This section describes the methodology used to derive the optical and microphysical properties of wildfire smoke from the BRAMS-simulated PM_2.5 fields. Although BRAMS–SFIRE internally represents aerosol–radiation interactions, the corresponding optical diagnostics are not available in the analysis output fields. For this reason, all radiative quantities examined in this study were obtained through offline post-processing. A set of microphysical assumptions, including particle size distribution, hygroscopic growth, refractive index, and ageing state, was adopted to translate PM_2.5 mass into wavelength-dependent extinction and absorption. These choices (summarised in Table 3 described in the Sect. 2.3.6) ensure physical consistency with observed biomass-burning aerosols and provide a coherent framework for the multi-stage optical analysis presented in Sect. 2.3.1–2.3.8.

To enable a consistent comparison between SOD and AOD, it is also necessary to clarify the spectral behaviour of smoke aerosols and the physical interpretation of column optical depth, both in the BRAMS framework and in the MERRA-2 reanalysis.

The evaluation of SOD against MERRA-2 products requires three methodological components: (i) a spatial harmonisation procedure enabling point-to-point comparison; (ii) a background correction that isolates the fire-induced optical contribution; and (iii) sensitivity tests that quantify how microphysical uncertainties affect the optical diagnostics.

This section describes these three elements, highlighting their interconnections and significance. First, the regridding strategy used to place the BRAMS grid and the MERRA-2 reanalysis fields onto a common spatial framework is described. Next, the methodology adopted to remove the contribution from regional background aerosols present in MERRA-2, which is due to its representation of large-scale sources, but absent in the BRAMS–SFIRE simulation that only contains fire-emitted smoke, is defined. Finally, the diagnostic scenarios applied in the optical post-processing to assess the robustness of the results and their stability with respect to microphysical uncertainty, are outlined.

Smoke aerosols modify the surface radiation budget through scattering and absorption of solar radiation. To quantify these effects, differences between simulations without and with fire emissions were evaluated using the accumulated surface downwelling shortwave (W m⁻²) and longwave (W m⁻²) radiative fluxes. This comparison provides a first-order estimate of how the presence of smoke alters the net radiative energy reaching the surface.

In addition to these broadband radiative diagnostics, the smoke-absorbed shortwave flux at 550 nm, Fabs550nm (W m⁻²), was estimated by relating the vertically integrated aerosol absorption to the heating potential of the plume. The column absorption optical depth at 550 nm is computed as: 34τabs550=∫z0zt∫Qabsmeffz,550,grdry,550πgrdry2ndryrdry,zdrdrydz where Qabs is the Mie absorption efficiency for the effective refractive index meff, g is the hygroscopic growth factor, rdry the dry particle radius, and ndry the dry-number distribution.

A simplified diagnostic estimate of smoke-absorbed radiant energy is then obtained as: 35Fabs550nm=Hsenτabs550 where Hsen is the surface sensible heat flux associated with fire. Although not a full radiative-transfer computation, this formulation provides a physically interpretable proxy linking absorption optical depth to the potential radiative heating within the plume.

Building on the radiative impacts examined in Sect. 2.5, this section investigates how smoke-induced heating interacts with the vertical thermodynamic structure of the atmosphere. The analysis focuses on the coupling between aerosol optical properties, namely extinction and absorption and atmospheric stability in smoke-affected layers.

Atmospheric stability was evaluated using vertical profiles of potential temperature (θ) and relative humidity (RH), extracted along a longitudinal cross-section intersecting the plume. These profiles were displayed in paired panels, showing the differences between fire and no-fire conditions relative to the corresponding no-fire reference fields, thereby isolating smoke-induced thermodynamic modifications. Thermal inversions were identified from layers where θ increases with height (dθ/dz>0), indicating inhibited vertical mixing. Although RH is not used as a formal inversion criterion, its vertical structure provides a supplementary indicator: sharp RH transitions or inflection points often coincide with inversion layers and therefore help delineate the vertical extent of stratified regions.

To relate thermodynamic modifications to aerosol–radiation interactions, vertical cross-sections of the smoke extinction coefficient (SEC) and smoke absorption coefficient (SAC) at 550 and 700 nm were analyzed together with CAPE and CIN diagnostics. CAPE and CIN were computed and displayed in paired panels for both the smoke and no-fire states. This layout provides a direct visual comparison between: (i) the underlying thermodynamic structure; (ii) the radiative perturbations induced by smoke (via SEC and SAC); and (iii) the corresponding stability response under convectively favorable (CAPE) or suppressed (CIN) conditions.

This combined framework, thermodynamic profiles, RH structure, and smoke optical properties, provides the basis for interpreting how absorption-driven heating may alter stratification and vertical mixing. The resulting cross-sections used in the analysis are presented in Figs. 17–21 of Sect. 3.5 (Results and Discussion), where their implications are examined in detail.

Significant differences in the simulated burned area for the same Sertã wildfire event, as depicted in Fig. 6, were observed between BRAMS-SFIRE and the WRF-SFIRE results previously reported by Menezes et al. (2024). BRAMS-SFIRE simulated a broader westward spread, while WRF-SFIRE produced a more confined footprint that extended further north. These discrepancies arise from differences in model configuration, fire propagation schemes, and spatial resolution.

First, although WRF-SFIRE resolves fire spread at a finer horizontal resolution (20 m) than BRAMS-SFIRE (200 m), BRAMS exhibits stronger sensitivity to terrain-induced effects due to its topographic assimilation and interaction with the atmospheric fields at coarser grid scales (2 km in diagnostics). This can amplify upslope spread, especially on west-facing slopes, where heating of unburned fuels by the flame front is intensified. Additionally, BRAMS employed a default rate of spread approximately ten times greater than that used in WRF, resulting in more aggressive lateral fire expansion, particularly in regions where wind-driven constraints are weaker.

Second, resolution differences in output diagnostics further influence the apparent burned area. WRF's higher resolution (20 m) output preserves more spatial detail, limiting the smoothing of the burned footprint. Conversely, the BRAMS burned area (200 m resolution) was accumulated and plotted in the BRAMS 2 km grid resolution, which introduces spatial averaging and may artificially enlarge the apparent extent of fire-affected regions.

Moreover, the delayed onset of burned area in BRAMS-SFIRE when using lower propagation rates is not solely due to ignition misalignment, but rather a result of the interaction between the fire spread rate and the coarser fire grid resolution (200 m). Unlike WRF-SFIRE, which operates at a much finer 20 m resolution for fire propagation, BRAMS requires the flame front to traverse larger grid cells to ignite neighbouring areas. This means that if the default rate of spread is reduced without compensating for grid size, the fire may fail to reach adjacent cells in a timely manner, leading to an underrepresentation of burned area in the early stages. Therefore, a balanced calibration is needed between the fire spread rate and grid resolution in BRAMS to ensure that the ignition front progresses consistently with observed timelines. While finer resolution in WRF (20 m) inherently provides more accurate spatial detail in fire development (Menezes et al., 2024), BRAMS relies on a more approximate representation, making it sensitive to how rate-of-spread parameters are tuned relative to its mesh scale.

In summary, while WRF-SFIRE offers finer-scale detail in fire perimeter shape, BRAMS-SFIRE exhibits stronger coupling between fire dynamics, topography, and atmospheric feedbacks. For prognostic simulations of fire spread and smoke transport in complex terrain, BRAMS-SFIRE may provide enhanced physical realism, provided its rate of spread is properly calibrated to avoid over propagation, particularly in orographically enhanced directions.

Following the assessment of burned area estimates, the results for the optical properties of the smoke plume are presented and discussed, starting with a comparison between simulated SOD and MERRA-2 AOD.

