Forest Fire Aerosol – Weather Feedbacks over Western North 1 America Using a High-Resolution , On-line Coupled , Air-Quality 2 Model 3

The influence of both anthropogenic and forest fire emissions, and their and subsequent chemical and 12 physical processing, on the accuracy of weather and air-quality forecasts, was studied using a high resolution, on-line 13 coupled air-quality model. Simulations were carried out for the period 4 July through 5 August 2019, at 2.5-km 14 horizontal grid cell size, over a 2250 x 3425 km domain covering western Canada and USA, prior to the use of the 15 forecast system as part of the FIREX-AQ ensemble forecast. Several large forest fires took place in the Canadian 16 portion of the domain during the study period. A feature of the implementation was the incorporation of a new on 17 line version of the Canadian Forest Fire Emissions Predict ion System (CFFEPSv4.0). This inclusion of 18 thermodynamic forest fire plume-rise calculations directly into the on-line air-quality model allowed us to simulate 19 the interactions between forest fire plume development and weather. 20 Incorporating feedbacks resulted in weather forecast performance that exceeded or matched the no-feedback forecast, 21 at greater than 90% confidence, at most times and heights in the atmosphere. The feedback forecast out -performed 22 the feedback forecast at 35 out of 48 statistical evaluation scores, for PM2.5, NO2 and O3. Relative to the 23 climatological cloud condensation nuclei and aerosol optical properties used in the no-feedback simulations, the on24 line coupled model’s aerosol indirect and direct effects were shown to result in feedback loops characterized by 25 decreased surface temperatures in regions affected by forest fire plumes, decreases in stability within the smoke plume, 26 increases in stability further aloft, and increased lower troposphere cloud droplet and raindrop number densities. The 27 aerosol direct and indirect effect reduced oceanic cloud droplet number densities and increased oceanic rain drop 28 number densities, relative to the no-feedback climatological simulation. The aerosol direct and indirect effects were 29 responsible for changes to the near-surface PM2.5 and NO2 concentrations at greater than the 90% confidence level 30 near the forest fires, with O3 changes remaining below the 90% confidence level. 31 The simulations show that incorporating aerosol direct and indirect effect feedbacks can significantly improve the 32 accuracy of weather and air quality forecasts, and that forest fire plume rise calculations within a on-line coupled 33 model changes the predicted fire plume dispersion and emissions, the latter through changing the meteorology driving 34 fire intensity and fuel consumption. 35

of forest fire hot-spots, climatological data on the extent of area burned by land use type, databases of fuel type linked 153 to emission factors, and an a priori weather forecast to provide the meteorological inputs required to predict forest 154 fire plume rise. The latter point is worthy of note in the context of the direct and indirect feedback studies noted above 155 both climate and weather simulations with prescribed forest fire emissions have consistently resulted in large 156 perturbations of weather patterns in the vicinity of the forest fires. However, their approaches for predicting forest 157 fire plume rise and fire intensity and fuel consumption in operational regional scale forecasts up until now have relied 158 on weather forecast information provided a priori and hence lacking those meteorological feedback effects. 159 The connection of the ADE and AIE within a regional air-quality and weather forecast model context is referred to as 160 "coupling", with such a model being described in that body of literature as "on-line coupled" (Galmarini et al., 2015) 161 or "aerosol-aware" (Grell and Freitas, 2014). However, several researchers have examined aerosol-radiative coupling 162 along with fire spread and growth (as opposed to fire intensity and fuel consumption). The latter work employs very 163 high-resolution forest fire spread and growth models , and due to their very high resolution, an additional level of 164 coupling, that of interaction of dynamic meteorology with the heat released by the fire, may be included. However, 165 the resolution requirements for these models (and their need for a relatively small computational time step) constrains 166 their application to a relatively small region. A requirement for these approaches is the use of a very high resolution 167 fire growth model imbedded within the air-quality model. At these resolutions, the simulated local-scale meteorology 168 determines fire spread on the landscape, which in turn modifies the temperature and wind