TCEQ has O3 and other pollutants routinely measured at the Continuous Ambient Monitoring Station (CAMS) across the Houston region. Some of these stations also observe meteorological variables, such as wind speed and direction, temperature, and relative humidity (RH). These data can be downloaded from the Texas Air Monitoring Information System (TAMIS, https://www17.tceq.texas.gov/tamis/index.cfm?fuseaction=home.welcome, last access: 27 October 2023) website. A commercial shrimp boat and a pontoon boat owned by the University of Houston (UH) were operated mainly on the east and west sides of Galveston Bay, respectively. Another commercial boat, the Red Eagle, was docked to the north of Galveston Island and typically traveled up to 90 km offshore in the Gulf of Mexico and occasionally northward through the Ship Channel to the port of Houston. Automated O3 sampling instruments were installed on the three boats, with a compact weather station measuring temperature, pressure, RH, and wind conditions. The sample inlet was attached to an elevated location on the boats to avoid titration from the exhausts of the boats. Details of these devices can be found in Griggs et al. (2023). In addition, ozonesondes were launched from the pontoon and Red Eagle boats on selected days and locations to investigate the vertical O3 profiles. All the campaign data can be found on the TRACER-AQ website (https://www-air.larc.nasa.gov/cgi-bin/ArcView/traceraq.2021, last access: 27 October 2023).

During the offshore operational period of July to October, hourly averaged O3 mixing ratios exceeded 100 ppb several times. We identified O3 exceedance days, when offshore boat O3 observations registered a daily maximum 8 h average (MDA8) O3 in exceedance of 70 ppb, which is the current criterion of the NAAQS for O3. Six episodes with high O3 were obtained for 26–28 July, 25 August, 6–11 September, 17–19 September, 23–26 September, and 6–9 October. These episodes are accompanied by at least one CAMS site exceeding the 70 ppb MDA8 O3 threshold, indicating an extensive land–water air mass interaction.

This study used the WRF model v3.9.1.1. We set up three domains with different horizontal resolutions that cover the contiguous United States, southeastern Texas, and the Houston–Galveston–Brazoria region and referred to as domains d01, d02, and d03, respectively, as shown in Fig. 1. The corresponding horizontal resolutions and grid numbers for d01–d03 are 12 km × 12 km (373 × 310 grids), 4 km × 4 km (190 × 133 grids), and 1.33 km × 1.33 km (172 × 184 grids), respectively. All domains have identical vertical resolutions, with 50 hybrid sigma–eta vertical levels spanning from the surface to 10 hPa. Boundary conditions of the two inner domains were generated from the outer domain.

To select the WRF configurations that best represent the monitoring data, we designed eight model experiments with different initial and boundary condition (IC/BC) inputs, microphysics options, PBL schemes, data assimilation methods (e.g., observation nudging), and reinitializing techniques. Details of the design and evaluation of each experiment can be found in Liu et al. (2023). Based on the campaign-wide evaluation of the modeled meteorology, the best simulation was used to drive the CAMx model. The model configuration of the best simulation includes the hourly High-Resolution Rapid Refresh (HRRR) meteorological data as IC/BC inputs, the local closure Mellor–Yamada–Nakanishi–Niino (MYNN) PBL option (Nakanishi and Niino, 2009), and the Morrison double-moment (2M) microphysics scheme (Morrison et al., 2009), with no nudging and reinitializing techniques applied. Other settings used for the WRF simulation include the Monin–Obukhov similarity surface layer (Foken, 2006), the Noah land surface scheme (Chen and Dudhia, 2001), the rapid radiative transfer model (RRTM) longwave and shortwave radiation schemes (Iacono et al., 2008), and the new Tiedtke cumulus parameterization (Zhang et al., 2011).