Understanding the spectral behaviour of smoke aerosol optical properties is essential for interpreting optical depth measurements. In this context, both the simulated SOD at 550 nm and the MERRA-2 AOD at the same wavelength represent the column-integrated attenuation of solar radiation due to the presence of aerosols in the atmosphere. However, there are conceptual differences that affect their comparison. SOD corresponds specifically to the fraction of AOD attributed to smoke particles generated by wildfires and biomass burning, while MERRA-2 AOD includes all aerosol types present in the atmosphere, such as mineral dust, sulfates, nitrates, sea salt, and soot.

When SOD ≤ AOD, only a fraction of the total aerosol burden can be attributed to smoke. When SOD ≈ AOD, the aerosol population is dominated by biomass-burning particles, as expected during intense fire activity with minimal influence from other sources. In contrast, situations where SOD ≪ AOD indicate that non-smoke aerosol types, such as mineral dust, industrial sulfate pollution, or marine aerosols (e.g., Saharan intrusions), contribute substantially to the observed optical depth.

These physical considerations provide the basis for interpreting the spatial correspondence between simulated SOD and MERRA-2 AOD, shown in Fig. 7.

Figure 7 presents the spatial distribution of SOD under the Baseline configuration, for both dry and wet conditions, compared with MERRA-2 AOD at 550 nm at 18:00 UTC on 15 October 2017, during the t1-fresh phase. Four fixed base classes between 0 and 4 (unitless) were defined for AOD, as this range is most representative of intense fire plumes.

At this time, both SOD and AOD show a pronounced increase relative to earlier afternoon conditions, reflecting the intensification and vertical deepening of the smoke plume over central Portugal. The dry SOD contours highlight the plume core, whereas the wet SOD field extends further downwind and attains higher values, indicating enhanced optical thickness driven by hygroscopic growth under humid conditions. The overall spatial pattern compares well with MERRA-2 AOD, with the model capturing the main plume structure and gradients, albeit with a somewhat narrower extension compared to the reanalysis. These results reinforce the strong aerosol load and plume development characteristic of the t1-fresh phase. The highest SOD values fall within the expected range for intense wildfire plumes, although some amplification is likely due to plume retention within the limited model domain, which restricts northward dispersion.

Beyond the spatial structure at 18:00 UTC, the temporal evolution of SOD and AOD provides further insight into model–reanalysis consistency. In the Baseline configuration, the comparison between simulated SOD and MERRA-2 AOD from 13:00 to 17:00 UTC reveals a persistent mismatch in both magnitude and spatial organisation, despite the progressive strengthening of the modelled plume. At 13:00 UTC, SOD remains confined and weak, whereas MERRA-2 depicts a broader, already well-developed smoke layer. This early underestimation reflects a timing offset, with the model delaying the onset of plume intensification relative to the reanalysis. From 14:00 to 16:00 UTC, both dry and wet SOD increase substantially, but the rate of increase is disproportionate relative to AOD. The simulated fields exhibit a much sharper spatial gradient and a narrower plume core, while MERRA-2 maintains a smoother, more diffused structure. This behaviour indicates that SOD responds primarily to the intense, locally concentrated core of the simulated plume, whereas MERRA-2 retains a broader regional aerosol contribution that is not present in BRAMS. Although an upwind background subtraction was applied to MERRA-2 AOD, it specifically removes only the inflow background and does not eliminate the wider regional AOD structure inherent to the reanalysis. As a result, MERRA-2 continues to exhibit a spatially smoother, regionally extensive optical depth, while SOD reflects only the fire-driven aerosol load. Consequently, the SOD–AOD relationship remains weakly constrained, with MERRA-2 showing elevated AOD even where SOD remains low or absent. By 17:00 UTC, SOD reaches its maximum values, particularly in the wet configuration, with magnitudes that fall within the physical range of dense wildfire smoke but exceed the AOD variability in the corresponding area. Part of this amplification is attributable to plume confinement within the limited model domain, which prevents northward dispersion and leads to artificial accumulation. In contrast, MERRA-2 retains a more spatially extensive plume, reflecting long-range transport that the regional simulation cannot capture.

The very high SOD values, arising from the intense fuel burning confined to a subset of SFIRE grid cells, fall within the physical range of extreme wildfire plumes and accentuate the contrast with the smoother and lower AOD structure in MERRA-2. This discrepancy further weakens the SOD–AOD correspondence at this stage of the plume evolution. Part of the discrepancies between SOD and AOD across the domain arise from temporal offsets, differences in plume confinement, and the distinct spatial scales represented by the model and the reanalysis, which collectively limit their correspondence during the plume's evolution.

Visualisation of the SOD–AOD scatter relationships (Fig. 8a–b) representing the Baseline optical configuration (test 1), provides insight into how the simulated plume evolves relative to the MERRA-2 reanalysis throughout the afternoon. Between 15:00 and 19:00 UTC, the statistical coupling between SOD (wet configuration) and MERRA-2 AOD improves, although substantial differences remain in their absolute magnitudes and spatial structure. At 15:00 UTC, the regression fit derived from the scatterplot (slope = 0.56, R2 = 0.03) shows weak correspondence, primarily because the AOD field retains a high regional background level (intercept ≈ 2.08) not represented by the localised smoke plume in BRAMS. By 19:00 UTC, the relationship strengthens somewhat (slope = 0.30, R2 = 0.36), but a clear shift becomes apparent: the simulated SOD reaches much higher values than the MERRA-2 AOD across most of the range. This indicates that the model generates a more optically intense core of the plume, whereas MERRA-2 maintains a smoother and less concentrated aerosol structure. Hygroscopic growth in the wet configuration further amplifies the SOD, enhancing the optical depth relative to the reanalysis. The Conservative (test 2) and Lower Envelope (test 3) optical configurations exhibit regression behaviour very similar to the Baseline test, with only minor quantitative differences. At 15:00 UTC, both tests 2 and 3 display weak statistical coupling between SOD (wet) and MERRA-2 AOD, with slopes near 0.55, intercepts around 2.1, and negligible explained variance (R2 ≈ 0.02–0.03). These metrics indicate that, as in the baseline case, MERRA-2 retains a substantial background AOD contribution that is not represented by the simulated plume, and that early-phase SOD variability accounts for only a small fraction of the observed optical depth. By 19:00 UTC, the coupling strengthens in both configurations: slopes decrease to roughly 0.29–0.30, intercepts fall to ∼ 0.65 (as in test 1), and R2 increases to 0.35–0.36. This convergence across the three tests demonstrates that modifying κ, rg,dry, or narrowing σg leads to only modest changes in the regression behaviour. The similarity of the results confirms that the optical response is relatively insensitive to plausible microphysical perturbations within the tested parameter space.

Table 4 provides a time-resolved evaluation of the statistical relationship between SOD and AOD during the Sertã wildfire period, highlighting the contrasting behaviour of the fresh and mixed phases. During the fresh phase (13:00–18:00 UTC), both dry and wet SOD differ markedly from the MERRA-2 AOD. Across 13:00–17:00 UTC, AOD remains relatively stable (Mean AOD ranges from ∼0.9 to ∼1.7, with a peak of ∼2.24 at 15:00 UTC), whereas SOD increases sharply, especially in the wet configuration, rising from values near zero at 13:00 UTC to about 7 (Mean SOD ≈ 6–7) by 17:00 UTC. This produces a transition in the bias defined as SOD - AOD from strongly negative (-1.9 to -0.9) to strongly positive (+5 to +6). Correspondingly, MAE and RMSE increase substantially, and correlations remain extremely low (r≈0.02–0.06, R2 ≈ 0.00), indicating that SOD explains virtually none of the spatial variability present in the AOD field. Physically, this reflects limited spatial covariance between the developing, highly localised simulated plume and the smoother regional aerosol structure represented by MERRA-2.

From 19:00 UTC onward, the plume enters the mixed phase (19:00–23:00 UTC), during which SOD increases from values below 10 in the fresh phase to Mean SOD values between 35 and 65, while AOD remains nearly constant around 0.8–0.9. This leads to very large positive biases and correspondingly large error metrics. Linear correlations improve modestly (up to r≈0.58), and R2 increases to at most 0.34, indicating a partial but still limited recovery of spatial coherence between SOD and AOD. These diagnostics are consistent with progressive plume confinement and accumulation within the regional domain, yielding SOD magnitudes far exceeding the column aerosol load represented in MERRA-2.

