WorksPhysics and Astronomy
2D and 3D unified RHD atmospheres and wind models of early B-supergiants including approximate NLTE effects
Abstract
B supergiants are evolved massive stars with low surface gravities and lower effective temperatures, with substantial radiation-driven winds. Generally, they have been studied assuming 1D spherical symmetry. We present the first multi-D unified atmosphere and wind models for B supergiants, especially investigating the effects of approximate NLTE ionization and subsurface turbulence on the atmosphere and wind dynamics. We computed multi-D box-in-a-star RHD unified atmosphere and wind simulations of 2D/3D early B supergiant models using a hybrid opacity approach. We calculated line opacities from approximate NLTE populations, including line scattering. We calculated the flux-, energy-, and Planck-mean line opacities separately from our large line database using Sobolev escape probabilities, assuming two-level atoms. The models develop a highly turbulent atmosphere with time-averaged turbulent velocities of ~ 40 - 55 km/s at the photosphere in the 2D simulations, and ~ 25 km/s in 3D. In the cooler B1, B1.5 models, the wind acceleration is more gradual than in the hotter B0 model, leading to a significantly higher maximum wind speed in the B0 model. Further, the approximate NLTE formalism strongly alters the ionization stratification in the wind compared to LTE, modifying the line force. Mass-loss rates obtained in the models are of order 10^-6 - 3 x 10^-6 solar masses per year. Including turbulent pressure in hydrodynamically consistent NLTE CMF steady-state 1D simulations with comparable parameters can produce higher mass-loss rates and lower terminal wind velocities than in corresponding models without turbulent velocity, in overall qualitative agreement with our multi-D simulations. Our calculations suggest that in the lower surface-gravity B supergiants, in addition to modifying the photospheric density scale heights, turbulent pressure also influences global properties of the wind.
Introduction
B supergiants (BSGs) are evolved post-main-sequence massive stars characterized by high luminosities and relatively low surface gravities ([10, 37]). Their atmospheres are coupled to strong radiation-driven outflows, resulting in substantial mass loss ([60] (Pauldrak et al. 1986)). Their effective temperatures, which are lower than those of O stars, place them in a regime where changes in ionization balance can effectively modify the line ensemble responsible for line driving ([35, 54, 10, 38, 44, 60, 70]). At their high luminosity-to-mass ratios, these stars are susceptible to dynamical motions associated with subsurface opacity zones ([8, 21, 78] (Ma et al. 2026)). Furthermore, observations show large non-rotational line broadening and spectroscopic variability ([44]) in both wind and photospheric lines ([72, 23, 13, 71]), pointing to a highly dynamical atmosphere and outflow. Moreover, photometric analysis using space-based instruments such as TESS ([61]) reveals a stochastic low-frequency variability (SLF) ubiquitously present in BSGs ([6] (Ma et al. 2024; Kourniotis et al. 2025)).
A long-standing uncertainty concerning line-driven winds in this regime is the so-called bi-stability jump. Early theoretical calculations suggested that luminous blue variables (LBVs) can have either a faster wind ( km/s) or a slow-moving wind ( km/s) ([53]). This transition was attributed to a change in the slope of the line-strength distribution, , around 20 kK (see [9, 18, 59] for the definition of ). This was later observed in terminal velocities around 21 kK ([36]). Monte Carlo and ionization calculations by Vink et al. [84] provided an explanation based on a change in ionization stage. As the effective temperature decreases, iron recombines from Fe-IV to Fe-III, increasing the number of driving lines, with a subsequent increase in the mass-loss rate [85, 86]. Subsequent studies have suggested that this transition in terminal velocity may be more gradual, with large scatter [10, 44, 24]. By contrast, the predicted increase in mass-loss rate at the bi-stability region has not been observed in recent empirical studies [13, 82, 2, 83]. Theoretical global NLTE wind models using the METUJE code [33] by Krticka et al. [34] do not predict an abrupt discontinuity in mass-loss rates and terminal wind speeds at the bi-stability, but rather a broader transition. The model grid by Björklund et al. [4], based on the dynamically consistent NLTE comoving-frame (CMF) method using FASTWIND by Sundqvist et al. [74], also does not generally predict an increase in mass-loss rate in that regime. However, recent hydrodynamically consistent model calculations by Vink et al. [87] using PoWR do recover a bi-stability jump. At higher Eddington ratios, the dynamically consistent OBA hypergiant PoWR models of Bernini-Peron et al. [3], which include turbulent pressure, show complex mass-loss behavior across the temperature regime. These global wind predictions, such as mass loss and terminal velocities, are also quite sensitive to changes in the ionization balance, which can alter the line ensemble. In particular, can change substantially depending on the dominant ionization stage [59, 55, 75]. Such ionization stratification is well established in NLTE calculations of Wolf-Rayet (WR) winds [41, 20, 64]. With their low effective temperatures, ionization stratification in the outflowing regions may become increasingly important in BSGs.
At the same time, previous multi-D radiation-hydrodynamical (RHD) simulations of O stars [28, 68, 15, 78, 16, 47] and quiescent LBVs [66] have shown significant velocity fields around the photosphere, except when very strong magnetic fields are present (ud-Doula et al. 2025). Such turbulent velocity fields provide additional pressure support, modifying the density stratification around the photosphere [15, 19, 81]. Recently, Ma et al. (2026) computed a 3D RHD envelope simulation of a BSG and found that instabilities associated with the subsurface iron-opacity bump generate a velocity field consistent with observed macroturbulent velocities [72, 71]. This trend agrees with previous multi-D RHD O star modeling efforts shown in [67, 15, 16, 47]. Similar to O stars, atmospheres of BSGs are tightly coupled to the outflowing line-driven wind, and both regimes should be evolved in a unified way. Furthermore, we include line scattering and approximate NLTE (aNLTE henceforth) effects while calculating the opacities that control both momentum and thermal coupling between gas and radiation.
We present the first multi-D unified atmosphere with line-driven wind models of early BSGs. In Sect. 2 we summarize the RHD setup; Sect. 3 introduces and discusses the models. Sect. 4 describes how the line-force parameters alter between the LTE and aNLTE treatment. Sect. 5 describes the velocity structure, mass-loss rates, and temporal variations. In Sect. 6 we compare averaged multi-D wind properties with hydrodynamically consistent steady-state 1D CMF models. Finally, Sect. 7 discusses observational implications, limitations, and concludes the work.
