Particulate matter with an aerodynamic diameter of 2.5 µm or smaller (PM2.5) is a key contributor to poor air quality in the United States, posing significant risks to public health and the environment, and contributing to substantial loss of life (Cohen et al., 2017; Colmer et al., 2020; Huang et al., 2025). Over the past few decades, poor air quality in the US has contributed to over 100 000 premature deaths annually, far exceeding fatalities from all other weather-related causes combined, which average around 500 per year (Huang et al., 2025). Given its public health significance, PM2.5 is one of the important pollutants used in calculating the Air Quality Index (AQI) – a standardized system designed to communicate daily air pollution levels to the public at the US Environmental Protection Agency (EPA). Elevated PM2.5 concentrations frequently result in “unhealthy” AQI ratings, triggering health advisories and public warnings.

PM2.5 in the United States originates from a range of both anthropogenic and natural sources. Anthropogenic sources include agricultural activities and combustion processes, such as emissions from motor vehicles, power plants, industrial facilities, and residential heating systems. Among natural sources, wildfires are a particularly significant contributor, especially in the western United States, where their frequency and intensity have escalated dramatically over the past two decades (Wen and Burke, 2021). According to the US Environmental Protection Agency (EPA), wildfires account for approximately 15 % to 30 % of total PM2.5 emissions nationwide (EPA, 2017). While national seasonal averages of PM2.5 have generally declined, summer PM2.5 concentrations in the western US have remained persistently high, primarily due to wildfire smoke (O'Dell et al., 2019). In addition to degrading air quality, wildfires have caused widespread property loss. Since 2005, more than 129 000 homes, businesses, and other structures have been destroyed by wildfire-related events (https://headwaterseconomics.org/natural-hazards/structures-destroyed-by-wildfire, last access: 18 February 2026), underscoring the urgent need for more effective strategies in air quality monitoring, forecasting, and wildfire management.

The National Oceanic and Atmospheric Administration (NOAA) has developed an advanced regional Air Quality Modeling (AQM) prediction system within the Unified Forecast System (UFS) framework to enhance the accuracy of air quality forecasts across the United States, particularly during wildfire events (Huang et al., 2025). The National Air Quality Forecast Capability (NAQFC), operated by NOAA's National Weather Service (NWS), has been providing operational air quality forecast guidance for over 20 years, with continuous inclusion of new capabilities. Under NAQFC, the AQM version 7 was implemented and became operational on 14 May 2024. The system features online coupling of atmospheric and chemical models, allowing dynamic interactions between meteorology and atmospheric chemistry. This integration improves the representation of emissions and ensures real-time feedback of meteorological fields that influence chemical transformations and the transport of pollutants in the atmosphere. The UFS-AQM online system has consistently shown improved performance in simulating major wildfire events, including the significant wildfires in the northwestern coastal regions of the US in September 2020, and widespread smoke transport from Canadian wildfires in the summer of 2023. This system was officially implemented on 14 May 2024 as NOAA's operational air quality prediction system (AQMv7), replacing the previous offline-coupled Global Forecast System using the Finite Volume Cube-Sphere dynamical core (GFS-FV3) version 15 with the Community Multiscale Air Quality modeling system (CMAQv5.0.2) (Chen et al., 2021).

PM2.5 data assimilation (DA) has proven effective in reducing errors in air quality forecasts (e.g., Pagowski et al., 2010, 2014; Schwartz et al., 2012; Wu et al., 2015; Robichaud, 2017; Sun et al., 2020; Lee et al., 2022; Chen et al., 2022; Ha, 2022; Park et al., 2022; Vogel et al., 2025, among others). Pagowski et al. (2010) demonstrated that fine aerosol forecasts benefit from AirNow PM2.5 DA, showing improved verification scores for a period of at least 24 h. Schwartz et al. (2012) found that assimilating AirNow PM2.5 observations significantly improved surface PM2.5 forecasts over the CONUS compared to forecasts without DA. Wu et al. (2015) reported that incorporating ground-based PM2.5 observations notably enhanced 24 h forecasts during a severe pollution episode in Shanghai. Similarly, Chen et al. (2022) showed that assimilating multi-source PM2.5 data significantly improved WRF-Chem PM2.5 forecasts with benefits lasting up to 48 h. Lee et al. (2021) highlighted the effectiveness of assimilating ground in-situ surface PM2.5 observations in improving the short-term PM2.5 predictions in Northeast Asia.

