Numerical simulations are performed using version 4.6.1 of the Weather Research and Forecasting (WRF) numerical model . Numerical simulations are conducted in a single square domain with horizontal grid spacing of 1 km, consisting of 250 × 250 grid points. The same 64 vertical levels used in are employed, as they were shown to accurately reproduce severe storm dynamics in a real urban setting. Periodic Boundary Conditions (PBCs) are implemented. A sufficiently large domain is chosen to ensure that spurious wave interactions do not occur during the period analyzed in this study.

The simulations are initialized at 10:00 local time (LT), and the storm is triggered at 16:00 LT, a typical initiation time of convection (e.g., ). The supercell is initiated using a thermal bubble released at 16:00 LT near the bottom-left corner of the domain, at grid point (40,40). The bubble is a thermal perturbation of +3 K released at 1800 m a.g.l. The perturbation relaxes to zero with a vertical radius of 1800 m and a horizontal radius of 10 km. This setup allows the storm to develop for approximately two hours before reaching the city. This numerical configuration is similar to that used by to investigate interactions between supercells and idealized orography.

The microphysics scheme used is , with hail as the prognostic rimed ice species. Longwave radiation is parameterized using the Rapid Radiative Transfer Model , while shortwave radiation is parameterized using the Dudhia scheme . Radiative forcing corresponds to conditions at a latitude of 45° N (the same as Milan) on 31 May. Noah-MP is used as the land surface model , while the Revised MM5 scheme is employed to resolve the surface layer processes.

Several schemes are tested for the representation of PBL processes, coupled with the 1.5 TKE closure (km_opt = 2 in the WRF namelist), including the Boulac , Yonsei University (YSU, ), or Mellor–Jamada–Janjic schemes. Since no remarkable differences are observed among the PBL schemes considered, the YSU scheme is preferred because it is widely used in severe storm studies at mid latitudes and also in urban settings (e.g., ).

The domain is entirely flat, with a surface soil moisture

Lower values of soil moisture caused the development of an excessively dry PBL that suppressed deep moist convection.

A spin-up time is needed in the simulations to permit the generation of the urban boundary layer. During this spin-up time, deep moist convection is not desired. To achieve the desired controlled experiment, the Weisman–Klemp supercell sounding profile is used. Since the radiative heating during the day quickly triggers deep moist convection, an artificial thermal inversion of 4 K is added in the profile to coincide in height with the lifting condensation level (LCL, Fig. a). Through trial and error, an increase of 4 K is chosen as the threshold that completely suppresses the development of moist convection. The same issue was addressed similarly by and . At the LCL, the relative humidity is also reduced by a factor of 1.5, to resemble the presence of an Elevated Mixing Layer (EML, see Fig. 1 of ), a frequent feature during severe storm events both in the US and in Europe . The denominator 1.5 is chosen in order to decrease the mid-level moisture, without drying the troposphere excessively. A thermal inversion of 4 K is compatible with real EML profiles

Note that the thermal inversion and moisture decrease associated with an EML are usually located at an elevation higher than the LCL. We place them at the LCL because the daytime growth of the PBL lifts the hygrothermal discontinuity to a more realistic elevation.

The vertical profile of relative humidity rh(z) is defined as: 2rh(z)=1-34zztr34z≤zlcl[1-34zztr34]/1.5zlcl≤z≤ztr0.25/1.5z≥ztr

An exponent of 0.75 is used instead of the 1.25 employed in the original version of , in agreement with . The moisture in the boundary layer is defined by fixing a maximum value for specific humidity qbl = 14.5 g kg⁻¹.

Six hours after the start of the simulation, the solar heating deepens the PBL (Fig. b), eroding the thermal inversion and creating high Convective Available Potential Energy (CAPE) values, as expected by the conceptual model of . At this time, the convective inhibition is negligible, and convection can grow once a lifting mechanism is provided. In the previous 6 h, deep moist convection is suppressed by the thermal inversion, permitting the generation of the urban boundary layer.