RHD formulation and model construction
Following Moens et al. [49], we solve the RHD equations on a finite volume mesh in multiple dimensions, using a Cartesian box-in-a-star setup including corrections for spherical divergence (Sundqvist et al. [76]; Moens et al. [49]) using the general-purpose hydrodynamic code MPI-AMRVAC2 (Xia et al. [89]; Keppens et al. [31]). We compute the mass, momentum, and both gas and radiation energy conservation equations, including effects of the radiation field in both momentum transfer and energy exchange. Gravitational force is included assuming a point mass. The system of equations is closed with the ideal gas law, assuming a constant adiabatic index and the flux-limited diffusion approximation (FLD, Levermore & Pomraning 1981; Moens et al. [49]). Following Debnath et al. [14], we separately calculate the flux-weighted mean that couples momentum transfer between radiation and matter and the energy-and Planck-weighted mean, which controls radiation-matter energy exchange. To evaluate the energy-exchange opacity means, we distinguish between line scattering and true absorption, taking aNLTE populations into account as outlined in Debnath et al. [14]. Line opacities are calculated using the Munich Atomic Database (Pauldrach et al. [52]; Puls et al. [59]), consisting of million lines for all the relevant elements (H-Zn) up to their eighth ionization stage.
For both the Planck-, energy-, and flux-mean opacities, we use the hybrid opacity technique introduced by Poniatowski et al. [56], where we obtain the total opacities by summing the Rosseland means obtained from OPAL tables (Rogers & Iglesias [62]; Iglesias & Rogers [26]) with the calculated line opacities in a supersonic medium, assuming the Sobolev (1960) approximation. Identical to Debnath et al. [14], for the energy and Planck mean calculations we subtract electron scattering from the OPAL opacities, as coherent scattering does not contribute to energy exchange (Mihalas [45]). We then fit the line-opacity multipliers using fit parameters first introduced by Gayley [18]. For the flux means, we follow Poniatowski et al. [55], and for the energy and Planck means we follow Debnath et al. [14]. We use metal mass fraction and helium mass fraction for computing both line opacities and Rosseland means. We kept the initial and boundary conditions identical to Debnath et al. [15] and Moens et al. [47]. For further details, we refer to Moens et al. [49], Moens et al. [48], Debnath et al. [15] and Debnath et al. [14].
We construct a small set of BSG models, namely B0, B1, B1.5, and B1, with the adopted stellar parameters summarized in Table 1. Each 2D model spans radially, and later-
| Model | Dim | [km/s] | [kK] | [/yr] | [km/s] | |||||
| B0 | 2D | 30 | 22.0 | 3.23 | 39 | 30.4 | 5.57 | 0.33 | 802 | |
| B1 | 2D | 30 | 32.2 | 2.89 | 55 | 25.6 | 5.60 | 0.35 | 470 | |
| B1.5 | 2D | 25 | 36.9 | 2.70 | 50 | 22.7 | 5.51 | 0.34 | 359 | |
| B1 | 3D | 30 | 29.5 | 2.97 | 25 | 26.5 | 5.59 | 0.34 | 419 |
Table 1. Fundamental parameters for the BSG models studied in this work
ally, the B0 model ranges over . Following Debnath et al. [15], we define as the radius where the radiation temperature kK. To ensure we adequately resolve atmospheric structures, we preemptively increased the lateral domain size compared to the previous O star simulations shown in Debnath et al. [15], Moens et al. [47] and Delbrœk et al. (2025), since the characteristic size of atmospheric structures generally correlates with the local scale height for O stars (Van der Sijpt et al. in prep.). The cooler BSGs have low surface gravity and extended envelopes; this may lead to a large atmospheric pressure scale height and larger structures. To allow structure to develop laterally, we have extended the lateral range to for the B1 and B1.5 models 3. For computational feasibility the 3D B1 model spans radially, and laterally. The 2D models resolve in both radial and lateral directions throughout. At minimum, one total pressure scale height is resolved by six grid cells, with similar resolution for the other 2D models. For the 3D model, we resolve to , for which at least one grid cell length resolves one total pressure scale height.
Characteristics of the early BSG models
After an initial relaxation phase, the models settle to a quasi-stationary regime similar to Moens et al. [48], Debnath et al. [15] and Moens et al. [47]. As in O stars, structure appears in deep subsurface layers around the iron-opacity bump (Jiang et al. [28]; Debnath et al. [15]; Van der Sijpt et al. [78]) and then propagates outward to the photosphere and outflowing regions, as shown in Figure 1 (left panel) for the B1 model. This creates highly turbulent layers around the photosphere. On average, the 2D models have higher laterally and temporally averaged photospheric turbulent velocities than the 3D model. Such high turbulent velocities result in turbulent pressure of the same order as radiation pressure in layers above the iron-opacity bump and the photosphere. Qualitatively, the radiation, gas, and turbulent pressure behave similarly to the O stars shown in Debnath et al. [15] and Moens et al. [47].

Figure 1. The left figure illustrates the volume rendering of the density for the B1 model. The right panel shows the emergent monochromatic optical continuum intensity at 5000 Å for four snapshots separated by 10 000 s, viewed from the local normal surface (). The colorbar shows the emergent radiation temperature calculated by . An animation of this simulation is available at the following online video.
Following Delbroek et al. [16], Figure 1 (right panel) plots the emergent monochromatic continuum intensity at 5000 Å for a few representative snapshots for the B1 model. The emergent specific intensity, , is computed by solving along rays perpendicular to the stellar surface (), assuming a source function given by the Planck function , where is the local gas temperature. We use the local Rosseland mean to compute the opacities, . The emergent intensities illustrate layers around the photospheric optical depths of order unity. We then express the radiation temperature, as . Figure 1 displays spatial variations of the emergent intensities with large-scale bright and dark patches, with the local radiation temperature varying by a few thousand Kelvin around the mean. However, compared to the O4 model in Delbroek et al. [16], the structures here are much larger, with typical scales visible in the figure of order .
Due to their higher turbulent pressure, the 2D models are more inflated than the 3D model. The photospheric radius is smaller in the B1 model compared to the 2D B1 model. As a result, the 3D model has a higher effective temperature than the 2D model. However, the 3D simulation has a smaller spatial resolution than the 2D model. Previous resolution tests in 2D have shown that atmospheric inflation slightly decreases as the resolution is increased. This suggests that a higher 3D resolution would likely exhibit a slightly smaller photospheric radius.
At maximum, the radiative enthalpy (radiation advection) (see Debnath et al. [15], for definition) carries only 10%, 20%, and 23% of the total energy for the 2D B0, B1, and B1.5 models, respectively. Like its 2D counterpart, the B1 model also carries a maximum of 20% of the total energy by radiative enthalpy, which is of a similar order of magnitude as the O supergiant models in Moens et al. [47]. This is much lower than in transitional cool-plate models with higher luminosity-to-mass ratios in Schillemans et al. [66], which find that radiation enthalpy carries of the total energy at maximum. For both 2D and 3D B1 models, the region where radiation enthalpy contributes at least 1% of the total flux extends from 230 kK to 80 kK. This is somewhat broader than for the O4 model in Debnath et al. [15] and Delbroek et al. [16], which extends 220–130 kK.