fields, in turn affecting 169 future fire spread. The seminal work on this topic was carried out by Clark et al. (1996), and Linn et al. (2002). More 170 recent work includes the development of the WRF-FIRE model (Mandel et al., 2011;Coen et al., 2013), with full 171 chemistry added in the WRFSC model (Kochanski et al., 2016). Examples of the resolution required for these models 172 cloud condensation nuclei and ice nuclei, the microphysics needs to be double-moment (predicting both mass and 220 number) for at least cloud droplets and ice crystals, respectively. Until recently, detailed BMSs were only used at 221 cloud resolving scales, hence requiring these relatively high resolutions to be recommended in feedback modelling. 222 In recent years, multi-moment BMSs have been used in operational NWP for model grid spacings of 2-4 km (e.g. 223 Seity et al., 2010, Pinto et al., 2015, Milbrandt et al., 2016. Further, condensation schemes with detailed microphysics 224 are starting to use non-binary CF components (e.g. Chosson et al., 2014, Jouan et al., 2020, thereby allowing detailed 225 microphysics to be used at larger scales, and hence allowing the same indirect feedback parameterizations to be used 226 at multiple scales. Nevertheless, the expectation is that detailed parameterization will provide a more accurate 227 representation of cloud formation at the near cloud-resolving scales, without the complicating aspect of a diagnostic 228 CF, motivating the use of km-scale grid spacing for feedback studies. 229 The formation of secondary aerosols from complex chemical reactions are another key consideration in feedback 230 forecast implementation, given the impact of aerosol composition on aerosol optical and cloud formation properties, 231 as described above. 232 In the sections which follow, we describe our high resolution, on-line coupled air-quality model with on-line forest 233 fire plume rise calculations, which was created as part of the FIREX-AQ air-quality forecast ensemble 234 (https://www.esrl.noaa.gov/csl/projects/firex-aq/), to address the following questions: 235 (1) Will a on-line coupled model of this nature provide improved forecasts of both weather and air-quality, using 236 standard operational forecast evaluation tools, techniques and metrics of forecast confidence? That is, despite the 237 uncertainties in the literature as described above, are these processes sufficiently well described in our model that 238 their use results in a formal improvement in forecast accuracy? 239 (2) Are the changes in forest fire plume rise associated with implementing this process directly within a on-line 240 coupled model sufficient to result in significant perturbations to weather predictions and to chemistry? What are 241 these perturbations? 242 We employ our on-line coupled model with 2.5-km grid cell size domain covering most of western North America, 243 and compare model results to surface meteorological and chemical observations, and to vertical column observations 244 of temperature and aerosol optical depth (AOD), in order to quantitatively evaluate the effect of feedback coupling of 245 the ADE and AIE on model performance. We then compare feedback and no-feedback simulations to show the 246 impacts of the ADE and AIE feedbacks on cloud and other meteorological predictions, and on key air quality variables 247 (particulate matter, nitrogen dioxide, and ozone). We begin our analysis with a description of our modelling platform. 248 2 Model Description 249

GEM-MACH 250
The Global Environmental Multiscale -Modelling Air-quality and CHemistry (GEM-MACH) model in its on-line 251 coupled configuration has been described elsewhere (Makar et al., 2015a,b;Gong et al., 2015Gong et al., , 2016 Girard et al., 2014) with gas and particle process representation using the most of western Canada and the USA (Figure 1). The meteorological boundary conditions for the simulation were a 258 combination of 10-km resolution GEM forecasts updated hourly (themselves originating in data assimilation analyses 259 of real-time weather information; Figure 1(a)), and 2.5-km GEM simulations (Figure 1(c)) employing, in the northern 260 portion of this 2.5-km domain, the Canadian Land Data Assimilation System (Carrera et al., 2015), to better simulate 261 surface conditions. Both "feedback" and "no feedback" simulations were carried out on a 30-hour forecast cycle 262 ( Figure 2). Following the usual practice for weather forecasts, the analysis -driven meteorological forecasts at 10 km 263 resolution were updated operationally every 24 hours at 12 UT (Figure 2(a)). These 10 km resolution weather forecasts 264 were used to drive a 30-hour, 10-km resolution GEM-MACH forecast (Figure 1 Figure  271 2(d)). The two stages of meteorology-only simulations were carried out to prevent chaotic drift from the observed 272 meteorology, and to allow spin-up time for the cloud fields of that meteorology to reach equilibrium (6-hour 273 timeframe). Chemical initial concentrations for each consecutive forecast within the 2.5-km GEM-MACH model 274 domain were "rolled over" or "daisy-chained" between subsequent forecasts without chemical data assimilation. 275 Forecast performance scores presented here are for the inner 2.5-km domain from this set of linked 24 forecast 276 simulations, mimicking operational forecast conditions. 277

