The impact of observation nudging on simulated meteorology and ozone concentrations during DISCOVER-AQ 2013 Texas campaign

Accurate meteorological fields are imperative for correct chemical transport modeling. Observation nudging, along with objective analysis, is generally considered a lowcost and effective technique to improve meteorological simulations. However, the meteorological impact of observation nudging on chemistry has not been well characterized. This study involved two simulations to analyze the impact of observation nudging on simulated meteorology and ozone concentrations during the 2013 Deriving Information on Surface conditions from Column and Vertically Resolved Observations Relevant to Air Quality (DISCOVER-AQ) Texas campaign period, using the Weather Research and Forecasting (WRF) and Community Multiscale Air Quality (CMAQ) models. The results showed improved correlations between observed and simulated parameters. For example, the index of agreement (IOA) improved by about 9 % for surface temperature and 6–11 % for surface zonal (U-WIND) and meridional (V-WIND) winds when observation nudging was employed. Analysis of a cold front event indicated that nudging improved the timing of wind transition during the front passage. Observation nudging also reduced the model biases for the planetary boundary layer height predictions. Additionally, the IOA for CMAQ simulated surface ozone improved by 6 % during the simulation period. The high-ozone episode on 25 September was a post-front ozone event in Houston. The small-scale morning wind shifts near the Houston Ship Channel combined with higher aloft ozone early morning likely caused the day’s ozone exceedance. While observation nudging did not recreate the wind shifts on that day and failed to reproduce the observed high ozone, analyses of surface and aircraft data found that observation nudging helped the model yield improved ozone predictions. In a 2 h period during the event, substantially better winds in the sensitivity case noticeably improved the ozone. The average IOA for ozone in the period increased from just over 0.4 to near 0.7. Further work on improving the capability of nudging to reproduce local meteorological events such as stagnations and wind reversals could enhance a chemical transport model’s skill for predicting high-ozone events.


Reviewer 2
Original comments by reviewer 2: Specifically, the sentences are ambiguous, incomplete, and awkward throughout the text. One has to go over several times for many sentences to guess what the authors are trying to say. The amount of corrections needed is beyond what a reviewer can suggest in details. I am listening a few typical issues below. *** It should have been clearly stated early on what exactly the study is trying to accomplish; what variables they "nudged" exactly, and what results they examined. A reader should not have to read through all the details of the model setups to find out what variables they actually "nudged". Statements such as ". . .the impact of OA on the simulated meteorology and ozone concentrations. . ." or ". . . indicated that OA improved the timing of wind transition . . .", are throughout the paper without indicating OA on what, or nudging what.
Another issue is that this manuscript was not written for more general readers, terminologies were used without providing background. They never explicitly explain the connection between WRF and CMAQ before using WRF-CMAQ. The terms nudging and OA were used interchangeably without explaining the differences. **************************

General response:
We thank the reviewer for his/her input. As for scientific significance, we think the paper has several findings not seen before in the previous studies (regarding more detailed impact of objective analysis on meteorology including temperature, winds and PBL height and chemical concentrations of ozone). And these findings are important for later works in their effort to improve the WRF's nudging process. In the revised manuscript, we worked to clearly explain about how the objective analysis improved the performance of WRF and WRF-CMAQ simulations. For example, we showed how the nudging process improved the meteorology and chemical concentrations on the 25 September (Our group worked as air quality forecasting group for DISCOVER-AQ project performed in Houston in September of 2013. As we authors acknowledged, none of the previous forecasting/modeling exercises from a few modeling groups couldn't make a reasonable simulation of ozone on September 25).

In this study, we showed that objective analysis approach significantly improved meteorology for September and it also improved chemical concentrations of ozone, but the order being improved is smaller than that of meteorology. Further, we discussed what would be another cause for the uncertainty in the ozone simulation.
On the language issue, we acknowledge the paper was heavy and occasionally hard to follow.