Figure 2 displays the laterally and temporally averaged density profile for the B1 model, along with the initial 1D conditions and a 1D model calculated from 3D-informed turbulent velocities and radiative enthalpy transport. To obtain a shallower density stratification in the inner layers around the iron-opacity bump, following Debnath et al. [15], we included convection using a simple Schwarzschild [69] criterion with mixing-length , since it provides a better fit to 3D data compared to previously used (Moens et al. [47] for O stars, together with a turbulent pressure support characterized by a turbulent velocity km/s. We kept constant throughout the photosphere and subphotospheric layers until kK. At higher temperatures, we gradually reduce the turbulent velocity until kK, beyond which a substantially stronger inward taper is applied. As in the 3D model, the calibrated 1D model shows turbulent pressure exceeding gas pressure in the subphotospheric layers. Although radiation pressure remains the dominating contributor in magnitude, turbulent pressure provides a substantial additional contribution to the atmospheric pressure stratification. Further, in layers around the iron-opacity bump, the 1D model shows convection carrying of the total energy flux at maximum.
Figure 2. The figure displays the temporally and laterally averaged 3D density as a function of , along with the initial 1D structure and the modified density structure obtained using a turbulent velocity km/s inferred from the 3D simulation and . The vertical dashed line indicates the effective temperature of the star at the photosphere.
In the dense atmosphere, and are approximately equal to each other. Further out, and begin to differ, and the gas and the radiation temperatures begin to decouple. On average, in these regions, the gas stays cooler than the radiation temperature, as shown in Figure 3 for the B1 model. However, small pockets of hot gas are present where the material is compressed at the shock fronts, heating the gas to temperatures higher than the local radiation temperature. The volume of gas in the outflowing regions that stays hotter than the local radiation temperature is 3%, and 4.9% for the B1 and B1 models, respectively. However, for both B1 and B1 models, the volume of gas that stays above the effective temperature of the star is 0.2%, which is lower than that of the O stars shown in Debnath et al. [14].

Figure 3. Probability density distribution of the gas temperature, with yellow indicating high probabilities and blue vice versa for the B1 model using one hundred 3D snapshots. The black solid curve and red dashed curve indicate the temporally and laterally averaged gas and radiation temperature, respectively. The orange vertical line shows the average photosphere.
Ionization and line-force parameters
The line-driven radiative acceleration depends on the ionization state of the gas ([9, 18, 59, 64, 55, 75]). The aNLTE formalism modifies the ionization stratification, which alters the wind dynamics. In this section, we discuss the effects of changes in ionization on the line ensemble.
Line-strength distribution
The aNLTE treatment modifies the level populations and thus the line-force parameters entering the flux-mean opacity and radiative acceleration (see Debnath et al. [14]). To isolate the aNLTE effects, we calculate reference LTE populations using the local gas density, , while imposing . Using from aNLTE would simultaneously introduce two effects, namely, modifications to the level populations and the thermal structure. In aNLTE, the level populations depend on both and , , and the dilution factor. For details, we refer to Debnath et al. [14] and references therein. In such a parameterization, the line-force multiplier is fitted using three parameters. provides its maximum value in the optically thin limit. sets the upper cutoff of the line-strength distribution. determines the slope of the power law ([18, 55]).
The 2D and the 3D models produce qualitatively similar line-force parameter behavior over the overlapping radial domain. However, due to its larger radial extent, we focus on the 2D model in the following analysis. We find that near the photosphere of the B1 model (), the line-force parameters behave qualitatively similarly between the two methods; farther out, they differ. In dense wind pockets, becomes substantially larger in the aNLTE treatment; simultaneously, decreases and increases. Thus, aNLTE changes both the effective strength of the line ensemble and the shape of the line-strength distribution through and . The dynamical importance of the line-force parameters also changes depending on the local characteristic depth ([55, 75]). In the force-multiplier formulation, enters through the combination , informing whether the line ensemble is approaching the optically thin or saturation limit. Guided by this, we examine the laterally averaged for the 2D B1 model in Figure 4. At the photosphere, remains large in both methods and naturally decreases at larger radii, although values obtained for the present method are smaller than LTE. Nevertheless, over the entire computational domain we find that , indicating that the wind on average does not approach the effective optically thin limit, ([73, 9, 18, 55]). However, applies only to the average structure; in local pockets of rarefied gas, may become small. Further, we point out that not every contributing line is optically thick. A line of strength with Sobolev depth would be optically thin when and saturated when . On average over the entire computational domain, the ensemble remains in the optically-thick regime.

Figure 4. Laterally and temporally averaged as a function of radius for the B1 model. The LTE and aNLTE calculations are shown by the blue dashed and red solid curves, respectively.
To understand how changes in and affect the line force, we examine how different line strengths contribute to the total force multiplier. Following Gayley [18]; Poniatowski et al. [55], the statistical distribution of the lines can be written as
Here, provides a standard statistical frequency distribution of lines ([17, 59, 55]), and is the Gamma function ([18, 55]). Also,
Eq. (1) describes the number of lines within a given line-strength and frequency interval. The contribution from all lines to the force multiplier within some is,
where is the contribution from a single line. Substituting the value of the line distribution from Eq. (1) and the force multiplier shown in Poniatowski et al. [55] and then integrating Eq. (2) over frequency, we arrive at the contribution to the force multiplier per logarithmic interval in line strength,
Eq. 3 describes how lines within a logarithmic line strength interval contribute to the total force multiplier. When and ,
As , the contribution from strong lines to the force multiplier increases. In the opposite optically thick limit, as and ,
and with , the contribution to the force multiplier decreases with increasing line strength. The force contribution therefore peaks between the optically thin and thick transition, with typically in orders of unity. The exact peak value, however, depends on and the exponential cutoff . Once a line is highly saturated, its contribution only weakly depends on line strength; therefore, extending the line distribution towards larger will not significantly increase the force multiplier.
Figure 5 shows the distribution of across different line strengths for various radii for a laterally averaged B1 snapshot. The dotted line shows the line strength at which , corresponding to Sobolev optical depth . Lines to the left of the dotted line have and are optically thin, whereas lines on the right have and are optically thick. Near the photosphere, the left panel of Figure 5 (at ) shows that LTE and aNLTE distributions are almost identical in shape and magnitude, agreeing with our previously found identical line-force parameters in these regions. At and , these low- and intermediate-strength lines contribute substantially more to the total force multiplier in aNLTE than in LTE, which in turn produces a larger line-force. At the same time, aNLTE at extends to a smaller high- value than LTE. However, these high- lines contribute almost negligibly to the total force multiplier. Therefore, the present aNLTE treatment in the outer wind shifts the effective line-strength distribution to a larger contribution from lower- and intermediate-strength lines, resulting in an increased total integrated line force.

Figure 5. The contribution of the logarithmic line-strength interval to the force multiplier at various radii for the B1 model for a laterally averaged singular snapshot. The LTE and the aNLTE distributions are shown by the curves indicated in the legend. The vertical dotted line shows the median value of , for which the Sobolev optical depth is unity, and the shaded regions show the lateral variations for a snapshot.