This study also used the CAMx model v7.10. The three CAMx domains aligned with the WRF grids but had smaller spatial coverage. The corresponding horizontal resolutions and grid numbers for domains 1–3 are 12 km × 12 km (372 × 244 grids), 4 km × 4 km (156 × 126 grids), and 1.33 km × 1.33 km (153 × 162 grids), respectively. All domains have identical vertical resolutions, with 30 vertical levels from the surface to ∼ 100 hPa. The IC/BC inputs for the outmost domain are from the GEOS-Chem (v12.2.1) global simulation, with the National Emissions Inventory (NEI) 2011 nitrogen oxide (NOx) emissions scaled down to 2021. The Carbon Bond version 6 revision 5 (CB6r5) was used for gas-phase chemistry, including the inorganic iodine depletion of O3 over oceanic water (Burkholder et al., 2019). The first-order eddy viscosity (K-theory) diffusion scheme was selected for vertical mixing within the PBL, in which the vertical diffusion coefficients (Kv) were supplied from WRF outputs. Dry deposition is based on the Wesely scheme (Wesely, 1989).

Emission files with 12 and 4 km spatial resolutions from the preliminary 2019 SIP modeling platform provided by TCEQ are used in the simulation. These emissions include anthropogenic emissions, biogenic emissions generated from the Biogenic Emission Inventory System (BEIS), wildfire emissions based on the Fire INventory from NCAR (FINNv2), and ship emissions estimated from the Gulfwide Emissions Inventory (GWEI). No lighting emissions are included in the model. Since our domains are smaller than those in the SIP modeling, the original emission files were cropped to match the grid boundaries for CAMx to read properly. In addition, we redistributed the on-road emissions from 4 to 1.33 km over the Houston area. The 4 km emission fluxes were first disaggregated evenly to the 1.33 km grids and then collected onto major roads, using a 1 km rasterized road shapefile to produce on-major-road 1.33 km emissions. Some 1.33 km grid points off the major roads had missing values, which were filled using a smoothing method that averaged eight nearby grid points. The scaling factors for on- and off-major-road emissions were kept in order to maintain the on-road emission budget consistent before and after the spatial redistribution. Finally, total emissions were calculated by adding the 1.33 km on- and off-major-road emissions. The emissions for other sectors were also similarly interpolated to 1.33 km, without separating into no- or off-major-road temporary emissions. The redistributed emissions were tested to perform better in capturing the on-road distributions than using the Flexi-nesting function in CAMx (Fig. S1 and Table S1 in the Supplement), which can regrid the emissions on the fly.

The simulation was performed for two periods, 20–30 July and 20 August–13 October, to cover the six high-O3 episodes defined in Sect. 2.1. A 10 d spin-up before each period was applied. Other days in the two periods are considered clean scenarios, with low-O3 concentrations. Process analysis, including integrated process rate analysis (IPR), integrated reaction rate analysis (IRR), and chemical process analysis (CPA), was turned on when running the model. IPR contains O3 change rate from several chemical and physical processes, such as chemistry (CHEM), horizontal and vertical advection (ADV) and diffusion (DIF), and deposition (DEP). IRR provides detailed information about the reaction rate of all the chemical reactions in the CB6r5 scheme. CPA improves upon IRR by computing parameters useful for understanding O3 chemistry, such as the O3 production rate and regime. The abundance and reactivity of ozone precursors determine the ozone production regime, which can be indicated by the loss of HOx radicals (HOx = OH + HO2) through the termination of ozone chain reactions. Under low-NOx conditions, the most important HOx loss is the self-reaction of the hydroperoxyl radical (HO2), producing hydrogen peroxide (H2O2), which is used to represent NOx-limited ozone production. In urban areas with high-NOx concentrations, the dominant sink for HOx radicals is the oxidation of nitrogen dioxide (NO2) by hydroxyl radical (OH), resulting in the production of nitric acid (HNO3). Thus, the O3 formation regime in the model is determined based on the ratio of H2O2 production rate from HO2 self-reaction to HNO3 production rate from OH reaction with NO2, in which P(H2O2)/P(HNO3) < 0.35 indicates a VOC-limited regime, and ≥ 0.35 indicates a NOx-limited regime (Sillman, 1995). There is no transition scheme available in this method.