CFFEPS Version 4.0: On-line forest-fire plume rise calculations 278
In addition to the above algorithm improvements relative to GEM-MACH implementations, this model system setup 279 has incorporated the first on-line calculation of forest-fire plume -rise by energy balance driven using on-line 280 meteorology, in a new version of the Canadian Forest Fire Emissions Prediction System (CFFEPS). The algorithms 281 of CFFEPSv2.03 are described in detail and evaluated elsewhere (Chen et al., 2019), but will be outlined briefly here, 282 as well as subsequent modifications to this forest fire emissions processing module. 283 CFFEPS combines near-real-time satellite detection of forest fire hotspots with national statistics of burn areas by 284 Canadian province and by specific fuel type across North America. CFFEPS assumes persistence fire growth in the 285 subsequent 24-to 72-hour forecasts with hourly fuel consumed calculated (kg m -2 ), based on GEM forecast 286 meteorology and predicted fire intensity and fuel consumption in grid cells representing fire locations. The modelled 287 fire fuel consumption is then linked with combustion-phase specific emission factors (g kg -1 ) for fire specific emissions 288 and chemical speciation. Fire energy associated with the modelled combustion process is also estimated, and is used 289 in conjunction with a priori forecasts of meteorology within the column to determine plume rise. In its off-line/non-290 coupled configuration (Chen et al., 2019), CFFEPS carries out residual buoyancy calculations at five pres et pressure 291 levels (surface, 850, 700, 500, 250 mb). CFFEPS predicts plume injection heights, which are in turn used to 292 redistribute the mass emissions below the plume top to the model hybrid levels. This approach employed in 293 CFFEPSv2.03 provided a substantial improvement in forecast accuracy relative to the previous approach employing 294 modified Briggs (Briggs, 1965, Pavlovic et al., 2016 plume rise formulae in the offline GEM-MACH forecast system 295 (Chen et al., 2019). A recent evaluation of the plume heights predicted by CFFEPS was carried out utilizing MISR 296 and TROPOMI satellite retrieval data (Griffin et al, 2020 stability. These findings imply that plume rise calculations employing an a priori weather forecast lacking the impact 306 of fire plumes via the ADE and AIE may not accurately predict the weather conditions critical to subsequent forest 307 fire plume rise prediction. In order to study this possibility, and to allow forest fire plumes to influence weather and 308 hence subsequent fire spread/growth, several changes were made to CFFEPS implementation, resulting in to be updated during model runtime. When GEM-MACH is run in on-line coupled mode, the ADE and AIE 318 implementations allow model-generated aerosols to modify the predicted meteorology, in turn influencing predicted 319 fire emissions and plume rise, closing these feedback loops. The on-line implementation of CFFEPSv4.0 thus allows 320 us to investigate the effects of meteorology on subsequent forest fire plume development, the changes to modelled 321 aerosol compositions, and, ultimately, the feedbacks to weather. 322 The formation of particles from forest fires affects meteorology on the larger scale via the ADE and AIE, in turn 323 modifying the regional scale atmospheric features affecting fire growth, such as the temperature profiles below forest 324 fire plumes. However, we note that CFFEPSv4.0 employs forest fire heat to determine plume rise as a subgridscale 325 thermodynamic process parameterization rather than a very high resolution explicit fire growth parameterization; the 326 very local scale weather modifications due to the addition of forest fire heat to the atmosphere are not incorporated 327 weather forecast lapse rates. To the best of our knowledge, this is the first implementation of a dynamic forest fire 333 plume injection height scheme incorporated into a on-line coupled high-resolution, operational air quality forecast 334 modelling system. The impact of this feedback on both weather and air-quality can be substantial, as we show in the 335 following sections. 