The Cloud Model 1 (CM1) version 21.1 is utilized in LES mode. CM1 is an idealized, fully compressible, non-hydrostatic, cloud-resolving numerical atmospheric model. The simulation starts at 09:00 local time (LT) (LT corresponds to the Beijing standard time, UTC+8), shortly after local sunrise, and runs until 20:00 LT (11 h). The first two hours of the simulation are considered spin-up for the model to achieve a reasonable atmospheric state, and are discarded from the analysis. The simulations are conducted under dry conditions without any microphysics scheme, and latent heat fluxes are excluded. Subgrid-scale turbulence is parameterized by the 1.5-order TKE closure from .

Advection and momentum transport are handled by the Weighted Essentially Non-Oscillatory (WENO) scheme , with improved smoothness indicators included . The scheme guarantees positive-definiteness and scalar mass conservation of passive tracers. For time integration, the Klemp–Wilhelmson time-splitting method is used, which keeps acoustic and gravity waves separated from the meteorologically relevant modes, thereby increasing computational efficiency.

The model domain is defined with a 50 m horizontal grid spacing and a size of 25km×25km×14km, consisting of 500×500×216 grid points. The vertical grid spacing is set to 20 m below 300 m altitude, then increases linearly from 20 to 100 m at 10 km, and then maintains constant spacing until 14 km. Previous studies have demonstrated that the CM1 model performs well for both idealized and semi-idealized setups when studying mountain-induced turbulence . The model applies a semi-slip condition at the lower boundary and a free-slip condition for winds at the top boundary, where the semi-slip condition allows tangential wind to respond to surface drag while still permitting some slip (i.e. the near-surface wind is reduced but not forced to zero). The lateral boundary conditions are periodic representative of extensive mountainous terrain.

To simulate the high ABL of 9515 m m.s.l. reported on 4 March 2008 , the model uses an idealized sounding as the initial condition: a 300 m surface mixed layer with constant potential temperature θ=306.8 K, followed by a stable layer with a weak lapse rate of 0.1 K(100m)-1 until the top of the domain. The weak stability above the initial mixed layer was chosen to represent specific pre-monsoon conditions over the TiP, where a more southerly jet position is associated with enhanced upper-level potential vorticity and a nearly dry-adiabatic free troposphere . The relative humidity profile is constant at 0 %, as well as the background wind speed profile (u=1ms-1). These initial conditions ensure favorable conditions for deep CBL growth. A Rayleigh damping layer is introduced above 11.2 km altitude to the top of the model domain, with a damping time of 300 s. The Coriolis parameter is set to zero.

The surface exchange coefficients are constant, based on Monin–Obukhov similarity theory using a prescribed surface roughness length of 0.01 m. The diurnal cycle of surface sensible heat flux is prescribed as follows: starting at zero at 09:00 LT, it follows a sine function peaking at 180 Wm-2 after six hours (15:00 LT) and returning to zero after 11 h (20:00 LT). These prescribed flux values are adopted based on the observed surface fluxes obtained from the authors of .

To study mixing sources and to help in boundary layer height diagnosis, two passive tracers are introduced into the model. One is a surface-based tracer emitted with a constant surface flux Jsurf.=4×10-5kgm-2s-1 with no initial concentration (c0,surf.=0). The other is an upper-level tracer, initialized with a mixing ratio of c0,up.=1000 ppm above 4 km height above plain level (APL) and has no flux (Jup.=0).

The study focuses on the western Tibetan Plateau near the Gerze station (32.17° N, 84.03° E; Fig. a). The selected region (Fig. b) is intended to represent the influence of heterogeneous terrain, with mountains rising by approximately 1400 m, on boundary-layer evolution. For early March reference period, local sunrise at this location occurs at approximately 08:47 LT. Terrain data were obtained from the Shuttle Radar Topography Mission (SRTM) at a resolution of ∼30 m , cropped to the model domain, reprojected to UTM Zone 45N, and interpolated to the CM1 Cartesian grid. To ensure smooth lateral transitions, the outer 4 km of the domain (outside the ellipse in Fig. b) were tapered to the minimum elevation of 4200 m m.s.l. (≡0 m APL) using a super-Gaussian mask (see Appendix ). A single Gaussian filter with σ=75 m was then applied (Fig. c) to reduce small-scale roughness while retaining the dominant terrain features. The resulting smoothed surface (Fig. d) serves as the basis for the simulations, including the two subdomains R-FLAT and R-TOPO used later in the statistical analysis.

The small-scale orography contains slopes up to about 36° and characteristic ridge–valley structures typical of the central Tibetan Plateau (Fig. a), with elevation differences up to ∼1.4 km and ridge widths of 5–10 km. Four experiments are conducted: REAL, which uses the realistic orography; FLAT, which represents an elevated plateau at 4200 MSL without local orography; and the corresponding experiments REALu10 and FLATu10, which include an imposed large-scale background wind. In REALu10 and FLATu10, the horizontal wind speed increases linearly above the mountain crest, reaching u=10ms-1, in order to introduce a non-zero plateau-scale flow comparable to that reported by their Fig. 5, while keeping near-surface winds weak to avoid triggering strong orographic waves during initialization. The base elevation in FLAT and along the smoothed edges of REAL and REALu10 corresponds to the minimum elevation within the realistic terrain (Fig. a). The full list of experiments and their descriptions is provided in Table .