Many operational regional air quality prediction systems around the world use some form of data assimilation to initialize the forecasts. These approaches vary in complexity, ranging from simple optimal interpolation to full variational or ensemble Kalman filter methods (e.g. Robichaud et al., 2016; Wei et al., 2024; Colette et al., 2025). In NOAA's current regional air quality model (AQM) operations, aerosol and chemical initial conditions are “warm-started” using 6 h forecasts from the previous model cycle. The implementation of an aerosol data assimilation system can further enhance short-term air quality forecasts by providing more accurate spatial analyses of initial aerosol distributions.

To establish aerosol data assimilation capabilities for NOAA's regional operational AQM system, we employ the Joint Effort for Data assimilation Integration (JEDI) (Trémolet and Auligné, 2020). JEDI is a flexible, agnostic, and modern data assimilation system applicable to a wide range of forecasting systems (e.g. Liu et al., 2022; Huang et al., 2023; Sluka, 2024). JEDI offers a platform that supports efficient scientific development and facilitates the transition from research to operations. As part of a broader strategic shift, NOAA and partner agencies are transitioning their data assimilation systems to JEDI, opening the door for rapid integration of new scientific advancements, greater consistency across modeling systems, and enhanced collaboration across research communities and operational centers.

This study aims to develop and evaluate an initial aerosol analysis capability for the NOAA's regional AQM system by assimilating PM2.5 observations using the JEDI three-dimensional variational (3D-Var) data assimilation framework. Compared to previous PM2.5 data assimilation studies, this research adopts the NOAA's regional operational AQMv7 system and incorporates a new PM2.5 transform in JEDI for assimilating PM2.5 observations. In addition to evaluating the impact of assimilating AirNow PM2.5 measurements on air quality prediction, this study also examines the impact of assimilating low-cost PurpleAir observations. Although PurpleAir data are valuable for PM2.5 analysis (White et al., 2026) and real-time air quality monitoring, their potential impact on numerical air quality prediction remains insufficiently explored. To the authors' best knowledge, this is the first study to demonstrate the value of PurpleAir observations for air quality prediction during the wildfires of September 2020 using the AQMv7 system.

The paper is organized as follows: Sect. 2 provides a description of Methodology including the NOAA's AQM system, 3D-Var approach, and JEDI PM2.5 assimilation. Experimental setup is presented in Sect. 3 including case description, AQM configuration, AirNow and PurpleAir PM2.5 observations and background errors setup. Results are described in Sect. 4. A summary and discussion are presented in the final section.

This section provides an overview of the impact of DA on PM2.5 forecasts. A total of 80 forecasts – initialized four times daily from 00:00 UTC on 2 September to 18:00 UTC on 21 September 2020 – are used to evaluate model performance. AirNow PM2.5 observations are used to verify the forecast. Forecast errors are assessed using bias, mean absolute error (MAE), and root mean square error (RMSE). Forecast performance is evaluated using box plots, which illustrate the distribution, spread, and central tendency of forecast errors. Time series of PM2.5 at various forecast hours are presented to examine the temporal evolution of forecast performance. Additionally, spatial distributions of PM2.5 including observations, forecasts, forecast errors, and forecast differences are analyzed to evaluate the spatial impact of data assimilation on PM2.5 predictions.