Taken together, the hourly statistics confirm that part of the discrepancy between SOD and AOD arises from temporal offsets, plume confinement, and the fundamentally different spatial scales resolved by the model and the reanalysis. These structural differences limit the correspondence between the two datasets throughout the plume evolution.

Supplement Tables S1–S3 extend the SOD–AOD evaluation by quantifying how the Baseline, Conservative, and Lower Envelope microphysical configurations affect the model–reanalysis agreement during the mixed phase. Across all configurations, the results are highly consistent: mean SOD is one to two orders of magnitude larger than the corresponding MERRA-2 AOD (e.g., SOD ≈ 18–30 for the ALL mask, increasing to 90–200 for the P70–P90 masks, while AOD remains near 1). Consequently, bias values (SOD - AOD) are strongly positive and increase with the percentile mask (e.g., +16 to +30 for ALL, rising to +180–200 for P90). MAE and RMSE follow the same pattern, reaching 80–110 for the ALL mask and 250–345 for P90. These statistics indicate that the simulated smoke column burden greatly exceeds the column optical depth represented by MERRA-2, regardless of the assumed microphysical configuration.

Background subtraction affects only the reanalysis field, reducing AOD by approximately 0.6 (1.17 to 0.57). As a result, bias becomes slightly larger, but MAE and RMSE change only marginally. Correlation coefficients remain extremely weak for the ALL mask (r≈0.00–0.06; R2≈0.00), demonstrating that the discrepancies are not due to a removable background offset but rather reflect fundamental structural mismatches in spatial representativeness between SOD and AOD.

The percentile masks (P70–P90) highlight a systematic deterioration in skill as increasingly high-AOD grid cells are selected. Bias rises to +180–200, RMSE exceeds 300, and correlations become increasingly negative (r≈-0.30 to -0.50), while R2 remains ≤ 0.24. This behaviour indicates that the regions of strongest AOD in MERRA-2 correspond to aerosol components and vertical layers not represented by SOD, including smoothed background mass, coarse-mode contributions, and long-range transported aerosol that the regional plume simulation cannot reproduce. The regression slopes remain near zero (|slope| < 0.002), confirming that variations in SOD magnitude have essentially no explanatory power for the spatial variability of AOD.

Differences across the microphysical perturbations (ABSORBING, WIDE_SIGMA, SMALL_RG, LARGE_RG) are minor, typically altering correlation by ≤ 0.05 and MAE/RMSE by ≤ 2–5 %. Similarly, the near-identical results across the three microphysical frameworks (BASE, CONSERVATIVE, LOWER ENVELOPE) demonstrate that reasonable adjustments to κ, rg,dry, and σg have only a secondary effect on the SOD–AOD comparison. The dominant source of discrepancy therefore lies not in microphysics, but in the excess plume mass and the spatial confinement of the smoke plume in the model domain relative to the broad, regionally smoothed optical depth represented in MERRA-2.

The statistical findings can be interpreted considering several physical and representativeness factors that fundamentally limit the correspondence between SOD and MERRA-2 AOD. Physically, the persistent decoupling between these variables reflects fundamental representativeness differences between the instantaneous fine-mode SOD fields produced at 2 km resolution and the hourly, 50–70 km column-integrated AOD from MERRA-2. SOD primarily reflects the fire-emitted smoke plume resolved by the regional model, whereas AOD also includes coarse-mode background aerosols and vertically extended layers. These structural differences explain the very low covariance and consistently large intercepts. Additional phase shifts arise from temporal mismatch between instantaneous model outputs and the assimilation-centred AOD, as well as plume-scale variability that is difficult for a coarse reanalysis to reproduce. Taken together, Tables S1–S3 show that BRAMS–SFIRE strongly overestimates SOD magnitude relative to MERRA-2 AOD, while underrepresenting the broader, regionally distributed optical-depth structure seen in the reanalysis. This behaviour reflects a plume that is too concentrated and optically intense locally, but less spatially extensive than the smoothed regional aerosol field in MERRA-2. A primary contributor to this mismatch is the contrast in spatial and temporal resolution between the high-resolution model output and the coarse MERRA-2 reanalysis. The BRAMS–SFIRE runs at 2 km horizontal resolution and produces instantaneous outputs, whereas MERRA-2 provides ∼ 0.5° × 0.625° (∼ 50 × 70 km) hourly means centred near hh:30. As a result, sharp smoke gradients and localized peaks captured by the model are heavily smoothed in MERRA-2, attenuating peak AOD values and shifting their apparent timing relative to the simulated plume. The smoothing also blends smoke with background aerosols, reducing the spatial contrast between plume and non-plume regions.

Differences in aerosol composition further amplify this representativeness gap between the model and the reanalysis. The SOD evaluated here is derived from fine-mode PM_2.5 (fresh smoke), while MERRA-2 AOD represents total extinction, integrating coarse-mode dust and sea salt. In regions where coarse-mode aerosols are abundant, total AOD can diverge from, or even anticorrelate with, fine-mode SOD in both space and time. This compositional difference is consistent with elevated regression intercepts, i.e., a background optical load not associated with the fire plume.

Additional discrepancies arise from the way MERRA-2 assimilates observational AOD and represents rapidly evolving wildfire plumes. MERRA-2 assimilates satellite AOD observations (e.g., MODIS, VIIRS), subject to retrieval/sampling constraints; its smoke fields depend on emission inventories and the analysis cycle. Rapid, pyro-convective injections from intense wildfires, characterized by sub-hourly variability and strong vertical transport, are difficult to capture in reanalysis products. This can produce phase shifts in both timing and vertical structure relative to the model, especially during the most active burning period (13:00–17:00 UTC).

Humidity-driven growth constitutes another factor influencing the magnitude and spatial alignment of extinction between the model and MERRA-2. The wet SOD uses modelled RH to allow growth in particle size and extinction efficiency. MERRA-2 may employ coarser meteorology and simplified κ parameterizations. Discrepancies in RH and κ directly affect the scattering enhancement and extinction efficiency, contributing to systematic differences in optical depth amplitude. The better agreement obtained under wet conditions supports the physical importance of this process.

Temporal pairing between the model and the reanalysis also introduces structural differences, especially during rapid plume evolution. The comparison pairs instantaneous model fields at H:00 with MERRA-2 hourly means representing roughly [H-30, H+30 min]. During periods of rapid-fire intensification or plume advection, this temporal mismatch further smooths and delays AOD peaks relative to modelled SOD maxima, manifesting as the apparent decoupling observed in the early afternoon hours.

Some of the discrepancy may also stem from simplifying assumptions in the optical calculations applied to the BRAMS-SFIRE output. The Mie-based SOD assumes spherical particles with a prescribed log-normal size distribution derived from bulk PM_2.5 mass concentration, using fixed microphysical parameters. Under extreme fire conditions, this simplification tends to overrepresent the number of small particles per unit mass, enhancing extinction near the fire front. This effect increases simulated SOD peaks but has a minor influence on broader plume statistics, as confirmed by the microphysical sensitivity tests.

The optical diagnostics quantified shortwave extinction, whereas wildfires emit intense longwave radiation that alters buoyancy and turbulence, promoting lofting and secondary aerosol formation. Such small-scale feedbacks are not explicitly resolved in MERRA-2, yet they modulate plume height and density and thus the spatial alignment of SOD and assimilated AOD.

The optical diagnostics presented in Sect. 3.2 revealed substantial differences between simulated SOD and MERRA-2 AOD, particularly during the early stages of plume development. To understand these discrepancies, it is necessary to examine how the mass distribution and composition of the smoke plume relate to its optical properties and radiative behaviour.