The environmental wind profile is defined by the quarter-circle hodograph proposed by . The profiles of the zonal u(z) and meridional v(z) wind components used for the current study, and plotted in the hodograph of Fig. a, are defined as: 3u(z)=u1-u1cos⁡(πz2h1)-u24z<h1u1+(z-h1)u2-u1h2-h1-u24h1≤z<h234u2z≥h2 and 4v(z)=v1sin⁡(πz2h1)z<h1v1z≥h1 where h1 = 1500 m, h2 = 6000 m, u1 = 7 m s⁻¹, u2 = 40 m s⁻¹, v1 = 10 m s⁻¹ are chosen to produce a right-moving supercell ( used similar parameters). The diurnal mixing in the boundary layer and the surface drag changes the imposed wind profiles, as can be inferred by comparing the hodographs at 10:00 and 16:00 LT in Fig. . This is also a consequence of the lack of a meridional thermal gradient in the domain, which hinders the presence of a thermal wind capable of maintaining the curved hodograph in long-lasting simulations.

The transient evolution of the simulated profile suggests that the optimal timing for triggering the supercell should be carefully estimated. In the first part of the simulation thermodynamic and kinematic atmospheric parameters change (as shown in Fig. S1 in the Supplement); in particular, CAPE and LCL increase, Convective Inhibition (CIN) and Storm Relative Helicity (SRH) decrease. Between 17:00 and 20:00 LT, these parameters reach a minimum and a quasi-steady state, before nocturnal radiative cooling again significantly modifies the atmosphere. Therefore, triggering the storm at 16:00 LT and analyzing its interaction with the urban area between 17:00 and 19:00 LT is an appropriate choice. Even if at the beginning the supercells are moving into a slightly less favourable environment, any difference between the numerical experiments (that includes the city) and the no_urban simulations (Table ) observed in that time span can be reasonably attributed to the urban morphological parameters rather than to changes in the broader environment.

Values of selected parameters calculated from the simulated sounding at 16:00 LT, when convection is initiated, are shown in Table . At 16:00 LT, the atmosphere is conducive to supercell storms, since the substantial – albeit not extreme – Surface Based (SB) CAPE (2740 J kg⁻¹, virtual correction is considered), strong bulk shear 0–6 km (32.3 m s⁻¹) and moderate 0–3 km SRH (200 J kg⁻¹). The highest Θe in the low-levels is observed at the surface, suggesting that storms are surface-based with this profile.

The BEP-BEM model requires several urban morphological parameters to run. The thermal features of the city are kept constant: the default values (reported in Table ) of Local Climate Zone 2 are used . The sensitivity of the supercell to some of the BEP-BEM parameters is investigated in this work: street width (s), building width (b), building height (hb), and urban fraction (uf). In Table , a summary of the numerical experiments conducted is provided, while in Fig. , a graphical summary is furnished. The range of parameters explored is selected by studying the mean morphological parameters of some European and US cities frequently affected by severe storms (see, for example, Table S1 and Fig. S2 in the Supplement), using the GLobal Building heights for Urban Studies (GLOBUS) dataset . In that dataset, the plan area fraction (λp) and the building surface-to-plan area ratio (λb) are reported. These are defined as: 5λp=bs+b6λb=2hbs+b+λp. λp is the fraction of ground covered by buildings, it quantifies the building density in a city. Instead, λp quantifies the total urban area (roofs, walls, streets) exposed to the atmosphere.

First, a sensitivity analysis of city size is conducted, with city diameters ranging from 0 km in the no_urban scenario to 30 km (numerical experiments city size). Table  presents the parameters used to define the idealized city uniformly in the control runs. These are based on the Milan, Italy, GLOBUS data (Table S1). Milan, Italy, is chosen as the reference city both because it was object of investigation in a preceding case study analysis , and also because its morphological parameters are relatively typical of European cities. It is characterized by urban fraction of 0.63, mean building heights of 20 m and (λp,λb) = (0.24,0.78).

