Introduction

Wildfires are described as unplanned, uncontrolled, and unpredictable combustion events occurring in areas influenced by diverse topographical and meteorological factors1,2,3,4. Given their adverse impacts on human life5,6 and the environment, wildfires are a subject of extensive research across multiple scientific fields7,8. Consequently, the development of strategies and models focused on studying large fire events and understanding their impacts is of critical importance. The record-breaking 2023 Canadian wildfires are an iconic recent example, as approximately 5% of the total forest area of Canada was lost following fires that took place between May 2023 and September 2023 (2023 Canada Report. Canadian Interagency Forest Fire Centre. https://ciffc.ca/sites/default/files/2024-03/03.07.24_CIFFC_2023CanadaReport%20%281%29.pdf) (The State of Canada’s Forests: Annual Report 2022. Canadian Minister of Natural Resources. https://natural-resources.canada.ca/forest-forestry/much-forest-does-canada-have), which is seven times higher than the average annual area burned of the previous 40 years9. This resulted in the largest Canadian wildfire emissions ever recorded (Copernicus: Emissions from Canadian wildfires the highest on record – smoke plume reaches Europe. Atmosphere Monitoring Service, Copernicus. https://atmosphere.copernicus.eu/copernicus-emissions-canadian-wildfires-highest-record-smoke-plume-reaches-europe) and were also quantitatively comparable to the annual fossil fuel emissions of large nations10. The content of the emissions included extreme amounts of aerosols along with reactive and non-reactive gases detrimental to humanity in terms of air quality and climate change10,11, as for example the amount of emitted carbon dioxide was 25% higher than emissions from all primary tropical tree cover loss in 202311. A more detailed spatial analysis of the Canadian wildfire event can be found in12, while for the purposes of this study we present Fig. 1 which contains plots of the difference between 2023 wildfire-season wildfire flux of black carbon and organic carbon over Canada and an average of the same measures over 2003–2020. This data has been extracted from the Copernicus Atmosphere Monitoring Service (CAMS) Global Fire Assimilation System (GFAS) (https://atmosphere.copernicus.eu/global-fire-emissions)13.

Fig. 1
figure 1

Difference (2023 wildfire season–2000 to 2003 average wildfire season) in wildfire flux of black carbon and organic carbon over Canada.

Full size image

Rising global mean temperatures combined with reduced regional humidity, both driven by the ongoing climate change, are identified as key factors contributing to this unprecedented series of wildfires14. The high temperatures and low humidity levels led to significantly elevated fire risk, with 2023 recognized as the warmest and driest year since at least 198015. Understanding the atmospheric impacts of such large wildfire events is essential, since a drastic increase in fire activity across many parts of the globe is expected in the coming decades16,17,18. This increase arises from warmer temperatures as seen in climate projections, which indicate that 2023 conditions are likely to become the norm in the 2050's, even under moderate climate mitigation scenarios19. Increasing our understanding of these impacts can be achieved by employing various modelling techniques of varying complexity20,21,22. The fact that the high latitudes warm faster than the rest of the globe23,24 makes the issue of understanding the impacts of high-latitude fires on the atmosphere even more pressing. Even in the present day, the high latitudes have been shown to be a world region experiencing particularly strong year-to-year variations in wildfire emissions25,26. Understanding the implications of these varying levels of emissions for large-scale weather evolution and its predictability is a topic that has not been approached substantially by the scientific community so far21.

This study employs the EC-Earth3 Earth system model27 to investigate the large-scale atmospheric impacts of the 2023 Canadian wildfire emissions, both in terms of atmospheric composition effects and in terms of the evolution of meteorological conditions across the entire NH. Specifically, Canadian wildfire aerosol spread in the NH atmosphere has been studied, modelled aerosol optical depth (AOD) values over Canada have been compared to observational data, and the impact of the emissions on meteorological conditions and atmospheric composition in the NH is investigated.

