The SCQ fire started on 24 July 2021 at 14:00 UTC (LT-2) in Catalonia, Spain, with the ignition point located at 41.52329° N, 1.369071° E. It spread eastward under a predominantly westerly wind and directly developed a convective plume (Fig. a). Occasional overshooting created short-lived pyrocumulus clouds between 17:00–19:00 UTC, termed overshooting pyrocumulus (oPyroCu) by . After 19:00 UTC, no further pyrocloud formation was observed . The change from moist convection (i.e. pyrocloud formation) to dry convection (i.e. no pyrocloud formation) coincided with the arrival of the sea breeze at 19:00 UTC and sunset (19:41 UTC). The sea breeze caused the advection of moisture combined with backing of the wind from 300 to 240–270° (Fig. d). The moisture advected by the sea breeze, combined with the cooling around sunset, caused an increase in relative humidity (Fig. c). Nevertheless, observations indicate that the moistening did not result in renewed pyrocloud formation (Fig. a). Simultaneously, despite the cooling and moistening of the atmosphere around sunset (Fig. c), the SCQ fire maintained a dry convective plume until midnight (22:00 UTC).

At 19:51 UTC, a radiosonde was released at the right flank of the fire (yellow star; Fig. a) into the convective plume of the SCQ fire that measured vertical profiles of temperature, specific humidity, wind speed and rising speed. A description of the equipment and measurement techniques is provided by . The radiosonde drifted predominantly eastward during its ascent (grey and red circles; Fig. a), with measurements starting at 80 m a.g.l. (above ground level; Fig. b). In ambient conditions, the maximum rising speed of a radiosonde is ±2 ms-1. Rising speeds above 2 ms-1 indicate a vertical acceleration of the radiosonde by the convective motions of the wildfire plume . Hence, from the rising speed during the ascent of the radiosonde (Fig. b), we conclude that the radiosonde measured inside the convective plume of the SCQ fire between 0.3–1.5 km altitude a.g.l. In addition to measurements within the convective plume, the radiosonde recorded a partial profile of the ambient temperature and relative humidity during its descent.

Consequently, we will use the radiosonde measurements to evaluate how well the simulated pyro-convection and the ambient conditions compare to reality. For the evaluation, we will focus on comparing the observed and simulated temperature profiles. Furthermore, we will also compare the measured wind speed in the convective plume with the simulated wind speed. The measured wind speed during the descent of the radiosonde in the surroundings of the plume and the measured specific humidity are not used for the comparison. We consider the wind speed measurements during descent unreliable due to the high descent speed of the radiosonde (8–9 ms-1). The specific humidity measurements, on the other hand, are reliable but redundant for evaluating our simulation, as both observations and the simulation indicate a dry convective plume (i.e. no pyro-clouds). Hence, the simulation was sufficiently dry.

The radiosonde measurements were taken using the Windsonde S1H2 system, which has a sensor accuracy for temperature and pressure of 0.3 °C and 1.0 hPa . Comparison of the S1H2 sondes to regular meteorological Vaisala RS41-SG sondes has shown two challenges with their measurements. Firstly, the S1H2 sonde shows a slow humidity response at cloud tops . Secondly, the S1H2 sondes measure relatively noisy wind profiles in turbulent conditions . The first concern is irrelevant to our study as we focus on a dry convective plume. The second challenge results indeed in a relatively noisy wind pattern inside the atmospheric boundary layer (see Fig. c, Sect. ). Nonetheless, it is sufficient to provide a range of realistic wind speeds observed inside the atmospheric boundary layer. Furthermore, despite the noisy signal, it provides a clearly distinguishable signal of the capping inversion. Moreover, an exact match of the radiosonde measurements with an LES simulation is always impossible inside an atmospheric boundary layer, as radiosondes provide instantaneous measurements of the turbulence, whereas LES focuses on producing similar turbulence on average. Hence, the goal of the comparison is not a point-by-point match, but rather an evaluation of how well our LES setup reproduces the overall characteristic of the observed plume and atmospheric boundary layer structure in terms of mixed layer potential temperature, wind speed and boundary layer height.

