Forest Fire Aerosol – Weather Feedbacks over Western North 1 America Using a High-Resolution , Fully 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, fully 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 on17 line version of the Canadian Forest Fire Emissions Prediction 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 improvements in most metrics of both air-quality and meteorological model 21 forecast performance, through comparison of no-feedback and feedback simulations with surface, radiosonde, and 22 satellite observations. For the meteorological simulations, these improvements occurred at greater than the 90% 23 confidence level. Relative to the climatological cloud condensation nuclei and aerosol optical properties used in the 24 no-feedback simulations, the fully coupled model’s aerosol indirect and direct effects were shown to result in feedback 25 loops characterized by increased surface temperatures, decreased lower troposphere temperatures, and increased lower 26 troposphere cloud droplet and raindrop number densities. The aerosol direct and indirect effect reduced oceanic cloud 27 droplet number densities and increased oceanic rain drop number densities, relative to the no-feedback climatological 28 simulation. The aerosol direct and indirect effects were responsible for changes to the aerosol concentrations at greater 29 than the 90% confidence level throughout the model domain, and to NO2 and O3 concentrations within forest fire 30 plumes. 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 fully coupled model 33 changes the predicted fire plume dispersion and emissions, the latter through changing the meteorology driving fire 34 behaviour and growth. 35 https://doi.org/10.5194/acp-2020-938 Preprint. Discussion started: 7 October 2020 c © Author(s) 2020. CC BY 4.0 License.

The connection of the ADE and AIE within an air-quality and weather forecast model is referred to as "coupling", 159 with such a model being "fully-coupled". However, we note that in the field of very high resolution forest fire 160 behaviour modelling, coupling of biomass burning with the atmosphere has also been defined as the interaction of 161 dynamic meteorology with the heat released by the fire, where the initial meteorology determines fire spread on the 162 landscape. This in turn, modifies the temperature and wind fields, in turn affecting future fire spread (Clark et al., 163 1996, Linn et al., 2002. The coupling presented in the present paper refers to that between the aerosols released by 164 fires and other sources to meteorology through the ADE and AIE, with the resulting changes in meteorology in turn 165 influencing fire behaviour (fire intensity, fuel consumption, etc.), in turn influencing emissions height and distribution, 166 closing this feedback loop. 167 A key consideration in parameterizing the AIE (via aerosol-cloud interaction) is the manner in which the cloud 168 condensation process is represented in the meteorological component of the modelling system. In numerical weather 169 prediction (NWP) models, clouds and precipitation are represented by a combination of physical parameterizations 170 that are each targeted at a specific subset of moist processes. These include "implicit" (subgrid-scale) clouds generated 171 by the boundary layer and the convection parameterization schemes (e.g Sundqvist, 1988), and "explicit" clouds from 172 the grid-scale condensation scheme (Milbrandt and Yau, 2005(a,b), Milbrandt, 2015, Milbrandt and173 Morrison, 2016). Depending on the model grid these "moist physics" schemes vary in their relative importance. 174 However, regardless of the horizontal grid cell size, the grid-scale condensation scheme plays a crucial role in 175 atmospheric models, though to different degrees and using different methods, depending on the grid spacing and the 176 corresponding relative contributions of the implicit schemes. A grid-scale condensation scheme will in general consist 177 of the following three components: 1) a subgrid cloud fraction parameterization (CF, or cloud "macrophysics" 178 scheme); 2) a microphysics scheme; and 3) a precipitation scheme (Jouan et al., 2020). The cloud fraction (CF) is the 179 percentage of the grid element that is covered by cloud (and is saturated), even though the grid-scale relative humidity 180 may be less than 100%. The microphysics parameterization computes the bulk effects of a complex set of cloud 181 microphysical processes. If precipitating hydrometeors are advected by the model dynamics, the precipitation is said 182 https://doi.org/10.5194/acp-2020- 938 Preprint. Discussion started: 7 October 2020 c Author(s) 2020. CC BY 4.0 License. to be prognostic; if precipitation is assumed to fall instantly to the surface upon production, it is considered diagnostic. 183 The precipitation "scheme" is not a separate component per se, since it simply reflects the level of detail in the 184 microphysics parameterization, but it is a useful concept to facilitate the comparison of different grid-scale 185 condensation parameterizations. 186 With a wide range of grid cell sizes in current NWP models, there is a wide variety of types of condensation schemes 187 and degrees of complexity in their various components. For example, cloud-resolving models (with grid spacing on 188 the order of 1 km or less) have typically used detailed bulk microphysics schemes (BMSs), with prognostic 189 precipitation, and no diagnostic or prognostic CF component (i.e. the CF is either 0 or 1). Large-scale global models 190 use condensation parameterizations, sometimes referred to as "stratiform" cloud schemes, typically with much simpler 191 microphysics and diagnostic precipitation, but with more emphasis on the details of the CF. However, with continually 192 increasing computer resources and decreasing grid spacing (both in research and operational prediction systems), the 193 distinction between schemes designed for specific ranges of model resolutions is disappearing and condensation 194 schemes are being designed or modified to be more versatile and usable across a wider range of model resolutions 195 (e.g. Milbrandt and Morrison, 2016). 