This section, therefore, shifts the focus from column-integrated optical depth to the underlying aerosol mass fields, PM_2.5, black carbon (BC), and organic carbon (OC), and investigates how their spatial structure governs the extinction properties diagnosed in the previous section. Because wildfire smoke is dominated by carbonaceous aerosols with distinct emission pathways and optical roles, analysing their spatial patterns provides essential physical context for interpreting SOD–AOD differences.

While SOD explicitly quantifies shortwave extinction at 550 nm, the radiative environment of the plume is simultaneously influenced by longwave emission from the fire, which enhances buoyancy, contributes to plume-rise, and facilitates the formation of secondary aerosols. Although longwave processes do not enter the SOD calculation directly, they modify the thermodynamic structure of the plume and thus affect the dispersion and vertical redistribution of the carbonaceous particles that control shortwave extinction. In this sense, longwave heating indirectly influences the spatial and vertical distribution of scattering and absorbing aerosols.

Figure 9 shows the spatial distribution of column-averaged PM_2.5 concentration and the corresponding MERRA-2 surface BC and surface OC mass concentrations at 19:00 UTC, during the mixed phase when the plume becomes more coherent. The clear spatial coherence between PM_2.5 and the carbonaceous species is consistent with a dominant carbonaceous contribution to the simulated fine-mode mass associated with biomass burning. As typically observed in wildfire plumes, OC dominates BC by roughly an order of magnitude, consistent with smouldering/mixed-combustion conditions that generate large amounts of semi-volatile organic compounds which rapidly condense into fine-mode particles. BC remains confined to a narrower core associated with flaming combustion, while OC extends farther downwind, forming a broader envelope around the PM_2.5 plume. This contrast reflects the two-source nature of carbonaceous smoke: BC forms primarily under high-temperature, oxygen-limited flaming conditions, while OC arises from lower-temperature pyrolysis and subsequent condensation processes that disperse more broadly downwind. This structural contrast, narrow BC ridge, broad OC plume, is clearly captured at 2 km model resolution but becomes smoothed at the MERRA-2 grid spacing (∼ 50–70 km). Such smoothing dilutes gradients and contributes to the weaker early-phase agreement between SOD and AOD noted in Sect. 3.2. By 19:00 UTC, however, the plume's organisation improves and the spatial correspondence among PM_2.5, BC, and OC becomes more coherent, aligning with the improved SOD–AOD coupling reported for the mixed phase.

To further connect aerosol mass with optical extinction, Fig. 10 compares the simulated SOD with the MERRA-2 BC and OC extinction AOD fields at 550 nm for 18:00 UTC. The spatial patterns reaffirm the behaviour already identified in Fig. 9: BC extinction remains weak and confined to a narrow region near the fire core, whereas OC extinction extends over a much broader downwind area, mirroring the wider PM_2.5 and OC mass envelope characteristic of smouldering and mixed-combustion emissions.

The simulated SOD field shows a much stronger spatial correspondence with OC extinction than with BC, confirming that OC is the dominant contributor to shortwave extinction at 550 nm in this plume. This is physically expected, as 550 nm is strongly sensitive to fine-mode scattering aerosols, and OC constitutes the bulk of the fine carbonaceous mass. In contrast, BC, although strongly absorbing, represents only a small fraction of the total aerosol mass and thus contributes minimally to visible-band extinction.

Some spatial mismatches remain, primarily due to the coarse horizontal resolution of MERRA-2 (50–70 km), which smooths gradients that the 2 km BRAMS–SFIRE simulation resolves sharply. This smoothing reduces peak extinction values in the reanalysis and broadens plume boundaries, producing slight offsets between SOD maxima and the corresponding features in the MERRA-2 extinction fields.

Figure 11 provides a quantitative assessment of how SOD relates to the carbonaceous extinction components of MERRA-2 during the mixed phase (19:00 UTC, 550 nm). Approximately 10 % of the data lie outside the plotting limits, but this level of clipping does not affect the statistical interpretation.

In panel (a), which compares SOD with BC extinction AOD, the ordinary-least-squares (OLS) fit shows a very small slope (≈ 0.06), a small positive intercept (≈ 0.03), and a modest coefficient of determination (R2 ≈ 0.28). These values indicate that only a limited fraction of the spatial variability in BC extinction is explained by SOD. This weak coupling is physically consistent with BC representing only a minor fraction of the total carbonaceous mass at this stage and with the coarse ∼ 50–70 km MERRA-2 grid smoothing the sharp BC gradients resolved by the 2 km model. The horizontally elongated scatter distribution relative to the 1:1 line further highlights this weak covariance and the strong representativeness differences between the two datasets.

In panel (b), the comparison between SOD and OC extinction AOD yields a substantially stronger statistical relationship. The regression slope (≈ 0.38) is roughly six times larger than for BC, the intercept is higher (≈ 0.16), and the coefficient of determination increases to R2 ≈ 0.31. These values reflect the dominant contribution of OC to smoke extinction at 550 nm and the closer correspondence between SOD patterns and the spatial structure of OC in MERRA-2. The slight flattening of the scatter at large SOD values likely results from the coarse resolution of the reanalysis, the limited dynamic range of the OC extinction field, and differences in humidification growth or mixing-state assumptions between the model's κ-Köhler/Mie closure and the GOCART/MERRA-2 framework. Taken together, these results confirm that OC extinction retains a much clearer imprint of the plume diagnosed by SOD than BC extinction, consistent with OC dominating visible-band extinction in biomass-burning smoke. The positive intercepts in both relationships indicate a persistent background extinction component present even where plume–model covariance is strong. No additional attribution (e.g., combustion phase variability or cloud effects) can be inferred from these diagnostics alone without independent observational constraints.

To extend the analysis beyond the scatter relationships in Figs. 10–11, Table 5 provides domain-wide statistical metrics for black carbon (BC) and organic carbon (OC) extinction under all microphysical configurations, quantifying bias, variance explained, and sensitivity to microphysical perturbations. These diagnostics reinforce the contrasting behaviour already identified in the spatial and scatter comparisons: while OC extinction exhibits a coherent and physically interpretable relationship with SOD, BC extinction remains only weakly coupled to the modelled smoke field. For BC, all microphysical configurations (Baseline, Conservative, Lower Envelope) produce nearly identical metrics, demonstrating that the SOD–BC relationship is largely insensitive to perturbations in κ, rg,dry, or σg. Mean SOD values range from 7.6–8.6, overwhelmingly exceeding the corresponding MERRA-2 BC extinction (≈ 0.06), yielding very large positive biases (7.5–8.5) and RMSE values between 52–58. These large errors are expected: BC represents only a minor fraction of carbonaceous aerosol mass in biomass-burning plumes and contributes weakly to visible-band extinction at 550 nm. Correlation coefficients remain low but positive (Pearson r≈0.14; Spearman ρ≈0.75), indicating that although BC extinction contains a monotonic spatial ordering relative to SOD, it explains only about 2 % of the SOD variance (R2=0.02). The regression slopes for BC (≈0.000) confirm that spatial variability in BC extinction barely responds to changes in SOD, while the small intercept (∼ 0.058) is consistent with a persistent background BC extinction present in the reanalysis, likely representing non-smoke contributions and the coarse effective resolution of MERRA-2.

In contrast, the OC extinction field shows a stronger and more physically meaningful connection to SOD. Although mean SOD (≈ 7.6–8.6) still greatly exceeds mean OC extinction (≈ 0.34), the bias is smaller than for BC, and the correlations are modestly higher (r≈0.16; ρ≈0.75). The OC regression slopes (0.001–0.002), while still small in absolute terms, are one order of magnitude larger than the near-zero slopes obtained for BC, mirroring OC's dominant contribution to carbonaceous aerosol mass and its stronger role in visible-band scattering. The regression intercepts (∼ 0.325) indicate a substantial baseline OC extinction in MERRA-2, consistent with the presence of regionally distributed secondary organic aerosols and with the smoothing inherent to the coarse-grid reanalysis. As with BC, the near-identical statistics across the three microphysical experiments reflect that microphysical uncertainties play only a secondary role in controlling the SOD–OC relationship. Overall, the diagnostics confirm that OC dominates the optical signature of the plume at 550 nm, acting as the principal contributor to scattering and thus to the radiative impact of the smoke layer, while BC contributes a much smaller, spatially confined absorbing component.