The same set of numerical experiments is also performed using an urban fraction (uf) of 1.0 instead of 0.63 (numerical experiments city size uf1, Table ). An urban fraction of 1.0 means the entire urban surface is built-up (i.e., no croplands within the city radius). However, 20 % of the street surface remains covered by trees in these experiments, as a city without any trees is unrealistic. Although a mean urban fraction of 1.0 is not observed in European or U.S. cities (Table S1), this experiment is conducted to explore the impact of city size under the extreme scenario of a fully urbanized city.

A third set of numerical experiments, called urban fraction, is conducted to examine the supercell's sensitivity to urban fraction. Values ranging from 0.4 to 1.0 are investigated for an idealized city with a diameter of 20 km (Table ).

The building height experiments, investigate the effect of building height on the simulated storm (Table ). The building heights explored range from 10 to 60 m, representing cities from the Chicago metropolitan area to downtown New York.

Furthermore, the building density experiments explore the effect of building density on the simulated storms (Table ). Since the BEP-BEM scheme uses street and building widths as input parameters (rather than the λ parameters), the building height and the building width are kept constant at 20 m, while the street width is varied from 106 to 11.5 m. This corresponds to a range of (λp,λb) values from (0.1,0.3) to (0.4,1.2), representing a transition from a very dispersed to a highly compact city.

Lastly, some additional experiments are conducted to investigate other specific questions. An experiment is carried out using the settings of numerical experiment uf1_30km, but with vegetation placed over all the buildings of the idealized city , to investigate the effect of this nature-based solution on storm dynamics (this experiment is called green roofs; see Table ). A second experiment, called no_thermo, is conducted with the settings of the uf1_30km experiment and the same urban roughness, but with the urban energetic balance replaced by the rural one. This experiment serves to disentangle the dynamical (i.e., building barrier-effects) and thermodynamics (i.e., UHI and UDI) effects of the city on the storm and the pre-storm environment. An experiment, called storm_shifted (Table ), is also run to test the sensitivity of the results to whether a different part of the urban plume is crossed by the storm. In this experiment, the thermal bubble is shifted from grid-point (40,40) to (30,60), while keeping the same parameters of experiment uf1_30km (Table ).

An ensemble approach is adopted to better assess the physical signal of urban-induced storm modifications across different flow realizations, given the intrinsic chaotic variability of the atmosphere. For each numerical experiment, eleven simulations are conducted: one is the control run, and the other ten are generated using the Stochastic Kinetic Energy Backscatter Scheme (SKEBS, ). SKEBS generates stochastic streamfunction perturbations through autoregressive processes in spectral space and injects them into regions where the model exhibits excessive systematic kinetic energy loss. In the present study, the parameters used to generate the ensemble with SKEBS are set following , who investigated ensemble dispersion in deep moist convective events. The only difference from their setup is the value of the maximum wavenumber of the potential temperature and streamfunction perturbation, which is set to match the number of grid points in the current simulations.

The ten ensemble members are generated using ten distinct random seeds, which are kept constant across all numerical experiments. As a result, the same members across different experiments are using the same seeds and are not fully independent (this is considered in the statistical analysis). This is done intentionally to better evaluate the influence of each parameter over the different flow evolutions (i.e., members).

The Statistical and Programmable Objective Updraft Tracker (SPOUT, ) algorithm is used to track the simulated updrafts. Fine-tuning of some user-defined parameters is necessary to achieve satisfactory tracking of the supercell updrafts. Table S2 lists these parameters and compares them with those originally employed by .

Further filtering of the SPOUT-detected tracks is necessary to isolate the right-moving supercell track, as the algorithm frequently splits it. This occurs because SPOUT tracks updrafts individually, and supercells may cycle, producing new updrafts.

To filter out these updrafts, only the track that begins near the location where the thermal bubble is released is considered. If the track terminates at x < 150 (i.e., before 3 h after the bubble release), a new track is searched for within a 10 km radius. If found, the new track is connected to the previous one. This process is repeated until the track reaches x = 150 or the storm dissipates.