Results

Modelled aerosol analysis and comparison with observations

In this section, we examine the modelled aerosol differences between the FIRE and noFIRE simulations, and the results of both ensembles are compared with measured AOD over the Canadian region. Figure 2 shows the monthly global AOD differences between the FIRE and noFIRE simulations, averaged across May-September.

Fig. 2: Aerosol optical depth (AOD) differences between FIRE and noFIRE.
figure 2

a Global May-September average differences (FIRE – noFIRE) for the parameter: AOD. b Averaged NH regional longitudinal analysis of anomalies for the parameter: AOD. c Demonstration of chosen sub-regions for longitudinal analysis. Hatching in Fig. 2a represents the application of a typical Student’s t-test for statistical significance. Main 2023 wildfire sites over Canada (black spots in Fig. 2a) obtained via National Burned Area Composite dataset of the Canadian Wildland Fire Information System, available as open data under the Creative Commons Zero at https://cwfis.cfs.nrcan.gc.ca/datamart.

Full size image

The 5-monthly averaged global difference plots show the spreading of wildfire aerosol emissions throughout a large part of the NH (Fig. 2a, b). There are precedents for such extended propagations after intense wildfire events, for example, the 2017 Northern American event which formed vast aerosol plumes over Canada, a portion of which were transported by the polar jet stream and circled the NH within a few weeks28,29, or the case of the 2019 Australian wildfires28. However, the amount of emissions resulting from the 2023 wildfires was generally significantly higher than previously observed, with a carbon emissions magnitude of 647 TgC10. Despite the substantial long-range transport of pollution, the amounts of wildfire aerosols transported beyond CNA are small, due to the short lifetime of particulate pollution. The largest AOD differences between the simulations are found over WC&US and EC&US and in the neighbouring downwind area of CNA, owing to the location of the emission sites, while more remote differences are small.

For an enhanced representation of longitudinal diversity in regional values, in Fig. 2b each region is assigned a coloured triangle which changes orientation and colour with longitude (from left-pointing to right-pointing and from blue to red when moving from the westernmost region to the easternmost region, respectively). It is apparent from the month-by-month longitudinal analysis (Fig. 2b) that the remote impact of the wildfire emissions was not immediate, as the anomalies outside of the Canadian-adjacent regions are mainly seen starting from the month of June onwards, while the fires broke out in May, first in the region of Alberta. While the AOD anomalies are naturally highest in the WC&US and EC&US, the influence on AOD throughout the NH steadily increases until July, then steadily decreases (Fig. 2b). A longitudinal gradient is apparent even in the general vicinity of Canada, as the A&NEP region features very low AOD anomalies, and a higher AOD anomaly is present in the EC&US than in the WC&US region for most of the analyzed months (Fig. 2a, b).

A comparison between measurements and model output AOD was performed, over the area where the bulk of the aerosol pollution is located, i.e. Canada. First, 10 AERONET sites30 within the Canadian borders have been selected and their AOD averages for May-September 2023 have been compared with model AOD (Fig. 3). These stations have been selected based on having data for at least approximately half of the period of interest (which is the case of, for example, the “Churchill” station), and on the basis of covering Canada’s territory as evenly as possible. The black numbers represent correlation and differences between FIRE values and AERONET values, and the red numbers represent correlation and differences between noFIRE values and AERONET values. A blue highlight signifies that more than half of the month is covered by the data (marked in legend as “Good”), yellow suggests that half or more than half of the month is missing from the data (marked in legend as “~1/2 data missing”), and red indicates an almost complete lack of data wherein three quarters or more of the month is missing from the data and less than 5 consecutive days of data exist (marked in legend as “Almost no data”).

Fig. 3: Comparison between model AOD and AERONET AOD.
figure 3

a Correlation between model AOD and AERONET AOD. b Bias of model AOD compared to AERONET AOD. Legend indicates the respective “monthly data quality” for each station. Map of Canada in the background obtained via the OpenStreetMap project, available as open data under the Open Data Commons Open Database License at openstreetmap.org. Main wildfire sites in background (pink spots) obtained via National Burned Area Composite dataset of the Canadian Wildland Fire Information System, available as open data under the Creative Commons Zero at https://cwfis.cfs.nrcan.gc.ca/datamart.