196 Aerosol-cloud interactions and feedback mechanisms are difficult to represent in grid-scale condensation schemes 197 with very simple microphysics components. For example, to benefit from the predicted number concentrations of 198 cloud condensation nuclei and ice nuclei, the microphysics needs to be double-moment (predicting both mass and 199 number) for at least cloud droplets and ice crystals, respectively. Until recently, detailed BMSs were only used at 200 cloud resolving scales, hence requiring these relatively high resolutions to be recommended in feedback modelling. 201 In recent years, multi-moment BMSs have been used in operational NWP for model grid spacings of 2-4 km (e.g. 202 Seity et al., 2010, Pinto et al., 2015, Milbrandt et al., 2016. Further, condensation schemes with detailed microphysics 203 are starting to use non-binary CF components (e.g. Chosson et al., 2014, Jouan et al., 2020), thereby allowing detailed 204 microphysics to be used at larger scales, and hence allowing the same indirect feedback parameterizations to be used 205 at multiple scales. Nevertheless, the expectation is that detailed parameterization will provide a more accurate 206 representation of cloud formation at the near cloud-resolving scales, without the complicating aspect of a diagnostic 207 CF, motivating the use of km-scale grid spacing for feedback studies. 208 The formation of secondary aerosols from complex chemical reactions are another key consideration in feedback 209 forecast implementation, given the impact of aerosol composition on aerosol optical and cloud formation properties, 210 as described above. 211 In the sections which follow, we describe our high resolution, fully coupled air-quality model with on-line forest fire 212 plume rise calculations, which was created as part of the FIREX-AQ air-quality forecast ensemble 213 (https://www.esrl.noaa.gov/csl/projects/firex-aq/), to address the following questions: 214 (1) Will a fully coupled model of this nature provide improved forecasts of both weather and air-quality, using 215 standard operational forecast evaluation tools, techniques and metrics of forecast confidence? That is, despite the 216 uncertainties in the literature as described above, are these processes sufficiently well described in our model that 217 their use results in a formal improvement in forecast accuracy? 218 https://doi.org/10.5194/acp-2020-938 Preprint. Discussion started: 7 October 2020 c Author(s) 2020. CC BY 4.0 License.
(2) Are the changes in forest fire plume rise associated with implementing this process directly within a fully coupled 219 model sufficient to result in significant perturbations to weather predictions and to chemistry? What are these 220 perturbations? 221 We employ our fully coupled model with 2.5-km grid cell size domain covering most of western North America, and 222 compare model results to surface meteorological and chemical observations, and to vertical column observations of 223 temperature and aerosol optical depth (AOD), in order to quantitatively evaluate the effect of feedback coupling of 224 the ADE and AIE on model performance. We then compare feedback and no-feedback simulations to show the 225 impacts of the ADE and AIE feedbacks on cloud and other meteorological predictions, and on key air quality variables 226 (particulate matter, nitrogen dioxide, and ozone). We begin our analysis with a description of our modelling platform. In GEM-MACH's climatological coupling configuration, prescribed, invariant climatological values for aerosol 235 optical properties and cloud condensation nuclei (CCN) are employed within the model's radiative transfer and cloud 236 microphysics modules. In the full coupling configuration, the ADE is simulated using GEM-MACH's predicted 237 aerosol loading and Mie scattering using a binary water-dry aerosol homogeneous mixture assumption, at the 4 238 wavelengths employed by GEM's radiative transfer algorithms, and at additional wavelengths for diagnostic purposes. 239 The full coupling also includes the AIE by simulating aerosol-cloud interaction via explicit droplet nucleation using Grabowski, 2008). The prognostic cloud droplet number and mass mixing ratios from the P3 microphysics are then 246 transferred back to the chemistry module for using in cloud processing of gases and aerosols (cloud scavenging and 247 chemistry) calculations, completing the AIE feedback process loop (Gong et al., 2015). 248 The chemistry modules of GEM-MACH also include process representation for gas-phase chemistry (ADOMII 249 mechanism, 42 gas species, Stockwell et al., 1989), cloud processing including aqueous chemistry, scavenging of 250 gases and aerosols, below-cloud removal and wet deposition (Gong et al., 2015), particle microphysics employing a 251 sectional size distribution and 8 chemical species (Gong et al., 2003), particle inorganic thermodynamics (Makar et 252 al., 2003), the formation of secondary organic aerosols using a modified yield approach (Stroud et al., 2018), process 253 https://doi.org/10.5194/acp-2020-938 Preprint. Discussion started: 7 October 2020 c Author(s) 2020. CC BY 4.0 License. to be updated during model runtime. When GEM-MACH is run in fully-coupled mode, the ADE and AIE 317 implementations allow model-generated aerosols to modify the predicted meteorology, in turn influencing predicted 318 fire emissions and plume rise, closing these feedback loops. The on-line implementation of CFFEPSv4.0 thus allows 319 us to investigate the effects of meteorology on subsequent forest fire plume development, the changes to modelled 320 aerosol compositions, and, ultimately, the feedbacks to weather. 