Direct and Indirect Effects of Aerosols During the 2023 Canadian Wildfires

Abstract

This modeling study investigates the impact of the 2023 Canadian wildfire aerosols (primarily black carbon and organic aerosol) on weather forecasts, concluding that incorporating real-time aerosol forcing improves model performance over using climatology. Experiments without real-time data severely underestimated aerosol optical depth (AOD), an error mitigated by including the forcing or using the coupled atmosphere–chemistry model. The aerosols exerted a strong direct radiative effect, reducing surface downward shortwave (SW) flux and generating corresponding surface cooling over the wildfire region. Furthermore, including aerosol–cloud interactions amplified this cooling and led to an increase in the overall cloud fraction and precipitation, illustrating complex indirect effects. While these physical improvements enhanced the representation of the atmosphere, the positive impact on overall medium-range forecasting performance (5–10 days) was modest, suggesting that the benefits of accurately representing wildfire feedback on the coupled Earth system are achieved through relatively slow processes, such as radiation feedback.

1. Introduction

The 2023 Canadian wildfire season was marked by 6551 fires that burned approximately 184,961 square kilometers (e.g., [1]). The resulting smoke traveled south into the United States, reaching as far as the state of Virginia, and even crossed the Atlantic Ocean to Europe. Wildfires cause direct damage to both properties and human life [2], and they also generate short-term disruptions of weather forecasts (e.g., [3,4]) and have long-term impacts on climate (e.g., [5,6]).

Aerosols, particularly black and organic carbon emitted by wildfires, influence the atmosphere in several ways. The primary direct effects arise through aerosol–radiation interactions (ARIs), including the extinction, scattering, and reflection of solar radiation (e.g., [7]). Indirectly, aerosols can reduce precipitation efficiency in clouds (e.g., [2,8,9,10]) by increasing the number concentrations of cloud condensation nuclei (CCN) and ice nuclei (IN), which in turn decreases the size and coalescence of cloud droplets and ice crystals through aerosol–cloud interactions (ACIs [11,12,13,14]). The semidirect effect [15] occurs when black carbon absorbs radiation and heats the surrounding atmosphere, a process that can suppress cloud formation and stabilize the lower atmosphere if the absorbing aerosol is located at low altitudes [16].

Wildfires also have long-term effects on climate. Kaufman and Nakajima (1993) [6] demonstrated that smoke particles can reduce cloud droplet size from 15 to 9 μm and decrease reflectance from 0.71 to 0.68 due to absorption of sunlight by biomass-burning aerosols. These insights were derived from Advanced Very-High-Resolution Radiometer (AVHRR) analysis during Amazon Basin biomass-burning season in the 1980s. Using the Community Earth System Model with prescribed daily fire aerosol emissions, Jiang et al. (2020) [17] estimated the total global fire aerosol radiative effect (RE) to be approximately 0.78 W m⁻²; it was primarily attributed to ACI. Similarly, Xu et al. (2021) [18], using the Energy Exascale Earth System Model (E3SM) with the Global Fire Emissions Database (GFED) and ground-based Aerosol Robotic Network (AERONET) aerosol optical depth (AOD) observations from 1997 to 2016, found that fire aerosols significantly increased global AOD by 7–14% and reduced net shortwave radiation on the surface by about −2.3 W m⁻². Fire-induced direct and indirect aerosol effects were found to lower the annual mean global land surface air temperature by approximately 0.17 K.