336 The locations of the daily forest hotspots detected during the study period, and the corresponding magnitude of the 337 daily PM2.5 emissions generated by CFFEPS for each hotspot are shown in forecast, while those above the zero line indicate superior performance of the no-feedback forecast. Here, the feedback 383 forecast was statistically superior at forecast hours 3, 6, 15, 18 and 24 at the 90% confidence level at these forecast 384 hours, and both simulations were at par (differences below the 90% confidence level) at hours 12 and 21, with the 385 no-feedback forecast being superior at 90% confidence at hour 9. The feedback forecast thus has superior 386 performance, at greater than 90% confidence, over half of the forecast hours evaluated within the domain, equivalent 387 performance at two hours (hours 12 and 21, both within 90% confidence limits), and inferior performance at one hour 388 (hour 9), during the simulation period. 389 All of the metrics for which surface temperature forecast performance differed at the 90% confidence level are shown 390 in Figure 6. In addition to MB, the scores for MAE, and RMSE showed superior forecast performance for the feedback 391 relative to the no-feedback case at the 90% confidence level for hours 15 and 18, while the improvement for the 392 correlation coefficient was only reached the 90% confidence level at hour 18. 393 The meteorological forecast performance metrics with statistically significant differen ces for surface pressure, 394 dewpoint temperature, and sea-level pressure are shown in Figures 7, 8, and 9 respectively. The model performance 395 differences in these three Figures show a similar pattern: a degradation in performance with the use of feedbacks at 396 hour 3, with the differences between the two forecasts either dropping below the 90% confidence level, or the feedback 397 forecast showing an improvement by hour 9, followed by several hours in which the feedback forecast has a superior 398 performance, usually at greater than 90% confidence. The duration of this latter period varies between the metrics, 399 from up to 18 hours for MAE for surface pressure (Figure 7(b)) to 3 hours for the correlation coefficient of dew-point 400 temperature (Figure 8(d)). 401 The initial loss of performance for the feedback forecast may represent a form of "model spin -up" that may be unique 402 to on-line coupled models, but may be affected or improved with further adjustments to the forecast cycling setup for 403 the chemical species. As noted earlier (Figure 2), in order to prevent chaotic drift from observed meteorology, we 404 made use of a 30-hour 2.5-km resolution analysis-driven weather forecast to update our on-line coupled model's initial 405 meteorology at hour zero of each 24 hour forecast. The cloud fields provided as initial conditions at hour zero include 406 observation analysis for the 6 hours prior to hour zero -these have reached a quasi-equilibrium in the high-resolution 407 weather forecast (Figures 2(b  Precipitation forecast performance from the two simulations varied depending on the metric chosen ( Figure 11). The 424 metrics in this case were based on the number of coincident precipitation "events" versus "non-events" as shown in 425 contingency Table 2. and/or forecast events that were correctly predicted. Following standard practice at Environment and Climate Change 431 Canada, the HSS is used as a measure of total precipitation accumulated over a 6-hour interval, with no lower limit 432 on the amount of precipitation defining an "event", while FB and ETS define precipitation "events" as being those 433 with greater than 2mm / 6 hoursconsequently FB and ETS have a smaller number of data points for comparison 434 than HSS. 435 Figure 11 shows improvements to the on-line coupled precipitation forecast at the 90% confidence level were seen for 436 the HSS 6-hour accumulated metric at hours 12 and 24, while the frequency bias index of 6-hour accumulated 437 precipitation showed degradation at hours 6 and improved performance at hour 12, and the equitable threat score of 438 6-hour accumulated precipitation showed significant differences at 90% confidence between the two simulations. As 439 is noted above, the latter two metrics employed a minimum 6-hour precipitation threshold of 2 mm prior to 440 comparisons (this is the reason for the reduced number of points available for comparison in Figure 11(b,c) relative 441 to Figure 11(a)). These findings suggest that the on-line coupled model's improvements for total precipitation ( Figure  442 11(a)) are the result of slightly improved performance for relatively light precipitation events (< 2mm 6hr -1 ).. 