ratios relative to the base run (incorporating Grell 3-D ensemble scheme), which would only add up to biases in the original runs. Therefore, our conclusions are not affected.

The net photochemical O3 production rate (ppbvh-1) from sunrise to noontime (06:30–12:30 IST), when most of the photochemical buildup of ozone takes place leading to its peak noontime mixing ratio, has been calculated utilizing the chemical tendencies in WRF-Chem (Barth et al., 2012; Girach et al., 2017). A comparison of monthly average O3 production rates among the three inventories is shown in Fig. 4. As seen also from the O3 mixing ratios (Fig. 3b), the HTAP emissions result in faster O3 production (∼9 ppbvh-1) throughout the IGP region. The highest O3 production rates for INTEX-B and SEAC4RS inventories are simulated only in the east Indian regions including the eastern parts of the IGP. It is noted that the rate of O3 production is lower (4–8 ppbvh-1) over most of the south-western IGP for the INTEX-B and SEAC4RS inventories. Differences are also found over the southern Indian region with stronger ozone production in HTAP, followed by INTEX-B and SEAC4RS.

Figure 5 provides insight into the spatial distribution of O3 production regimes estimated through the CH2O/NOy ratio (Geng et al., 2007; Kumar et al., 2012b), calculated during 06:30–12:30 IST, to help explain the differences in modelled ozone mixing ratios among the three simulations. The metric CH2O/NOy, as described by Sillman (1995), is suggested to be a useful diagnostic to determine the ozone production regime. Sillman (1995) evaluated the correlation between O3–NOx–VOC sensitivity predicted by photochemical model and CH2O/NOy ratio. The correlation has been derived combining results from serial computations with the model by varying the anthropogenic and biogenic emissions, and meteorology. The method has been successfully employed in investigating ozone distribution over south Asia (Kumar et al., 2012b), east Asia (Geng et al., 2007; Tie et al., 2013) and Europe (Mar et al., 2016). Tie et al. (2013) reported similarities between the results based on the CH2O/NOy ratio and those following another method described by Kleinmann et al. (2003) over Shanghai. A value of 0.28 for the CH2O/NOy ratio is suggested to be the transitional value from the VOC-limited regime to NOx-limited regime. The spatial distribution of regimes in all simulations in the present study is largely consistent with the findings of Kumar et al. (2012b) although the latter performed the analysis for afternoon hours (11:30–14:30 IST). The S4RS-RADM2 simulation predicts the entire IGP to be VOC sensitive, whereas in the HTAP-RADM2 and INTEX-RADM2 simulations, though the north-west IGP and eastern IGP are VOC sensitive, the central IGP is mostly NOx limited. The coastal regions are also predicted to be VOC limited in all the three simulations. With the north-western IGP being VOC limited in all simulations, the noontime ozone mixing ratios are found to be higher in this region in the HTAP-RADM2 simulation because of high NMVOC emissions in the HTAP inventory as evident from Fig. 2 and Table 2. Similar differences are also apparent in southern India.

Some of the ozone precursors (NOx/NOy, ethane and ethene) are also compared between the model and recent measurements over few stations (Table S1). Significant differences are seen in model-simulated NOx mixing ratios among different emission inventories (e.g. 6.5–30.5 ppbv at Kanpur) over the urban stations in the IGP. The model typically overestimated NOx mixing ratios at Delhi except in the INTEX-RADM2 simulation, which showed an agreement with observations within 1 SD. Total nitrogen oxides (NOy) showed relatively similar levels among different inventories (2.7–3.2 ppbv) at a high-altitude station (Nainital) in north India and were only slightly higher than the observed mean (1.8±1.6 ppbv). In contrast with the stations in northern India, NOx levels over a rural station (Udaipur) in western India are underestimated by a factor of 4. On the other hand, modelled ethane mixing ratios are underpredicted by a factor of about 2, whereas modelled ethene mixing ratios agree relatively well with observed values at Nainital in INTEX-RADM2 and S4RS–RADM2 as compared to HTAP-RADM2. More in situ observations, especially of ozone precursors, may provide better insights into the performance of the numerical models and employed emission inventories over this region.

In summary, these results show similar 24 h average ozone distributions but large differences in the ozone buildup until noon. The net photochemical ozone production in the morning hours (06:30–12:30 IST) is shown to be sensitive to the different inventories over this region, which is attributed to differences in total NOx and/or NMVOC emissions. We therefore suggest that a focus on 24 h averages only would be insufficient to evaluate the ozone budget and implications for human health and crop yield. Next, we compare the modelled diurnal ozone variations from three inventories with in situ measurements over 15 stations across south Asia.