Ionization stratification
The differences in the line-strength distribution can be understood by examining the ionization structure in the wind. Figure 6 compares the laterally averaged iron ion fractions for a single snapshot with the overlapped line-force parameters and for LTE and aNLTE formalisms. A qualitatively similar ionization stratification is found in the B1 model; however, as discussed earlier, its more limited radial extent does not allow us to capture the full transition in the outer wind, and thus we focus on the B1 model.

Figure 6. Colors showing laterally averaged Fe ion fractions for the B1 model for a single snapshot. The ionization stages are indicated in the labels. The left and right columns show the LTE and aNLTE calculations, respectively. The red dashed curves show the corresponding laterally averaged , and is shown in black solid curves.
As can be seen from Figure 6 in LTE, Fe-iii remains the dominant ionization stage throughout the wind. Similarly, in the present method, Fe-iii dominates around the photosphere () and the wind-launching region; however, further out, it recombines to Fe-ii. Consequently, in LTE, is substantially lower in the outer regions, whereas in the present aNLTE method stays large throughout the outflowing regions of this simulation domain. The regions where and diverge between the two methods coincide with the region where we find substantial recombination to Fe-ii in the present method, indicating that recombination to Fe-ii could alter the line ensemble, resulting in different and values for the present method. Although CNO ions also undergo significant recombination in the aNLTE, it occurs further inwards compared to where the line-force parameters diverge.
Atmospheric and wind dynamics
We now examine the dynamical properties of the BSG outflows, focusing on the velocity structure, mass-loss rates, and their temporal variations.
Velocity structure
In Figure 7, we show the laterally and temporally averaged and density-weighted radial velocities for the BSG models. The hotter B0 model has a rapidly accelerating wind. The mean radial velocity rapidly reaches a few hundred km/s outside the photosphere. The wind is still accelerating at the edge of the computational domain; however, the acceleration is much more gradual compared to the initial steep rise in velocity. On average, the wind reaches velocities up to 700 km/s. In contrast, both cooler B1 and B1.5 models develop a qualitatively different velocity structure. The wind accelerates more gradually than the B0 star and reaches much lower maximum wind speeds, only a few hundred km/s at the edge of the simulation domain. This may indicate a transition from a fast B0 wind to a substantially slow-moving outflow in the cooler B1 and B1.5 stars. The B1 model has a faster acceleration at the base of the wind; at maximum it reaches km/s, similar to the B1 and B1.5 stars. This slowly accelerating wind may be related to the classical change in wind velocity across wind recombination zones [53, 84], or perhaps to a different solution branch that under certain circumstances can be obtained within modified CAK wind theory [11, 12, 79, 80]. In any case, the emergence of our slower winds suggests an origin related to ionization changes in the wind. Interestingly, unlike its 2D counterpart, the B1 model does not exhibit a pronounced negative mean radial velocity around the photosphere; rather, the mean remains closer to zero, indicating that the strong photospheric infall found in the 2D models is significantly reduced in 3D. Additionally, Figure 7 also shows the systematic difference between the straight averages and their density-weighted counterparts. The density-weighted velocities remain below the ordinary mean velocities throughout the outflowing regions. The relatively low density-weighted velocities indicate that a large fraction of wind mass is slowly accelerating.
Figure 7. Dashed curves present laterally and temporally averaged radial velocities, and solid curves denote density-weighted radial velocities, as a function of over 60 snapshots for various models. The vertical lines indicate the average photosphere at optical depth .
Figure 8 shows the radial velocity distribution, with colors representing the density contrast for the B1 model. Although the mean flow, shown with black curves, in the outer line-driven regions moves at km/s, the rarefied medium moves much faster, reaching velocities exceeding 1000 km/s. This anti-correlation of line-driving with density has also been observed in previous multi-D simulations of both WR stars [48] and O stars [15]. The rarefied medium reaches escape velocity at a much smaller radius than the mean flow, which only reaches escape velocity beyond . In contrast, the 2D O4 model presented in Debnath et al. [15] shows that the mean flow already reaches escape speed at . In the present B1 models, the dense wind component barely drags along over an extended radial range, while the fast-moving material may have already attained escape velocity. Even at the outer edge of the simulation domain at , the dense medium stays much below the local escape speed. For the larger radially extended 2D B1 model, even at the dense material moves at velocities lower than the local escape velocity. By contrast, in the O4 model in Debnath et al. [15], at the outer edge of their simulation domain at , even the dense medium attains the local escape speed.

Figure 8. Radial velocity distribution of the B1 model as a function of , using 100 snapshots. Colors represent the density contrast, with yellow representing the high-density medium and blue the rarefied parts. The red dashed-dotted curve is the local escape velocity. The orange line is the laterally and temporally averaged photosphere . The black solid line is the laterally and temporally averaged density-weighted radial velocity, and the blue dash-dotted line is the laterally and temporally averaged radial velocity. The magenta dashed line is the turbulent velocity calculated by subtracting the density-weighted mean radial velocity, whereas the purple dotted line is constructed by subtracting the simple mean from the radial velocity. The cyan star provides the effective sonic point, , calculated using , and the golden cross provides the effective sonic point, , obtained using . The inset provides a zoomed view of the inner regions.
We further examine how the ionization stratification discussed in Sect. 4.2 relates to the dynamics of the wind. Figure 9 shows the velocity distribution of the B1 model for a single snapshot, with colors showing the ionization fractions of Fe-II, Fe-III, and the total line opacity for aNLTE and LTE. As shown in the middle panel, the present treatment qualitatively restructures the iron ionization balance throughout the wind. Compared to the reference LTE model, the present method is strongly recombined to Fe-II, especially in the slower, denser component. As discussed previously, ionization stratification modifies the line ensemble, clearly visible in the resulting effective line opacities. The comparison between the aNLTE and LTE line opacities shown in the bottom-right panel of Figure 9 illustrates this difference in the denser regions, with the being relatively higher in such regions due to the restructuring of the ionization species in the present method. Similar to the 2D model, the B1 model shows similar qualitative radial behavior, with Fe-II also becoming increasingly important in the outer wind. Due to the smaller radial domain, the region where Fe-II becomes important is less extended in the B1 model.

Figure 9. The plots show the distribution of the radial velocities as a function of radius for a single snapshot for the B1 model. The asterisk here denotes the LTE calculations. The colors represent different quantities. In the top panel, from left to right, the LTE , , and the total line opacity in LTE, . The middle row from the top, from left to right, provides the aNLTE , , and the total line opacity in aNLTE, . The bottom row, from left to right, compares , and . The black curve provides the laterally averaged radial velocity.