Figure 4 presents the bias, mean absolute error (MAE), and root mean square error (RMSE) for the 1–24 h forecast of domain-averaged PM2.5. Domain averages are computed over all states in the mainland United States. From the bias statistics (Fig. 4a), it is seen that the bias in the control run follows an upward trend initially, then reverses into a downward trend. The data assimilation runs show a similar trend, as data assimilation primarily corrects the model state and does not fully resolve inherent model bias. The control run underpredicted surface PM2.5 throughout the 24 h forecast period by about 1 µgm-3. This underprediction was improved in the data assimilation experiments. The two assimilated experiments, DA_AN and DA_ANPA, reduced the 1–24 h mean bias to -0.2 and -0.3 µgm-3, respectively. The PurpleAir PM2.5 assimilation experiment (DA_PA) also slightly improved the 1–24 h mean bias.

It is seen that DA_PA produces a near zero bias at 1 h forecast, however this is due to large positive and negative biases canceling out. Therefore, it should be used together with norms like Mean Absolute Error (MAE) and Root Mean Square Error (RMSE), which quantify the actual magnitude of forecast errors.

In terms of MAE and RMSE, all the three data assimilation experiments reduced the surface PM2.5 forecast error throughout the 24 h forecast period. The added value of assimilating PurpleAir PM2.5 data alongside AirNow observations is evident in the consistent MAE reduction (Fig. 4b). Its impact on RMSE (Fig. 4c) is also positive, though relatively small. Overall, all the data assimilation experiments show improved forecast skill compared to the control run.

Figure 5 shows box-and-whisker plots of PM2.5 forecast bias. Across all forecast hours, the interquartile range (IQR) – represented by the height of the boxes – is consistently smaller for the DA experiments compared to the control run. This indicates reduced forecast error spread between the 25th and 75th percentiles and suggests more consistent forecasts in the DA experiments. Although the median forecast bias in the control run is sometimes closer to zero, the DA_ANPA experiment performs comparably in terms of central tendency while showing clear improvements in reducing the mean forecast bias, as also reflected in Fig. 4a. Among the DA experiments, DA_AN and DA_ANPA show the most consistent improvement at 24 h lead times, with DA_ANPA slightly outperforming others during the early forecast hours (e.g., hour 1 to 12). This suggests that assimilating PurpleAir observations in addition to AirNow helps reduce bias and brings the forecasts closer to observed PM2.5 values in the short term.

Figure 6 shows time series of PM2.5 averaged over CONUS at forecast hours 1, 12, and 24, respectively. Consistent with the evaluations in Figs. 4 and 5, all DA experiments generally improve PM2.5 forecasts. Notably, all DA experiments help correct underpredictions during 2–9 and 14–17 September. In addition, the substantial overprediction during 9–13 September observed in the control run is partially mitigated by the DA experiments. Among the DA experiments, DA_AN and DA_ANPA show comparable performance and both outperform DA_PA.

While we have investigated the impact of DA on PM2.5 forecasts in terms of temporal evolution, it is also important to examine the spatial distribution of forecast fields, associated errors, and the impact of PM2.5 data assimilation.

Figure 7 presents the spatial distribution of temporally averaged PM2.5 forecasts at forecast hour 1, based on 80 forecasts initialized four times daily (00:00, 06:00, 12:00, and 18:00 UTC) from 2–21 September. The effects of wildfire events are clearly seen in Fig. 7a, c and e across California, Oregon, and Washington – where the fires occurred – as well as in downstream regions such as Montana, Wyoming, Utah, and Colorado impacted by smoke advection and transport.

1 h PM2.5 forecast errors in the control run are evident in Fig. 6a but are more clearly highlighted in Fig. 7b, which shows the difference between the control run and AirNow observations. Significant overpredictions appear along the California coast, as well as in parts of the Midwest and Northeast US, including Tennessee, Kentucky, West Virginia, and Virginia. Conversely, notable underpredictions are found over Colorado, New Mexico, much of Texas and Oklahoma, and several Gulf Coast states.