443 The amalgamated observations and model pairs of vertical temperature profile data from 39 radiosonde sites in western 444 North America are shown in Figures 12 and 13. Improvements in the forecasted temperature vertical profile with 445 increasing forecast time are evident at 250, 300, 400, 500, and 850 hPa in the 12 th hour forecast, with degradations at 446 200 and 700 hPa ( Figure 12). Improvements at 300, 925 and 1000 hPa may be seen in the 24 th hour ( Figure 13) 447 forecast; it is also worth noting the entire region at and below 300 hPa has improved temperature forecasts (mean 448 values to the left of the vertical line), albeit not always at >90% confidence. There are larger differences between the 449 1000 hPa forecasts, though these also have the least number of contributing stations (i.e. only those located close to 450 sea-level contribute to the lowest level temperature biases). Other levels of the atmosphere showed no statistically 451 significant change at the 90% confidence level in temperature profile forecast performance with the use of feedbacks. 452

Chemistry Evaluation 453
Improvements to air quality model performance metrics have been a focus for research since the 1980's starting with 454 dispersion model evaluation (Fox, 1981), and the identification of mean bias and normalized mean square error as 455 potentially useful metrics to complement the Pearson correlation coefficient (Hanna, 1988 Table 3, with boldface values indicating the better score 479 for the given simulation case. With respect to this table, we note that: 480 (a) The feedback simulation generally outperforms the no-feedback simulation (more bold-face scores in the 481 "feedback" rows, for 35 out of 48 metric comparisons). 482 (b) Feedback forecast score improvements occurred were more noticeable for PM2.5 (usually first to second digit), 483 followed by O3, with the NO2 scores often being the same for the first few digits. 484 (c) We note that the boundary conditions employed for our 2.5km simulations had a strong impact on model air-485 quality performance. As described above, these boundary conditions originated in a 10-km resolution simulation 486 making use of ECMWF global reanalysis values on its own lateral boundaries. The magnitudes of the statistics 487 of Table 3  American AOD across all mixing state assumptions, with European AOD negative biases ranging from unbiased to a 536 factor of 2. These earlier findings along with overestimates at forest fire plumes with our current homogeneous 537 mixture approach at 550nm suggest that the hygroscopic growth may be overestimated for forest fire particles, in turn 538 overestimating forest fire AODs locally, while external mixing assumptions may be required to improve model AOD 539 performance elsewhere in the model domain. 540

Model Evaluation Summary 541
Overall, the incorporation of feedbacks in this study has resulted in improvements in weather and air-quality forecast previously unexpected spin-up issue specific to on-line coupled models was noted: the impact of on-line coupled 546 particulate matter on cloud variables was sufficiently strong that cloud field adjustment in the first 6 hours of the 547 forecast was required prior to some weather forecast variable improvements to be apparent (surface pressure, dewpoint 548 temperature, sea-level pressure). While the current forecast cycling duration was constrained by operational 549 requirements, this suggests that forecast cycling should include both air-quality and meteorological variables during In this section, we compare time averages of the entire study period for the two simulations, both at the surface and in 562 vertical cross-sections through the model domain, to illustrate some of the changes in both weather and air-quality 563 associated with the incorporation of feedbacks. We have found differences at greater than 90% confidence between 564 the predicted meteorological and chemical forecasts in the vicinity of the Alberta/Saskatchewan forest fires, as well 565 as in contrasting changes between land and sea. We note again here that the "no-feedback" simulation makes use of 566 time and spatially invariant aerosol CCN and optical properties, within the meteorological portion of the model. The 567 comparisons thus show the differences associated with the use of climatological constant aerosol properties, and the 568 on-line coupled model-generated aerosols. 569 As in the meteorological evaluation, we have made use of 90% confidence levels in order to gauge the level of 570 significance of the differences between the feedback and no-feedback simulations in the following analysis. 571 The approach for representing model grid value 90% confidence levels is described in detail in SI Appendix A2. The 572 differences in the mean grid cell values between the simulations for which the above quantit y is greater than unity 573 differ at or greater than the 90% confidence level. Differences in the mean values, as well as the value of the above 574 ratio, are thus reported in the following section. 575