As discussed in the introduction, we aim to isolate the effect of the fires on the atmosphere, specifically focusing on the fire-induced circulation during dry pyro-convection. Hence, in our MicroHH simulation setup , we prioritise atmospheric realism while significantly simplifying the simulated fire behaviour to isolate the fire's impact on the atmosphere.

MicroHH is configured with a domain of 38.4km×25.64km×12km (x×y×z). The domain is elongated eastward (i.e. in the positive x direction) as the observed plume from the SCQ fire developed eastward, and MicroHH uses periodic boundaries. To prevent the unwanted recirculation of the simulated plume over the eastern border, we found that a domain of 38.4 km in the eastward direction was sufficiently large (see Appendix ). The simulations use an equidistant horizontal resolution of 25 m and a stretched vertical grid starting at 10 m above the surface (Table ).

Each simulation covers a 4-h period (16:00–20:00 UTC) on 24 July 2021. For the analysis (Sect. ), we used the last hour of the simulations (19:00–20:00 UTC) as it matches the timing of the sounding (19:51 UTC) and the period during which local firefighters reported the reversal of the surface winds. Hence, this is the period in which we expect a fire-induced circulation to develop. The first two hours serve as spin-up time, after which the fire is initialised. This provides an additional hour of spin-up time (18:00–19:00 UTC) to allow the convective plume to develop, ensuring a dry convective plume between 19:00–20:00 UTC.

Detailed implementations of the fire and meteorological boundary conditions are described in Sects.  and , respectively.

Table  lists the four LES simulations using in this study. The first two simulations, ref-run and fire-run, follow exactly the simulation setup described in Sect. . The only difference is that for the fire-run, a fire is initialised at 18:00 UTC, whereas for the ref-run, no fire is initiated. Consequently, the difference between the fire-run and the ref-run shows the impact of the fire on the atmosphere. Hence, using these two simulations, we can determine the impact of the fire-induced circulation on the thermodynamic structure of the atmosphere, our first objective.

The last two simulations, fire-run:no-u-advec and fire-run:pg, are modified versions of the fire-run designed to investigate whether the directional wind shear, as hypothesised, governs the development of a fire-induced circulation. In these two simulations, we modified the large-scale forcings of the zonal and meridional wind in two different ways to create increasingly less directional wind shear compared to the fire-run.

In MicroHH, three large-scale forcings are applied to the zonal and meridional wind components (Eqs.  and , respectivelly): (1) large-scale advection prescribed directly from ERA5 (∂u∂tlsa and ∂v∂tlsa), (2) the pressure gradient implemented through the geostrophic wind (ug and vg), and (3) nudging defined by the nudging velocities (un and vn) and a nudging timescale (τn) set at 10 800 s. 4∂u∂tls∼∂u∂tlsa︸Large scaleadvection+f0(v-vg)︸Pressure gradient-u-unτn︸nudging5∂v∂tls∼∂v∂tlsa︸Large scaleadvection-f0(u-ug)︸Pressure gradient-v-vnτn︸nudging, where f0 represents the Coriolis parameter of 9.67×10-5s-1 based on the center of the SCQ fire (latitude: 41.51775°, longitude: 1.494428°).

Figure  shows the large-scale forcing terms for the zonal and meridional wind components based on the fire-sim. Near the surface (<500 m), the large-scale advection term (a) is dominant for the zonal wind component (solid lines), whereas for the meridional wind component (dashed lines), the pressure gradient term (b) is the largest. Hence, to reduce the directional wind shear, we modified the fire-run in two ways: (1) turning off the large-scale advection of zonal momentum after 18:00 UTC (fire-run:no-u-advec) and (2) turning off both large-scale momentum advection and nudging after 18:00 UTC, resulting in a simulation with only a pressure gradient forcing (fire-run:pg). The subsequent changes in vertical wind profiles are discussed in Sect. .