Mass-loss rates
Using a few radial cells at the outer edge of the simulation domain, we calculate a temporally and radially averaged mass-loss rate for the models presented in this work, as listed in Table 1. Comparing the mass-loss rates with the mass-loss fitting formula provided by Björklund et al. [4], we find a mass-loss rate
of , , , and M/yr for the B0, B1, B1.5 and B1 model respectively, which are lower than in the 2D and 3D models presented in this work. To match the metallicity adopted in our multi-D simulations, we evaluate the Björklund et al. [4] mass-loss prescription at .
We obtain similar mass loss predictions from LIME4, where we find , , and M/yr for the B0, B1, B1.5 and B1 model, respectively. In LIME, the continuum contribution to the radiative acceleration is represented using electron scattering, while line acceleration is calculated using a force multiplier formalism under the Sobolev approximation (Gayley [18]; Poniatowski et al. [55]). However, around the photosphere of our models, the total continuum contribution is substantially higher than the electron-scattering opacity, resulting in a larger reduction of effective gravity due to continuum opacity sources. Accounting for the total continuum opacities, we obtain the photospheric continuum Eddington factors of approximately 0.45, 0.47, 0.52, and 0.52 for the B0, B1, B1.5, and B1 models, respectively, significantly exceeding the corresponding electron-scattering Eddington factor, listed in Table 1.
To approximately account for this enhanced continuum radiative support in LIME, we increase the luminosity-to-mass ratio so that in LIME matches the corresponding photospheric total continuum Eddington parameter obtained from the multi-D models. The resulting mass-loss rates increase to approximately , , , and M/yr for the B0, B1, B1.5 and B1 models respectively. The revised prediction matches the B0 model, but remains lower by factors of approximately 3.5, 2.5 and 3 for the B1, B1.5 and B1 models, respectively. Another ingredient that may contribute to such discrepancies is the turbulent pressure associated with substantial velocity fluctuations around the photosphere. This is neither included in the Björklund et al. [4] prescription nor in LIME. We investigate possible implications of such inclusions in Sect. 6.
Temporal variations
In simulations presented here, both regions around the photosphere and the outflow are temporally dynamic. We plot the laterally averaged radial velocities for various radial depths for the different models in Figure 10. The plot only displays data after the models have relaxed from the initial conditions. The B0 star already has high velocities near the stellar photosphere. Further out, the B0 model on average varies between 600 and 1000 km/s. For the B1 model, around the photosphere (1.5 or 3 in Figure 10), shows even more significant temporal variations. The laterally averaged velocity experiences intervals of small outward velocity of a couple of hundred km/s, followed by almost negligible average outflow speeds. Further out, wind speeds typically range from 100 – 400 km/s. Qualitatively, the B1.5 model behaves similarly to the B1 model. The 3D model also exhibits substantial temporal variability; at 3 (or 1.8 ) the velocity fluctuates approximately between 200 and 400 km/s, whereas at 5 (3 ) they typically range between 300 and 450 km/s. Compared to the 2D B1 model, the 3D model shows a more persistent outward flow in the inner wind, without extended time intervals of nearly zero mean radial velocities. Nevertheless, both B1 and B1, B1.5 models exhibit time-dependent, slowly accelerating outflow compared to the hotter B0 model.
Figure 10. Laterally averaged radial velocities for three different radii (, and ) shown as a function of time in days. For the B0 model, these radii correspond to approximately , and , and for the B1 model, these correspond to approximately , and , and finally for the B1.5 model, these correspond to approximately , and . For the B1 model, it corresponds to approximately and . The colors here represent the different models shown in the legend.
We further investigated temporal variability in the radiative luminosity at 3 for the models presented in this work. We constructed a normalized light curve and found the root-mean-square (RMS) flux variations were 9.54%, 8.56%, 9.85% and 7.79% for the B0, B1, B1.5, and B1 models, respectively. We also calculated Lomb-Scargle power spectra (Lomb [40]; Scargle [65]) of these luminosity variations and found an excess of power at low temporal frequencies in all our models. Frequencies below which 80% of the total power remains are approximately 28, 11, 6 and 23 Hz for the B0, B1, B1.5 and B1 models, respectively. For the 2D set of models, we find that power is concentrated at lower frequencies as we move towards the cooler models. In contrast, the 3D model shows a broader power distribution toward intermediate frequencies than its 2D counterpart.
Comparison to hydrodynamically consistent 1D models
As discussed in Sect. 5.2, the mass-loss rates from the 2D and 3D calculations are substantially higher than the ones calculated using the Björklund et al. [4] prescription. However, the underlying models used to create this prescription didn’t account for the high turbulent velocities that arise in the multi-D simulations of these stars. Motivated by this, we calculate new 1D, stationary, hydrodynamically consistent wind models following the method introduced by Sundqvist et al. [74]. These models use the 1D atmosphere code FASTWIND (Puls et al. [58]) to solve NLTE radiative transfer in CMF to obtain the radiative acceleration, using the same Munich Atomic Database (Pauldrach et al. [52]; Puls et al. [59]) as for the Sobolev-based line-force calculations in our multi-D simulations, but we adopt the standard solar metallicity (Asplund et al. [1]). We iteratively couple the calculated radiative acceleration to the stationary, spherically symmetric equation of motion, allowing us to compute both the mass-loss rate and velocity structure self-consistently. Introducing turbulent pressure in the equation of motion, (Jiang [27]; Debnath et al. [15]; Moens et al. [47]) the equation of motion is modified as (see also, Verhamme [81]),
where, is the gas pressure, with as the gas sound speed, is the radiative acceleration, and is the gravitational acceleration. We can rewrite the equation using an effective sound speed such that (see also, Kee et al. [30]),
Figure 11 shows the density and velocity structure obtained using the method described above for the B1 model parameters with km/s and without it. The model with km/s shows a much shallower density stratification than the model without any . The outcome mass-loss rates for the model with km/s is . For the model without any turbulent velocity, the mass-loss rate is , which is close to the mass-loss rate obtained using the Björklund et al. [4] prescription, with a small difference reflecting the inherent scatter that exists when fitting these models to translate them into an analytic prescription formula and the smaller used for this scaling test. Figure 11 also demonstrates the difference in the velocity structure, with the unmodified model showcasing a large radial velocity with terminal wind speeds km/s, whereas the model including a large turbulent pressure shows a more gradual wind acceleration comparable to the B1 model presented in the paper. The terminal wind for this model is merely km/s. The global wind properties of this CMF simulation thus resemble the corresponding 2D B1 model well.
Figure 11. Density and radial velocity structure of the hydrodynamically consistent FASTWIND models as a function of the modified radius. The colors indicate the two models, including a km/s model and one without any turbulent velocities.