Despite their substantial impacts, little research has examined how wildfires influence weather forecasting—an area that remains challenging. Peace et al. (2015) [4] reported that, in a coupled fire–atmosphere model, the sea-breeze frontal line propagated faster, and vertical motion within the frontal zone was enhanced. Hulstrom and Stoffel (1990) [3] observed that the optical depth of smoke clouds from the Yellowstone wildfire was up to 6.8 times greater than that under clear-sky conditions. As a result of this increased optical depth, Robock (1991) [19] reported daytime cooling amounting to 1.5–7 °C under wildfire smoke conditions, with negligible nighttime effects. Potter and McEvoy (2021) [20] compared daily weather patterns during large and small fires and found that the greatest response differences occurred in wind speed—reflecting variations in fire growth response rather than in the meteorological conditions themselves. Conrick et al. (2021) [21] investigated the influence of wildfire smoke on cloud microphysics during the September 2020 Pacific Northwest wildfires and found that thermodynamic factors were the primary driver of enhanced cloud lifetimes, with microphysical effects playing a secondary role.

The objective of this study is to evaluate the influence of wildfires on short-range weather predictions. The remainder of this paper is organized as follows: Section 2 describes the experimental design, Section 3 presents the results, and Section 4 provides the summary and conclusions.

2. Experimental Design and Model Description

The NOAA experimental atmospheric Global Forecast System (GFS) version 17 prototype 8 (GFS.v17_p8) [22], which has a horizontal resolution of approximately 26 km and 128 vertical levels extending to the mesopause (C384L128 configuration), was used in this study. The Thompson microphysics scheme [23,24], a double-moment parameterization, was employed to predict both the mixing ratios and number concentrations of condensates and hydrometeors. The activation of ice nuclei (IN) and cloud condensation nuclei (CCN) was performed using the approach described by Cheng and Yang (2025) [25]. Further development is underway under the guidance of new research (e.g., [22]). The Rapid Radiative Transfer Model for General Circulation Models (GCMs) (RRTMG) [26,27] was used for radiative transfer calculations involving the atmosphere, aerosols, and clouds.

To assess the direct effects of aerosols from the 2023 Canadian summer wildfires on GFS forecasts, two main experiments were performed. In the control experiment (CTL), climatological mean aerosol fields from 2014 to 2023 were used, while in the sensitivity experiment (RTF), three-hourly Real-Time Forcing (RTF) data from MERRA2 (Modern-Era Retrospective analysis for Research and Applications, Version 2, [28,29]) were used. Both experiments were conducted with the C384L128 configuration of the GFS for 240 h free forecasts, initialized every three days from 00 UTC 1 June 2023 to 00 UTC 1 September 2023. The direct effects of aerosols from the wildfires were evaluated by comparing the CTL and the RTF.

Additionally, we conducted a third experiment (EXP_GOC) using instantaneous aerosol data predicted by the coupled GFS and GOCART (Goddard Chemistry, Aerosol, Radiation, and Transport) model [30,31]. Within this framework, 15 aerosol species—including sulfate, five size bins each for dust and sea salt, and hydrophilic/hydrophobic organic and black carbon—are transported as passive tracers [32]. While the GOCART scheme handles most chemical and physical processes, two key exceptions exist: aerosol convective scavenging is integrated into the GFS scale-aware mass flux scheme, and dust is modeled via the Fengsha scheme [33]. Biomass-burning emissions were derived from the GBBEPx (Global Biomass Burning Emission Product—Extended) multi-satellite inventory [33,34]. This experiment was designed to allow evaluation of the performance of GOCART against the MERRA-3 hourly aerosol forcing.

The first three experiments (CTL, RTF, and EXP_GOC) included only aerosol–radiation interactions (ARIs). To further examine aerosol–cloud interactions (ACIs), a fourth experiment (EXP_RTACI) was conducted, which was identical to RTF except that ACI is activated—allowing aerosols to act as CCN and IN and enabling full coupling among aerosols, radiation, and clouds.

A summary of all the experiments is provided in

Table 1. Experiments performed.

Aerosol	ARI	ACI
MERRA2 2014-2023 (EXP CTL)	X
MERRA2 three-hourly real-time forcing (EXP RTF)	X
aerosols forecasted from GOCART (EXP GOC)	X
MERRA2 three-hourly real-time forcing (EXP RTACI)	X	X

. The comparison between EXP_RTF and EXP_RTACI represents the effects of ACI on numerical weather prediction (NWP) forecasts.