To evaluate the simulated pyro-convection between 19:00–20:00 UTC, we compare the simulated plume shape and vertical profile of the virtual potential temperature (θv) and wind speed (U) with visual observations (Fig. ) and the in-plume radiosonde measurements (Fig. ). The observed plume of the SCQ fire between 19:00–20:00 UTC is presented in Fig. a–c for 19:13, 19:34, and 19:59 UTC. Throughout the hour, a well-developed convection column was observed (1) with occasional overshooting (2) and a horizontally dispersed smoke layer, also called the dispersion layer (3). Here, the distinction between the convection column and the dispersion layer is made visually based on the dominant axis of development: vertical (i.e. convection column) or horizontal (i.e. dispersion layer). The average simulated plume shape between 19:00–20:00 UTC based on the inert tracer is shown in Fig. d (grey outline). Similar to the observed plume, the average simulated plume also shows a convection column and a dispersion layer. Occasional overshooting also occurs in the simulation (not shown), but not consistently enough to affect the average plume shape. Additionally, the simulation shows, on average, downward transport of smoke below the dispersion layer, which suggests subsiding motions ahead of the fire front. Although the downward transport of smoke does not become apparent in Fig. a–c, it does match descriptions of the plume behaviour provided by local fire fighters.

For the remainder of this comparison, we will compare the radiosonde measurements to the simulation, which has two challenges. Firstly, the location of the radiosonde relative to the plume during the measurements is unknown, as we do not have observations on the exact location of the plume during the sounding. Secondly, we cannot replicate the exact shape of the turbulent plume at the time of the sounding, since turbulence is a chaotic process. Instead, we can replicate the average shape of the plume between 19:00–20:00 UTC, the period in which the sounding was released. Despite these challenges, we know that the radiosonde moved predominantly eastward (Fig. a) and that it measured inside the convective core of the plume between 0.3–1.5 km altitude (Fig. b). To capture the convective core of the plume in the simulation, we extract a xz cross-section at y=12.8125 km. This xz cross-section represents an east-west plane through the centre of the fire, matching the predominant eastward movement of the radiosonde. Hence, we will compare the radiosonde measurements with the averaged xz cross-section of the plume in the fire-run between 19:00–20:00 UTC through the centre of the fire.

The measured rising speed and horizontal displacement of the in-plume radiosonde (black triangles; Fig. d) provide a quantification of the observed plume shape. The horizontal displacement of the radiosonde within the convective core of the plume (0.3–1.5 km altitude) shows that the averaged simulated plume shape is less tilted than the observed one. Above 1.5 km, the rising of the radiosonde effectively halts, resulting in approximately 3 km of primarily horizontal displacement between 1.5–1.7 km altitude before continuing its ascent (Fig. d). This suggests that the radiosonde exited the convection column between 1.5–1.7 km, but whether the exit occurred at the top of the plume cannot be derived from the rising speed and horizontal displacement of the radiosonde.

To further evaluate the simulated pyro-convection, we compare the simulated in-plume θv and U, with the radiosonde measurements during its ascent (black triangles; Fig. a and c). Additionally, we evaluate the simulated ambient conditions by comparing the observed θv during the descent of the radiosonde. The measurements near the surface (<0.3 km) during the ascent show an increase in θv (Fig. a). This increase coincides with the near-surface acceleration of the rising speed (Fig. b), which indicates that the increase in θv reflects the transition from the ambient atmosphere into the convective plume. Between 0.4–1.4 km, a relatively constant θv of 308 K is observed, which coincides with the observed convection column (i.e. where rising speed >2 ms-1). Above the well-mixed layer, an inversion is found (1.7–1.9 km) with a stable layer on top (>1.9 km), which explains the predominantly horizontal displacement of the radiosonde for 3 km between 1.5–1.7 km altitude (Fig. d). The inversion is also visible in the measured U profile inside the plume (Fig. c). The U is relatively constant (5–10 ms-1) up to 1.7 km above which it increases, indicating the transition from the atmospheric boundary layer into the free troposphere. The same inversion and stable layer are captured in the surroundings of the plume during the descent of the radiosonde (Fig. a).