Itemized response:
Since the reviewer's comments are not with line numbers, we try to respond as best as we can.
1. One of the major concerns is the contribution of the paper to science. Here we elaborate a few points. a. Meteorology is the foundation for emission and chemistry studies. Without a good set meteorology when studying real-world air quality, one can hardly draw conclusions with. As we addressed above, a good example is the 09/25 episode discussed in the paper, we still cannot say whether an emission event or unknown large-scale transport played major roles because meteorology is not well simulated. It was our motivation for this study in the beginning. b. FDDA is a critical tool in improving model performance in meteorology modeling. A large portion of performance gain in the last 30 years came from FDDA. Objective analysis (OA) and observation nudging are critical methods in FDDA. c. We have not found any paper providing a quantitative sensitivity study on the possible performance gain from OA and observation nudging on both meteorology and chemistry. Ngan et al. (2012) is the closest in it evaluating both meteorology and chemistry yet it is not a sensitivity study on obs-nudging. d. We have not found any other air quality study performing 1-hr observation nudging. e. More importantly, the paper provided an example showing the failure of model to replicate the high ozone even after OA and observation nudging are performed. Implicitly, this means that trying different physical schemes is unlikely to solve the issue (we did test a few cases although this statement largely came from our years experience in air quality modeling). This begs for the question: what can we do to recreate the right meteorology so that we can produce high ozone in chemistry model? As we addressed in the discussion of the revised version, we have been working to develop a novel technology (a new nudging process) to address the issue (preliminary results are surprisingly encouraging). f. We evaluated model PBL and ozone aloft to study the model sensitivity to obs-nudging and OA -which we have not seen before.

About nudging technique and related terminology.
a. The authors expected the readers to have basic knowledge on FDDA and nudging, i.e., what is nudging, what it for is, etc. We didn't provide further explanations on certain details due to the large size of the revised manuscript. However, we added some explanation in the revised version, e.g., FDDA in line 49-50, Objective Analysis in line 70-72. b. We have updated the manuscript to clarify the terminology, especially what we did in addition to standard grid nudging. We assimilated observations from various sources into the simulation. WRF has a separate program "OBSGRID" to perform objective analysis (OA) and create input for observation nudging. Therefore, performing OA generally means doing observation nudging simultaneously to make the most use of observations. However, the term "OA and observation nudging" is too cumbersome. Therefore, in revised manuscript, we used "observation nudging" and explained that it means the combined "OA and observation nudging" -line 72-75. 3. WRF-CMAQ and OA/observation nudging a. We changed "WRF-CMAQ" to "WRF and CMAQ", line 21. b. We explained the difference between OA and observation nudging. OA is to improve first guess meteorology analysis by assimilating observations, which is to improve the quality of analysis nudging. Observation nudging is to nudge model state toward observations at measurement locations -line 65-72.

The language issue
We have done a major overhaul on the language part to improve the readability, with input from several co-authors. As one can see from the track-change version, many changes were made. This is in addition to the modification suggested by the handling editor, reflected in the ACPD version.
We also made better plots to assist readability. We agree that the paper is not a light read and it has much information, especially on the 25 September, the high ozone day. We think most of them are quite relevant but did trim some contents, such as the ozone description on the 26 th .  Seaman (1990, 1994), and implemented in the Fifth-62 Generation PSU/NCAR Mesoscale Model (MM5). Not intended for optimal adjustment, nudging 63 is less computationally intensive andbut needs special care for the nudging coefficients. Nudging 64 involves adding an artificial tendency term to one or more model prognostic equations that 65 reflect the difference between the best estimate of the observed state and the model state at a 66 given location and time. In short, the goal is to "nudge" model state towards observed state. 67 There are several types of nudging such as 3D analysis nudging, surface analysis nudging, and 68 observation nudging. (obs-nudging). In the case of analysis nudging, the model state is nudged 69 toward gridded analysis. The difference between 3D and surface analysis nudging is that 3D 70 analysis (at all model levels except for surface) data are used to improve 3D fields while surface 71 analysis data are used to improve surface fields. In observation nudging, the model is perturbed 72 such that its predictions are nudged to match better with observations at individual locations, 73 both on surface and aloft. The MM5 nudging codes were later improved and incorporated into 74 the Weather Research and Forecasting (WRF) model by Liu et al. (2005Liu et al. ( , 2006. The 75 enhancements enable observation nudging to assimilate a large variety of direct or derived 76 observations. obs-nudging to assimilate a large variety of direct or derived observations. In 77 WRF, the inputs for obs-nudging are generated by WRF OBSGRID program. This program also 78 performs Objective Analysis (OA) to improve the quality of analysis nudging files. Objective 79 Analysis updates first guess meteorology analysis by incorporating observational data. Since 80 obs-nudging is usually performed along with OA (as in this study) to maximize the benefits of 81 assimilating observations, we also use OA to denote the combined Objective Analysis and obs-82 nudging processes in case names. Although not elaborated here, the WRF-CMAQ sensitivity to different observation nudging 112 frequencies was also explored. In theory, higher frequency of observation obs-nudging input 113 should have a higher probability to capture small scale events, such as local wind shifts. These 114 events may only slightly impact local weather, yet they can have a largemarked effect on 115 chemistry since it. This is well-known thatbecause local stagnation and wind convergence/reversals can contribute to the pollutant build-up (e.g.,, as indicated by Banta et al. 117 (1998 in-situ surface and aircraftaloft measurements. We evaluated model-measurement performance 157 and calculated statistics for both WRF and CMAQ. Meteorological fields critical to ozone 158 chemistry were examined to explore the model sensitivity to OA. The paper is structured as 159 following: Section 1 is introduction; Section 2 describes the measurement data and the modeling 160 system; Section 3 covers the evaluation protocols; Section 4 discusses the general meteorological 161 conditions that occurred during the campaign period; Section 5 presents the modeling results, 162 and Section 6 provides discussions and conclusions. output. 163 164 input,is downloadable from: http://rda.ucar.edu/datasets/ds608.0/. The NARR 228 http://rda.ucar.edu/datasets/ds608.0/. The data were based on an Eta 221 grid at 29 pressure 229 levels. Its horizontal resolution was 32-km and the frequency was 3-hourly. The initial and 230 boundary conditions were generated from the NARR analysis by WRF. An alternative to NARR 231 was the Eta-NAM analysis data. However, the data temporal frequency was lowered from 3-232 hourly to 6-hourly starting 2013. Our testtests showed that it was not as good as NARR for WRF 233 input,dataset -likely because of lower temporal resolution. 234