Full size image

The AERONET data used in this study is Level 1.5, which means they are cloud-cleared and have had quality controls applied. Level 2 data was not used because of data scarcity. All stations have more than 50 days of data for the 5 months of interest. As metrics of performance, Fig. 3a, b shows the temporal correlation coefficient (R) and mean absolute bias values between model AOD and AERONET AOD, for both the FIRE and noFIRE simulations. Most sites (six) show favourable (positive) R values in the case of the FIRE simulation (Fig. 3a). Four locations, all in southern Canada, show less favourable correlations. The “Fort McMurray” location likely shows negative correlation because of improper or insufficient AERONET data, the “Univ. of Lethbridge” and “Wakesiu” locations show negative correlation because of a clear discrepancy between model data and measurements, and the “Pickle Lake” location has almost null correlation. Overall, the improvement in model performance at most sites is noteworthy, especially given the discrepancy between the atmospheric initialization and the emissions year (2015 vs 2023).

Re-examining the AERONET data, it is possible to see that all the stations with data that correlate poorly with the model FIRE data feature their highest emissions in May. This is unexpected, since emission data and model data generally indicate that May had low emissions. The reason for this discrepancy is that, in May, a series of wildfires took place which are southwards of the main emission focal point, as can be seen in the 5-monthly data of Fig. 3. This southward subset of wildfires can be clearly seen in Fig. 3 of ref. 12. The same paper also details how some of these initial May emissions were deflected even further south, guided by a Midwestern USA trough. This is not apparent in the model data, and this discrepancy explains the poor correlation for those stations.

The Canada-wide average Pearson correlation coefficient between FIRE AOD and AERONET AOD is 0.282, whereas for noFIRE the value is 0.11 (Fig. 3a). The average bias between FIRE AOD and AERONET AOD is 0.084, whereas for noFIRE the value is 0.276 (Fig. 3b). The FIRE/noFIRE correlation ratio is 2.563, while the noFIRE/FIRE difference ratio is 3.285. These measures indicate that the FIRE simulation yields much better results than the noFIRE simulation, which confirms that including realistic fire emissions in the model leads to a better simulation of atmospheric aerosol pollution. Similar conclusions are drawn from comparisons with MODIS data, and these comparisons (Fig. 2-S) along with the associated discussion, are added in the supplementary material.

Hemispheric aerosol impacts on weather

With the establishment of the modelled propagation of the Canadian wildfire emissions, the following segment will present the large-scale meteorological effects, both direct and indirect, of these emissions.

Strong temperature decreases in the NH are found in the FIRE simulation (Fig. 4a), when compared to noFIRE. As expected, the cooling is particularly pronounced over Canada (up to −5.44 °C in August), where the emissions are located, but are also strong and significant over remote areas of the NH such as Eastern Europe and even Siberia (up to −2.63 °C in July), even though pollution increases from the Canadian wildfire emissions in those regions are minimal. The average hemispheric temperature anomaly is equal to −0.91 °C, and this highlights the potential for large regional emissions from wildfires to perturb weather conditions for weeks across a whole hemisphere. The temperature anomaly is most intense in the EC&US region, even though AOD differences are higher in the WC&US regions. For more information on the local responses across Canada, the supplementary material contains further discussions. An increase of AOD and cloud cover (Fig. 4c) tracks with a decrease in net downward radiative flux (Fig. 4b), which then is linked to a decrease in surface atmospheric temperature over central and eastern Canada, and throughout the northern Atlantic. Smaller but still sizeable and statistically significant changes in cloud cover are also found throughout Europe and Siberia. Correlations between AOD changes and cloud cover changes with temperature changes can be found in Fig. 5. Additionally, global atmospheric temperature changes at four different pressure heights are shown in the supplement (Fig. 3-S), wherein it is seen that cooling anomalies persist yet become more diffuse at higher altitudes, and mild atmospheric heating is seen at the 200hPa level.