We compared the observations with the median θv and U profile (blue line) inside the average simulated convection column (Fig. ). The medians of θv and U are determined per height inside the green-shaded area based on the averaged xz cross-section at y=12.8125 km. The spread in θv and U is indicated with the first and third quantiles (blue dotted lines). For consistency, the same approach is used to compare the measured and simulated ambient conditions surrounding the fire-induced plume, except for using the ref-run instead of the fire-run and the median based on the entire xz cross-section instead of the green-shaded area.

Figure b shows the median profile of θv inside the average convection column up to 2.3 km, which is the average height of the convection column between 19:00–20:00 UTC (Fig. d). At the surface (z=10 m), the median θv reaches 384 K (not shown), which quickly decreases due to mixing with the colder ambient air (304 K). This pattern is opposite to the sounding (black triangles), which shows a decrease in θv near the surface, because the simulated profile starts inside the flaming zone, whereas the radiosonde was launched outside the flaming zone.

Above the rapid decrease in θv, we find a well-mixed layer with a θv varying between 307–310 K in the simulation, which matches the observed mixed layer θv of 308 K (black triangles). However, the inversion capping the well-mixed layer is simulated ±0.3 km higher than observed (1.7 km). A similar overestimation is visible for the U profile. The simulated U is relatively constant up to 2 km altitude before increasing, indicating the presence of an inversion in the simulation at 2 km, while the observed U indicates an inversion at 1.7 km altitude.

The same difference in inversion height is present in the surroundings of the plume (Fig. a). The ambient conditions in the simulation are based on the ERA5 boundary conditions (Fig. ), which also overestimate the inversion height by 0.3–0.4 km. Hence, the inversion height mismatch between the observations and the simulation (Fig. a) is created by ERA5 and does not reflect the performance of MicroHH. The overestimation by ERA5 is consistent with measurements in the same region and period of the SCQ fire presented by . They showed a consistent overestimation of the convective boundary layer by ERA5 due to its inability to capture local surface heterogeneity. At night, this would result in an overestimation of the height of the residual layer, which we find in Fig. a.

Due to the overestimated inversion height, we conjecture that the simulation also overestimates the plume height. A higher plume could amplify the impacts of wildfire-induced pyro-convection on the surrounding kinematic and thermodynamic structure of the atmosphere. Nonetheless, it is not expected to significantly alter the outcomes of this study. The simulation setup was inspired by the SCQ fire to ensure atmospheric realism and to capture the fire-induced circulation. As we demonstrate below, the simulation successfully replicates a fire-induced circulation (Sect. ) as expected based on the observed surface wind reversal by the local fire fighters. Furthermore, the factors that govern the presence of a fire-induced circulation are located in the lowest 1 km above the surface (Sect. ), far below the observed and simulated injection height of 1.7 and 2 km, respectively. Therefore, while the simulation does not perfectly replicate the meteorological conditions observed during the SCQ fire, it suffices for the objectives of this study: (1) quantifying the impact of a fire-induced circulation on the thermodynamic structure of the atmosphere and (2) identifying the factors governing the development of a fire-induced circulation.