Physics and FDDA Options 235
Major physics options were used in the model are listed in Table 1 namelist variable in WRF's namelist file. For example, the "1" after YSU is the value of the 245 namelist variable "bl_pbl_physics" in WRF's namelist file. For both of the simulations, we 246 performed standard grid nudging for both of the cases using NARR analysis. For grid nudging 247 options, we generally followed the recommendations in the WRF's User Guide. For example, the 248 mass fields (temperature and moisture) were nudged only at layers above the PBL while wind 249 fields were adjusted at all layers including the surface layer. 250

Observation Nudging with MADIS and CAMS data in WRF 251
As mentioned in the introduction, observation nudging is regarded as a low-cost and effective 252 method for improving meteorological model performance, but it requires additionalAdditional 253 observational data. In this study, we acquired are required to implement obs-nudging and OA. To 254 generate the input files for the OBSGRID program, we processed the observation data and generating files in "little_r" format using similar procedures found inthe approach of Ngan et al. The "little_r" files from previous stepprocessed input observation data were fed into WRF 266 OBSGRID module to update the domain analyses ("met_em" files), and, generate additional 267 surface analyses ("sffdda") and text nudging files ("OBS_DOMAIN").. Actual observation obs- Theoretically, observation obs-nudging updating at a higher frequency should enhance the 273 model's performance. A typical frequency of input analysis data is 3-hourly while the frequency 274 for observational data is hourly. The 3-hourly frequency of input analyses may be the reason for 275 the default 3-hour time-interval in WRF's OBSGRID settings for generating the observation 276 obs-nudging files. Since there were few existing OAobs-nudging studies related to air quality 277 and we are not aware of any reference to the adoption of 1-hour input frequency, we assume that 278 all the existing studies used the default 3-hour interval. As the WRF model allows the interval to 279 be set to 1-hour or smaller when corresponding observational data were available, we tested both 280 1-hour and 3-hour scenarios. The results indicated that 1-hour OAobs-nudging had slightly better 281 performance than the 3-hour one. As a result, this study adopted 1-hour temporal frequency for 282 observation nudging. The quantities that were nudged were temperature, moisture, and the two 283 wind components (U-WIND and V-WIND). Obs-nudging for moisture was not performed in 284 this study. This was based on our past experiences since performing moisture nudging sometimes 285 trigger excessive artificial thunderstorms which disrupted model flow fields. 286 It should be noted that the default time interval for modified gridded analyses, i.e., the "metoa-287 em" and "sgfdda" files have to match input analysis data in OBSGRID. The namelist variable 288 was called "interval", with a default value of "10800" seconds. The time-interval for output 289 nudging files was set by namelist variable "int4d", with the same default value of "10800" 290 seconds. To output the observation nudging files hourly, "int4d" should be set to "3600" 291 seconds. This means that the OBSGRID output files, "metoa_em" and "OBS_DOMAIN", did 292 not have the same interval in our study. 293 In WRF, there were a few namelist variables controlling the frequency of grid nudging and 294 observation nudging. The first one was "interval_seconds", which should match the interval of 295 input grid nudging files ("met-em"). The second one was "sgfdda_interval_m", matching the 296 interval of surface grid nudging files ("sgfdda"). In our simulation, both intervals were equal to 297 3-hours. The third one was "auxinput11_interval", controlling the updating interval for 298 observation nudging files ("OBS_DOMAIN"). The last one, "obs-ionf", determined the nudging 299 frequency relative to internal integration time-step. For example, if the integration time-step for 300 the coarse domain is 30 seconds, setting "obs_ionf" to 1 means performing OA every 30 301 seconds, while setting "obs_ionf" to 3 means performing OA every 90 seconds. In our 302 simulation, "obs_ionf" is set to 1. 303 One departure from the default OA setting in WRF was that the moisture OA was turned off with 304 "obs_nudge_mois" set to 0. This was based on our past experiences since performing moisture 305 OA sometimes trigger excessive artificial thunderstorms which disrupted model flow fields. 306