3. Results

3.1. Effects on Aerosol Optical Depth

Wildfire emissions are characterized primarily by black and organic carbon aerosols as well as gaseous precursors such as volatile organic compounds and nitrogen oxides [35]. During the summer of 2023, the influence of wildfire emissions was largely confined to the Northern Hemisphere, as indicated by the summer mean aerosol optical depth (AOD) on forecast day 1 (Figure 1). All experiments underestimated AOD over the Northern Hemisphere compared with Moderate Resolution Imaging Spectroradiometer (MODIS) observations. It should be noted that MERRA-2 assimilates multiple observational datasets, including MODIS, Advanced Very-High-Resolution Radiometer (AVHRR) instruments, Multi-Angle Imaging SpectroRadiomete (MISR), AERONET, and numerous other additions and bias-corrected systems, which likely contributes to its higher accuracy.

As expected, EXP_CTL shows the strongest AOD underestimation (Figure 1a), since the climatological aerosol forcing does not include information about the 2023 wildfires. The region of underestimation extends from the northwest to the southeast, consistent with the spatial orientation of the major wildfire events. In addition, EXP_CTL and EXP_GOC also underestimate AOD near the west coast of Southern Africa, the Indian monsoon region, and Northeastern Asia, but they overestimate AOD over part of Northern Africa, Northern Asia, and the storm-track regions in the Southern Hemisphere. EXP_RTF and EXP_RTACI also exhibit similar biases, though their magnitudes are generally smaller.

The sources of aerosols in these regions differ substantially: near the west coast of Africa, aerosols are predominantly organic carbon; over northern Africa and Asia, they are primarily dust; and in the Southern Hemisphere storm-track regions, they mainly consist of sea salt. Although these regional biases are not directly related to the Canadian wildfires, they interact within the coupled Earth system and influence the mean biases and root-mean-square errors (RMSEs) in the analyses that follow.

The biases generally increase with forecast lead time (Figure 2 and Figure 3), particularly in EXP_CTL and EXP_GOC, which do not include real-time aerosol forcing. In addition to the natural increase in forecasting errors over time, positive feedback processes may also make a contribution. For instance, an underestimation of AOD can lead to enhanced surface warming, as more shortwave radiation reaches the surface. This surface heating can destabilize the lower atmosphere, promoting deep convection that transports aerosols vertically and out of the source region, further amplifying the forecasting bias.

Quantitatively, the AOD root-mean-square error (RMSE) and bias from the four experiments, evaluated against MODIS and (Visible Infrared Imaging Radiometer Suite) VIIRS observations and averaged between 65° N and 65° S, are summarized in

Table 2. Global AOD RMSE and bias calculated for the four experiments against MODIS observations, averaged over the latitude range from 65° S to 65° N.

	RMSE				Bias
	CTL	RTF	GOC	RTACI	CTL	RTF	GOC	RTACI
Day1	0.136	0.116	0.1277	0.118	−0.051	−0.044	−0.04	−0.042
Day5	0.138	0.119	0.144	0.121	−0.049	−0.043	−0.037	−0.038
Day10	0.139	0.122	0.159	0.126	−0.045	−0.036	−0.028	−0.03

and

Table 3. Global AOD RMSE and Bias calculated for the four experiments against VIIRS observations, averaged over the latitude range from 65° N to 65° S.

	RMSE				Bias
	CTL	RTF	GOC	RTACI	CTL	RTF	GOC	RTACI
Day1	0.122	0.105	0.114	0.105	−0.045	−0.036	−0.032	−0.035
Day5	0.123	0.108	0.13	0.109	−0.043	−0.033	−0.029	−0.03
Day10	0.124	0.111	0.147	0.114	−0.04	−0.029	−0.023	−0.025

, respectively. Overall, both the RMSE and bias from all the experiments show better agreement with VIIRS than with MODIS, with differences of approximately 10%. For example, the day-1 bias of EXP_CTL is −0.051 relative to MODIS and −0.045 relative to VIIRS, corresponding to a 13% difference (0.006/0.045 = 13%).