The weak sensitivity of the statistical metrics to the microphysical perturbations further demonstrates that the observed behaviour is governed primarily by species composition and representativeness differences between BRAMS–SFIRE and MERRA-2, rather than by uncertainties in aerosol size distribution or hygroscopic growth.

Beyond extinction behaviour, the single-scattering albedo (SSA) offers a complementary measure of the radiative character of the smoke plume by expressing the relative contributions of scattering and absorption. Accordingly, SSA fields derived from BRAMS–SFIRE were compared with their MERRA-2 counterparts to assess whether the model correctly captures the scattering–absorption balance within the plume.

Figure 12 contrasts the simulated SSA at 550 nm with two MERRA-2 products at 19:00 UTC: a BC+OC proxy (left panel) and the total-aerosol SSA (right panel). SSA represents the ratio of scattering to total extinction; hence low SSA values correspond to strongly absorbing aerosol (typically BC-rich fresh smoke), while high SSA indicates more scattering-dominated mixtures such as aged organics or background sulphates.

In the BC+OC comparison (Fig. 12, left), BRAMS–SFIRE yields SSA values mostly between 0.88 and 0.90, forming smooth horizontal gradients across the plume. This behaviour is consistent with moderately scattering aerosols and limited absorption. The corresponding MERRA-2 BC+OC proxy, however, exhibits sharper gradients and pockets of markedly lower SSA (< 0.82) near the fire core, reflecting stronger BC absorption in the reanalysis than in the model. These contrasts partly arise from MERRA-2's coarse grid (∼ 50–70 km), which simultaneously smooths plume edges and preserves mesoscale absorption signatures.

The right panel compares simulated SSA against total MERRA-2 SSA, which includes additional scattering species (sulphates, dust, sea salt). This total-aerosol SSA is higher and more horizontally uniform than the BC+OC proxy, reflecting the dominant contribution of non-carbonaceous scattering aerosols in the reanalysis. Their inclusion brings MERRA-2 SSA closer to the scattering-dominated regime simulated by BRAMS–SFIRE, though spatial differences remain due to the finer plume structure resolved by the 2 km model.

The discrepancies highlighted in Fig. 12 arise from several structural factors: (i) dilution and enhanced mixing in BRAMS–SFIRE, which reduce the local BC mass fraction and therefore increase the SSA (i.e., shift the aerosol population toward more scattering-dominated conditions); (ii) coarse spatial averaging in MERRA-2, which smooths sub-grid gradients but retains regional absorption patterns; and (iii) fundamentally different mixing-state assumptions. In this study, SSA is computed offline from PM_2.5using an externally mixed Mie scheme with a prescribed log-normal size distribution which does not account for absorption enhancement arising from internally mixed BC encapsulated within organic or sulphate material. By contrast, GOCART/MERRA-2 treats fine-mode aerosols as internally mixed, allowing lensing effects that increase BC absorption.

Consequently, this PM_2.5-based aerosol representation using a prescribed log-normal size distribution tends to overestimate SSA relative to MERRA-2 in the vicinity of the plume core. Because the PM_2.5-based optical calculation adopts fixed microphysical parameters within each phase scenario, the resulting SSA exhibits limited spatial variability, even across regions with strong concentration gradients.

To complement the spatial fields, Fig. 13 presents scatter-density comparisons between simulated SSA and MERRA-2 SSA for BC+OC (panel a) and total aerosols (panel b). In both cases, the relationships exhibit weak covariance and low explanatory power (R2≤0.15). The simulated SSA spans a narrow dynamic range, whereas MERRA-2 displays substantially larger spatial variability, particularly for the BC+OC proxy. This contrast leads to limited point-wise correspondence between datasets. Although the total-aerosol comparison shows slightly closer absolute values, correlations remain weak, reflecting structural differences in aerosol representation and mixing-state assumptions.

Taken together, Figs. 12–13 indicate that BRAMS–SFIRE captures the plume's spatial extent and its predominantly scattering-dominated character, consistent with the OC-rich composition shown in Figs. 9–11, yet it underrepresents the absorbing component associated with fresh BC. The persistence of SSA biases across all microphysical configurations further demonstrates that these discrepancies are structural, arising from resolution limitations, mixing-state assumptions, and representativeness differences rather than from uncertainties in aerosol microphysics.

Table 6 quantifies the SSA behaviour for both the BC+OC and total-aerosol configurations and confirms the spatial patterns identified in Figs. 12–13. When compared against the MERRA-2 BC+OC SSA proxy, the model exhibits a large positive bias of ∼ 0.36–0.37, consistent across the Baseline, Conservative, and Lower Envelope configurations. This bias reflects a systematic overestimation of scattering in the simulated plume: the mean model SSA is ∼ 0.90, whereas the BC+OC SSA in MERRA-2 is only ∼ 0.54 (“Mean simu” and “Mean MERRA-2”). These values corroborate Fig. 12, where BRAMS–SFIRE produces a more scattering-dominated (optically brighter, less absorbing) plume than the reanalysis.

The correlations remain weak, Pearson r ranges from -0.42 to 0.24, and Spearman ρ from -0.27 to 0.29, indicating little pointwise correspondence between model and reanalysis fields. The OLS slopes are large and variable (≈+14.1, −15.6, and −11.2 depending on configuration), and the intercepts span similarly large values (from about −12 to +15). These coefficients are not physically meaningful and primarily reflect the narrow dynamic range (i.e., low variance) of SSA combined with weak covariance between the datasets, rather than a physically meaningful relationship, and therefore should not be interpreted mechanistically.

In contrast, the comparison with MERRA-2 total-aerosol SSA yields a markedly better agreement. The bias nearly vanishes (+0.01) and both MAE and RMSE reduce to ∼ 0.01–0.02. Correlations remain weak and configuration-dependent (Pearson r=−0.35 to 0.47, Spearman ρ=−0.37 to 0.34). This improvement results from the fact that total-aerosol SSA in MERRA-2 includes additional scattering species, notably sulphates, dust, and secondary organics, which raise the reanalysis SSA to values close to those in the model (0.89–0.90, “Mean MERRA-2”). Thus, although BRAMS–SFIRE overestimates SSA relative to carbonaceous-only extinction, its bulk radiative behaviour aligns more closely with the full multi-species aerosol environment represented in MERRA-2.

Overall, these results confirm that the simulated plume is dominated by moderately scattering, OC-rich smoke, while the absorbing BC component is underrepresented relative to the reanalysis. The near-identical SSA statistics across all microphysical experiments show that the discrepancies are structural rather than parametric, arising from simplifications in the optical treatment (e.g., monodisperse, externally mixed particles), the model–reanalysis resolution gap, and differences in mixing-state representation, rather than from uncertainties in κ, rg,dry, or σg.

Figure 14 shows the frequency distributions of domain-wide near-surface PM_2.5 concentrations ( µg m⁻³) at 13:00, 16:00, 19:00, and 22:00 UTC on 15 October 2017. All panels use identical binning and a logarithmic y-axis to emphasize the occurrence of rare but extremely high concentrations. At every time step, the distributions are highly right-skewed: most grid cells fall within the lowest concentration bins, while a persistent heavy tail extends toward very large PM_2.5 values, characteristic of a spatially confined but optically and chemically intense smoke plume.