EmissionEmissions Processing 307
For anthropogenic sources we utilized the National EmissionEmissions Inventory of 2008 308 performance using NEI2008 appears reasonable. 320

CMAQ Configurations 321
The USEPA's CMAQ (Byun and Schere 2006) version 5.0.1 was adopted for this study,

407
To evaluate the WRF simulation, we calculated statistics for surface temperature and winds in 408 the 4-km domain. For PBL heights, we chose to plot out the time-series for the one site we had 409 observations due to significant amount of missing data (data coverage is about 50%). For CMAQ 410 evaluation, we calculated the surface ozone statistics for the whole month. Also, we plotted 411 vertical ozone profile and calculated biases for aloft ozone aloft on 09/25.the 25 th . 412

Temperature 414
The comparison of regional averaged daily temperaturesaverage hourly temperature for the 415 analyzed timesimulation period is shown in Figure 3. The regional observed averaged 416 dailysurface temperature was calculated by averaging the hourly temperature from ~60 CAMS 417 sites in the 4-km model domain. Despite the differences inThe base case temperature was too 418 high compared to the in-situ measurements. For example, the days with more 419 clouds/precipitation, No-OA maximum temperature for the simulated averaged temperatures 420 tracked21 st was 30 o C compared to 25 o C for the in-situ data very well. It was also evident that. 421 The statistics of hourly surface temperature are presentedlisted in Table 3. With higher IOA and 424 lower mean biases (MB), the "1Hr-OA" case was clearly better than the base case "No-OA". The 425 IOA of "1Hr-OA" was about 9% higher than the base case. showing weak southerly at 00 CST while all the others had mostly weak northerly. Starting from 508 01 CST, winds in the entire HGB area turned northerly to northeasterly and continued gaining 509 strength in the next few hours, indicating cold air had taken over the region. 510 Both cases performed reasonably well on 09/21 and the timing of wind shift was captured quite 511 accurately; although "No-OA" lagged about an hour.behind by ~ 1 hr. The winds turned weak 512 northerly at 00 CST for most sites andbut the "No-OA" case still showed the wind direction to be 513 all southerly. Besides the timing, OA also helped moderate the winds as the northeasterly winds 514 in "No-OA" case sometimes were too strong; obs-nudging helped moderate the winds. The 515 reduced V-wind bias in "1Hr-OA" was also evident in the wind model-measurement statistics on 516 09/21 is reduced from -2.5 m/s to -0.6 m/s after OA was performed. The performance of the OA 517 case during cold front passage was consistent with our past simulations.September. 518 were seen for some days. 533 Although model had high biases for majority of the days, biases were consistently lower for the 534 OA case during two periods: 09/07a reasonably good job on capturing the timing of intra-day 535 variations. However, both cases tended to 09/09overpredict the daily highs and 09/17 to 09/21. 536 The reduced biases were likely due to daily lows, especially in the first 8 days and between 15 537 and 21 September. An obvious departure is the 25 thboth cases missed the daily high. During 538 the model high bias period, the OA case usually did better in reaching the daily low although it 539 overpredicted the high a bit more than the base case. The night time biases were reduced likely 540 because the lower southerly winds in the OA case since model had higher backgroundtransported 541 less ozone originated from the Gulf. to the land. 542 In Figure 4, the first three orange circles showed the days with high model biases. The first two 543 circles consisted of days with lower Our results suggested that the modeled ozone concentrations 544 were likely higher in the Gulf than "normal" background ozoneactual. However during the 2 nd -545 4 th and 7 th -8 th of September, the incoming ozone from the Gulf was markedly lower. Since the 546 model ozone had fixed boundary values, the model was unable to capture the daily ozone 547 There were a few days with elevated ozone due to post-front meteorology conditions. The only 554 exceedance happened on 09/25, which was likely caused by meteorological events in Houston 555 and the Galveston Bay. AveragedThe overall ozone on 09/26 September was slightly higher after 556 southerly winds transported back the ozone from the Gulf, raising the ozone level in the entire 557 region. A more detailed analysis of model predictions on 09/25 and 09/26 will be presented in 558 following subsection of 5.2.3. 559