However, computing a CMF model using the average 3D model parameters yields a somewhat different result, as for this model km/s. In the CMF models, this results in an in-between case yielding and km/s, whereas in the 3D simulation we still obtain the higher mass-loss rate and lower wind speeds characteristic of the 2D models with higher turbulent velocities. A likely cause for this discrepancy is the large increase of in layers just above the photosphere observed in the 3D model, resulting in values higher than 25 km/s in regions where the line-driven wind is launched, especially when accounting for the fact that the high-density material that carries most of the total wind mass gets accelerated in significantly higher layers than the mean wind (Moens et al. [48]; Debnath et al. [15]). However, first tentative attempts to include a radially increasing also in the stationary 1D CMF models have not resulted in a significant increase in mass loss (and corresponding lower terminal wind speed), as the radial dependence also induces a counteracting force component due to the turbulent velocity gradient. This difference in sensitivity of the predicted mass-loss rate to the precise value and stratification of between the multi-D and 1D CMF models needs to be further investigated in future work.
To further illustrate how turbulent pressure modifies the density stratification, in the approximately quasi-hydrostatic layers below the wind-launching region, we may assume that the advective acceleration term to be small. We can then substitute the gas and the turbulent pressure terms in Eq. (7) to obtain,
where is the turbulent velocity-modified density scale height, denotes scale height without turbulent pressure and . Because the opacity is higher than electron scattering at the photosphere, we use the total flux-mean opacity (see for definitions Poniatowski et al. [55]; Moens et al. [49]), for these calculations. From the above equations, it is clear that as effective gravity decreases, a similar turbulent pressure yields a more geometrically extended atmosphere. For the B0, B1, B1.5, and B1 models presented here, we find to be 0.05, 0.08, 0.07, and 0.01, respectively, using turbulent velocity values at the photosphere, compared to when are 0.008, 0.0126, 0.0098, and 0.0049. The corresponding density scale height is modified by a factor of approximately 6.9, 6.3, 6.9, and 2.2. Therefore, for models with similar turbulent velocities, effects on the atmospheric density scale height become increasingly important as surface gravity reduces.
We consider an atmospheric layer of fixed thickness with approximately constant , , and . Assuming the same density at the base of the layer, the density enhancement relative to the case without any turbulent velocity can be approximated as
where . As , the factor approaches unity and the density increase becomes less sensitive to further increases in . At the photosphere, for the 2D B1 model we obtain , whereas using the 3D values it becomes 0.55. However, at the effective sonic point, becomes 0.97 and 0.96 for the B1 and B1 models, respectively. Despite the lower photospheric turbulent velocities in the 3D model, the 3D value becomes comparable to the 2D model near their respective effective sonic points. However, this saturation applies only to the simplified fixed-layer density-enhancement approach; in reality, mass losses depend on the radial stratification of the turbulent pressure, radiative acceleration, and the location of the effective sonic point.
Discussion
In this section, we discuss observational implications of the results presented in this work, along with limitations and future improvements, and finally conclude this work.
Observational implications
Turbulent velocities and macroturbulence. The models presented in this work show large velocity dispersions around the photosphere. These turbulent velocities are on the same order as the non-rotational broadening commonly inferred from BSG photospheric absorption lines (Ryans et al. [63]; Simón-Díaz et al. 2017; de Burgos et al. 2024; Serebriakova et al. 2024). For a 3D O star, Delbroek et al. [16] showed that broad photospheric spectral lines can naturally emerge without ad hoc extra broadening terms. Turbulent velocities found here will likely correspond to characteristic non-rotational line broadening of order several tens of km/s, comparable to observed trends. Future 3D spectral synthesis calculations of these stars are required to confirm whether this photospheric line-broadening due to subsurface turbulence seen in O stars extends towards the BSG regime. Further, BSGs have been observed to display significant wind line-profile variability [57, 44, 23]. As such, the temporal variations seen in this work may indeed point towards give rise to line-profile variability. Our models also suggest that velocity inferred from spectral diagnostics may sample different components in the outflow and therefore vary between different observed epochs.
Mass-loss rates and spectroscopic masses. Recent empirical studies of the Galactic, LMC, and SMC BSGs do not find such high mass-loss rates [13, 82, 2, 83]; rather, the empirically found mass losses overall align with the unmodified Björklund et al. [4] prescription. For determining the inferred spectroscopic masses for these stars, the turbulent pressure support predicted by multi-D simulations was not included. Previous comparisons between 1D and multi-D O stars have shown that including the turbulent pressure may change the spectroscopically inferred surface gravity, leading to a larger spectroscopic mass [19, 81]. In a case where such trends continue towards the BSG regimes, spectroscopic masses derived without turbulent pressure could be systematically underestimated. As such, an upward revision of spectroscopic mass would correspond to a lower inferred classical Eddington . As a consequence, the observed BSGs in the empirical studies may in reality occupy a somewhat lower luminosity-to-mass once turbulent pressure is treated consistently in the empirical studies.
Stochastic low-frequency variability. A direct observational consequence of the models presented here is the strong temporal variability. As shown in Sect. 5.3, all our models exhibit stochastic luminosity fluctuations with an excess of power at lower frequencies. These amplitude variations are of a similar order to the bolometric luminosity variability of found in the recent 3D BSG envelope simulations by Ma et al. (2026). However, the simulations presented here are unified atmosphere and wind models, whereas Ma et al. (2026) considers a 3D envelope model without a line-driven wind outflow. Additionally, Ma et al. (2026) finds the 80% power threshold to be approximately below in their synthetic BSG light curves. As shown in Sect. 5.3, we find a similar excess of power at low frequencies. Such SLF variability has been detected in BSGs using high-precision space-based photometry [6] (Ma et al. 2024; Kourniotis et al. 2025). The physical origin of the observed SLF is still debated. Several mechanisms have been proposed, such as internal gravity waves generated by core convection [5, 7] and turbulence associated with the iron-opacity bump [68]. The calculations presented in this work indicate that for BSGs in the stellar parameter range considered here, turbulence generated in the subsurface iron-opacity bump may at least qualitatively explain the observed SLF variability (see also, Ma et al. 2026). However, since we do not model the stellar core, we can’t rule out an internal gravity wave contribution to the SLF variability. Recent 3D hydrodynamical simulations by Pathak et al. [50], which include both core and envelope convective regions, suggest that simulations with only the outer envelope convection zone produce less power at low frequencies compared to their full star configuration. Work is currently underway to compute observable signatures from such simulations to compare with high-resolution spectroscopic observations [e.g. 29].
Limitations and next steps
The calculations presented in this chapter demonstrate that the aNLTE treatment can substantially modify the ionization balance in the wind, influencing the line-force parameters and affecting the dynamical properties of BSG winds. However, a few aspects of the BSG models presented in this chapter need deeper understanding and further investigation.
First, the current models in this chapter are limited to a small subset of the BSG parameter space. A natural extension would be to construct a larger set of 3D models covering a broader range of temperature (extending towards the colder side of bistability), surface gravity, luminosities, and metallicity. Such a grid would allow us to make systematic predictions of how the ionization behavior influences the dynamics of these stars.