Both DA_AN (Fig. 7c and d) and DA_PA (Fig. 7e and f) show similar spatial correction patterns across California, Oregon, and Washington, particularly in reducing overpredictions along the California coast. They also produce comparable large-scale adjustments across the Northeast, Midwest, and Southern US, with their spatial patterns (Fig. 7d and f) largely opposite in sign to those in the CTR–AirNow difference (Fig. 7b). This indicates that both DA experiments effectively mitigate the control run's over- and underpredictions.

However, the magnitude of the corrections is generally smaller in DA_PA than in DA_AN. DA_PA shows its strongest impact over Nevada, northern Utah, Colorado, and southwestern New Mexico, where it helps alleviate regional underpredictions (Fig. 7e and f). These improvements are also observed in DA_ANPA, whose spatial pattern closely resembles DA_AN except over these few states (figures not shown).

The above analysis shows that data assimilation effectively corrects the 1 h forecast errors in the control run, particularly in the DA_AN and DA_ANPA experiments. It is therefore of interest to examine how data assimilation affects longer forecast lead times. Figure 8 presents the MAE from the CTR experiment and the percentage change in MAE (%) between the data assimilation (DA) experiments and the CTR experiment at the 24 h forecast lead time. Negative values in Fig. 8b–d indicate a reduction in MAE.

It is seen from Fig. 8a that the largest MAE values occur in California, Oregon, Washington, Idaho, Montana, Wyoming, Utah, Colorado, and Arizona. The PM2.5 MAE in these states is generally greater than 10 µgm-3, with maximum values exceeding 150 µgm-3 at certain stations in California, Oregon, and Washington. In contrast, regions less affected by wildfire smoke exhibit MAE values below 10 µgm-3.

All DA experiments (Fig. 8b–d) show overall reductions in MAE at most stations. MAE reduction varies by location, with substantial improvements observed in regions with large forecast errors, such as California, Oregon, Washington, and Idaho, where MAE is reduced by approximately 20 %. On average, assimilation of AirNow PM2.5 observations alone significantly improves the 24 h forecast skill, with MAE reductions of 6.6 % based on MAE percentage changes averaged over all stations. In contrast, assimilation of PurpleAir data alone reduces the MAE by only 1.7 %. Assimilating both AirNow and PurpleAir PM2.5 observations reduces the MAE by 7.6 %, indicating that PurpleAir data provides complementary value to AirNow in improving forecast skill.

In the DA_PA experiment, a few eastern coastal states from Georgia to Virginia exhibit large percentage increases in MAE. These increases are associated with relatively small absolute errors, as MAE values in the CTR experiment over these states are typically only 2–4 µgm-3 (Fig. 8a). Increased MAE is also present at several stations at the 1 h forecast lead time in the DA_PA experiment in these areas (figure not shown). In contrast, the MAE increase at the 24 h forecast lead time is much less pronounced in the experiment assimilating only AirNow PM2.5 observations. This behavior suggests potential quality issues in the assimilated PurpleAir data in those areas that require further investigation, although a contribution from model errors cannot be ruled out.

To further examine model performance across different regions, error statistics (bias, MAE, and RMSE) for the averaged 1–24 h PM2.5 forecasts from the control and data assimilation experiments over the 10 EPA regions are analyzed. Figure 9 presents the EPA Regions 1–10 and Averaged 1–24 h PM2.5 forecasts errors statistics in each EPA region.

In terms of bias (Fig. 9b), the DA_AN and DA_ANPA forecasts show improved performance relative to the control run. The DA_PA experiment also improves bias over most EPA regions, except for increased overprediction in EPA regions 8 and 9. For MAE and RMSE, the DA_AN and DA_ANPA experiments generally show improved performance across all EPA regions. The DA_PA experiment also exhibits improved, or at least comparable, MAE relative to the control run. However, DA_PA shows slightly increased RMSE over EPA regions 1–4, while reduced RMSE is found over other EPA regions.