To explore the spatial impact of the simulated dry pyro-convection on the thermodynamic structure of the atmospheric boundary layer in the surroundings of the flaming zone, we calculated the boundary layer height (h) for the full 3D domain (Fig. ). We know the ambient atmospheric boundary layer between 19:00–20:00 UTC consists of a neutral boundary layer (0<z<0.5 km) topped by a warmer and drier residual layer (0.5<z<2 km), which in turn is topped by an even warmer and drier free troposphere (Fig. a and b). Hence, we calculated h using the minimum dqdz between 0.1–1.6 km altitude (Fig. ). The gradient in q identifies the interface between the relatively moist neutral layer and the drier residual layer (Fig. b), consistent with the mixed-layer framework described by . The altitude range is chosen to exclude the surface layer (z<0.1 km) and the capping inversion on top of the residual layer (2<z<2.5 km).

Figure  shows a decrease in h at the rear and the northern and southern flanks of the flaming zone (black dotted contour), whereas ahead (i.e. east) of the flaming zone, h increases within the first 2–3 km before it decreases. To investigate how these changes in h are reflected in the vertical structure of the atmosphere, we visualised the average vertical profile of θv between 19:00–20:00 UTC as a function of the eastward distance (xed; red hexagons in Fig. ) from the fire (Fig. ) for both the fire-run (orange line) and ref-run (blue dashed line). Here, the ref-run profiles represent the ambient conditions without a fire.

Directly above the fire (xed=0), the heating by the fire deepens the boundary layer from 0.5–2 km (Fig. b). West of the fire (xed=-0.8 km), we find a slight decrease of the neutral boundary layer height (Fig. a). East of the fire, two opposite effects become apparent before the impact of the fire on the atmospheric boundary layer structure becomes negligible at xed=7.5 km (Fig. e). Near the fire (xed=1.5 km), the neutral boundary layer deepens (Fig. c), while further eastward (xed=3.5 km) the neutral boundary layer height is lowered from 0.5 to 0.25 km (Fig. d).

To understand the mechanisms behind these thermodynamic changes of the atmospheric boundary layer (Figs.  and ), we analyse the modifications of the wind patterns by the fire in Sect.  and subsequently their connection to the thermodynamic changes in Sect. .

Figure  shows the vertical profiles of the zonal wind (u), meridional wind (v), wind speed (U) and wind direction (Udir) for the fire-run (a-d) and the two sensitivity experiments: fire-run:no-u-advec (e-h) and fire-run:pg (i-l). As intended (see Sect. ), the profiles show a decrease in directional wind shear due to reduced backing of the wind in the lowest ∼500 m. In both sensitivity experiments, the reduced backing is caused by an increase in u near the surface. The largest increase in u, and consequently the largest decrease in directional wind shear, occurs in fire-run:pg. The profiles of v and U in the sensitivity experiments remained relatively similar to the fire-run.

In the following sections, we investigate how these changes in wind profiles modify the fire-induced circulation (Sect. ) and subsequently whether these modifications alter the impact of the circulation on the thermodynamic structure of the atmospheric boundary layer (Sect. ).

Figure  shows the two distinct scenarios regarding the fire-induced circulation found in our LES experiments: one core simulation (fire-run) aimed at simulating the observed fire-induced circulations in realistic meteorological conditions and two sensitivity experiments designed to investigate the impact of the directional wind shear on the development of the fire-induced circulation. The core experiment produces a clearly defined circulation (Fig. a) comprised of three components: updrafts within the plume (1), downdrafts ahead of the plume (2), and reversed surface winds connecting the downdrafts and updrafts into a circulation (3). The sensitivity experiments generate the same three components; however, no coherent circulation develops (Fig. b). Despite the kinematic differences, all experiments produce identical thermodynamic impacts. Inside the plume (grey outline), all simulations show heating (red shading) near the surface and cooling at the top (blue shading). Ahead of the flaming zone, two opposite patterns are visible: cooling in the first 2 km followed by heating further eastward.

The circulation in our results (Fig. a) matches the observations by in both size and location. Based on Doppler measurements of the horizontal wind, suggested the presence of a persistent circulation ahead of the fire, which, similarly to our results, was observed in between the convective plume and downdrafts 2–3 km ahead of the fire. Furthermore, our simulations and the observations by both show downdrafts directed away from the circulation.