Second, in the FLD closure relation, the flux is determined by the gradient of the radiation energy density, and lateral gradients in the energy density may generate a lateral radiative flux and radiation force [49]. Since FLD constrains the flux direction to follow the gradient of the radiation energy density, the approximation becomes less reliable as the radiation field becomes more anisotropic [46], which is likely the case around the photosphere. Work is currently underway to replace the FLD with a two-moment radiation transport scheme, M1, for the massive RHD simulations, which evolves both the radiation energy density and the flux and does not involve constraining the flux from the gradient of the energy density.
Third, the pseudo-planar geometry adopted in these simulations. As the lateral coordinate remains locally Cartesian, the pseudo-planar geometry can be regarded only as a local patch of the stellar atmosphere. This approximation becomes less applicable when the lateral structures are comparatively large.
Lastly, the present calculations employ a diluted frequency-integrated radiation field and, thereby, integrated mean opacities. Although this provides an efficient way to incorporate aNLTE effects in multi-D calculations, the ionization balance is intrinsically a frequency-dependent problem. The BSGs have a rich spectrum, and using an integrated radiation field may over- or underestimate how different parts of the spectrum interact with the gas. Work is currently underway to develop a frequency-dependent NLTE opacity formalism including line-driving for multi-D RHD simulations.
Summary and Conclusion
In this work, we presented the first unified multi-D atmosphere and wind models of early BSGs. We included aNLTE effects in the ionization balance and in calculating the line opacities. We conducted our study with three 2D and one 3D model spanning –30.4 kK. Similar to the multi-D O stars (Jiang et al. 2015; Debnath et al. 2024; Van der Sijpt et al. 2025; Delbroek et al. 2025; Moens et al. [47]), we find large velocity fluctuations originating deep within the stellar envelope, which propagate to the optical photosphere. Compared to the 2D model, the 3D model shows reduced photospheric turbulent velocities, lower atmospheric inflation, and less pronounced fall-back of the mean radial velocity at the photosphere. These differences suggest that while 2D models capture qualitatively the mechanisms governing the atmosphere and wind, 3D simulations are necessary for quantitative characterization of the atmospheric dynamics. Due to their low surface gravities, turbulent pressure modifies the density scale height around the photosphere to such an extent that it influences the global mass-loss rates. As a result, for the same parameter, we find higher mass-loss rates than the Björklund et al. [4] prescriptions, which are based on 1D hydrodynamically consistent FASTWIND models (Sundqvist et al. [74]). When computed using turbulent pressure obtained from multi-D, the mass-loss rates from 1D hydrodynamically consistent FASTWIND are higher, and the terminal wind speeds are lower. Further, we found that in our aNLTE treatment, the ionization stratification in the wind significantly differs from the reference LTE ionization. Moreover, the wind dynamics also differ between the models, with the hotter B0 model exhibiting a faster wind and the cooler B1 and B1.5 displaying a gradual and slower wind. Finally, we also found significant temporal variations in our models; the photospheric radiative luminosity varies by an order of 10%. The simulations also show an excess of power at low temporal frequencies similar to observed trends.
Acknowledgements. The computational resources used for this work were provided by Vlaams Supercomputer Centrum (VSC), funded by the Research Foundation-Flanders (FWO) and the Flemish Government. DD, JS, and NM acknowledge support from the European Research Council (ERC) under the Horizon Europe grant agreement 101044048. JS further acknowledges support from the Belgian Research Foundation Flanders (FWO) Odysseus program under grant number G0H9218N, from FWO grant G077822N, and from KU Leuven C1 grant BRAVE C16/23/009. The authors also acknowledge the fruitful discussions, suggestions, and comments from the present and past members of the EQUATION team. Data analysis and visualisation in this project relied on WindSimTools (Moens et al. [47]), NumPy (Harris et al. [22]), SciPy (Virtanen et al. [88]), and Matplotlib (Hunter [25]).
References
- [1]Asplund, M., Grevesse, N., Sauval, A. J., & Scott, P. 2009, ARA&A, 47, 481
- [2]Bernini-Peron, M., Sander, A. A. C., Ramachandran, V., et al. 2024, A&A, 692, A188
- [3]Bernini-Peron, M., Sander, A. A. C., Sabhahit, G. N., et al. 2026, A&A, 708, A206
- [4]Björklund, R., Sundqvist, J. O., Singh, S. M., Puls, J., & Najarro, F. 2023, A&A, 676, A109
- [5]Bowman, D. M., Aerts, C., Johnston, C., et al. 2019a, A&A, 621, A135
- [6]Bowman, D. M., Burssens, S., Pedersen, M. G., et al. 2019b, Nature Astronomy, 3, 760
- [7]Bowman, D. M., Burssens, S., Simón-Díaz, S., et al. 2020, A&A, 640, A36
- [8]Cantiello, M., Langer, N., Brott, I., et al. 2009, A&A, 499, 279
- [9]Castor, J. I., Abbott, D. C., & Klein, R. I. 1975, ApJ, 195, 157
- [10]Crowther, P. A., Lennon, D. J., & Walborn, N. R. 2006, A&A, 446, 279
- [11]Curé, M. 2004, ApJ, 614, 929
- [12]Curé, M., Cidale, L., & Granada, A. 2011, ApJ, 737, 18
- [13]de Burgos, A., Simón-Díaz, S., Urbaneja, M. A., & Puls, J. 2024, A&A, 687, A228
- [14]Debnath, D., Sundqvist, J. O., Moens, N., et al. 2026, A&A, 714, A418