321 The formation of particles from forest fires affects meteorology on the larger scale via the ADE and AIE, in turn 322 modifying the regional scale atmospheric features affecting fire growth, such as the temperature profiles below forest 323 fire plumes. However, we note that the local scale weather modifications due to the addition of forest fire heat to the 324 atmosphere are not yet incorporated into fire spread or GEM microphysics. Specifically, when the feedback version 325 https://doi.org/10.5194/acp-2020- 938 Preprint. Discussion started: 7 October 2020 c Author(s) 2020. CC BY 4.0 License. of GEM-MACH incorporating CFFEPSv4.0 is used in its fully coupled configuration, CFFEPSv4.0 calculates forest 326 fire plume rise using the meteorological predictions which include the ADE and AIE generated by forest fire aerosols 327 on atmospheric stability from the current fully-coupled model timestep. The resulting added aerosol mass due to the 328 fire in turn affects the meteorology through ADE and AIE, closing this feedback loop. To the best of our knowledge, 329 this is the first implementation of a dynamic forest fire plume injection height scheme incorporated into a fully coupled 330 high-resolution, air quality forecast modelling system. The impact of this feedback on both weather and air-quality 331 can be substantial, as we show in the following sections. 332 The locations of the daily forest hotspots detected during the study period, and the corresponding magnitude of the 333 daily PM2.5 emissions generated by CFFEPS for each hotspot are shown in

Feedback and No-Feedback Simulations 343
Two simulations were carried out for the period July 4 th through August 5 th 2019; a "feedback" (ADE and AIE 344 feedbacks enabled -fully coupled model) and a "no-feedback" simulation (ADE and AIE make use of GEM's 345 climatological aerosol radiative and CCN propertiesthe decoupled model). During this period, five large forest fires 346 took place in the northern portion of the modelling domain. The two parallel combined meteorology and air-quality 347 forecasts in the fully coupled model with/without ADE and AIE coupling were evaluated using the US EPA AIRNOW 348 data (https://www.airnow.gov) and Environment and Climate Change Canada's EMET and ARCAD operational 349 forecast evaluation systems, respectively. Following evaluation, the simulation mean values of hourly meteorological 350 and chemical tracer predictions were compared to analyze the impact of fully coupled ADE and AIE feedbacks on 351 both sets of fields. 352

Meteorology Evaluation 354
Surface meteorological conditions were evaluated at three-hour intervals from the start of both of the two sets of paired 355 24-hour forecasts using standard metrics of weather forecast performance including mean bias (MB), mean absolute 356 error (MAE), root mean square error (RMSE), correlation coefficient (R) and standard deviation (). In all 357 comparisons, a 90% percent confidence level assuming a normal distribution was used to identify statistically different 358 results between forecast simulations. Note that 90% confidence levels are commonly used in meteorological forecast evaluation, with values of 80% to 85% recommended (Pinson and Kariniotakis, 2004) and up to 90% used (Luig et 360 al., 2001) for variables such as wind speed, rather than the 95% or 99% confidence levels in other fields, in recognition 361 of the difficulties inherent in prognostic forecasts of the chaotic weather system. Here, the confidence range feedback scores, such that scores below the zero line indicate superior performance of the feedback forecast, while 379 those above the zero line indicate superior performance of the no-feedback forecast. Here, the feedback forecast was 380 statistically superior at forecast hours 3, 6, 15 and 18 at the 90% confidence level at these forecast hours, and both 381 simulations were at par (differences below the 90% confidence level) at hours 9, 12, 21 and 24. The feedback forecast 382 thus has superior performance, at greater than 90% confidence, over half of the forecast hours evaluated within the 383 domain, during the simulation period. 384 All of the metrics for which surface temperature forecast performance differed at the 90% confidence level are shown 385 in Figure 6. In addition to MB, the scores for MAE, and RMSE showed superior forecast performance for the feedback 386 relative to the no-feedback case at the 90% confidence level for hours 6 through 15, while the improvement for the 387 correlation coefficient was only reached the 90% confidence level at hours 6 and 12. 388 Model 10-m windspeed forecasts were also improved with the incorporation of feedbacks for hour ranges between 389 hour 3 and hour 12, depending on the metric, with the longest duration improvement for MB, MAE, and RMSE, and 390 the shortest duration for correlation coefficient and standard deviation (Figure 9). A marginal performance 391 degradation of the feedback forecast at close to 90% confidence at hours 15-18 can also be seen for root mean square 392 error, correlation coefficient, and standard deviation in this Figure.  