Among all the experiments, EXP_CTL exhibits the largest RMSE and bias, while EXP_RTF shows the smallest RMSE, indicating that incorporating real-time wildfire aerosol forcing improves global forecasting performance. Including aerosol–cloud interactions (ACIs) in EXP_RTACI generally increased both the RMSE and bias, likely due to the added complexity of feedback processes. Interestingly, EXP_GOC displays the smallest overall bias despite having a larger RMSE than EXP_RTF and EXP_RTACI, which may result from partial error cancellation within its interactive processes.

3.2. Effects on Radiation Fluxes

In the following sections, we focus on the results from day 5 of the forecasts. Day 5 represents a balance between forecasting reliability and sensitivity: it does not exhibit the large forecasting errors seen on day 10, yet the differences between the experiments remain more pronounced than on day 1. This choice is consistent with several previous studies that also emphasized day-5 results (e.g., [36,37]).

We examined summer 2023 mean top-of-atmosphere (TOA) upward shortwave (SW) fluxes, outgoing longwave radiation (OLR), and surface downward longwave fluxes. However, no distinct spatial patterns directly attributable to the wildfires were identified, likely due to the complex feedback processes involved. Moreover, the differences between the model simulations and Clouds and the Earth’s Radiant Energy System (CERES) and Energy Balanced and Filled (EBAF [38]) observations are generally larger than the differences between the experiments themselves.

The downward surface shortwave (SW) flux from EXP CTL with climatological forcing is characterized by stronger fluxes in the south and weaker ones in the north (Figure 4a). This pattern can mainly be attributed to the effects of aerosols and clouds, as discussed in the following sections. Signals associated with wildfires can also be identified in the surface downward SW flux (Figure 4c,d). The reduction in downward SW flux, caused by the reflection, scattering, and attenuation of radiation by aerosols and clouds, is evident over northwestern Canada, with a small region showing statistically significant changes (above the 95% confidence level based on the Student’s t-test) in the RTF, GOC, and RTACI experiments. The affected region extends primarily in a northwest–southeast orientation, consistent with the spatial pattern of the wildfire. Notably, EXP RTACI shows extensive areas of reduced downward SW flux over the Arctic, the North Pacific, and the eastern coast, where moisture is abundant.

The total aerosol optical depth (AOD; Figure 5a) can partially explain the spatial pattern of the surface downward shortwave (SW) flux shown in Figure 4a. For instance, the region with the lowest AOD between 180–100° W and 30–40° N corresponds well with the area of strongest downward SW flux. However, the reduced downward SW flux near the Arctic does not coincide with particularly high AOD values between 45° N and 70° N, suggesting that clouds may play a more dominant role in modulating SW radiation in this area.

The results of the three sensitivity experiments—RTF, GOC, and RTACI—indicate that wildfire-produced aerosols generated two AOD centers (Figure 5c,d): one over Northwestern Canada and another over the southeast. These regions correspond closely to areas of reduced surface downward SW flux (Figure 4c,d). EXP RTF and EXP RTACI show very similar AOD distributions, implying that aerosol–cloud interactions exert limited influence on AOD. In contrast, EXP GOC exhibits the largest AOD values, especially over Eastern Canada.

Black carbon (BC) and organic aerosols (OAs)—including primary and secondary organic carbon—are the dominant carbonaceous species produced by wildfires. In a typical year without major wildfire events, BC and OAs account for approximately half of the total aerosol optical depth (AOD) (Figure 6a). During the wildfire period, the BC and OA AOD increased by about 50%, with EXP RTACI showing the largest enhancement. Comparing Figure 5c,d and Figure 6c,d reveals that the regional mean AOD increases in EXP RTF and EXP RTACI are about 0.018 and 0.047, and they are primarily attributable to BC and OAs. In contrast, EXP GOC exhibits an AOD increase of approximately 0.041, some of which originates from aerosol species other than BC and OAs in Northeast Canada.