Regarding other simulation studies, our results align with those of , who simulated the same fire. They show a circulation of similar spatial extent ahead of the fire, despite using a different LES framework. Furthermore, the simulations of a different fire by also revealed a fire-induced circulation, albeit at the flank rather than at the head of the fire as seen in our simulation. Although the location differs from our simulation, the size of the circulation, approximately 1.5 km horizontally and 1 km vertically, matches our results.

Other observations and simulations of dry pyro-convection do not explicitly show a fire-induced circulation, but they do show the reversal of the surface wind. The reversed surface winds reported by these studies could indicate a fire-induced circulation (Fig. a); however, our sensitivity experiments show that reversed surface winds are not always part of a larger vertical structure (Fig. b).

In the sensitivity experiments, two factors changed simultaneously: directional wind shear decreases while zonal wind speed increases. We hypothesised that the directional wind shear is the primary driver of the circulation. However, based on the results, we expect that, contrary to our hypothesis, the zonal wind speed determines whether the circulation can develop for two reasons. Firstly, the directional wind shear decreased by 25 % and 50 % in the sensitivity experiment. If the directional wind shear were the primary driver of the fire-induced circulation, a proportional decrease in the magnitude and spatial extent would be expected. Instead, the circulation disappeared completely in both sensitivity experiments. Secondly, the main changes in the fire-modified wind patterns underlying the disappearance of the circulation were the reduction in spatial extent and magnitude of the reversed surface winds. The other two components of the circulation, the up- and downdrafts, were mostly unaffected in the sensitivity experiments. This reduction in reversed surface winds can be explained by the increase in zonal wind speed. The higher the zonal wind speed, the harder it becomes to create reversed surface winds, thus the smaller the area and magnitude of the reversed surface winds become.

This explanation also matches the purely theoretical discussion of dry convective plumes by , who start their reasoning with a simple case: a well-mixed atmosphere without ambient wind and a stationary heat source. Without wind, a plume rises vertically directly above the fire. To satisfy mass conservation, ambient air is drawn inward from all sides to compensate for this rising mass. This continuous updraft and corresponding lateral inflow trigger surrounding downdrafts, creating circulations on all sides of the fire illustrated from a 2D perspective in Fig. 2 of

When wind is present, argues the plume is tilted downwind, and with strong enough winds, the downdrafts and the reversed surface winds are disconnected from the flaming zone. With the sensitivity experiments, we capture this argument exactly. In fire-run:no-u-advec, the zonal wind speed increases to around 2 ms-1 (Fig. e); the horizontal and vertical extent of the reversed surface winds decreases, but they still reach the fire. In contrast, with even higher zonal winds in fire-run:pg (4–5 ms-1; Fig. i), the area with reversed surface winds is pushed eastward, causing a disconnect between the flaming zone and the reversed surface winds.

Besides providing a mechanism that governs the existence of the fire-induced circulation, our results also provide an explanation for the difference in location of the fire-induced circulation between our simulation (i.e. ahead of the fire) and the simulation by (i.e. at the flank of the fire). In the simulation of , the wind was predominantly directed in the direction of the fire propagation. Consequently, the wind speed was higher along the direction of the fire propagation than across the fire front, the exact opposite of our core simulation (fire-run; Fig. ). Hence, the formation of a fire-induced circulation at the flank in fits our explanation; the relatively low wind speeds over the flanks allowed for sufficient reversal of the surface winds, thereby causing a circulation.

Similar to , also shows the development of fire-induced circulations at the flanks, although their simulation includes pyro-cumulus formation, whereas our simulations and those of do not. Nonetheless, the moment at which the circulations are shown by coincides with the moment there is hardly any wind speed (∼0 ms-1) aligned with the circulation, thus supporting our reasoning that the wind speed component aligned with the fire-induced circulation determines whether a circulation can develop. However, contrary to our arguments above regarding the theoretical discussion by , argues that their findings do not align with as their simulation includes a significant wind speed. We agree with that there is significant wind (∼15 ms-1 at the surface) in their simulations, but predominantly in the direction of the fire propagation. The circulations, on the other hand, occur at the flanks, in the direction without wind, which exactly fits the results from our sensitivity analysis and the theoretical discussion by .