Fig. 4: Global May-September average differences (FIRE – noFIRE) for key meteorological variables.
figure 4

a surface atmospheric temperature, b net downward radiative flux at top of atmosphere, c total cloud cover, and d surface pressure and wind at 850 hPa.

Full size image
Fig. 5: Month-by-month spatial cross-correlation between the changes (FIRE – noFIRE) for key meteorological variables.
figure 5

a surface atmospheric temperature and AOD, b surface atmospheric temperature and net downward radiative flux, c surface atmospheric temperature and cloud cover percentage, d cloud cover percentage and AOD.

Full size image

The influences of wildfires on temperature are found to be sizeable across most of the NH, while the remote influences on AOD beyond North America and the North Atlantic are relatively small. It has been demonstrated in past studies that regional aerosol perturbations can impact temperatures across a whole hemisphere31,32,33,34. Here, the observed cooling spreads longitudinally because of advection from the regions of greatest temperature anomaly to neighbouring regions, through the atmospheric circulation. It is worth noting that the regions of Eastern Siberia and Europe exhibit cooling which is even stronger than other areas in the NH, remote to the emissions region. Inspection of Fig. 4a and Fig. 4d yields that cooling anomalies in Eastern Europe and Eastern Siberia are collocated with induced cyclonic patterns (surface negative pressure anomalies). These anomalies point to an atmospheric circulation pattern that is similar to the Arctic Oscillation (AO). A negative AO index features weakening of the polar jet stream, which means that colder Arctic air is allowed to enter mid-latitude areas35. The conditions of the AO under a negative AO index manifest themselves as increases in average surface pressure at polar latitudes and decreases in average surface pressure at mid-latitudes35, which is largely the same pattern of surface pressure anomaly seen in the difference between model experiment and model counterfactual in our case. Thus, it is likely that the modelled temperature response to the wildfire emissions projects onto the existing mode of variability of the AO, which led to the emissions cooling the Eastern Siberia and Eastern Europe regions more strongly. In past work of ours32 we have demonstrated such a projection of temperature response to mid-latitude anthropogenic aerosol emissions onto the AO, and here we show that this mechanism also holds for wildfire emissions.

To investigate the mechanisms behind these impacts further, spatial cross-correlations between the differences (FIRE – noFIRE) of surface atmospheric temperature and AOD, surface atmospheric temperature and net downward radiative flux, surface atmospheric temperature and cloud cover percentage along with cloud cover percentage and AOD are shown (Fig. 5). Most of the strongest correlations are found in the month of July, and while most of the average regional correlations are roughly lumped (except for the ones representing WC&US, EC&US and CNA, which are stronger due to being closer to the emission region), one can observe a clear increase in all of them towards a July peak (Fig. 5a–d). Individual aerosol species show similar tendencies to AOD, and these are shown in the supplement (Fig. 4-S). Correlations between the surface temperature change and the radiative flux change are less pronounced, except for the regions of WC&US, EC&US and CNA where stronger positive correlations are found, which is to be expected as the strong radiative effects are mainly local to the emissions and do not spread far; there is still a positive correlation for other regions, albeit a very small one (Fig. 5b). Correlations between surface atmospheric temperature anomalies and cloud cover anomalies are high only in the EC&US area, and lower yet still expectedly negative values are found in regions further east (Fig. 5c).