Figure  shows the evolution of domain-averaged potential temperature (θ^) and u^-wind profiles. The θ^ profile shows the atmosphere heating from the surface, with a sustained unstable surface layer, followed by a deep and growing adiabatic, but statically unstable layer. The ABL height extends up to more than 9 km a.m.s.l. at 19:00 LT for all experiments, but slightly higher in REALu10. The FLAT wind profiles remain close to the imposed background wind, except within the lowest 0.5 km where surface drag reduces the flow. In contrast, the REAL experiments exhibit more nuanced wind structures due to the combined effects of surface drag, complex orography, thermal winds, plain-to-mountain circulations, and upper-level recirculation. Nevertheless, their late-day structure becomes broadly similar. The REALu10 and FLATu10 profiles are highly comparable: they begin with strong vertical shear, but downward momentum transport gradually produces a finite mean wind throughout the boundary layer and a shear layer across the entrainment zone.

However, the domain-averaged profiles alone do not clearly capture structural differences in the CBL. To address this, Fig.  depicts the spatial structure of convective flow at 15:00 LT, when the CBL is fully developed, showing the ensemble mean vertical velocity (u‾) and potential temperature (θ‾) fields, which reveal distinct patterns among the experiments. The spatial distribution of the thermally-driven flow differs slightly between the FLAT and REAL terrain configurations due to underlying surface heterogeneity. In FLAT (Fig. a and e), the quasi-stationary horizontal structures of updrafts (positive w‾) and downdrafts (negative w‾) are visible, reflecting the homogeneity of the surface. Convective thermals develop through turbulent interactions in the absence of fixed surface features, leading to temporally and spatially variable convective structures. In contrast, the thermal flow in REAL (Fig. b and f) is slightly more organized. The updrafts are aligned with orographic features, indicating a mesoscale flow signal. Elevated terrain serves as a main factor in the response to surface heating, and generates persistent, thermally-driven updrafts. This results in a more structured and spatially confined flow organization, with enhanced activity concentrated over and adjacent to mountainous regions.

In stark contrast is FLATu10 (Fig. d and h), where convection results in elongated streaks in the direction of the mean wind. These streaks reflect the formation of roll-like vortices, a property of shear-dominated turbulent boundary layers. The result is a transition from isolated quasi-stationary updrafts to a more continuous, shear-aligned convection structure. In REALu10 (Fig. c and g), the complex orography modifies the convection layout into na orography-modulated irregular shear-aligned convection.

Figure  shows the instantaneous turbulent flow field in terms of the vertical velocity perturbation w′ at 15:00 LT. The simulations share identical surface sensible heat fluxes and initial conditions, which leads to broadly similar turbulent structures across all cases. In the horizontal cross-sections (Fig. a–d) the turbulent field exhibits an irregular cellular pattern, with eddy structures ranging from a few hundred meters to over 5 km. In FLAT and REAL, the eddies are more or less randomly distributed throughout the domain, consistent with the prescribed homogeneous surface sensible heat flux. REALu10 and FLATu10 (Fig. c, d, g, and h) have slightly larger and more elongated horizontal eddy structures.

In addition to horizontal sections shown in Fig. , we examined instantaneous vertical velocity perturbation fields at lower levels (see Appendix ). At these heights (500 m a.g.l.), the turbulent structures appear as irregular, fine-scale cells. They are in contrast to the flow at 6 km a.m.s.l., which has larger, more coherent eddies and characteristic of fully developed convection.

We examine turbulence intensity in terms of turbulence kinetic energy (TKE), 5TKE=12(u′2+v′2+w′2)‾, where u2‾,v2‾,w2‾ are given in Eq. (): 6u′2‾=u2‾-u‾2v′2‾=v2‾-v‾2w′2‾=w2‾-w‾2

During the day, vigorous turbulence efficiently mixes heat, momentum, and scalars, producing a well-developed CBL. At midday (15:00 LT), the TKE profiles of FLAT and REAL are similar, while the shear cases (REALu10 and FLATu10) show systematically higher values throughout and above the ABL (Fig. a). Since all runs start from identical conditions, this enhancement is expected: the added wind shear directly contributes to turbulence shear production (Fig. b). An increased TKE above the ABL, particularly in REALu10, also reflects weak gravity-wave activity generated by overshooting thermals over the ridge. Such waves are a physically expected response in a deeply convective, weakly stratified environment and do not materially affect the diagnosed ABL structure. Shear also organizes turbulence into roll structures, in contrast to the isolated plumes seen in the no-shear cases (see REAL and FLAT in Fig. ).