The spatial pattern of 2 m temperature (Figure 7a) closely resembles that of the surface downward shortwave (SW) flux (Figure 4a), with warmer temperatures in the south corresponding to stronger downward SW fluxes and cooler temperatures in the north associated with weaker fluxes. This indicates that the 2 m temperature largely reflects the influence of surface SW radiation. The differences in 2 m temperature between the three sensitivity experiments and CTL (Figure 7c,d) also correspond well to the respective SW flux differences. The negative temperature anomalies induced by high aerosol loading and elevated AOD from wildfires are distributed along a northwest–southeast axis, with the strongest cooling over Northwestern Canada, where the changes exceed the 95% confidence level based on the Student’s t-test. Among all the experiments, EXP RTACI produced the largest area of negative 2 m temperature anomalies, likely due to the combined effects of abundant cloud condensation nuclei (CCN), ice nuclei (IN), and water vapor in this region.

In addition to aerosols, clouds (Figure 8a and Figure 9a) also play an important role in shaping the spatial patterns of the surface downward shortwave (SW) flux. The lower cloud fraction in the south, which increases toward the north, leads to stronger downward SW fluxes in southern regions and weaker fluxes in the north (Figure 4a). Mid-level clouds are much less abundant than high and low clouds and follow a similar spatial pattern.

The direct effects of aerosols on high clouds are relatively small. A slight increase in high cloud cover near 60° N and 115° W (Figure 8b,c) is evident in EXP RTF and EXP GOC, likely resulting from cooling caused by enhanced SW reflection by aerosols. In contrast, the indirect aerosol effects lead to an overall 5% increase in the total cloud fraction across the domain, primarily due to the enhanced availability of ice nuclei (IN) and cloud condensation nuclei (CCN) from the increased aerosol loadings.

The response of low-level clouds differs from that of high-level clouds. A reduction in low cloud cover was observed over Central Canada (45–65° N, 110–80° W) across all three sensitivity experiments, despite an overall regional increase of about 2% in EXP RTACI. This reduction is likely associated with the warming effect of black carbon, which absorbs shortwave radiation and reduces cloud formation locally.

The regional distribution of cloud water (including both liquid and ice components) (Figure 10a) is generally consistent with the cloud fraction patterns shown in Figure 8a and Figure 9a, except for the large cloud water concentrations over the North Pacific and along the eastern coast of Canada. The two experiments that include only aerosol–radiation interactions (ARIs)—EXP RTF and EXP GOC—did not produce distinct spatial patterns of cloud water (Figure 10b,c). In contrast, EXP RTACI generated a substantially greater liquid water content over the ocean and the eastern coast of Canada, where both moisture and IN/CCN concentrations are high. The limited cloud water over Central Canada may be attributed to two factors: (1) competition for available moisture among aerosols to form IN and CCN, and (2) the heating effect of black carbon, which absorbs shortwave radiation and suppresses cloud formation.

To quantitatively assess the impact of wildfires on radiation fluxes, the root-mean-square errors (RMSEs) of the top-of-atmosphere (TOA) upward shortwave (SW) fluxes against CERES-EBAF observations are listed in

Table 4. RMSE of TOA upward shortwave flux (W m⁻²) against CERES-EBAF.

TOA_UP_SW	Global	North America
CTL	17.3856	23.127
RTF	17.3221	22.9259
GOC	17.1588	22.6171
RTACI	23.1889	22.0543

for both the global and North American domains. The two experiments that included wildfire aerosol effects, RTF and GOC, performed better than the CTL globally and regionally, as expected, with EXP GOC showing slightly better performance. For North America, EXP RTACI produced the smallest RMSE among all the experiments, although it performed the worst globally. It should be noted that the UFS model was not retuned after aerosol–cloud interactions (ACIs) were included. The feedback among clouds, radiation, and aerosols is highly complex, making consistent improvements in RMSEs and anomaly correlation (AC) scores—discussed later—particularly challenging. The performance of the three sensitivity experiments for surface downward SW fluxes (

Table 5. Same as Table 4, except it corresponds to the surface downward shortwave (W m⁻²).