Substantial reductions in MAE and RMSE are observed over EPA regions 8–10, where fire events occurred and/or where regions were most influenced by transported smoke. This is consistent with the large 1 h MAE reduction shown in Fig. 8. Notably, the DA_PA experiment, which assimilates PA observations alone, reduces both MAE and RMSE over EPA regions 8–10. Assimilating PurpleAir data together with AirNow data results in slightly smaller MAE and RMSE over these regions, suggesting that PurpleAir observations nonetheless provide complementary information to AirNow in these regions.

The latest version of NOAA's regional AQM system became operational on 14 May 2024. This system is based on the online coupling of the FV3 atmospheric model with the CMAQ model within the UFS framework. To improve initial conditions for AQM and enhance predictions of wildfire impacts on air quality, the capability to assimilate PM2.5 observations into AQMv7 was developed within JEDI and evaluated using its 3D-Var assimilation component in this study. The developed assimilation scheme can also be used to generate analysis (reanalysis) dataset for other applications, for example, providing data for training artificial intelligent models used in air quality prediction.

Data assimilation experiments were conducted for the September 2020 Western US wildfire episode, using 3-hourly cycling with observations from the AirNow and PurpleAir networks. In the data assimilation experiments, the location- and time-dependent background error standard deviations of an aerosol variable are specified as proportional to its background values, using a diagnosed scaling factor. In addition to updating the analyzed aerosol variables in each mode, the particle number concentration and surface area of each mode are also updated. The results show that assimilating AirNow PM2.5 observations improves 1–24 h forecast skill. Assimilating PurpleAir data alone yields modest improvements in MAE. Combining PurpleAir with AirNow observations provides additional benefit by slightly further reducing MAE relative to AirNow-only assimilation, indicating that PurpleAir observations nonetheless provide complementary information to AirNow. The AirNow data assimilation alone or with PurpleAir data generally show reduced MAE and RMSE across all EPA regions, whereas the largest reductions in MAE and RMSE are observed over regions affected by fire events and/or strongly influenced by transported smoke. The positive impact of PurpleAir data assimilation on smoke prediction during the September 2020 wildfires has also been demonstrated in an experimental Rapid Refresh Forecast System coupled with the Smoke and Dust Model (Wang et al., 2023), where it significantly reduced the model's 24 h underprediction of surface PM2.5. Given that PurpleAir data coverage has improved since September 2020, the results of this study further highlight its potential to complement AirNow observations by filling spatial gaps and improving PM2.5 analysis and forecast skill.

In this initial development and evaluation of aerosol data assimilation in JEDI for AQMv7, the control variables are defined as individual forecast aerosol variables. In previous work on aerosol data assimilation for an earlier version of AQM using the GSI system (Wang et al., 2021), one option for the control variables was to define them as the total aerosol mass in each of the three modes, resulting in just three control variables. A control variable transform was then applied to partition the analysis increments across these modes to individual aerosol species, based on the ratio of each species' mass to the total mass within the corresponding mode. The use of total aerosol mass in the three modes as control variables – thereby reducing the number of control variables from 70 to 3 – is planned for a future phase of development. The use of total masses as control variables also reduces the cost of the background error statistics calculation and iterative minimization (Kumar et al., 2019).

This study focused on surface-level PM2.5 and did not incorporate vertical profile constraints with satellite-based aerosol optical depth (AOD) retrievals, which could further enhance forecast skill. A key challenge is the need for a robust forward operator in the CRTM AOD module – specifically, the creation and validation of lookup tables (LUTs) for AOD calculations with AQM. As an intermediate solution, existing LUTs in CRTM, such as the GEOS-5 LUTs, have been tested by grouping and mapping AQM aerosol species to those used in GEOS-5 (Wang et al., 2025). AOD assimilation also depends on an accurate vertical distribution of aerosols in the background field so that the CRTM AOD operator can provide meaningful gradient information at the correct vertical levels to constrain the analysis update. However, AQM models have shown deficiencies for the September 2020 fire events in representing smoke concentrations at and above plume rise levels, largely due to how fire emissions are injected into the model. This will be improved in the next update of the operational AQM.