Effects of Feedbacks on Time-Averaged Meteorology 576
The feedbackno-feedback differences in the simulation-period average cloud droplet number density (number kg -1 577 of air) and mass density (g water kg -1 of air) along centred cross-sections spanning the length and width of the 2.5-km 578 resolution model domain are shown in Figure 15 (cross-section locations are shown in Figure 1). The "Ocean", these cross-sections. Figure 15 also shows the confidence ratio values as described aboveregions where the 581 predicted mean values differ at or above the 90% confidence level are shown in red, while those differences below 582 the 90% confidence interval are shown in blue. Feedbacks increase the cloud droplet number density over the northern 583 part of the domain, including the region impacted by the Alberta/Saskatchewan forest fires, from the surface up to 584 about 500 mb (roughly equivalent to hybrid level 0.500), and decrease at higher elevations further to the south and 585 along the length of the model domain into the western USA (Figure 15(a)). Cloud droplet numbers also decrease over 586 the ocean, but increase eastwards over the land (Figure 15(b)). The latter is unrelated to the forest fires; this is an 587 indication that the modelled aerosol number concentration over the ocean is much lower than the single climatological 588 aerosol population assumed in the no-feedback run, resulting in lower cloud droplet number concentrations. The 589 changes are significant at the 90% confidence level from the surface up to hybrid level 0.60 in the northern region 590 which is most impacted by forest fire smoke, and in isolated regions further aloft along the south to north cross-section 591 ( Figure 15(c)), and over the regions of the ocean in the west to east cross-section (Figure 15(d)). Higher-than-592 climatology aerosol loadings, a large portion of which are due to the forest fires, resulted increased cloud droplet 593 number densities in the lower troposphere, while decreasing them in the mid-to-upper troposphere (Figure 15(a)). 594 confidence level increase south-westwards. Given the local and episodic nature of rainfall events, the high level of 625 significance in this case probably results from the presence or absence of individual rainfall events between the two 626 simulations affecting the local average and standard deviations. 627 Several systematic changes in the average values of the model's meteorological output fields were noted due to the 628 use of feedbacks relative to aerosol property climatologies (Figure 18), although all fall below the 90% confidence 629 level for the difference in the mean values between the two simulations ( Figure 19). Specific humidity increased in 630 the region most affected by fires (Figure 18(a), surface air temperature decreased below the smoke plumes while 631 increasing further south (Figure 18(b)), while dewpoint temperature decreased (Figure 18(c)), implying a decrease in 632 relative humidity with feedbacks. Surface pressure increased over the land (mostly east of the Rockies), particularly 633 in the region downwind of the Alberta / Saskatchewan fires while decreasing over the ocean (Figure 18 (Figure 20 (a,b)). Feedbacks thus increase near-surface temperatures, relative to the no-feedback 640 meteorological model's simple aerosol climatology, in regions far from the fires, decreasing them near the fires, 641 decrease temperatures in the lower free Troposphere, and increase temperatures further aloft. All of these differences 642 between feedback and no-feedback simulations, despite their large geographic range, fall below the local 90% 643 confidence ratio. However, when the differences in air temperature resulting associated with feedback and no-644 feedback forecasts are compared to observations across the entire domain (as opposed to at gridpoint locations as in 645 Figures 18 and 19) the 90% confidence level is exceeded both at the surface at specific forecast times (Figure 6(a)), 646 and at multiple heights aloft at the 12 th and 24 th forecast hours (Figures 12, 13). 647

Effects of Feedbacks on Time-Averaged Chemistry 648