Each point along the track is considered the storm center at a given time step. Storm diagnostics are extracted every 10 min, and the extracted variables are listed in Table . For each time step, the mean (or maximum or minimum, depending on the variable) is calculated within an 8 km radius around the storm center, and this value is assigned to the storm at that time step.

Then, three regions are defined using a temporal criterion to compare the storms at the same developmental stage across different numerical experiments. Those regions are defined with respect to a zero, defined as the time step when the supercell is above the city, corresponding to grid point (90,76) in the control run of the no_urban experiment. This time step is at 18:00 LT (see, for instance, Fig. d). Then, three 40 min windows are defined (sketch in Fig. ). The first one, from -20 to +20 min, defines the city region. The upstream region is defined from -60 to -20 min. Lastly, the downstream region is defined from +20 to +60 min. In these windows, the variables previously computed for each time step are averaged. Thus, considering the 10 min output frequency, four time steps are averaged for each window.

Given a set of numerical experiments, a statistical analysis is conducted for each region and for each variable. The variables are normalized with respect to a reference zero value (defined case-by-case in Table ). The normalized zij variable for the ensemble member j and numerical experiment i is given by: 7zij=xij-x0jx0j where xij is the original variable and x0j is the reference value for the ensemble member j. This approach is adopted to understand how storm diagnostics vary relative to a reference case (e.g., how updraft speed changes with increasing city size compared to the no_urban case). A linear fit is performed using the curve_fit function from SciPy on the normalized variables sharing the same SKEBS seed. Any null values are excluded: specifically, if a storm dissipates, the corresponding ensemble member is omitted from the fit. This exclusion is necessary because, for some variables (e.g., Θ), including a null value in the linear fit lacks physical meaning. In general, excluding such values has been observed to increase the fit's statistical significance.

We focus solely on the slope parameter of the linear fit, particularly its sign. The statistical significance of the slope relative to a null hypothesis of a zero slope is assessed using a 5 %–95 % bootstrap confidence interval, computed from 1000 iterations using the bootstrap function of SciPy.

In the numerical experiments city size, the effect of increasing the diameter of the city with the same characteristics as Milan on the supercell is investigated. The normalized mean value of the storm's UH in the city region is shown in Fig. a for each ensemble member as a function of city diameter (the same plots for the upstream and downstream regions are displayed in Fig. S3). In the city region, there is a clear decreasing trend in UH, even with considerable ensemble spread. A similar pattern is observed in the upstream region (Fig. S4a), whereas in the downstream region a more consistent decreasing trend is seen across all ensemble members (Fig. S3c).

Table  presents the mean slopes for each storm diagnostic variable in the three regions. A general negative trend is observed, though it is statistically significant only for some of them

Θ shows a positive trend. This is because the cold pools weaken as the storm weakens, leading to less pronounced negative Θ anomalies associated with downdrafts – i.e., a relative increase in Θ.

In the urban fraction experiments, most storm diagnostic variables decrease as a function of the urban fraction (Table ). However, only some of these trends are statistically significantly lower than zero in the upstream region. By contrast, no statistically significant trends are observed in the city and downstream region (except for Θe(500 hPa) in the downstream region). In the upstream region, the ensemble spread is relatively small (Fig. S5a), which leads to statistically significant trends – even though the slope of the mean linear fit is smaller than in the other two regions. Nevertheless, examining the ensemble mean of the UH swath in the uf0.4 and uf1.0 experiments (Fig. e, f) reveals a visible weakening of the storm. We conclude that the supercell weakens with increasing urban fraction, although this trend is not statistically significant lower than zero.