Between 13:00 and 16:00 UTC, the distributions broaden substantially. Although the mode remains anchored in the lowest bin, indicating that the majority of the domain experiences only weak smoke influence, the gradual infilling of intermediate bins and the emergence of a long upper tail reflect the onset of plume intensification and expansion during the early afternoon. By 19:00 UTC, the histogram exhibits its widest dynamic range, with sustained occupancy across intermediate and high concentration classes and peak values approaching 107 µg m⁻³. This behaviour signals maximum plume loading near the surface, consistent with the period of strongest fire activity and turbulent mixing. By 22:00 UTC, the distribution narrows slightly: the extreme upper tail becomes less populated, suggesting partial dilution and reduced near-surface entrainment following the evening transition. Nevertheless, the heavy tail does not vanish, implying that localized pockets of very high PM_2.5 persist despite weaker turbulence and declining emissions. Intermediate concentration bins retain notable counts, indicating that the plume continues to influence a non-negligible fraction of the domain even as the intensity of near-surface impacts diminishes.

From a physical standpoint, the coexistence of an overwhelmingly low-concentration mode with a pronounced heavy tail underscores the highly intermittent nature of smoke exposure: large portions of the domain remain only weakly affected, while a small number of grid cells experience extreme concentrations that dominate the regional burden. The afternoon broadening of the distributions reflects plume growth, lateral spread, and enhanced turbulent transport, whereas the partial evening contraction corresponds to stabilization and reduced mixing. Such behaviour highlights the operational relevance of intense, spatially localized hotspots, which evolve rapidly through the afternoon and can persist into the evening despite overall decreases in plume activity.

Whereas Fig. 14 documents the evolution of PM_2.5 distributions throughout the day, Fig. 15 isolates the histogram corresponding to 19:00 UTC on 15 October 2017, highlighting the structure of the heavy-tailed distribution at the hour of maximum plume loading. Figure 15 shows the domain-wide frequency distribution of simulated PM_2.5 concentrations at that time. The histogram reveals a pronouncedly right-skewed distribution, with the vast majority of grid cells occupying the lowest concentration bins, while a progressively smaller number populate the intermediate and upper ranges. The logarithmic vertical axis accentuates this structure: although extremely high PM_2.5 values occur infrequently; they form a distinct heavy tail that extends up to nearly 1 × 10⁷ µg m⁻³.

This statistical behavior reflects the intrinsic spatial heterogeneity of the plume. Most of the domain remains only weakly affected by smoke at this hour, whereas a limited set of grid cells, associated with the fire core and the most concentrated portions of the near-surface plume, exhibit exceptionally large PM_2.5 loadings. These extreme values are physically plausible in the context of intense wildfire activity, where strong pyroconvective injection, limited dilution, and local recirculation can sustain highly elevated concentrations.

The coexistence of a dominant low-concentration mode with a persistent heavy tail indicates a smoke field that is spatially sparse yet extremely intense where present, consistent with the plume morphology seen in the spatial plots. Consequently, domain-wide exposure at 19:00 UTC is governed disproportionately by a small fraction of high-PM_2.5 cells, underscoring the presence of sharp gradients and localized hotspots that play a central role in the air-quality and radiative impacts of the event.

While the model captures the broad spatial organisation and the characteristic heavy-tailed distribution of the plume, Figs. 14–15 show that a very small number of grid cells attain PM_2.5 and SOD values far exceeding those reported in observational studies. Surface and aircraft measurements in intense wildfire environments typically report maximum PM_2.5 concentrations of 103–10⁴ µg m⁻³ (Gili et al., 2025; Landis et al., 2017), and even the highest shortwave AOD values observed by AERONET, FTIR, or aircraft seldom exceed 10–17.5 (Daniels et al., 2024; Eck et al., 2019; Frausto-Vicencio et al., 2023; Kassianov et al., 2025; Paton-Walsh et al., 2005; Shinozuka and Redemann, 2011). Concentrations in the 10⁶–10⁷ µg m⁻³ range or SOD magnitudes of several thousand therefore have no documented observational analogue.

Such extremes likely arise from the interaction between very sharp emission gradients near the fire front, unresolved sub-grid variability in fuel consumption, and the behaviour of numerical transport schemes under steep concentration contrasts. Large tracer spikes have been documented in models employing Semi-Lagrangian advection, particularly in some global forecasting systems (e.g., ECMWF-type models or other spectral NWP frameworks), where negative-value clipping and subsequent mass-adjustment procedures can generate artificial local extrema (Aranami et al., 2015; Bermejo and Conde, 2002; Sørensen et al., 2013; Zerroukat and Shipway, 2017). However, BRAMS–SFIRE does not use a classical Semi-Lagrangian transport scheme. The atmospheric dynamics are solved in an Eulerian framework derived from the RAMS core, while fire spread is computed using a level-set formulation. In this context, the extreme PM_2.5 values observed here could arise from the interaction between intense near-source emission gradients, grid-scale representation constraints, including the spatial mapping of vegetation and fuel types, and transport–emission coupling under steep concentration contrasts.

Although these outliers occupy only a very small fraction of the domain and do not significantly affect the large-scale optical diagnostics presented earlier, their presence indicates that further refinement is required in the representation of near-source smoke production and its numerical transport. The underlying causes of these extremes, linked to the parameterisation of fuel consumption, emission intensity, and the FRP formulation, are examined in detail in Sect. 3.6.

This section examines how smoke–radiation interactions modulate atmospheric stability during the wildfire event, linking PM_2.5 loading, optical properties, and radiative forcing to changes in CAPE, CIN, and boundary-layer structure.

Atmospheric stability is primarily governed by the vertical temperature gradient or lapse rate. The behaviour of an air parcel depends on how rapidly it cools adiabatically compared to the ambient temperature profile. If a rising parcel cools at a rate lower than the environmental lapse rate, it becomes warmer and less dense than its surroundings, gaining positive buoyancy. Conversely, if it becomes denser than the surrounding air, it will descend. When vertical displacements of air parcels are amplified by buoyancy forces, the atmosphere is considered unstable, favouring convection, cloud formation, and potentially severe weather. In contrast, if the parcel is cooler and denser than its environment, vertical motion is suppressed, indicating a stable atmosphere.

Stability is evaluated by comparing the environmental lapse rate with the theoretical adiabatic lapse rates: the Dry Adiabatic Lapse Rate (DALR ≈ 9.8 K km⁻¹) and the Moist Adiabatic Lapse Rate (MALR, typically 4–7 K km⁻¹, depending on humidity). Strongly stable conditions occur during temperature inversions, where temperature increases with height. Surface-based inversions commonly form at night within the lowest few hundred metres, while subsidence inversions associated with high-pressure systems often appear between 1–3 km. Such inversions act as barriers to vertical mixing and are crucial for air pollution accumulation.

When moist air parcels rise, they initially follow the DALR until reaching the Lifting Condensation Level (LCL), where condensation begins. The release of latent heat during condensation (an exothermic process) partially offsets adiabatic cooling, increasing the parcel's buoyancy and shifting it to follow the MALR. If sufficient energy is available, the parcel reaches the Level of Free Convection (LFC) and may continue to ascend to the Equilibrium Level (EL), near the tropopause.

The atmospheric layer closest to the surface is the Planetary Boundary Layer (PBL), where turbulence and mixing are most pronounced due to surface heating and friction. The structure and height of the PBL vary diurnally and are strongly influenced by radiative forcing and atmospheric stability. During the daytime, especially in wildfire regions, intense surface heating and smoke-induced turbulence may erode stable layers, promoting deep mixing.

These vertical motions have significant implications for aerosol dispersion. In stable conditions, especially under inversions, particulate matter accumulates near the surface due to suppressed mixing. Under unstable, convective conditions, particles are efficiently mixed and vertically transported.

CAPE quantifies the potential energy that a rising air parcel can convert into kinetic energy if lifted adiabatically. Higher CAPE values suggest stronger updrafts and a greater likelihood of convective development, enhancing vertical mixing and pollutant dispersion. CIN, on the other hand, represents the energy barrier that must be overcome for a parcel to initiate free convection. Large CIN values suppress vertical motion and contribute to the accumulation of pollutants and aerosols near the surface.