The time series of the daytime (10:00–18:00 central daylight time, CDT) mean O3 at the three boats are shown in Fig. 2a, and the evaluation statistics are listed in Table 1. The evaluation excludes nighttime data to reduce the effects from land, as the boats stayed at the dock at night. Indeed, an hourly time series evaluation with nighttime data included (Fig. S2 and Table S2 in the Supplement) shows a larger bias between modeled ozone and boat observations. The spatiotemporal variability in the daytime O3 at the three boats is well captured by the model, with a correlation coefficient (R) value greater than 0.70. Overall, the model overestimates daytime O3 by 4.57 ppb (11 %), 7.82 ppb (22 %), and 4.35 ppb (9 %) for the pontoon boat, Red Eagle, and shrimp boat, respectively. On episode days, high-O3 mixing ratios can be found over Galveston Bay and the Gulf of Mexico (Fig. 2b). The model captures some of the variability (R = 0.42–0.51), with negative mean bias (MB) values of ∼ 4.5 ppb (8 %) for the pontoon and shrimp boats and a nearly unbiased simulation (MB = 0.05 ppb) for the Red Eagle boat. While the O3 variability is better predicted on clean days (R = 0.69–0.76), the model shows higher values of MB than those on high-O3 days, ranging from 9.15 ppb (29 %) to 11.28 ppb (41 %), which drives the overall model overestimation.

While we did not find any previous efforts modeling offshore O3 in the Houston area to compare our results, an evaluation against onshore measurements can help validate our model performance. The time series of the daytime mean O3 from simulations and observations from CAMS sites are displayed in Fig. 3, and the evaluation statistics are summarized in Table 2. The model captures the onshore O3 variability (R = 0.79), with an overall overestimation of 7.89 ppb (20 %), mainly due to the high positive bias of 10.93 ppb (34 %) on clean days. This result is comparable with the model performance from previous studies focusing on the same area (e.g., Xiao et al., 2010; Pan et al., 2015; Kommalapati et al., 2016), which further verifies the reliability of our model settings.

We also evaluated the modeled vertical O3 profiles against the afternoon (12:00–18:00 CDT) ozonesondes launched over Galveston Bay and the Gulf of Mexico. During the study period, there were five and nine afternoon launches over Galveston Bay on clean and O3 episode days, respectively, while the Gulf of Mexico only had five afternoon launches during high-O3 events. All the ozonesondes available for high-O3 conditions are from the two episodes of 6–11 and 23–26 September. The average O3 profiles from these launches are shown in Fig. 4. Free-tropospheric O3 with altitudes greater than 2 km is underestimated for both locations on both clean and O3 episode days, which indicates that the long-range transported O3 is underrepresented by the model. Over Galveston Bay, the overestimation of O3 within the mixed layer below 2 km on clean days changes to underestimation on episode days, and the underestimation increases from 5 ppb at the surface to 10 ppb near 1 km. The two high-O3 episodes in September are featured by a O3 plume between 2–3 km, as shown by the O3 lidar observations in Liu et al. (2023), which is missed by the model, as shown by the two example days on 9 and 24 September of each episode over Galveston Bay (Fig. S3 in the Supplement). The underestimation of O3 in the lower free troposphere and the mixed layer on episode days can be partly explained by the missing of the high-O3 plumes, which can be mixed down when the cap inversion is weak (Liu et al., 2023). There is an approximately 10 ppb underestimation across all altitudes below 4 km over the Gulf of Mexico. An ozonesonde from the Gulf of Mexico on 9 September recorded high ozone up to the top of the marine layer at 370 m, which is missed by the model and leads to the highest bias. This case will be discussed in the case study in Sect. 3.3.

To conclude, despite the overall model bias for vertical O3 distributions, the acceptable model performance for offshore and onshore O3 prediction at the surface indicates that the modeling system can be applied to conduct process analysis and help identify the processes influencing high-O3 concentrations over the water surface.