and/or forecast events that were correctly predicted. Following standard practice at Environment and Climate Change 401 Canada, the HSS is used as a measure of total precipitation accumulated over a 6-hour interval, with no lower limit 402 on the amount of precipitation defining an "event", while FB and ETS define precipitation "events" as being those 403 with greater than 2mm / 6 hoursconsequently FB and ETS have a smaller number of data points for comparison 404 than HSS. 405 Figure 11 shows improvements to the fully coupled precipitation forecast at the 90% confidence level were seen for 406 the HSS 6-hour accumulated metric by hours 18 and 24, while the frequency bias index of 6-hour accumulated 407 precipitation showed degradation at hours 6 and 24, and the equitable threat score of 6-hour accumulated precipitation 408 showed degradation at hour 24. As is noted above, the latter two metrics employed a minimum 6-hour precipitation 409 threshold of 2 mm prior to comparisons (this is the reason for the reduced number of points available for comparison 410 in Figure 11(b,c) relative to Figure 11(a)). These findings suggest that the fully coupled model's trend towards 411 improved total precipitation over time (Figure 11(a)) is the result of improved performance for relatively low-level 412 precipitation events (< 2mm 6hr -1 ), offsetting a degradation of performance for higher level precipitation events. 413 Precipitation events have thus become more frequent, but "lighter" with the use of the feedback parameterizations. 414 The meteorological forecast performance metrics with statistically significant differences for surface pressure, 415 dewpoint temperature, and sea-level pressure are shown in Figures 7, 8, and 10 respectively. The model performance 416 differences in these three Figures show a similar pattern: a degradation in performance with the use of feedbacks at 417 hour 3, with the differences between the two forecasts either dropping below the 90% confidence level, or the feedback 418 forecast showing an improvement by hour 6, followed by several hours in which the feedback forecast has a superior 419 performance. The duration of this latter period varies between the metrics, from up to 18 hours for MAE for surface 420 pressure (Figure 7(b)) to 6 hours for the standard deviation of dew-point temperature (Figure 8(d)). 421 We believe that the initial loss of performance for the feedback forecast may represent a form of "model spin-up" that 422 may be unique to fully coupled models, but may be affected or improved with further adjustments to the forecast 423 cycling setup for the chemical species. As noted earlier (Figure 2), in order to prevent chaotic drift from observed 424 meteorology, we made use of a 30-hour 2.5-km resolution analysis-driven weather forecast to update our fully coupled 425 model's initial meteorology at hour zero of each 24 hour forecast. The cloud fields provided as initial conditions at 426 hour zero include observation analysis for the 6 hours prior to hour zero -these have reached a quasi-equilibrium in 427 the high-resolution weather forecast (Figures 2(b,e)) by the time they are used as initial and boundary conditions in improvements in the 10, and 50 hPa 12 th hour temperatures and 50 hPa 24 th hour temperatures, while 500 hPa 24 th 446 hour temperature performance degraded slightly. There are larger differences between the 1000 hPa forecasts, though 447 these also have the least number of contributing stations (i.e. only those located close to sea-level contribute to the 448 lowest level temperature biases). Other levels of the atmosphere showed no statistically significant change at the 90% 449 confidence level in temperature profile forecast performance with the use of feedbacks. 450

Chemistry Evaluation 451
Chemistry forecast quality is usually evaluated using standard statistical metrics against hourly observations collected 452 from surface measurement stations. Both simulations' performance for ozone (O3), nitrogen dioxide (NO2) and 453 particulate matter with diameters less than 2.5 m(PM2.5) were evaluated using hourly AIRNOW data (USA: AQS 454 network: https://www.epa.gov/aqs; Canada: NAPS network: http://maps-cartes.ec.gc.ca/rnspa-naps/data.aspx). The 455 summary performance metric scores for the two simulations grouped, according to contributing measurement network, 456 are shown in Table 2, with boldface values indicating the better score for the given simulation case. With respect to 457 this table, we note that: 458 (a) The feedback simulation generally outperforms the no-feedback simulation (more bold-face scores in the 459 "feedback" columns, with a few notable exceptions, discussed below). 460 (b) In some evaluation metrics, the feedback simulation showed substantial quantitative improvements over the no-461 feedback simulation (e.g. feedback PM2.5 MB is reduced by over a factor of 3 relative to its no-feedback 462 counterpart over Western Canada, the region of wildfire activity). 463 (c) For cases when the no-feedback simulation outperforms the feedback simulation, the relative magnitude of the 464 performance difference is smaller than the feedback simulation's improvements (e.g. Western USA PM2.   Figure 14(a)). It 502 is likely that at least some of the biases in PM2.5 and NO2 and consequently in the production of secondary ozone, 503 reflect the absence of these sources. Scatterplots of all paired AOD values in Figure 15 (a,b) and for the northern 504 https://doi.org/10.5194/acp-2020-938 Preprint. Discussion started: 7 October 2020 c Author(s) 2020. CC BY 4.0 License. portion of Alberta (Figure 15(c,d)) show that the overall negative bias is due to a large number of underestimated 505 The local AOD positive biases could also be the result of the mixing state assumptions of the Mie code used here for 517 generating aerosol optical properties. These assumptions may also account for negative AOD biases over much of 518 the remainder of the model domain. We have used a mass-weighted homogeneous mixture approach, with the 519 complex refractive index values for the 8 particle species being calculated for pure water-dry component homogeneous 520 mixtures at each of the 12 particle size bins, followed by mass weighting to generate values for each of the model 521 components. As noted earlier, this overall negative bias