Surface DN SW	Global	North America
CTL	21.9616	28.6287
RTF	21.5634	27.4383
GOC	21.9064	27.9882
RTACI	27.7747	27.9135

) is generally consistent with that for the TOA upward SW fluxes.

3.3. A Case Study

Although the summer-mean RMSE and bias analyses presented above are essential for understanding the direct and indirect effects of aerosols, a process-level case study can provide deeper insight. A 10-day free forecast initialized at 00 UTC on 21 August 2023 was analyzed to examine these effects in more detail. Figure 11a shows the time series of large-scale precipitation (solid lines) and convective precipitation (dotted lines) for the four experiments and the NCEP analysis (denoted as OBS), averaged over the region between 40°–65° N and 130–65° W.

During the first seven days, all four experiments produced similar values of total and convective precipitation. However, after August 27, the simulations begin to diverge, with EXP CTL and EXP RTACI yielding higher total and convective precipitation. EXP GOC consistently produced the least precipitation, while EXP RTF exhibited a sharp increase in total precipitation primarily due to enhanced large-scale precipitation, as the convective component showed no comparable increase.

Correspondingly, EXP RTACI produced the largest amounts of low-, middle-, and high-level cloud cover during the final three days of the forecast, whereas EXP GOC produced the least (Figure 11b–d). These differences suggest that the indirect effects of aerosols play a significant role in EXP RTACI. An increase in the number of wildfire aerosols increases the activation of ice nuclei (IN) and cloud condensation nuclei (CCN), leading to greater cloud formation [11]. In contrast, the reduced cloud cover and precipitation in EXP GOC may be attributed to the semi-direct effects of aerosols: black carbon absorbs solar radiation, warms the cloud layer, and raises the threshold for water condensation [15,39].

The aerosol optical depths (AODs) of BC and OAs produced by wildfires increased by approximately 57% in EXP RTF and EXP RTACI and by nearly 105% in EXP GOC (Figure 12). The wildfires were primarily located in Northwestern Canada, which also corresponds to the region with the highest climatological BC and OA concentrations (Figure 12a). The substantially larger BC and OA AOD in EXP GOC may be linked to the reduced cloud cover and precipitation observed in that experiment, likely a result of the semi-direct effect of aerosols (Figure 11c). The AOD patterns and magnitudes in EXP RTF and EXP RTACI are very similar, as both use the same aerosol forcing; the minor differences between them are mainly a product of variations in relative humidity.

The direct effects of aerosols—namely, extinction, scattering, and reflection—largely determine the distribution of surface downward shortwave (SW) fluxes (Figure 13). In EXP CTL (Figure 13a), stronger downward SW fluxes can be found in the south, where AOD is low, while weaker fluxes occur in the north, where AOD is high. The elevated AOD in Northwestern Canada caused by wildfire emissions led to reduced downward SW fluxes in this region across all three sensitivity experiments (Figure 13b–d). The indirect effects of aerosols in EXP RTACI appear to amplify the direct effects, producing the lowest downward SW fluxes in Northwestern Canada, extending eastward, and the highest fluxes in the south. This pattern suggests that increased aerosol loading enhances cloud formation and reflection, while reduced aerosol concentrations are associated with fewer clouds and stronger surface SW fluxes.