A comparison of WRF-Chem simulated diurnal ozone variability with recent in situ measurements over a network of 15 stations in the south Asian region is shown in Fig. 6. WRF-Chem is found to successfully reproduce the characteristic diurnal ozone patterns observed over the urban (e.g. Mohali, Delhi, Kanpur, Ahmedabad, Bhubaneswar and Pune) and rural (e.g. Anantapur, Gadanki) stations, indicating strong ozone buildup from sunrise to noontime and the predominance of chemical titration (by NO) and deposition losses during the night. In general, WRF-Chem captures the daily amplitude of O3 changes at relatively cleaner and high-altitude stations, typically showing less pronounced diurnal variability, such as Nainital in the Himalayas, although with differences in timing when the model and observations attain minimum ozone mixing ratios, thus leading to relatively low correlation coefficients (see later in the text). The diurnal variability in O3 indicated by ΔO3, i.e. the difference between diurnal mean and hourly values, is further compared between the model and observations at all the stations (Fig. S6). This comparison intends to focus more on evaluating the model's ability to reproduce different diurnal patterns over urban, rural and clean chemical environments and minimizing the representation of absolute ozone levels. It is seen that model successfully captures the observed variability in ozone at most of the sites in this region. However, a limitation is noticed in resolving the stations well in the vicinity of complex terrain (such as in Jabalpur), attributed to the stronger spatial heterogeneity due to forests, hills and mountains within a small area (Sarkar et al., 2015).

To briefly evaluate the possible effects due to the difference in meteorological year between the model and observations, we repeated the HTAP-RADM2 simulation for a different year (2010) as shown in the Supplement (Fig. S4). The effect of changing the meteorological year in the model simulation is generally small (mostly within ±3 ppbv in 3 years), except at a few stations in the north (Nainital and Pantnagar) and east (Bhubaneswar). The effect is seen to vary from 4.8 to 6 ppbv (in 3 years) at these three stations. These differences are found to be associated with the interannual variations in the regional and transported biomass-burning emissions, as seen from MODIS fire counts and MOZART/GEOS5 boundary conditions (not shown here).

The model ability to reproduce diurnal variations at all stations is summarized using a Taylor diagram (Taylor, 2001) in Fig. 7. The statistics presented are normalized SD, normalized centred root mean square difference (RMSD) and the correlation coefficient. The normalization of both SD and RMSD is done using the SD of the respective observational data. The point indicated as “REF” represents the observational data against which the model results are evaluated. WRF-Chem simulations show reasonable agreement with observations showing correlation coefficients generally greater than 0.7 for most sites. The locations such as Nainital and Jabalpur for which r values are lower (0.3–0.7) are associated with unresolved complex terrain, as mentioned earlier. Note that the Taylor diagram has been used to present evaluation statistics for a general overview and intercomparison, i.e. how the model reproduces the diurnal variation at different stations, irrespective of the emission inventory.

The WRF-Chem simulated spatial distributions of 24 h average and noontime average surface ozone are compared in Fig. 8. The monthly values of the 24 h and noontime ozone mixing ratios from measurements are also shown. Overall, the average ozone mixing ratios over south Asia are simulated to be higher with the MOZART chemical mechanism compared to RADM2, which is consistent with the results of Mar et al. (2016) for the European domain. The 24 h average ozone mixing ratios over India simulated with MOZART chemistry are found to be higher than those with RADM2 chemistry, especially over the eastern Indian region (∼60 ppbv and more for MOZART compared to ∼40–55 ppbv for RADM2). Average ozone levels over the coastal regions are found to be similar between the two mechanisms (30–40 ppbv). MOZART chemistry also predicts high 24 h average ozone mixing ratios (55 ppbv and higher) over the Tibetan Plateau region, similar to RADM2. A striking difference between the two chemical mechanisms is found over the marine regions adjacent to south Asia (Bay of Bengal and northern Indian Ocean), with MOZART predicting significantly higher 24 h average ozone levels (35–50 ppbv) compared to RADM2 (25–40 ppbv). A comparison of noontime average ozone distributions between the two chemical mechanisms shows that MOZART predicts higher ozone concentrations than RADM2 over most of the Indian region by about 5–20 ppbv, except over western India. The differences are up to 20 ppbv and more over the southern Indian region, highlighting the impacts of chemical mechanisms on modelled ozone in this region. The mean bias (MB) values (model-observation) for 24 h and noontime average ozone at individual stations is provided in the Supplement (Tables S2 and S3).