In the previous meteorological impacts section, changes in aerosol loading relative to the climatology, dominated by 649 forest fires, were shown to have a significant impact on cloud formation and atmospheric temperatures through ADE 650 and AIE. These might be expected to in turn influence and be influenced by particulate matter emitted by the forest 651 fires, with the plume rise of the forest fires dependent on the meteorological changes. Air temperatures increase levels 0.90 to 0.70 from the impact of the smoke plumes is similar to the findings of Saponaro et al. (2017). These 656 changes air temperatures implies a decrease in near-surface atmospheric stability associated with feedbacks, given 657 that the overall temperature gradient from the surface has become more negative (that is, the ambient lapse rate has 658 increased). Rising air parcels will follow an adiabatic lapse rate; these increases in the ambient lapse rate imply that 659 rising air parcels will have an increasing tendency to be warmer than their environment. Feedbacks have thus reduced 660 atmospheric stability within the forest fire smoke in the lowest part of the atmosphere; the atmosphere there has 661 become more unstable. Meanwhile, the feedbacks decrease the environmental lapse rate further aloft above the forest 662 fire smoke, between hybrid levels 0.848 and 0.339. Rising air parcels in this region following an adiabatic lapse rate 663 will thus have an increasing tendency to be colder than their environmentthe atmosphere above the smoke plumes 664 has become more stable. This is echoed by the response of the concentration fields to the near-surface stability change, 665 as can be seen through comparisons of the PM2.5, NO2 and O3 surface concentrations changes ( Figure 21) and as 666 vertical cross-sections (Figures 22, 23, 24), respectively. 667 Changes above the 90% confidence level for PM2.5 and NO2 occur near the forest fires themselves (red regions, near 668 top of model domain, Figure 21(a,b)), though remain below 90% confidence for O3 (Figure 21(c)). 669 Feedbacks result in near-surface PM2.5 decreases in the regions downwind of the forest fires (Figure 21(a), Figure  670 22(a), note the large blue region and more intense blue region near surface in Figure 22 The feedback-induced changes in primary and secondary pollutants in the forest fire regions are consistent with the 687 decrease in atmospheric stability noted abovea greater proportion of the primary particulate matter and NO2 resulting 688 from near-surface forest fire emissions of NO remain aloft with the addition of feedbacks. The decrease in surface decreases in the rate of secondary O3 formation. Alternatively, the reduction in near-surface O3 concentrations may 693 reflect a decrease in light levels reaching the surface due to cloud attenuation (aerosol indirect effect), with the 694 resulting lower photolysis rates resulting in a reduction in surface photochemical ozone production. 695 Our analysis thus suggests a net enhanced upward transport occurs in forest fire plumes due to feedbacks, and that this 696 transport is linked to feedback-induced: 697 (1) Increases in local near-surface atmospheric stability, reducing downward mixing of particulate plumes 698 once aloft (Figure 22(a)); 699 (2) Increases in cloud droplet numbers throughout the lower troposphere (Figure 15(a)); and 700 (3) Increases in rain drop numbers aloft (Figure 16(a)). 701 This combination suggests the presence of an AIE feedback loopincreased lower atmosphere stability 702 results a greater proportion of particulate matter remaining aloft, in turn resulting in more particles remaining at higher 703 levels in the atmosphere where they may act as cloud condensation nuclei, increasing cloud droplets aloft ( Figure  704 15(a)). This in turn results in increased lower middle troposphere cooling, through the 1 st AIE (increase in cloud 705 droplet numbers aloft leading to increased cloud albedo and cooling of the atmosphere below the cloud tops) while 706 the corresponding decreases in particles and cloud condensation nuclei at lower levels results in a smaller near-surface 707 impact on the AIE and ADE, hence relatively minor changes on near-surface temperatures (Figure 20(a)). This 708 combination maintains a feedback-induced near-surface unstable temperature gradient, relative to the no-feedback 709 simulation employing aerosol property climatologies. We acknowledge that these changes in temperature fall below 710 the 90% confidence level for the averages over all times, though note that differences in mean bias relative to 711 observations for the two simulations became significantly different at specific times of day in the forecasts ( Figure  712 6(a), hours 3, 6, 15 and 18, corresponding to 15, 18, 3 and 6 UT, or 9 AM, 12 noon, 9 PM, and midnight MDT), 713 implying that the temperature changes at these specific times reach a higher level of sig nificance. Similarly, Figures  714   12 and 13 show reductions in the near-surface temperature biases with the use of feedbacks. 715

Summary, Differences in Forecast Simulation-Period Averages 716