An opposite trend is observed in the supercell's sensitivity to building height. In all three regions, the mean UH increases with increasing building height (Figs. d and S6). The positive trend in mean UH is statistically significantly larger than zero across all three regions studied (Table ). Other mid-level storm diagnostic variables exhibit a similar positive trend; however, these trends are statistically significantly larger than zero only in the upstream and city regions for some variables (e.g., w500 and w700). In agreement with this result, the ensemble mean of the UH swath in the h60 experiment (Fig. b) is higher than in the h10 experiment (Fig. a).

In contrast, low-level storm diagnostic variables show a decreasing trend with increasing building height (Table ), and this trend is statistically significantly lower than zero for some variables (e.g., w950 in the city region). This result is surprising and interesting, as it contrasts with the storm behavior aloft and also with the hypothesis that taller buildings should enhance vertical motions due to the building-barrier effect and mass conservation (e.g., ).

Variables describing the storm's mid-level characteristics show a general weakening trend in the upstream and city regions with increasing building density (Table ). However, this trend is statistically significantly smaller than zero only for some variables (e.g., w500 and ζ500, Table ). The mean UH does not exhibit a clear trend with increasing building density (Fig. S7). Low-level storm diagnostic variables do not show a consistent trend. In general, in the downstream region, no clear trend is observed either aloft or near the surface, and none of the studied variables exhibit statistically significant trends.

The ensemble mean of the UH swath shows slightly lower values downstream of the city in the λb1.2 experiment (Fig. d) compared to the λb0.3 experiment (Fig. c). Overall, we conclude that the storm does exhibit some sensitivity to building density; however, this parameter appears to be the least influential in determining storm dynamics among the previously studied parameters.

Comparing Fig. c with Fig. e, it is evident that the weakening of the storm as it passes over the city is much less pronounced when the roofs are covered with vegetation (green roof experiment). Furthermore, using the rural energetic balance in the no_thermo experiment (Fig. f) results in a UH swath that is very similar to that of the no_urban experiment (Fig. a). Thus, the greater the amount of vegetation in a city, the less effect the city has on the approaching storm.

The uf0.63_30km simulations produce a weak UHI of about 1.5 K over the city (Fig. a), in good agreement with typical daytime UHI magnitudes (e.g., ). A pronounced UDI is also reproduced (peak of -2 g kg⁻¹), corresponding to a relative humidity difference of about 13 % between the city and the surrounding rural areas, consistent with several other studies . A wake of higher temperatures and lower humidity extends downwind (with respect to the southeasterly surface wind) up to 80 km from the city. Within this wake, a wind convergence is induced by the urban land use, while winds over the city are significantly reduced by increased drag. In the uf1_30km simulations, the UHI and UDI are slightly stronger, but the spatial pattern is similar (Fig. b).

The impact of urban fraction on UHI and UDI is much smaller than that of city size, about 0.5 K and -1.0 g kg⁻¹, respectively (Fig. c). The downwind wind convergence is also weaker. A more compact city does not significantly alter the UHI or UDI but does reduce surface winds over the city (Fig. e).

An interesting result is that taller buildings reduce the UHI by about 1 K and increase moisture by 0.5 g kg⁻¹ (Fig. d). Wind speeds are weaker in simulations with taller buildings, as well as the wind convergence downwind of the city. The reduction in temperature and increase in moisture in the urban canopy are likely due to reduced ventilation and consequent moisture stagnation . Our simulations begin with high moisture in the PBL, which decreases during the day due to mixing. In deeper urban canopies, drying is slower due to building shade, resulting in lower temperatures and higher moisture .

Finally, in the no_thermo experiment, the UHI and UDI are eliminated, as expected, while maintaining the wind reduction over the city (Fig. f). This experiment can be interpreted as preserving the city's dynamic perturbation on the atmosphere (the building-barrier effect) while suppressing its thermodynamic effect (the UHI and UDI). Overall, city size is the factor with the most decisive influence on the UHI, consistent with previous studies .