The correlation between the cloud cover anomaly and the AOD anomaly (Fig. 5d) is relevant given the fact that clouds have a known relationship to aerosols, though the quantification of aerosol–cloud interactions is commonly recognized as perhaps the most complex and uncertain issue in atmospheric climate science36. Koren et al.37 found a double regime, with low aerosol AODs associated with cloud cover increases and high aerosol AODs linked to cloud cover decreases. These effects are owed to various microphysical factors and semi-direct effects of aerosol particles, and they have been previously confirmed by observational evidence38. More recent work by Koren et al. 39 finds similar results over mid-latitudes as well. The regime for which there would be an increase in cloud cover is for values of AOD lower than 0.337. All NH average AOD differences in our study are smaller than 0.3 (Fig. 3b), and even though areas in and around Canada exhibit higher values, the results of the cross-correlation analysis (Fig. 5d) indicate that it is mainly cloud increases that result from Canadian wildfire emissions throughout the NH. These cloud cover enhancements are by far larger over the areas local to the fire plume (WC&US and EC&US), as the microphysical effects at play are local by definition.

Since the modelled longitudinal transport of the emissions was comprehensively discussed, considerations regarding the modelled vertical transport must also be raised. Regarding the modelled vertical transport of the wildfire aerosols, past research has shown that atmospheric chemistry models such as TM5 underestimate BBA plume altitudes when compared to CALIOP data40. It is currently theorized that the bias is due to wet deposition being too strong in these models41. TM5 can also present errors in modelled aerosol lifetime, however, with regards to the CALIOP data comparison TM5 presents a median performance relative to the rest of the models40. The fact that our configuration of EC-Earth3 lacks a plume rise model might also contribute to a lower vertical propagation. Upper troposphere distribution over the 40 to 70° N latitude band in the 2023 burning season shows strong evidence of the vertical transport of wildfire smoke to the upper troposphere due to high pyrocummulonimbus activity42. Without a plume rise model component, such activity cannot be modelled in our experiments. In any case, the model does produce a realistic, but decreased, vertical evolution – and there are indications of modelled aerosol intrusions even in the lower stratosphere according to Fig. 3-S.

It could be argued that altitude does not play a very important role in terms of the aerosol radiative effects. How important this role is can be difficult to quantify, particularly given the fact that different types of aerosols play different roles. A study43 has found that the aerosol radiative effect shows maximum sensitivity to the aerosol layer height in the thermal-infrared region, followed by the ultraviolet and visible regions – meaning that most of the radiation responsible for incident radiative effects plays a reduced role in the relation between aerosol radiative effects and the aerosol layer height. Furthermore, dust aerosols are the most sensitive to altitude, followed by polluted dust and pollution; thus, wildfire aerosols would fall in the category of least sensitive to altitude43. Also, under cloudless skies, changing the altitude of an artificial aerosol layer has a minor impact on the levels of shortwave radiation at the top and bottom of the atmosphere, even for high aerosol loads44.

Discussion

Modelled monthly changes in radiative effects, cloud cover, large-scale circulation, and temperature patterns throughout the NH are found as a result of the 2023 Canadian wildfire emissions. The mechanisms underlying these changes are analyzed and explained. Key findings include the modelled long-range transport of wildfire pollutants within the troposphere, resulting in significant impacts on cloud cover and temperature. Additionally, modelled patterns of cooling over the Canadian region are attributed to the dual effects of aerosols via direct and indirect (cloud) influences. Large-scale modelled circulation anomalies extending as far as Eastern Siberia also contribute to regional cooling. Our results highlight that wildfire-generated aerosols can produce both local and remote hemispheric temperature anomalies that can persist for months.

In conclusion, in this study, we have employed a global Earth system model to determine the large-scale fast atmospheric impacts of the emissions from the 2023 Canadian wildfire event. The model results have also been compared with measured data over the region of Canada, with favourable results. This unprecedented event is shown to have impacted the atmosphere throughout the NH for several months, the most significant of these impacts being surface cooling. Further such studies will enhance our understanding of the large-scale atmospheric impacts of emissions caused by extreme wildfires, which can lead to improved weather prediction in the future.