Performance Statistics 560
The ozone statistics were displayedare listed in Table 4. Both cases had very close correlation of 561 0.72 and 0.73. However, the mean biases in the OA case were lower by 3.2 ppb, which helped 562 raise the IOA from 0.78 to 0.83. The model standard deviation increased in the OA case and 563 matched better with observation.that of the in-situ data. The improvement in IOA was slightly 564 less inas compared to that for temperature and winds. From theThe wind plots of Figure 9, we can see indicate that the winds in the HGB region at 8 592 CST were light northerly for sites located on the north side while windsthey were mostly 593 westerly for the sites in the middle and south. The base case winds were all northerly while the 594 OA case had northwest winds for north side and west winds for the middle and south. TheHence, 595 the model winds in OA case were muchare more realistic than the winds in base case. The 909 596 CST winds were similar to those of 808 CST. As a result, the ozone statistics in Table 5 showed 597 that the OA case had much better correlation and IOA than the base case during 8-908-09 CST. 598 This example demonstrated OA'sthe ability of obs-nudging to correct erroneous winds. 599 However, later events showed OAit may not always be able to perform consistently. 600 The bay breeze started to develop at 10 CST near the C556 site. The early onset was likely to be 601 related to the warming up on the previous afternoon on 09/24 (as indicated in Figure 3).. At 10 602 CST most other sites to the west of HSC experienced light northwest winds while windsthose at 603 HSC wereoriginated from the northeast. Combined with the easterly bay breeze, a convergence 604 zone was formed just below C556, where emissions from the HSC area stalled and accumulated. 605 At 13 CST, the whole region had light winds and the bay breeze was well developed. The 606 highest ozone indeed appeared in C556 and its vicinity. The rapid increase of ozone 607 concentration for C556 between 909-13 CST is shown in Figure 8. 608 It is important to note that both modeled cases missed the wind shifts in the HSC area, and the 611 resultedresulting convergence zone near C556. This could explain the model's inability to 612 recreate the sharp ozone increase at C556. Figure 9 shows that the ozone levelconcentrations 613 around HSC area isare quite low (~10 ppb) at 08 CST. A further examination showed that while 614 both model cases missed the wind shift and convergence, though the patterns were different. The 615 base case had flawed winds for most of the morning: instead of a weak northwesterlywesterly, it 616 had stronger northeasterlynorthwesterly to northerly. By 08 CST, winds were almost uniformly 617 northerly in the base case while they were weak west-northwesterly in the OA case (Figure 9). 618 The oval in Figure 9's top-left panel shows the mismatch of winds around C556 in the base case. 619 As a result, the NO x produced in the city was carried further to the southeast in the model in the 620 base case. Until 13 CST, base case winds did not shift directions by much. The OA case got the 621 early hour weak northwesterly right, but missed the bay breeze onset between 10 and 13 CST 622 (oval in Figure 10). The OA case could not reproduce the small-scale wind reversal near C556, 623 suggesting there is a limitation in the current WRF OA's capability. On the other hand, the OA 624 case did improve the spatial ozone pattern, as the high ozone area was closer to HSC after OA 625 ( Figure 10). 626 While it is easy to understand the improvements in temperature and winds after obs-nudging was 687 applied, it is more difficult to explain how other variables such as precipitation and clouds 688 reacted to obs-nudging. The indirect impact of these meteorological variables on ozone was 689 harder to assess. In our study, we did not evaluate clouds quantitatively as there were no 690 digitized cloud fraction data available for our modeling domains. A preliminary analysis on 691 convection showed that there were occasions in which model missed the convection or 692 precipitation and there were other occasions in which model created artificial convection. The 693 convection cells were usually visible as "star-burst" from surface wind vector plotsarrows 694 going out to different directions from a center. However, the mismatch in convection appeared to 695 be not a serious issue since only a few occurrences were observed in the month of September. 696 697 Figure 13. Vertical ozone profiles from 09/25_08 CST to 09/25_16 CST of 2013 for two cases 699 of No-OA and 1Hr-OA compared with corresponding observations. 700