The regions impacted by these two energy terms are bounded by two key levels: the LFC and the EL. Below the LFC, the parcel is cooler than its environment, requiring energy input to rise – this energy defines the CIN. Above the LFC, the parcel is positively buoyant, contributing to CAPE. These estimates often assume pseudo-adiabatic processes (no entrainment and removal of condensed water), allowing latent heat release to maintain saturation.

In wildfire scenarios, evaluating CAPE and CIN becomes particularly relevant for assessing pyroconvection. Wildfires inject substantial heat and moisture into the atmosphere, potentially reducing local CIN (by warming the PBL) and allowing parcels to reach their LFC. When background CAPE is significant, this can trigger the formation of convective clouds (flammagenitus or pyrocumulus), and in extreme cases, deep convective events such as pyroCb (pyrocumulonimbus). Studies have shown that intense fire heat fluxes increase CAPE and reduce CIN, facilitating the vertical transport of smoke and the development of fire-induced storms.

In addition, turbulent fluxes generated by rising air parcels and plume dynamics influence pollutant dispersion. Parcels moving upward often carry higher aerosol concentrations, resulting in well-defined plume cores. This distribution can often be approximated by Gaussian profiles in horizontal cross-sections, in agreement with Fick's law of diffusion, where the diffusion flux is proportional to the gradient in species concentration.

The interaction of aerosol particles with radiation is governed by two fundamental optical processes: scattering and absorption. Scattering occurs when incident electromagnetic radiation is redirected by particles. In the case of Mie scattering, relevant when the particle diameter is comparable to the wavelength of the incident radiation, light penetrates the particle, and the internal refractive index gradients cause the redirection of light. OC aerosols act primarily as efficient shortwave scatterers, although certain organic fractions (e.g., brown carbon) can contribute weak absorption in the UV–blue region.

Absorption involves the uptake of photon energy by electrons in the atoms or molecules of the aerosol particle. The absorbed energy excites electrons to higher energy levels, and as they relax to their ground states, the energy is primarily dissipated through non-radiative processes, increasing the internal kinetic energy of the particle. This manifests macroscopically as thermal heating of the particle, i.e., the conversion of radiant energy into thermal energy. BC is a strong absorber in the visible and near-infrared spectra and significantly contributes to atmospheric heating.

While both OC and BC interact with radiation in the UV–visible range (0.3–0.7 µm), for most fine-mode aerosols, LW radiative effects are weaker than SW effects, although dense, absorbing smoke layers can exert measurable LW heating through absorption and thermal re-emission.

The absorption of SW radiation by BC-rich aerosol layers elevates local temperatures, promoting thermal stability. This may lead to or enhance thermal inversions, thereby increasing CIN and inhibiting convection and cloud formation. In contrast, OC scattering reduces incoming SW radiation and cools the surface, reducing the surface–atmosphere temperature contrast and decreasing the lapse rate. This stabilisation of the lower troposphere lowers CAPE and diminishes the buoyant energy available for vertical motion.

To link these thermodynamic mechanisms to the radiative forcing exerted by the plume, Fig. 16a–d present the shortwave (SW) and longwave (LW) surface flux perturbations induced by smoke at 16:00 UTC (fresh phase, t1) and 19:00 UTC (mixed phase, t2). In all panels, the fields are expressed as No Fire – Fire, such that positive values denote a reduction in downwelling flux caused by the presence of smoke (i.e., the fire simulation receives less radiative energy at the surface than the no-fire control), while negative values correspond to an increase in downwelling flux in the fire run. This definition allows a direct attribution of radiative effects to smoke processes alone, isolating aerosol scattering, absorption, and thermal re-emission relative to a clean-air background. Figure 16 therefore provides a physically coherent depiction of how the plume modifies the surface radiative energy budget during the late-afternoon transition (16:00–19:00 UTC), linking these perturbations to the aerosol optical characteristics described in Figs. 7–12.

At 16:00 UTC, during the fresh and more concentrated phase of the plume, SW differences reach values above 183 W m⁻² within the plume core. These large positive differences indicate that the fire simulation receives substantially less solar radiation at the surface due to strong aerosol extinction. The spatial maxima of SW attenuation are tightly collocated with regions of enhanced column smoke absorption, highlighting the direct radiative impact of high aerosol loading.

In the longwave band, ΔLW values are predominantly negative (approximately -40 to -60 W m⁻²) within the same regions, indicating enhanced downwelling longwave radiation in the fire simulation relative to the no-fire control. This increase in surface LW flux is consistent with additional atmospheric emissivity associated with the warm, smoke-laden air mass and fire-induced moisture perturbations.

By 19:00 UTC, the smoke plume becomes more diluted, and its column-absorption contours stretch downwind, indicating transport and dispersion during the transition to the mixed phase. Consistent with this dilution, the shortwave perturbation remains evident but is weaker and more spatially diffuse, with the strongest SW reductions confined to the regions of highest column absorption. In the longwave, the perturbation persists and remains closely collocated with the absorption maxima, indicating enhanced downwelling LW in the fire within the plume core. Together, the SW reduction and LW increase show that the plume continues to modify the surface radiative budget across both phases, while the spatial structure of the signal increasingly reflects plume dilution and advection.

Beyond the instantaneous flux perturbations, the reduction in shortwave irradiance within the plume-affected boundary layer limits surface-driven buoyancy production, while the enhanced longwave emissivity of the smoke layer increases radiative heating aloft. This differential heating between the surface and the lower troposphere favours the development or strengthening of a temperature inversion, effectively acting as a radiative lid over the boundary layer. Such radiatively induced modifications in stability suppress turbulent mixing and promote boundary-layer stabilisation during periods of peak smoke loading, providing a dynamical pathway through which aerosol optical properties can influence boundary-layer evolution and help explain the persistence of high near-surface PM_2.5 concentrations observed in Sect. 3.4. Because the strength of this mechanism depends on aerosol absorption characteristics, the SSA biases identified earlier suggest that the model may underestimate the magnitude of smoke-induced radiative heating within the plume core, implying that the stabilising radiative lid could be even more pronounced under fully absorbing conditions.

Meridional cross-sections of extinction (SEC) and absorption (SAC) along 8.1° W (Figs. 17–19) reveal how spectrally dependent radiative heating interacts with lower-tropospheric stability during the evolution of the smoke plume. Colour shading denotes SEC and SAC at the relevant wavelength, while solid black and dashed red isolines represent the Fire and No-Fire CIN and CAPE, respectively. CAPE is absent within the 0–1.2 km layer and is therefore omitted; CIN is shown only where non-zero.

At 15:00 UTC and 550 nm (Fig. 17), extinction and absorption are concentrated within the lowest 300–500 m, forming a compact, optically intense layer centred near 40° N. The near-perfect co-location of SEC and SAC maxima indicates a fresh, BC/OC-rich plume with minimal dilution. CIN isolines in the Fire simulation lie systematically above those of the No-Fire run, evidencing a localized weakening of the low-level inversion. This displacement is consistent with shortwave heating by BC-containing aerosols, which warms the plume and begins to erode static stability, an early manifestation of the radiative-lid mechanism diagnosed from the surface forcing fields in Fig. 16.

By 19:00 UTC (Fig. 18), the plume has deepened and become more vertically diffuse, with SEC and SAC extending intermittently up to ∼ 1 km. CIN perturbations intensify and expand horizontally between 39.9–40.1° N, demonstrating persistent thermodynamic modification by aged, lofted smoke. The joint action of turbulent mixing and continued radiative heating suppresses surface–atmosphere coupling, favouring the accumulation and persistence of smoke near the surface. The spatial coherence between elevated CIN and enhanced SEC/SAC confirms that smoke-driven warming remains dynamically relevant even after the plume has become partially diluted.