of AOD predictions is a common problem in air-quality 522 models and may be due to assumptions regarding the model mixing state (Curci et al., 2015). That comparison of 523 multiple mixing state assumptions on AOD with observations for European and North American modelling domains 524 (Curci et al., 2015), showed a typical factor of two model under-prediction of 440 nm North American AOD across 525 all mixing state assumptions, with European AOD negative biases ranging from unbiased to a factor of 2. For the 526 latter group, those models employing an assumption of external mixing, with hygroscopic growth factors for sulphate 527 and nitrate assumed to be similar to those of sulphuric acid, had the highest AODs and hence closest values compared 528 to observations at 440nm. However, in that investigation, the latter method also sometimes resulted in AOD over-529 predictions by a factor of 2 at 870 nm. These earlier findings along with overestimates at forest fire plumes with our 530 current homogeneous mixture approach at 550nm suggest that the hygroscopic growth may be overestimated for forest 531 fire particles, in turn overestimating forest fire AODs locally, while external mixing assumptions may be required to 532 improve model AOD performance elsewhere in the model domain. 533

Model Evaluation Summary 534
Overall, the incorporation of feedbacks in this study has resulted in improvements in weather and air-quality forecast 535 accuracy, albeit with some caveats. Weather forecast variables showed improvements at the 90% confidence level 536 for several fields, and vertical profiles showed improvements, particularly close to the surface, and with increasing 537 forecast lead time. Total precipitation scores also improved. A previously unexpected spin-up issue specific to fully 538 coupled models was noted: the impact of fully coupled particulate matter on cloud variables was sufficiently strong 539 that cloud field adjustment in the first 6 hours of the forecast was required prior to some weather forecast variable 540 https://doi.org/10.5194/acp-2020-938 Preprint. Discussion started: 7 October 2020 c Author(s) 2020. CC BY 4.0 License. improvements to be apparent (surface pressure, dewpoint temperature, sea-level pressure). While the current forecast 541 cycling duration was constrained by operational requirements, this suggests that forecast cycling should include both 542 air-quality and meteorological variables during fully-coupled forecast spin-up periods. That is, the model tracer 543 concentrations 6 hours prior to the current forecast start-up could also be used during the initial meteorological spin-544 up period, thus allowing chemistry and cloud formation to spin-up simultaneously. Scores for surface PM2.5, NO2, 545 and O3 also generally improved with the incorporation of feedbacks, with some metrics showing large improvements. 546 In comparison to satellite-based AOD values, the current model's AOD values were generally biased low, with 547 exceptions being in the regions of Alberta and Saskatchewan with active forest fires where AOD was biased high. 548 The latter comparison also showed that large fires off-domain in Alaska likely had a large impact on AODs in the 549 eastern and northern section of the model domain due to missing boundary condition contributions. These sources 550 were missing due to operational limitations in the model simulations shown here. 551 552

Effects of Feedbacks on Selected Simulation-Period Average Variables 553
In this section, we compare time averages of the entire study period for the two simulations, both at the surface and in 554 vertical cross-sections through the model domain, to illustrate some of the changes in both weather and air-quality 555 associated with the incorporation of feedbacks. We have found differences at greater than 90% confidence between 556 the predicted meteorological and chemical forecasts in the vicinity of the Alberta/Saskatchewan forest fires, as well 557 as in contrasting changes between land and sea. We note again here that the "no-feedback" simulation makes use of 558 time and spatially invariant aerosol CCN and optical properties, within the meteorological portion of the model. The 559 comparisons thus show the differences associated with the use of climatological constant aerosol properties, and the 560 fully coupled model-generated aerosols. 561 As in the meteorological evaluation, we have made use of 90% confidence levels in order to gauge the level of 562 significance of the differences between the feedback and no-feedback simulations in the following analysis. At each 563 model grid cell the values of the standard deviation about the mean for each respective simulation was calculated. 564 The difference between the means becomes significant at a given confidence level c if the regions defined by (1) 571 The differences in the mean grid cell values between the simulations for which the above quantity is greater than unity 572 thus differ at or greater than the 90% confidence level. Differences in the mean values, as well as the value of the 573 above ratio, are thus reported in the following section. 574

Effects of Feedbacks on Time-Averaged Meteorology 575