Relative to the no-feedback simulation employing an aerosol climatology, the AIE feedback as simulated here is 717 associated with increases in near-surface stability over both ocean and forest-fire influenced land areas. Over oceans, 718 near-surface particulate matter is removed as cloud condensation nuclei, resulting in increased cloud droplet numbers, 719 maintaining the temperature gradient through the 1 st aerosol indirect effect. In the vicinity of forest fires, increases in 720 near-surface stability result in more PM2.5 remaining aloft, increasing the availability of cloud condensation nuclei 721 aloft, increasing cloud droplet numbers aloft, hence also maintaining the less stable near-surface temperature gradient 722 through the 1 st aerosol indirect effect. We note that the ADE may also play a weak role, particularly in the southern 723 part of the domain, where lower atmosphere temperature gradient increases are not accompanied by significant though small magnitude increases in PM2.5 in the lower atmosphere (Figure 22(a), southern half of cross-section), 726 and temperature profile changes (Figure 20) below the 90% confidence level. 727 728

Conclusions 729
The work carried out here suggests that the answers to our two research questions ("Can on-line coupled models chemistry?") are both a qualified "yes". Within the high resolution domain size employed here, improvements or 733 matching weather forecast performance was seen for most times and heights in the atmosphere, at greater than 90% 734 confidence. Improvements in model performance for surface PM2.5, NO2 and O3 were also found, across most 735 statistical measures (35 out of 48 statistical evaluation scores showed improvements). Comparing average vertical 736 cross-sections, the chemical concentration changes associated with feedbacks were the most significant close to the 737 forest fires in the northern portion of the domain. There, increased net vertical transport associated with decreased 738 near-surface stability lowered near-surface PM2.5 and NO2 concentrations and increased them aloft, and resulted in 739 reduced surface O3. 740 Our simulations suggest that aerosol optical depth in the region, as well as the overall chemical performance of the 741 model, was strongly influenced by upwind boundary conditions. AODs were biased low d espite PM2.5 positive 742 biases, suggesting that the homogeneous mixture approach for aerosol optical properties results in a general under-743 prediction of aerosol optical depths, in accord with Curci et al. (2015), and that obtaining better data for forest fire 744 aerosol optical properties should be a priority for future study, as well as an examination of external mixture 745 approaches. Positive AOD biases in the region affected by fires suggests that forest fire plumes have significantly 746 different optical properties, and may be less hygroscopic, than industrial aerosols of comparable size. Special / 747 separate treatment of forest fire CCN and optical properties are therefore also recommended in future work. 748 On-line coupling forest fire plume rise calculations with the weather parameters was shown to have a significant 749 impact on the height of primary pollutants reached by forest fires, the formation of near-surface ozone near the forest 750 fires, and on particulate matter. These changes were largely driven by the AIE, which maintains an increased lapse 751 rate (decreased near-surface stability) over the forest-fire-influenced and oceanic portions of the region studied. Weak 752 evidence for the influence of the ADE was shown in the southern part of the domain, where increas es in particulate 753 matter were also accompanied by decreases in stability between the surface and the lower-middle troposphere (the 754 differences were at a lower than 90% confidence level for these comparisons of temperatures averaged over all model 755 times). 756 Relative to the no-feedback aerosol climatology for CCN and aerosol optical properties, the simulations carried out 757 here suggested that in the vicinity of forest fires feedbacks significantly increase cloud droplet number densities near 758 the surface and aloft, and significantly increase rain drop number densities aloft, relative to forecasts driven by 759 climatological aerosol properties. Over the oceans, feedbacks decreased cloud droplet number density and increased 760 rain drop number density aloft, relative to the simulation employing invariant CCN properties. Oceanic cloud droplet mean differences for which for the most part remained below the 90% confidence level). This provides some evidence 763 for a shift in atmospheric water mass associated with feedbacks from cloud water to rain over the oceans relative to 764 the no-feedback climatology, though this shift occurred largely within the variability of the cloud fields within each 765 simulation. Longer simulations may be needed to achieve higher confidence in this finding. 766 767