- [15]Debnath, D., Sundqvist, J. O., Moens, N., et al. 2024, A&A, 684, A177
- [16]Delbroek, L., Sundqvist, J. O., Debnath, D., et al. 2025, A&A, 704, A104
- [17]Friend, D. B. & Castor, J. I. 1983, ApJ, 272, 259
- [18]Gayley, K. G. 1995, ApJ, 454, 410
- [19]González-Torà, G., Sander, A. A. C., Sundqvist, J. O., et al. 2025, A&A, 694, A269
- [20]Gräfener, G. & Hamann, W.-R. 2005, A&A, 432, 633
- [21]Grassitelli, L., Fossati, L., Langer, N., et al. 2015, A&A, 584, L2
- [22]Harris, C. R., Millman, K. J., van der Walt, S. J., et al. 2020, Nature, 585, 357
- [23]Haucke, M., Cidale, L. S., Venero, R. O. J., et al. 2018, A&A, 614, A91
- [24]Hawcroft, C., Sana, H., Mahy, L., et al. 2023, arXiv e-prints, arXiv:2303.12165DOI
- [25]Hunter, J. D. 2007, Computing in Science and Engineering, 9, 90
- [26]Iglesias, C. A. & Rogers, F. J. 1996, ApJ, 464, 943
- [27]Jiang, Y.-F. 2023, Galaxies, 11, 105
- [28]Jiang, Y.-F., Cantiello, M., Bildsten, L., Quataert, E., & Blaes, O. 2015, ApJ, 813, 74
- [29]Kalita, K. J., Bowman, D. M., Abdul-Masih, M., & Simón-Díaz, S. 2025, A&A, 703, A2
- [30]Kee, N. D., Sundqvist, J. O., Decin, L., de Koter, A., & Sana, H. 2021, A&A, 646, A180
- [31]Keppens, R., Teunissen, J., Xia, C., & Porth, O. 2020, arXiv e-prints, arXiv:2004.03275arxiv.org/abs/2004.03275
- [33]Krtička, J. & Kubát, J. 2017, A&A, 606, A31
- [34]Krtička, J., Kubát, J., & Krtičková, I. 2021, A&A, 647, A28
- [35]Kudritzki, R. P., Puls, J., Lennon, D. J., et al. 1999, A&A, 350, 970
- [36]Lamers, H. J. G. L. M., Snow, T. P., & Lindholm, D. M. 1995, ApJ, 455, 269
- [37]Langer, N. 2012, ARA&A, 50, 107
- [38]Lefever, K., Puls, J., & Aerts, C. 2007, A&A, 463, 1093
- [40]Lomb, N. R. 1976, Ap&SS, 39, 447
- [41]Lucy, L. B. & Abbott, D. C. 1993, ApJ, 405, 738
- [44]Markova, N. & Puls, J. 2008, A&A, 478, 823
- [45]Mihalas, D. 1978, Stellar atmospheres
- [46]Moens, N. 2023, Radiation hydrodynamics of massive star outflows
- [47]Moens, N., Debnath, D., Verhamme, O., et al. 2025, A&A, 704, A121
- [48]Moens, N., Poniatowski, L. G., Hennicker, L., et al. 2022b, A&A, 665, A42
- [49]Moens, N., Sundqvist, J. O., El Mellah, I., et al. 2022a, A&A, 657, A81
- [50]Pathak, P., Blouin, S., Herwig, F., & Woodward, P. R. 2026, The Astrophysical Journal, 1000, 89
- [52]Pauldrach, A. W. A., Lennon, M., Hoffmann, T. L., et al. 1998, in Astronomical Society of the Pacific Conference Series, Vol. 131, Properties of Hot Luminous Stars, ed. I. Howarth, 258
- [53]Pauldrach, A. W. A. & Puls, J. 1990, A&A, 237, 409
- [54]Petrenz, P. & Puls, J. 2000, in Astronomical Society of the Pacific Conference Series, Vol. 214, IAU Colloq. 175: The Be Phenomenon in Early-Type Stars, ed. M. A. Smith, H. F. Henrichs, & J. Fabregat, 626
- [55]Poniatowski, L. G., Kee, N. D., Sundqvist, J. O., et al. 2022, A&A, 667, A113
- [56]Poniatowski, L. G., Sundqvist, J. O., Kee, N. D., et al. 2021, A&A, 647, A151
- [57]Prinja, R. K., Massa, D., & Fullerton, A. W. 2002, A&A, 388, 587
- [58]Puls, J., Najarro, F., Sundqvist, J. O., & Sen, K. 2020, A&A, 642, A172
- [59]Puls, J., Springmann, U., & Lennon, M. 2000, A&AS, 141, 23
- [60]Puls, J., Vink, J. S., & Najarro, F. 2008, A&A Rev., 16, 209
- [61]Ricker, G. R., Winn, J. N., Vanderspek, R., et al. 2015, Journal of Astronomical Telescopes, Instruments and Systems, 1, 014003
- [62]Rogers, F. J. & Iglesias, C. A. 1992, ApJS, 79, 507
- [63]Ryans, R. S. I., Dufton, P. L., Rolleson, W. R. J., et al. 2002, Monthly Notices of the Royal Astronomical Society, 336, 577
- [64]Sander, A. A. C., Vink, J. S., & Hamann, W.-R. 2020, MNRAS, 491, 4406
- [65]Scargle, J. D. 1982, ApJ, 263, 835
- [66]Schillemans, P., Sundqvist, J. O., Debnath, D., et al. 2026, A&A, 713, A127
- [67]Schultz, W. C., Bildsten, L., & Jiang, Y.-F. 2020, ApJ, 902, 67
- [68]Schultz, W. C., Bildsten, L., & Jiang, Y.-F. 2022, The Astrophysical Journal Letters, 924, L11
- [69]Schwarzschild, M. 1958, Structure and evolution of the stars.
- [70]Searle, S. C., Prinja, R. K., Massa, D., & Ryans, R. 2008, A&A, 481, 777
- [71]Serebriakova, N., Tkachenko, A., & Aerts, C. 2024, A&A, 692, A245
- [72]Simón-Díaz, S., Godart, M., Castro, N., et al. 2017, A&A, 597, A22
- [73]Sobolev, V. V. 1960, Moving Envelopes of Stars
- [74]Sundqvist, J. O., Björklund, R., Puls, J., & Najarro, F. 2019, A&A, 632, A126
- [75]Sundqvist, J. O., Debnath, D., Backs, F., et al. 2025, arXiv e-prints, arXiv:2504.00936DOI
- [76]Sundqvist, J. O., Owocki, S. P., & Puls, J. 2018, A&A, 611, A17
- [78]Van der Sijpt, C., Sundqvist, J. O., Debnath, D., Driessen, F. A., & Moens, N. 2025, A&A, 694, A54
- [79]Venero, R. O. J., Curé, M., Cidale, L. S., & Araya, I. 2016, ApJ, 822, 28
- [80]Venero, R. O. J., Curé, M., Puls, J., et al. 2024, Monthly Notices of the Royal Astronomical Society, 527, 93
- [81]Verhamme, O. 2025, Spectroscopy of massive stars at low metallicity: Studying the atmospheres and outflows of B-stars
- [82]Verhamme, O., Sundqvist, J., de Koter, A., et al. 2024, A&A, 692, A91
- [83]Verhamme, O., Sundqvist, J., de Koter, A., et al. 2026, A&A, 710, A166
- [84]Vink, J. S., de Koter, A., & Lamers, H. J. G. L. M. 1999, A&A, 350, 181
- [85]Vink, J. S., de Koter, A., & Lamers, H. J. G. L. M. 2000, A&A, 362, 295
- [86]Vink, J. S., de Koter, A., & Lamers, H. J. G. L. M. 2001, A&A, 369, 574
- [87]Vink, J. S., Sabhahit, G. N., & Sander, A. A. C. 2026, A&A, 711, L4
- [88]Virtanen, P., Gommers, R., Oliphant, T. E., et al. 2020, Nature Methods, 17, 261
- [89]Xia, C., Teunissen, J., Mellah, I. E., Chané, E., & Keppens, R. 2018, The Astrophysical Journal Supplement Series, 234, 30