This section examines how the CAMx-simulated O3 processes change during high-O3 episodes relative to clean days. The process analysis is calculated over a subregion of the Gulf of Mexico, with high-O3 mixing ratios observed (black box in Fig. 2b) and integrated across the lowest five model layers that are comparable to the morning PBL heights over water. The diurnal average of each process on clean and O3 episode days is shown in Fig. 5a. Chemistry (CHEM) is the major O3 source during daytime and becomes the primary O3 sink after sunset. Advection (ADV) serves as a pathway for an O3 sink for most hours, especially during the day, while vertical diffusion (DIF) mostly contributes as an O3 source. Deposition (DEP) constantly removes O3 from the atmosphere at all hours, yet with a marginal value of 0.1 ppbh-1. Similar patterns can be found over the Houston urban area, with a much bigger magnitude (Fig. S4 in the Supplement). During high-O3 events, CHEM is the most important process causing higher-O3 levels over water relative to clean days, followed by vertical DIF (Fig. 5b). We examined the simulated O3 vertical profiles and PBL heights averaged over the process analysis region on clean and episode days in Fig. S5 in the Supplement. O3 across the entire profile is higher on episode days than clean days, indicating an elevated O3 background on high-O3 days. In addition, the profiles of potential temperature observed by the ozonesondes show an inversion layer at ∼ 1.5 km on episode days (Fig. S6 in the Supplement). More vertical diffusion can occur if high O3 in the inversion layer is mixed down from above the PBL when the capping inversion is weak (Liu et al., 2023).

The CPA analysis can provide more insights into the enhanced O3 production during high-O3 events. We first investigated the rates of HO2 self-reaction and OH reaction with NO2 in Fig. 6a and b, since they are used by the model to determine the O3 chemical regime. A region-wide increase in the HO2 self-reaction rate leads to the enhancement of PO3 under a NOx-limited regime (Fig. 6c). Similarly, the frequency of PO3 under a NOx-limited regime also increases regionally (Fig. S7 in the Supplement). The frequency at each grid cell is the ratio of the number of hours with a greater than zero NOx-limited PO3 to the total midday hours (11:00–15:00 CDT) during the study period. HO2 is formed, following the oxidation of VOCs by OH. Thus, we further compared the OH reactivity of VOCs averaged from 11:00 to 15:00 CDT on clean and episode days in Fig. 7. Isoprene has the highest contribution to the total VOC reactivity on the land, but its reactivity does not increase during high-O3 events. Instead, paraffin, formaldehyde, and acetaldehyde are the three VOCs experiencing the highest increase in the reaction rate with OH over both land, by 0.22 ppbh-1 (84 %), 0.19 ppbh-1 (45 %) and 0.15 ppbh-1 (73 %), and water, by 0.18 ppbh-1 (114 %), 0.15 ppbh-1 (44 %) and 0.11 ppbh-1 (82 %), respectively, which indicates a higher contribution from regional transport on episode days, as they are relatively long-lived VOCs capable of traveling long distances. Indeed, the paraffin IPR analysis shows that the ADV process dominates the increase in the paraffin during morning hours from 06:00 to 11:00 CDT over water (Fig. S8 in the Supplement). The trajectory analysis focusing on two O3 episodes in September shows that air masses were transported from the northern/central states (Soleimanian et al., 2023), consistent with the wind directions demonstrated in Fig. 6. Such wind conditions can also bring NOx emissions from the Houston Ship Channel downwind towards the western side of Galveston Bay and the Gulf of Mexico, causing a higher-OH reaction rate with NO2 (Fig. 6b) and enhanced PO3 under a VOC-limited regime (Fig. 6d) therein.

In summary, O3 chemistry is the major process responsible for the high-O3 mixing ratios over the Gulf of Mexico during the study period. The VOC species with a long lifetime advected from the northeastern increase over land and water, leading to a region-wide enhancement of PO3 under a NOx-limited regime. The downwind transport of NOx from the Houston Ship Channel also expands the VOC-limited area towards the west side of Galveston Bay and the Gulf of Mexico, contributing to the higher-than-normal PO3.

Although the above analysis reveals the general reasons responsible for the high offshore O3 events, the multiple-day average can miss out on some important aspects regarding the causes of these events. In this section, we selected 2 case days, 9 September and 7 October, to further demonstrate the development process of high O3 in detail.