The feedbackno-feedback differences in the simulation-period average cloud droplet number density (number kg -1 576 of air) and mass density (g water kg -1 of air) along centred cross-sections spanning the length and width of the model 577 domain are shown in Figure 16 (the locations of the cross-sections are shown in Figure 1). The "Ocean", "Land", and 578 "Forest Fire" regions identified are with reference to the approximate locations of these features along these cross-579 sections. Figure 16 also shows the confidence ratio values as described aboveregions where the predicted mean 580 values differ at or above the 90% confidence level are shown in red, while those differences below the 90% confidence 581 interval are shown in blue. Feedbacks increase the cloud droplet number density over the northern part of the domain, 582 including the region impacted by the Alberta/Saskatchewan forest fires, from the surface up to about 600 mb (roughly 583 equivalent to hybrid level 0.600), and decrease further aloft and along the length of the model domain into the western 584 USA (Figure 16(a)). Cloud droplet numbers also decrease over the ocean, but increase eastwards over the land ( Figure  585 16(b)). The latter is unrelated to the forest fires; this is an indication that the modelled aerosol number concentration 586 over the ocean is much lower than the single climatological aerosol population assumed in the no-feedback run, 587 resulting in lower cloud droplet number concentrations. In both cases, the differences are significant at the 90% 588 confidence level from the surface up to hybrid level 0.87 and in isolated regions at hybrid level 0.550 along the south 589 to north cross-section (Figure 16(c)), and over the ocean in the west to east cross-section (Figure 16(d)). Higher-than-590 climatology aerosol loadings, a large portion of which are due to the forest fires, resulted increased cloud droplet 591 number densities in the lower troposphere, while decreasing them in the mid-to-upper troposphere. This impact of 592 feedbacks is in accord with the satellite observations of Saponaro et al. (2017), and was also seen in Takeishi et al. 593 (2020). In contrast, cloud droplet mass density(i.e. cloud liquid water content) largely decreases across the domain 594 along the north-south cross-section (Figure 16(e)), as well as over the ocean, with a varying pattern over the land in 595 the east-west cross-section (Figure 16(f)). The magnitudes and significance levels for the average change in cloud 596 droplet mass are lower than for cloud droplet number, with the most significant differences occurring over the ocean 597 ( Figure 16(g,h)). 598 Consistent with the cloud droplet number changes, rain droplet numbers and mass mixing ratios increase with the 599 feedback simulation over both the forest region impacted by the forest fires (Figure 17(a,e)) and over the ocean ( Figure  600 17(b,f)), with a varying impact over the land and more distant from the forest fire sources (Figure 17(f)). The changes 601 are significant at the 90% confidence level for rain droplet number in these regions (compare Figure 17(a) with 17(c); significant increase in hydrometeor mass. In the forest fire-impacted region, the ADE and AIE in the feedback 606 simulation significantly increase the number of cloud droplets near the surface and decrease the number of cloud 607 droplets in the middle to upper troposphere (Figure 16(a,c)). The rain drop number in the middle troposphere ( Figure  608 17(a,c)) also increases significantly between hybrid levels 0.90 to 0.70 (Figure 17(e,g)). Near-surface rain drop 609 number and rain drop mass differences throughout the cross sections (Figure 17(e,f)) fall below the 90% confidence 610 level (Figure 17(g,h). 611 Over the oceans, water droplet number and mass both decrease (Figure 16(b,f)), and raindrop number and mass 612 increase (Figure 17(b,f)); more atmospheric water is converted to rain drops as a result of the feedbacks, relative to 613 the climatology in the no-feedback simulation. However, these changes are more significant aloft than at the surface, 614 with the difference in both rain drop number and mass falling below the 90% confidence level near the surface. We 615 interpret these changes as a shift in over-ocean liquid hydrometeor numbers and to a lesser degree the water mass aloft 616 from cloud droplets to rain drops due to the AIE in the feedback setup relative to the climatology of the no-feedback 617 simulation. The changes occur at the 90% confidence level aloft, but the near-surface changes are smaller and are 618 usually below the 90% confidence level. 619 Differences in the average precipitation flux and the confidence ratio values are shown in Figure 18. Changes in 620 average precipitation (Figure 18(a)) appear random, though locally these differences are significant at the 90% 621 confidence level (Figure 18(b)). Both the magnitude of the differences and the frequency in their reaching the 90% 622 confidence level increase westwards. Given the local and episodic nature of rainfall events the high level of 623 significance in this case probably results from the presence or absence of individual rainfall events between the two 624 simulations affecting the local average and standard deviations. 625 Several systematic changes in the average values of the model's meteorological output fields were noted due to the 626 use of feedbacks relative to aerosol property climatologies (Figure 19), although all fall below the 90% confidence 627 level for the difference in the mean values between the two simulations ( Figure 20). Surface air temperature generally 628 increased (Figure 19(b)) though less so in the region most affected by forest fires, dewpoint temperature decreased 629 (Figure 19(c)) implying a decrease in relative humidity with feedbacks. Surface pressure increased, particularly in 630 the region downwind of the Alberta / Saskatchewan fires (Figure 19(d)). Planetary boundary layer height increased 631 over the land (Figure 19(e)), consistent with decreased atmospheric stability. The friction velocity also increased with 632 the use of feedbacks (Figure 19(f)); this is consistent with a decrease in stability and an increase in turbulent energy 633 The air temperature increases are limited to the lowest part of the atmosphere (Figure 21 (a,b)), usually below hybrid 634 level 0.914 (approximately 1km above the surface). The feedbacks decrease temperatures between hybrid levels 0.914 635 and 0.721. Feedbacks thus increase or do not affect temperatures near the surface, and decrease temperatures in the 636 lower free Troposphere. However, all of these features, despite their large geographic range, fall below the 90% 637 confidence level, reflecting the large variability in surface temperatures contained within each simulation. Longer 638 time simulations than carried out here are required in order to improve confidence in the temperature predictions 639 across all forecast hours. However, these results, particularly for surface temperature, may be contrasted with Figures