Methods

Model configuration

The model configuration employed for this study is the standard EC-Earth3-AerChem GCM configuration, which includes an atmospheric component (“Integrated Forecasting System cycle 36r4”, IFS 36r4), an ocean component (“Nucleus for European Modelling of the Ocean 3.6”, NEMO 3.6), and a chemistry component (“Tracer Model 5”, TM5)27,45. The model components are coupled, and the initial conditions for all model components were contained in previous model restarts. This configuration was also one of the eight configurations of the EC-Earth3 model, which participated in the Coupled Model Intercomparison Project Phase 6 (CMIP6)27. Two sets of model experiments for the period May-December 2023 were performed, each consisting of 10 ensemble members: a) one including the Canadian wildfire emissions of that year (hereafter referred to as the “FIRE” simulation) as input, and b) one excluding them (“noFIRE”). Individual ensemble members are differentiated by initiating the model from consecutive days (for example, the first member is started on the 1st of January, the second is started on the 2nd January, and so on). The supplement contains the model temperature difference (FIRE – noFIRE) average and variance across the global domain, Northern Hemisphere (NH), and Canada, as a function of ensemble members Fig. (1-S). It is found that as the number of members increases, the signal of average differences stabilizes, thus the number of members is sufficient to overcome model uncertainties.

Regarding potential uncertainties in GFAS emissions, a comprehensive study46 has found that, in comparison with other emission inventories, GFAS data agreement with AERONET and MISR data is moderate, and that one of the areas of lower agreement is indeed boreal North America. It must also be mentioned that GFAS relies on satellite observations, which can have limitations in terms of spatial resolution and near-real-time availability. With this in mind, it is possible that small GFAS inventory deficiencies could also be a factor in AOD underestimation, and that the actual impacts of the 2023 Canadian wildfire emissions might be even higher than what has been modelled.

Data Sources

As wildfire emissions input, we used CAMS GFAS data for 2023 (https://atmosphere.copernicus.eu/global-fire-emissions)13. The emission data downloaded from GFAS includes both aerosols and reactive gases. For the nullification of the emissions in the noFIRE ensemble, the region over which this process is performed is a rectangle with the geographical coordinates 45N-70N by 55W-140W. All non-wildfire emissions are obtained as per the CMIP6 “Shared Socioeconomic Pathways” (SSP) 370 scenario47. For the initialization of the simulation, we used atmospheric variables for year 2015, due to model data availability restrictions. The fact that we are using initial conditions from a different year to the emissions is expected to have little impact, given the fact that the objective of this study is to establish the potential large-scale effects of emissions of the magnitude of those that occurred in Canada in 2023 in any given year, rather than to specifically model the exact atmospheric conditions of 2023.

Technical and geographic considerations

We mainly focus on two types of analysis in this paper, when it comes to investigating the impacts of the emissions: plots of the averaged values of the chosen model variables in given regions of the NH, and plots of the cross-correlations between pairs of key variables in these given regions. Longitudinally, the region analyzed covers the entire NH with the middle represented by the Greenwich prime meridian, and latitudinally it is bounded by the 35th and 70th parallels north, since the 35th is commonly considered to be the poleward fringe of the northern subtropics48 and the 70th is situated closely above the Arctic Circle while also being the northward edge of the noFIRE emission nullification rectangle, as previously mentioned. Choosing regional separations within the given area is based on examining averaged emission impacts in the regions of highest interest. Thus, we define the following sub-regions: “Alaska and Northeastern Pacific” as 180W–140W abbreviated “A&NEP”, “Western Canada and US” as 140W–100W abbreviated “WC&US”, “Eastern Canada and US” as 100W–55W abbreviated “EC&US”, “Central Northern Atlantic” as 55W-15W abbreviated “CNA”, “Western Europe” as 15W-20E abbreviated “WE”, “Eastern Europe and Western Russia” as 20E-60E abbreviated “EE&WR”, “Western Siberia and Central Asia” as 60E-100E abbreviated “WS&CA”, “Eastern Siberia and North China” as 100E–140E abbreviated “ES&NC”, and “Russian Far East and Northwestern Pacific” as 140E–180E abbreviated “RFE&NWP”. These regions are shown in Fig. 2c.