Figure 9a shows a comparison of the monthly average chemical O3 tendency (ppbvh-1) from 06:30 to 12:30 IST. In contrast with average O3 mixing ratios, which were found to be higher in HTAP-MOZ, the net O3 production rates at the surface are higher in HTAP-RADM2 over most of the domain, especially in the IGP and central India. The net O3 production rates at the surface with HTAP-RADM2 are found to be 6 to 9 ppbvh-1 and more over the IGP, whereas these values are generally lower in HTAP-MOZ (4–8 ppbvh-1), except in the north-eastern IGP (>9 ppbvh-1). Figure 9b shows the sum of the chemical tendency and vertical mixing tendency at the surface for the HTAP-RADM2 and HTAP-MOZ. Analysis of the vertical mixing tendency revealed that higher surface ozone mixing ratios in the MOZART simulation are due to mixing with ozone-rich air from aloft. In the HTAP-RADM2 simulation, vertical mixing dilutes the effect of strong chemical surface ozone production. Further analysis of vertical distributions of chemical O3 tendencies reveals stronger photochemical production of ozone aloft with MOZART compared to RADM2 (Fig. S7). This leads to higher ozone mixing ratios aloft in MOZART simulations. A sensitivity simulation is conducted using a different PBL parameterization (Yonsei University scheme) to examine its influence on our conclusions. Comparison of monthly average (in April) planetary boundary layer heights between the two PBL schemes revealed that the differences are mostly within ±150 m with Yonsei scheme generally resulting in higher PBL heights over India (Fig. S9). Nevertheless, the chemical tendencies combined with vertical mixing tendencies of surface O3 are found to be nearly similar with Yonsei scheme (Fig. S10) as in the base runs using the MYJ scheme (Fig. 9b) with MOZART still producing higher ozone aloft (not shown) as in the original runs. Thus, changing the PBL scheme still results in production of more ozone aloft in MOZART, which is getting mixed with near surface air, which corroborates that our conclusions are not affected.

Mar et al. (2016) showed that RADM2 exhibits greater VOC sensitivity than MOZART (i.e. producing higher changes in ozone given a perturbation in VOC emissions) under noontime summer conditions over Europe. This is consistent with our findings as well, that the net surface photochemical ozone production is greater for HTAP-RADM2 than for HTAP-MOZART, given the high VOC emissions in the HTAP inventory. At the surface, the MOZART mechanism predicts larger areas of VOC-sensitivity (as diagnosed by the CH2O/NOy indicator, Fig. 10) and lower net photochemical ozone production than RADM2. With increasing altitude, both the HTAP-RADM2 and HTAP-MOZART simulations show a general increase of CH2O/NOy over India; i.e. the chemistry tends to exhibit increased NOx sensitivity with increasing height (Fig. S11). At model levels above the surface, HTAP-MOZART shows greater net photochemical production of ozone than HTAP-RADM2 (Fig. S7), which is what Mar et al. (2016) have also reported for the surface O3 over Europe. When these effects are combined, mixing leads to higher surface ozone mixing ratios for HTAP-MOZART than for HTAP-RADM2. A sensitivity simulation using a different photolysis scheme (Madronich TUV photolysis scheme) with HTAP-RADM2 setup revealed similar surface ozone mixing ratios and chemical tendencies at various model levels with small differences (<5 %) over most of the Indian region (not shown). So our results would be similar if we use Madronich TUV scheme instead of Fast-J scheme with RADM2. Further, Mar et al. (2016) used Madronich TUV scheme with RADM2 and Madronich F-TUV scheme with MOZART chemical mechanism and reported that the two different Madronich photolysis schemes had only a small contribution to the differences in the predicted ozone by two chemical mechanisms. The major difference between the two chemical mechanisms was due to differences in inorganic reaction rates (Mar et al., 2016). Hence, we conclude that in our study too, the differences over Indian region are primarily due to the choice of the chemical mechanisms irrespective of photolysis scheme used. Also note that the aerosol radiation feedback is turned off, so that the calculated differences mainly result from the representation of gas-phase chemistry rather than of aerosols between MOZART and RADM2. Our analysis also shows the importance of chemical regime in understanding differences between the chemical mechanisms and highlights the significant effects of the employed chemical mechanism on modelled ozone over south Asia.

Figure 11 shows a comparison of WRF-Chem simulated ozone variations on diurnal timescales with recent in situ measurements over a network of stations across south Asia for the two chemical mechanisms (MOZART and RADM2); again with the same emission inventory (HTAP). Qualitatively, both simulations produce very similar diurnal patterns (see also Fig. S12); however, the absolute O3 mixing ratios are found to differ significantly (Fig. 11) between the two chemical mechanisms. Noontime ozone mixing ratios predicted by MOZART are either significantly higher (at 9 out of 15 stations) or nearly similar (at 6 stations). MOZART-predicted O3 at Dibrugarh, Kanpur, Jabalpur, Bhubaneswar, Gadanki and Thumba was found to be higher by ∼12 ppbv, 5, 8 ppbv, 10, 11 and 12 ppbv, respectively, compared to RADM2 (Table S3). Over several urban and rural stations in India (e.g. Delhi, Ahmedabad, Pune, Kannur and Thumba) MOZART is found to titrate ozone more strongly during the night while resulting in higher or similar ozone levels around noon. The contrasting comparison between noon and nighttime found at these sites suggests that evaluation limited to 24 h averages would not be sufficient and that model performance on a diurnal timescale should be considered to assess the photochemical buildup of O3.

The model performance of two chemical mechanisms in reproducing diurnal variation at all stations is summarized using a Taylor diagram in Fig. 12. Both chemical mechanisms show reasonably good agreement (r>0.7) at most of the sites, except one station associated with highly complex terrain (Nainital). On the Taylor diagram, most of the HTAP-RADM2 results are found to be closer to the “REF”, as compared to HTAP-MOZ results, suggesting that the RADM2 chemical mechanism is better suited to simulate diurnal variation of ozone over this region.