Effects of Feedbacks on Time-Averaged Chemistry 643
In the previous meteorological impacts section, changes in aerosol loading relative to the climatology, dominated by 644 forest fires, were shown to have a significant impact on cloud formation and atmospheric temperatures through ADE 645 and AIE. These might be expected to in turn influence and be influenced by particulate matter emitted by the forest 646 fires, with the plume rise of the forest fires dependent on the meteorological changes. Air temperatures increase 647 slightly in the model surface layer (Figure 19(b vertical gradient in temperature) implies a decrease in atmospheric stability associated with feedbacks, given that the 651 overall temperature gradient from the surface has become more negative. This is echoed by the response of the 652 concentration fields to the near-surface stability change, as can be seen through comparisons of the PM2.5, NO2 and 653 O3 surface concentrations changes ( Figure 22) and as vertical cross-sections (Figures 23, 24, 25), respectively. 654 For all three surface fields, changes above the 90% confidence level occur near the forest fires themselves (red regions, 655 near top of model domain, Figure 22(a,b,c)). The differences in particulate matter concentrations are also significant 656 at the 90% confidence level throughout the model domain (Figure 22(a)). Note that while the PM2.5 mean values are 657 significantly different at the 90% confidence level throughout the model domain, the magnitude of those differences 658 are sometimes small, particularly in the upper atmosphere, where the aerosol concentrations are relatively small. 659 However, the regions with the larger magnitude regional differences in PM2.5 concentrations also occur at greater 660 than the 90% confidence level (compare spatial locations of coloured regions in Figure 22(a,b) to red regions in Figure  661 22(c,d)). 662 Near-surface PM2.5 decreases in the regions downwind of the forest fires (Figure 22 blue region and more intense blue region near surface in 23(a)), suggesting less PM2.5 mass is present near the surface 664 due to the feedbacks. This could reflect a change in injection height of the plumes in addition to other transport 665 changes associated with the decrease in atmospheric stability. Lower troposphere decreases in PM2.5 over the ocean 666 and increases over the land at or greater than the 90% confidence level may also be seen (Figure 23(b,d)). 667 Feedbacks result in an increase in near-surface NO2 in several inland urban centers and less NO2 at surface level 668 downwind (Figure 22(b), though these differences are only significant at the 90% confidence level within the forest 669 fire plumes (Figure 24(a,c)). Ocean versus land NO2 differences remain below the 90% confidence level. 670 Feedbacks decreased surface O3 near the forest fires (Figure 22(c), Figure 25(a)), while decreasing increasing O3 aloft. 671 The forest fires are also the only area where the differences in between mean ozone forecasts have greater than the 672 90% confidence. 673 Overall, the most significant effects of the feedbacks were: (1) changes in PM2.5 concentrations throughout the model 674 domain, and (2) changes in NO2 and O3 within the forest fire plumes. 675 The feedback-induced changes in primary and secondary pollutants in the forest fire regions are consistent with the 676 decrease in atmospheric stability noted abovea greater proportion of the primary particulate matter and NO2 resulting 677 from near-surface forest fire emissions of NO are carried upwards with the addition of feedbacks. The decrease in 23(a))this implies that the changes associated with feedbacks occur in NOx-limited environments, i.e., with 680 relatively high VOC/NOx ratios. In such environments, decreases in NOx emissions may lead to decreases in the rate 681 of secondary O3 formation. 682 Our analysis thus suggests enhanced upward transport occurs in forest fire plumes due to feedbacks, and that this 683 transport is linked to feedback-induced: (1) decreases in local atmospheric stability (Figure 21(a)); (2) increases in 684 cloud droplet numbers near the surface (Figure 16(a)); and (3) increases in rain drop numbers aloft (Figure 17(a)). 685 This combination suggests the presence of an AIE feedback loopdecreased stability results in higher forest fire 686 plume rise, in turn lofting proportionally more particles to higher levels in the atmosphere where they may act as cloud 687 condensation nuclei, increasing cloud droplets aloft (Figure 16(a)). This in turn results in increased lower middle 688 troposphere cooling, through the 1 st AIE (increase in cloud droplet numbers aloft leading to increased cloud albedo 689 and cooling of the atmosphere below the cloud tops) while the corresponding decreases in particles and cloud 690 condensation nuclei at lower levels results in a smaller near-surface impact on the AIE and ADE, hence relatively 691 minor changes on near-surface temperatures (Figure 21(a)). This combination maintains a feedback-induced less 692 stable temperature gradient, relative to the no-feedback simulation employing aerosol property climatologies. 693 Similarly, over the oceans, the feedback-induced decrease in surface PM2.5 (Figure 23 which through the 1 st AIE, maintains the slightly less stable temperature profile. We acknowledge that these changes 698 in temperature fall below the 90% confidence level for the averages over all times, though note that differences in 699 mean bias relative to observations for the two simulations became significantly different at specific times of day in 700 the forecasts (Figure 6(a), hours 3, 6, 15 and 18, corresponding to 15, 18, 3 and 6 UT, or 9 AM, 12 noon, 9 PM, and 701 midnight MDT), implying that the temperature changes at these specific times reach a higher level of significance. 702