The spatial patterns of surface temperature (Figure 14) correspond closely to those of the surface downward shortwave (SW) fluxes. The region of higher surface temperatures between 40–50° N and 110–80° W (Figure 14a) aligns well with the areas of stronger downward SW flux in the control experiment (Figure 13a). In contrast, the three sensitivity experiments exhibit lower temperatures over Northwestern Canada and higher temperatures in the south (Figure 14c,d). Overall, the net effect of the wildfire aerosols is surface cooling, despite localized warming in the south, with regional mean temperature changes of −0.061 K for EXP RTF, −0.075 K for EXP GOC, and −0.17 K for EXP RTACI. The magnitude of cooling more than doubles when indirect effects of aerosols are included, a finding consistent with previous studies (e.g., [40,41]).

To analyze the impacts of aerosols, a core fire region (53–55° N and 122–125° W) characterized by peak AOD with clouds and precipitation was selected to investigate vertical profile structures (Figure 15). From these mean profiles, we calculated the Convective Available Potential Energy (CAPE) for the four experiments. The values were 149.8 J kg⁻¹, 155.3 J kg⁻¹, 145.1 J kg⁻¹, and 144.3 J kg⁻¹ for the CTL, RTF, GOC, and RTACI cases, respectively.

Overall, GOC exhibited the lowest CAPE, along with minimal cloud cover and precipitation. Interestingly, while RTACI produced enough precipitation and cloud cover to effectively deplete the available CAPE, GOC did not. Based on the mean profiles, three distinct scenarios were observed: (1) GOC was characterized by a warmer, drier environment with aerosol loading. In this scenario, aerosols competed for limited moisture, resulting in the lowest levels of cloud and precipitation formation. (2) RTF showed cooler low-level temperatures and higher moisture content. This led to increased cloud cover but reduced precipitation, consistent with the Twomey effect. (3) Regarding RTACI, with CCN activation in effect, this scenario produced the most significant amount of precipitation. This effectively reduced both CAPE and cloud cover, illustrating the second indirect effect.

The precipitation pattern (Figure 16) does not closely follow the surface temperature distribution (Figure 14a). However, it appears to correspond reasonably well with AOD, particularly over Eastern Canada (Figure 12a). Since the semi-direct effects of aerosols tend to warm the atmosphere, the direct cooling effects of aerosols may play a more significant role in modulating precipitation. EXP RTACI, which includes indirect effects of aerosols, produced more precipitation than the other two experiments, which include only direct effects of aerosols (Figure 16c,d). The precipitation anomalies predominantly extend in a west-to-east direction, suggesting that factors other than aerosols may be important in shaping this pattern.

3.4. Impact on Large-Scale NWP Forecasting Performance

Previous studies have reported mixed results regarding the impact of aerosols on the forecasting performance of numerical weather prediction (NWP) models. Incorporating more sophisticated aerosol–cloud–radiation interaction schemes does not always improve forecasting performance due to the complex feedback and inherent uncertainties (e.g., [37,41]). Using monthly varying aerosol climatology instead of a fixed climatology in the ECMWF IFS, Rodwell and Jung (2008) [40] found improved forecast performance at 5–10-day lead times in both tropical and extratropical regions. However, there has been little work examining the influence of wildfires on global and regional forecasting performance. In this section, we investigate the impact of wildfire aerosol forcing on overall medium-range weather-forecasting performance.

Figure 17 illustrates the differences in 500 hPa geopotential height anomaly correlation (AC) scores between the three sensitivity experiments and the CTL run for the summer of 2023. Scores are provided for the global domain, the Northern Hemisphere (20–80° N), and the Southern Hemisphere (20–80° S). Calculated against NCEP reanalysis, the AC score quantifies a model’s fidelity in representing synoptic-scale systems. While forecasting performance improved at 5–10-day lead times across all domains—consistent with Rodwell and Jung (2008) [40]—these differences do not exceed the 95% significance level (determined via Student’s t-test).