Although our simulations provide insight into the mechanism that governs the presence of fire-induced circulation, future studies using non-stationary fires are required to investigate whether the presence of fire-induced circulations impacts fire behaviour. For example, radar observations show that reversed surface winds, the surface component of fire-induced circulations, coincide with the formation of fire-generated vortices . This apparent correlation aligns with the operational experience of the Catalan Fire and Rescue Service, which suggests that reversed surface winds are typically associated with fire-generated vortices and long-range spot fires. To investigate these observations, two-way coupled fire-atmosphere simulations are required.

We studied the impact of the fire-modified winds (Fig. ) on the surrounding thermodynamic structure of the atmospheric boundary layer to investigate whether dry convective fires such as the SCQ fire can counteract the cooling and moistening of the atmosphere at night ahead of itself. Warming and drying would explain the continued burning throughout the night of the SCQ fire, the fire that inspired this study. We found no warming or drying of the atmosphere at the surface due to the descending motions (Fig. ). However, we did find two opposite changes in the boundary layer structure ahead of the fire (Fig. ). Within the first 2 km ahead of the fire, we found a 40 %–60 % deepening of the neutral boundary layer. Further ahead, between 2–4 km ahead of the fire, we observed a 50 % decrease in boundary layer height.

Although the boundary layer height is not directly linked to fire behaviour, it is connected to plume growth. Generally, a higher well-mixed boundary layer is beneficial to plume growth. This suggests that when a dry convective fire, such as the SCQ fire, increases the boundary layer height within 2 km ahead of the fire front, it makes it easier for a fire to maintain a convective plume while advancing. We know from observations e.g., that pyro-convection generally increases fire spread rates. Hence, we hypothesise that by modifying the boundary layer ahead, a fire can more easily sustain a convective plume, which subsequently promotes continued burning throughout the night, despite worsening burning conditions (i.e. cooling and moistening). The opposite applies when the boundary layer height is decreased; in that case, it becomes harder to sustain a convective plume, which is expected to decrease the ability of a fire to continue burning throughout the night.

This hypothesised impact on the fire behaviour only applies to fires that spread in the same direction as their plume, since that is the region where the boundary layer is modified. Hence, the hypothesis applies to the SCQ fire, since both the fire and plume spread eastward, which corresponds with the area where the boundary layer structure was modified in the simulation (Fig. ). Visual observations suggest that the hypothesis is a plausible explanation for the continued burning until midnight by the SCQ fire, as the fire not only burned until midnight (22:00 UTC) but also maintained a convective plume throughout the night (Fig. a). With this explanation, we assume that the SCQ fire would continually deepen the boundary layer ahead while advancing forward.

A further complicating factor for predicting the impact of the fire-modified boundary layer on the fire behaviour is the formation of a surface inversion throughout the night. The surface inversion is not yet present in our LESs as we focused on the period (19:00–20:00 UTC) surrounding the shift from a positive to a negative surface heat flux (19:41 UTC). Consequently, there has been insufficient time for the formation of a surface inversion. We expect that the formation of a surface inversion would limit the ability of the fire to maintain a convective plume and subsequently limit the extent to which the pyro-convection can alter the thermodynamic structure of the atmosphere ahead of the fire. Simultaneously, it is unknown how the fire-modified winds would affect the surface inversion ahead of the fire. Hence, to test the hypothesis, further studies using a two-way fire atmosphere coupling are needed to investigate the extent to which a modified boundary layer structure explains continued nighttime burning during dry convective fires such as the SCQ fire.