To understand the urban-induced flow perturbations, the sea level pressure field is investigated. A weak minimum (about -0.5 hPa) is generated downwind of the city (with respect to the surface winds) by the urban land use (Fig. a). The downwind pressure minimum coincides with the wind convergence. With a higher urban fraction, the pressure minimum is stronger (about -1.0 hPa; Fig. b), as also suggested by the comparison between the uf1.0 and uf0.4 experiments (Fig. c). On the other hand, with taller buildings, the pressure minimum is slightly weaker (about 0.2 hPa; Fig. d), while a more compact city plays only a marginal role in intensifying it (Fig. e). Lastly, in the no_thermo experiment, the pressure minimum is absent (Fig. f).

A relationship between the UHI and the mean sea-level pressure is evident: in simulations with a stronger UHI (e.g., uf1_30km experiment), the pressure minimum is also stronger, whereas when the UHI is weaker (h60 experiment) or absent (no_thermo experiment), the pressure minimum is weaker or absent. Thus, it appears that the pressure minimum is thermodynamically induced by the UHI (i.e., it is a consequence of the buoyancy and not of the dynamical modification of the flow by the buildings). To further support this hypothesis, an additional analysis is carried out. The dynamic pd′ and buoyancy pb′ pressure perturbations are computed for the ensemble mean of the uf1_30km simulations. The elliptic partial differential equations that define pd′ and pb′ are solved using a Gauss–Seidel iterative method. The reference p0 value is taken from the ensemble mean of the no_urban experiment. Figure  shows the results at the 10th vertical level (about 300 m a.g.l.). Buoyancy pressure perturbations pb′ clearly dominate the generation of the downstream pressure minimum, consistent with the previously noted sensitivity between the pressure minimum and the intensity of the UHI. The contribution of the dynamic pressure perturbations pd′ is negligible. This outcome is also valid for the h60 experiment (Fig. d, e, f): tall buildings have a marginal effect on the generation of the downwind pressure minimum with respect to the thermodynamic modification of the PBL induced by urban land use.

Surface Based Convective Available Potential Energy (SB CAPE) is markedly reduced by urban land use along a wake downstream of the city (Fig. a, b). The simulated storms cross the southern lobe of the urban plume with reduced SB CAPE (Fig. a). Note that the spatial spread of the supercell tracks over the city is about 5 km, and all tracks pass through areas with similar urban-modified environmental characteristics. The reduction in SB CAPE reaches up to 1000 J kg⁻¹ and coincides with the wake of higher temperature and lower moisture observed in Fig. . Urban fraction contributes to a decrease in SB CAPE of about 200 J kg⁻¹ (Fig. c), while taller buildings increase SB CAPE by up to 400 J kg⁻¹ (Fig. d). Building density has no significant effect on SB CAPE (Fig. e), consistent with the results previously discussed for other near-surface variables. In the no_thermo experiment, SB CAPE is unaffected by the presence of the buildings (Fig. f).

The same sensitivity of SB CAPE to urban morphological parameters observed for moisture suggests that the decrease in moisture more than offsets the temperature increase. Indeed, the Skew-T comparisons among the uf1_30km and no_urban experiments in Fig.  show that the drier, deeper and warmer PBL in the urban simulations (dashed lines) drives a reduction of CAPE area. While the UHI drives the generation of the downwind pressure minimum, the UDI drives the formation of the wake with reduced SB CAPE.

To assess the city's capacity to trigger convection, the differences in ensemble means of vertical velocity at 950 hPa are analyzed at 16:00 LT, before storm initiation. Upstream and over the city, no remarkable vertical velocities are induced by urban land use (Fig. a, b). In contrast, downstream of the city, a line of intensified vertical velocity extends up to 60 km from the city. In the uf1_30km experiment, the intensified vertical velocity reaches locally up to 1 m s⁻¹. This line coincides with the surface wind convergence. A higher urban fraction produces stronger vertical velocities (Fig. c), while taller buildings reduce their intensity (Fig. d). A more compact city slightly intensifies vertical velocities downstream (Fig. e), whereas vertical velocities are not produced in the no_thermo experiment (Fig. f).