At 700 nm (Fig. 19), extinction and absorption weaken systematically, SAC decreases by a factor of ∼ 2–3 and SEC by 30 %–40 % relative to 550 nm, yet the plume retains its vertical structure, and CIN isolines continue to be elevated in the Fire simulation. This attenuation reflects the wavelength dependence of mixed BC–OC aerosols: as organic chromophores (brown carbon) become optically less active in the red/near-infrared, BC-driven heating increasingly dominates the radiative response.

Conversely, at shorter wavelengths (e.g., 400 nm), SEC and SAC strengthen markedly (SEC ≈ 3.3 × 10⁻² m⁻¹; SAC ≈ 6 × 10⁻³ m⁻¹), remaining confined within the lowest 500–800 m, precisely where CIN differences are largest. This pronounced shortwave sensitivity, driven by OC and brown carbon absorption, reinforces the mechanism whereby visible/UV heating enhances plume buoyancy relative to its surroundings, modifies lower-tropospheric stability, and prolongs near-surface smoke residence.

Together, Figs. 17–19 show that radiative heating by BC and OC governs the vertical thermodynamic imprint of the plume, with 550 nm providing the strongest combined scattering–absorption signal, 700 nm capturing the persistence of BC-induced heating, and 400 nm emphasising the intense shortwave sensitivity of OC-dominated smoke. These spectral responses align with the radiative perturbations diagnosed at the surface (Fig. 16), collectively demonstrating how smoke modifies stability through a vertically structured and wavelength-dependent radiative forcing.

The inversion-layer analysis shown in Figs. 20–21 provides the dynamical counterpart to the radiative and optical diagnostics discussed in Figs. 16–19. Together, these results reveal how smoke functions not merely as a passive tracer but as an active radiative agent capable of reshaping lower-tropospheric stability during the late-afternoon evolution of the plume.

At 15:00 UTC (Fig. 20), a compact near-surface smoke layer, characterised by strong extinction and absorption at 550 nm, coincides with positive potential-temperature anomalies differences (θ′≈1–1.5 K) and reduced relative humidity (RH^′ < 0) beneath the inversion. These signatures represent the first thermodynamic imprint of radiative heating by the plume: the smoke-induced warming does not physically lift the inversion but erodes it from below, weakening the temperature gradient that sustains its stability. Immediately above the inversion, RH′>0 indicates moisture redistribution by modified turbulent mixing, consistent with a plume that is already altering the thermodynamic environment rather than simply responding to it. This early behaviour aligns with the shortwave-absorption patterns diagnosed in Fig. 16 and with the vertical concentration of extinction and absorption shown at 550 nm in Fig. 17.

By 19:00 UTC (Fig. 21), the perturbations intensify and extend vertically to 300–600 m, mirroring the deepened SEC/SAC fields documented in Figs. 18–19. The θ′ differences strengthen along the plume axis, and the accompanying RH^′ dipole (dry below, moist above) becomes more pronounced. These changes demonstrate that the inversion is becoming thermodynamically weaker and spatially irregular, not through geometric uplift, but through sustained modification of temperature and moisture stratification driven by radiative heating and plume-induced turbulence. This evolution is fully consistent with the radiative-lid mechanism inferred from SW/LW flux perturbations in Fig. 16: strong shortwave attenuation reduces surface heating, while longwave trapping and BC/OC absorption warm the lower troposphere, suppressing boundary-layer deepening and enabling the continued presence of lofted aerosol layers.

The co-location of FIRE–No Fire differences with regions of enhanced SEC and SAC further confirms that inversion modification is directly coupled to the plume's wavelength-dependent radiative forcing. Heating by dense smoke is particularly efficient at short wavelengths (400–550 nm), reinforcing the thermodynamic differences at low levels and supporting persistent vertical redistribution of smoke. In this way, the inversion cross-sections provide a dynamical validation of the radiative pathways previously identified, demonstrating how smoke alters both the strength and structure of stability regimes that ultimately govern plume dispersion and residence time.

Because the BRAMS–SFIRE simulations were performed over a limited regional domain that does not cover all of continental Portugal, a natural concern is that artificial aerosol accumulation might arise through numerical trapping near the downwind boundary. To assess this, we analysed the domain-integrated column PM_2.5 north of 40.35° N (Fig. 22). The resulting time series shows no spurious monotonic increase or retention unrelated to plume evolution; instead, column mass increases and declines in phase with the fire activity and plume advection. This confirms that boundary conditions do not dominate the simulated aerosol budget.

To diagnose the physical origin of the sharp local maxima inside the domain, we examined the joint evolution of vertically integrated PM_2.5 and surface sensible heat flux at three representative locations along the plume pathway (Source, Midpoint, Endpoint; Fig. 23). In all cases, abrupt increases in PM_2.5 coincide with intense pulses of sensible heat flux, a signature of vigorous combustion and rapid vertical injection. These maxima are confined to limited regions near the active fire front or within the nascent plume core, reflecting steep near-source gradients characteristic of high-intensity crown–surface fire propagation. Such behaviour differs fundamentally from prescribed burns or low-intensity wildfires (Adetona et al., 2016; Miranda et al., 2010), which exhibit broader plume cross-sections and more gradual dilution.

The alignment of the PM_2.5 spikes with heat-flux pulses strongly suggests that the extreme column PM_2.5 values originate from physically plausible, highly localised fire–atmosphere interactions. However, the sensitivity of these peaks to the local sensible heat flux also indicates that uncertainties in fuel load, vegetation composition, and canopy structure can amplify near-source emissions and confine the initial aerosol distribution more strongly than observed. In mixed Mediterranean forest systems, where fuel heterogeneity is pronounced, misrepresentation of understory and crown biomass can substantially bias the heat release rate and, consequently, the vertical and horizontal gradients of PM_2.5. Improved fuel mapping and characterisation would therefore likely reduce the magnitude of these extremes in future simulations.

This interpretation is consistent with the behaviour observed in SOD and extinction fields. Earlier sections showed that SOD exhibits a heavy-tailed distribution and extremely sharp local maxima, some of which exceed values supported by field and remote-sensing observations. These peaks align spatially with the PM_2.5 spikes identified here, suggesting a shared origin in over-intensified near-source injection rather than in domain-scale numerical artefacts. While numerical transport under very steep concentration gradients can modulate local extremes, the evidence from Figs. 22–23 indicates that the dominant contribution originates upstream, during the emission and injection stage, rather than during plume advection.

This conclusion is consistent with the discrepancies previously noted between BRAMS–SFIRE SOD and MERRA-2 extinction, SSA, and BC/OC partitioning. The strongest disagreements occurred in regions where BRAMS–SFIRE produced exceptionally concentrated plume cores. These differences likely stem not from the radiative or microphysical parametrisations themselves but from the fire source term, which controls the thermal and mass fluxes injected into the atmosphere. In particular, uncertainties in the formulation of the radiative power–to–mass conversion (Sect. “Implementation of FRP in SFIRE”) and its dependency on fuel type and combustion phase can introduce biases in the magnitude and vertical placement of smoke emissions. A systematic reassessment of these parameters, together with refined fuel maps for Iberian vegetation types, would help reduce the sharpness of near-source gradients and improve agreement with regional-scale observations.

Finally, consistent with previous observational and modelling studies (Carroll et al., 2025; Huff et al., 2021; Lareau and Clements, 2015; Reid et al., 2005), the extreme values occupy only a negligible fraction of the domain and do not characterise the broader plume. The large-scale structure agrees well with MERRA-2 AOD, indicating that the SOD, extinction, and SSA diagnostics analysed in earlier sections predominantly reflect the regional plume evolution rather than isolated near-source spikes. Taken together, the evidence in Figs. 22–23 demonstrates that the extreme PM_2.5 and SOD values arise from physically consistent but highly localised fire–atmosphere processes, intensified by uncertainties in fuel representation and emission parametrisation, rather than from numerical accumulation or domain-boundary artefacts.