Between 15:00 and 19:00 LT, turbulence is sustained and convective eddies deepen in all experiments (Fig. e–h). By 19:00 LT, each case develops a very deep, well-mixed boundary layer. An important distinction emerges in the late afternoon: despite the decline in surface sensible heat flux, TKE does not decay as quickly as in a typical CBL at lower altitudes, most notably in the REALu10 case. This sustained turbulence is consistent with the weak static stability in our simulations, under which mechanically and thermally driven production can maintain elevated TKE values even as the surface forcing weakens. Whereas TKE in classic CBLs normally peaks in mid-afternoon and then decreases as buoyant production diminishes see, e.g.,, the behaviour here instead indicates persistent thermally driven circulations under weak upper-level stability, consistent with the atmospheric structure described by .

Terrain plays a clear role in shaping turbulence. In FLAT (Fig. a and e), TKE exhibits the classic CBL structure, with maxima concentrated in the middle of the mixed layer. In REAL (Fig. b and f), terrain heterogeneity anchors convective eddies to slopes, producing a more continuous turbulent layer above the mountain that extends above 9 km a.m.s.l. by 19:00 LT.

The effect of added shear (FLATu10) and terrain forcing (REALu10) further amplifies turbulence and promotes vertical TKE growth (Figs. a and c, d, g, and h). The turbulent structures are also more spatially homogeneous than in FLAT and REAL.

By midday, convective turbulence mixes the surface-based tracer within the growing boundary layer, producing a well-mixed state (Fig. ). Columns of elevated tracer concentration coincide with persistent updraft cores (Fig. ), especially over complex terrain where quasi-stationary thermals continuously draw tracers upward. The localized tracer maxima and uneven mixing are more pronounced in REAL (second column in Fig. ) than in FLAT (first column), reflecting the orographic influence. In REALu10 and FLATu10, the added shear produces more homogeneous horizontal tracer fields (third and fourth column in Fig. ) because of the shear-organized rolls (Fig. ), leading to more uniform mixing. The vertical tracer extent in Fig.  provides a clear marker of the mixing depth, corresponding closely to turbulence maxima (Fig. ) and thermal perturbations (Fig. ). In REAL, tracer and turbulence extend higher over the ridge, while in REALu10 the tracer reaches greater heights over adjacent flat areas, consistent with shear-enhanced mixing. Entrainment from aloft further highlights these differences: in REAL, upper-level tracer mixes downward by 15:00 LT along the mountain flanks, whereas in FLAT entrainment is delayed until 17:00–19:00 LT and occurs more sporadically (Fig. i and m). In contrast, REALu10 and FLATu10 achieve more domain-wide and uniform entrainment and show the strongest entrainment fraction at 19:00 LT (Fig. m–p).

Tracer-based profiles and diagnosed ABL heights (Fig. ) confirm these contrasts. As shown in Fig. a and b, orography alone has only a small influence on the domain-mean ABL height (REAL–FLAT). By contrast, adding shear over flat terrain (FLATu10–FLAT) deepens the mean ABL height by roughly 12 % at 15:00 LT and 5 % at 19:00 LT, and similar increases occur when shear is added over real terrain (REALu10–REAL). The combined effect relative to the FLAT baseline (REALu10–FLAT) is about 13 % at 15:00 LT and 7 % at 19:00 LT, indicating that shear is the primary driver of domain-mean deepening, with orography providing a smaller but positive secondary contribution.

The upper tail of the ABL height distribution (99th percentile in Fig. ) reveals a much stronger orographic imprint. At the domain scale, REAL exceeds FLAT by about 10 % at 15:00 and 19:00 LT in the 99th percentile, whereas the shear-induced increases (FLATu10–FLAT; REALu10–REAL) are smaller, especially at 19:00 LT. The strongest orographic impact, however, appears in the REAL subdomains: at 15:00 LT the R-TOPO mean ABL height exceeds R-FLAT by about 18 %, and the 99th percentile is almost 20 % higher. By 19:00 LT, the differences decrease but remain distinct (5 % for mean; nearly 12 % for 99th percentile). Thus, while domain-mean metrics are dominated by shear, small-scale orography substantially amplifies the deepest convective columns, especially over the mountain, with implications for upper-level tracer exchange. These quantitative values are summarized in Table .

Although these results show that the ABL height (in m MSL) increases over the mountain, the ABL depth (in m a.g.l.) is actually shallower there. Figure  clarifies this distinction by relating the diagnosed ABL height to the underlying terrain elevation. Figure b shows that ABL heights increase with elevation in REAL, consistent with elevated tracer layers and terrain-anchored thermals. However, Fig. a shows that the ABL depth (a.g.l.) decreases with increasing terrain height, particularly in REALu10, because the ridge raises the reference plane from which depth is measured. This demonstrates that the influence of small-scale orography is more nuanced than implied by height (MSL) alone: the ABL can be deeper in an absolute sense over the mountains while simultaneously being shallower relative to the local ground surface.

Finally, our tracer-based ABL height detection method (described in Sect. ) proved robust: applying a 5 % concentration threshold yielded consistent and physically meaningful height distributions that aligned well with turbulence features (Fig. ). Importantly, similar distributions were obtained whether using instantaneous fields, time-averaged data, or ensemble means, indicating that the method is not overly sensitive to the choice of averaging.