The absence of short-term (1–5 day) improvements suggests that aerosols’ effects on rapid processes, such as deep convection and frontal systems, are negligible in this configuration. Instead, the delayed improvement at 5–10 days indicates that wildfire feedback operates primarily through relatively slow processes, specifically through the impact of aerosol optical depth (AOD) on radiative forcing and temperature distribution. This aligns with our case study, which suggested that aerosol loadings are not the dominant direct drivers of large-scale or convective precipitation. Despite being modest, these enhancements represent a meaningful advancement in the physical representation and predictive performance of the GFS.

4. Discussion and Conclusions

This paper presents a detailed analysis of the impact of real-time wildfire aerosol forcing on global weather-forecasting models during the summer 2023 Canadian wildfires using four different experimental setups (EXP_CTL, EXP_GOC, EXP_RTF, and EXP_RTACI). The results are broken down into three main sections: effects on aerosol optical depth (AOD), effects on radiation fluxes and surface variables, and overall impact on forecasting performance.

Our analysis confirms that the experiments without real-time wildfire aerosol forcing (EXP_CTL) significantly underestimated the aerosol optical depth (AOD) in the Northern Hemisphere compared to observations, an error that increased over the forecast lead time due to a positive feedback loop involving surface warming. The experiment including real-time forcing but without aerosol–cloud interactions (EXP_RTF) showed the best overall AOD performance (the smallest root-mean-square error), indicating that incorporating this forcing improves global forecasting. Black carbon and organic aerosols were identified as the primary wildfire aerosols, and increases in the levels of these aerosols led to reduced surface downward shortwave (SW) flux—a cooling effect—evident over Northwestern Canada, the core wildfire region. This reduction was consistent with two AOD centers generated by the wildfire aerosols, one over Northwestern Canada and one over the southeast.

The reduction in surface SW flux corresponded directly to negative 2 m temperature anomalies (surface cooling), with the largest cooling observed in the experiment that included both aerosol–radiation and aerosol–cloud interactions (EXP_RTACI). Clouds played a key modulatory role, with aerosols’ indirect effects in EXP_RTACI leading to an overall 5% increase in total cloud fraction by enhancing the available cloud condensation and ice nuclei. A case study further highlighted the differing impacts on precipitation: EXP_RTACI produced the most precipitation and cloud cover (due to enhanced nuclei and indirect effects), while EXP_GOC consistently produced the least, likely due to the semi-direct effect of aerosols wherein black carbon absorbs solar radiation, warms the cloud layer, and suppresses formation.

This study examines the impact on numerical weather prediction (NWP) performance using 500 hPa height anomaly correlation (AC) scores. Consistent with previous research, the forecasting performance of the sensitive experiments modestly improved over the control experiment at 5–10-day lead times across the global domain and both hemispheres, though the changes were generally not statistically significant at the 95% level. This delayed improvement suggests that the feedback from wildfire with respect to large-scale weather operates through relatively slow processes like AOD and radiation changes. Ultimately, the results demonstrate that incorporating real-time wildfire aerosol forcing, particularly with aerosol–cloud interactions (EXP_RTACI), represents a step forward in the physical representation and overall performance of the global forecast system (GFS) model. Significant improvements exceeding the 95% confidence level are evident across multiple regions, particularly for AOD, as demonstrated in Figure 4, Figure 5, Figure 6, Figure 7, Figure 8 and Figure 9.

From a global modeling perspective, there is significant room for process-level improvement in both modeling and data analysis (DA), even if mid-range numerical AC scores do not show large immediate gains. Currently, GOCART remains a relatively simple aerosol scheme—used in systems like MERRA2 and Global Climate and Atmosphere Forecasting System—that lacks internal mixing, secondary aerosol formation, and other complex reactions typical of anthropogenic environments and large-scale wildfires. Incorporating these processes could further reduce observational bias. Additionally, refining real-time forcing in MERRA2—specifically by narrowing the gap between MODIS and VIIRS data—presents a viable path forward. By improving physical processes within the UFS, the reductions in AOD bias observed during regional wildfires may eventually translate into statistically significant (95%) improvements in global forecasting performance.