Summary, Differences in Forecast Simulation-Period Averages 703
Relative to the no-feedback simulation employing an aerosol climatology, the AIE feedback as simulated here is 704 associated with decreases in stability over both ocean and forest-fire influenced land areas. Over oceans, near-surface 705 particulate matter is removed as cloud condensation nuclei, resulting in increased cloud droplet numbers, maintaining 706 the temperature gradient through the 1 st aerosol indirect effect. In the vicinity of forest fires, decreases in stability 707 result in increased transport of PM2.5 aloft, increasing the availability of cloud condensation nuclei aloft, increasing 708 cloud droplet numbers aloft, hence also maintaining the increased temperature gradient through the 1 st aerosol indirect 709 effect. We note that the ADE may also play a weak role, particularly in the southern part of the domain, where lower 710 atmosphere temperature gradient increases are not accompanied by significant changes in cloud droplet numbers 711 ( Figure 16

Conclusions 716
The work carried out here suggests that the answers to our two research questions ("Can fully coupled models improve 717 both air-quality and meteorological forecasts?" and "Are the changes in forest fire forecasts associated with 718 implementing forest fire emissions within a fully coupled model sufficient to significantly perturb weather and 719 meteorology?") are both a qualified "yes". 720 The simulations analyzed here were conducted in preparation for an experimental forecast carried out as part of the 721 FIREX-AQ campaign, and hence were limited by operational time constraints to a sequence of nested 24-hour 722 forecasts. However, the high resolution domain size employed was sufficiently large to result in improvements in 723 weather forecast performance for both surface and profile variables at or above the 90% confidence level. 724 Improvements in model performance for PM2.5, NO2 and O3 were also found, across most statistical measures. The 725 differences between feedback and no-feedback simulations occurred at or above the 90% confidence level throughout 726 the model domain for PM2.5, though were limited at the 90% confidence level for NO2 and O3 to the immediate 727 vicinity of the forest fires. There, increased vertical transport associated with feedbacks lowered near-surface NO2 728 concentrations and increased them aloft, resulting in reduced surface O3 in the NOx-limited regions of the forest fire 729 plume. 730 The simulations suggest that the homogeneous mixture approach for aerosol optical properties results in a general 731 under-prediction of aerosol optical depths, in accord with Curci et al. (2015), and external mixture approaches are 732 recommended in further study. However, this general negative bias in simulated AOD is locally offset by positive 733 biases in the vicinity of forest fires. This suggests that forest fire plumes have significantly different optical properties, 734 and may be less hygroscopic than industrial aerosols of comparable size. Special / separate treatment of forest fire 735 CCN and optical properties are therefore also recommended in future work. 736 Fully coupling forest fire plume rise calculations with the weather parameters was shown to have a significant impact 737 on the height of primary pollutants reached by forest fires, the formation of near-surface ozone near the forest fires, 738 and on particulate matter throughout much of the three-dimensional model domain. These changes were largely driven 739 by the AIE, which maintains an increased temperature gradient (reduced stability) over the forest-fire-influenced and 740 oceanic portions of the region studied. Weak evidence for the influence of the ADE was shown in the southern part 741 of the domain, where increases in particulate matter were also accompanied by decreases in stability between the 742 surface and the lower-middle troposphere (the differences were at a lower than 90% confidence level for these 743 comparisons of temperatures averaged over all model times). 744 Relative to the no-feedback aerosol climatology for CCN and aerosol optical properties, the simulations carried out 745 here suggested that in the vicinity of forest fires feedbacks significantly increase cloud droplet near the surface, 746 increase cloud droplet number aloft, and significantly increase rain drop number densities aloft, relative to forecasts 747 driven by climatological aerosol properties. Over the oceans, feedbacks decreased cloud droplet number density and 748 increased rain drop number density aloft. Cloud droplet mass increased to a lesser degree (with smaller regions above 749 the 90% confidence level), as did rain drop mass (the mean differences for which for the most part remained below 750 the 90% confidence level). This provides some evidence for a shift in atmospheric water mass associated with 751 feedbacks from cloud water to rain over the oceans relative to the no-feedback climatology, though this shift occurred largely within the variability of the cloud fields within each simulation. Longer simulations may be needed to achieve 753 higher confidence in this finding.