The intensification of vertical velocities is related to the intensity of the UHI. A stronger UHI leads to a deeper pressure minimum, more vigorous surface wind convergence, and stronger vertical velocities. For example, in the simulation with taller buildings, vertical velocities are weaker (Fig. d) because both the pressure minimum (Fig. d) and the UHI (Fig. d) are weaker. In the no_thermo experiment, vertical velocities are absent (Fig. f) because neither a pressure minimum (Fig. f) nor a UHI (Fig. f) develops. Furthermore, building heights do not induce significant vertical velocities. In the no_thermo experiment 20 m tall buildings are present but there are no vertical velocities (Fig. f). Moreover, in the experiment with 60 m tall buildings (h60) vertical velocities are weaker along the convergence and more positive, i.e., closer to zero, where there is subsidence in the edges of the urban plume (Fig. d). This result explains why low-level vertical velocities in the simulated supercell do not increase with increasing building height, as discussed in Sect.  (Table ).

In the uf0.63_30km and uf1_30km experiments, SRH 0–3 km is reduced by up to 100 J kg⁻¹ (about 50 % less than in the no_urban simulations) south of the convergence line discussed in Sect. , and increased north of it with a similar magnitude (Fig. a). The increase (reduction) of SRH north (south) of the convergence line is due to the greater (smaller) veering of the wind caused by a stronger (weaker) surface easterly wind component (Fig. a) and a stronger pressure minimum. Thus, the modifications of SRH are ultimately induced by the pressure minimum generated by the UHI. An higher urban fraction amplifies this SRH pattern (Fig. c). In contrast, taller buildings cause the opposite pattern, i.e., higher SRH values south of the convergence line, due to the weaker pressure minimum. In the building density and no_thermo experiments, the SRH is almost unperturbed, except for slightly higher values over the city (Fig. e, f). The increase of SRH over the city results from reduced surface winds (Fig. e, f), consistent with the findings of . The spatial patterns in SRH 0–1 km are similar to those exhibited in SRH 0–3 km.

The thermal bubble that initiates the storm was positioned to steer the storm toward the urban center (Fig. ). However, the wind profile pushes the urban plume northwestward, so that the strongest urban perturbations are not located over the city, but downwind of it (Figs. , , ), as previously observed. Thus, the storm is not interacting with the most intense urban perturbations in the simulations discussed so far.

To verify the effect of a more direct interaction between the storm and the urban plume, the storm_shifted experiment was run, in which the storm is shifted 22 km northwestward and the city considered is 30 km wide with an urban fraction of 1.0. The results are displayed in Fig. a. In that case, the storm dissipates even more quickly, and the generation of new convective cells is observed northeast of the city. Hence, the results are not different from those obtained in the uf1_30km experiment. Even though the storm interacts with a stronger positive perturbation of SRH, the negative perturbation of CAPE is even more pronounced, causing the storm to dissipate faster.

Another interesting observation is that, in some ensemble members (e.g., Fig. b), storm dissipation occurs before the storm actually reaches the negative SB CAPE perturbation, raising doubts as to whether this is the actual mechanism driving the dissipation. Investigating the urban-induced perturbations along a cross-section taken along the storm's path (line AB in Fig. a) reveals that the area of low equivalent potential temperature (Θe) over the city, driven by the moisture reduction, extends upstream (with respect to storm motion, i.e., toward point A) at the top of the boundary layer (around 1500 m, Fig. b). In that area, an increase in potential temperature Θ is also present (Fig. b), associated with regions of subsidence (Fig. c). These features suggest enhanced inhibition (due to higher Θ), entrainment of more stable air (due to lower Θe), and subsidence at the top of the boundary layer upstream of the city; this likely promotes early storm dissipation. The early dissipation occurs only in a few ensemble members, where CAPE was already reduced in the no_urban experiment due to ensemble stochasticity. Hence, storm dissipation is primarily driven by the UDI, which reduces CAPE and Θe, but regions of subsidence associated with UHI-induced vertical motions may favour this process.