Wind effects on mechanically depressurized indoor containment zones: Impact of uncertainties in computational fluid dynamics wind pressure on ventilation network model output

Indoor pollutant-containing zones require utmost care in sealing the containment particularly if the pollutant is hazardous to public and occupational health, as is the case for nuclear industries, asbestos abatement zones, and others. It should be ensured that no pollutant particles escape to the outdoor environment. However, these particles could escape through unintended openings in building envelopes such as leaks. Hence, it is advised to ensure a negative pressure within the containment zone with respect to the outer environment, so that any uncontrolled flow is directed always into the containment zone [e.g., Great Britain, Health and Safety Executive (HSE), 2006; KTA-Geschäftsstelle, 2017]. For example, asbestos abatement from building interiors has the risk of suspended fibers being airborne and escaping into the outdoor atmosphere in the case of reverse flow from the containment zone to the outdoors. In order to minimize such situations, different government authorities (e.g., World Health Organization Regional Office for Europe, 2012; Legifrance, 2013; Ministerie van Sociale Zaken en Werkgelegenheid (SZW), 2016; and Occupational Safety and Health Administration (OSHA), 2019 in the case of an asbestos abatement worksite) have laid out regulatory measures. This includes maintaining a negative pressure within the containment with respect to the outdoor environment by means of a mechanical ventilation system (D'Angelo , 1987). This mechanical ventilation system should ensure that the flow out of the containment occurs only through a negative pressure unit equipped with high-efficiency particulate air (HEPA) filters. The inflow of fresh air into the containment is achieved by means of specifically designed air inlets and airlocks with check valves. The depressurization requirement established by the different authorities is, however, laid out without considering explicitly the potential effects of atmospheric winds and its fluctuations, as evident from the lack of scientific publications, standards, and guidelines on this issue. However, as clearly demonstrated in the scientific literature, atmospheric wind can create high suction pressures on the building envelopes (e.g., Cermak, 1975; Simiu and Scanlan, 1996; ASHRAE, 2001; Stathopoulos, 2002; and Holmes, 2004) and, therefore, it is expected that they can be responsible for breaches during containment asbestos removal.

Ventilation network models (VNMs) are valuable tools for analyzing the airflow in mechanically ventilated buildings (e.g., Axley, 2007; Chen, 2009). Their advantages include fast simulations with minimal computational resource requirements, compared to other numerical methods such as computational fluid dynamics (CFD). Over the past few decades, several VNMs have been developed, such as COMIS (Feustel, 1999), CONTAM (Dols and Polidoro, 2015), and SYLVIA (IRSN, 2022). VNMs are zone-based airflow models, where the building equipped with a ventilation system is represented by nodes and branches. Nodes represent rooms, junctions, or an exterior environment (boundary condition) in which properties like pressure, temperature, and pollutant concentration are assumed to be uniform. The equations of conservation of mass, momentum, and energy are applied for the nodes. Branches represent components of the ventilation system, such as air inlets, airlocks, leaks, filters, and fans in the case of asbestos abatement worksites. The mass flow rate between nodes is calculated with the generalized Bernoulli's equation for the branches.

Using external wind pressure data as input, VNMs can effectively analyze and investigate the effects of wind on indoor airflow through mechanically ventilated buildings or compartments within buildings (Cóstola , 2009; Le Roux , 2013). In this research study, the code SYLVIA has been used. SYLVIA is a VNM developed by the Institut de Radioprotection et de Sûreté Nucléaire of France (IRSN, 2022) and it has been validated in the past years for ventilation, fire and airborne contamination in nuclear facilities equipped with mechanical ventilation systems to establish internal depressurization (Le Roux , 2013, 2012; Vaux and Prétrel, 2013; Coutin, 2015; and Le Dez , 2021). Additionally, Jayakumari (2025) validated the model's effectiveness in predicting internal pressure for mechanically ventilated asbestos containment zones, which operate at a lower depressurization range compared to nuclear facilities (e.g., −5 to −40 Pa). SYLVIA has the capability to perform transient simulations using time series of external pressures as input, to predict its influence on an initial depressurization established within a containment volume. The user-friendly interface includes data import/export and checks, minimizing the risk of setup errors made by users and enabling the design and simulation of complex systems.

However, reliable data of external wind pressures are imperative for an accurate indoor airflow prediction with VNMs. As discussed by Cóstola (2009), there are different sources of external wind pressures available for use in VNMs. This includes analytical models and provisions in codes/standards, field measurements, WT tests, and CFD simulations. Analytical models and provisions in codes and standards (e.g., ASHRAE, 2001; ASCE, 2002) typically focus on isolated buildings with simplified geometries. Furthermore, they lack the provision of time-series data for external wind pressures, which is essential for input in transient simulations. Field measurements are the most realistic and consequently also the most representative source of wind pressures on buildings (Levitan and Mehta, 1992; Richards , 2001; and Li , 2021). However, recording external pressure time series at different locations on building facades over a sufficiently long period as input for SYLVIA is challenging and very cumbersome and expensive. In addition, a field measurement campaign comes with the practical challenges, such as obtaining approvals for measurements in existing buildings, particularly when hazardous materials like asbestos particles are involved, making them not always feasible. Such measurements are typically used to gain insights into reality and for validation purposes (e.g., Blocken, 2014, 2015; Mateus , 2016). Atmospheric boundary layer wind tunnel (ABLWT) tests are still the most common source to obtain wind pressures for VNMs (Castro and Robins, 1977; Cheung , 1997; Stathopoulos, 2002; Cermak, 2003; Blocken, 2018; and Lamberti , 2020). The tests are typically performed on reduced-scale models, following the required similarity criteria to attain high-quality data under controlled conditions (e.g., Castro and Robins, 1977; Hunt, 1982; Lin , 1995; Richards , 2001; Tominaga and Blocken, 2015; 2016; and Jayakumari , 2023; 2024). In addition to WT tests, the rapid explosion of computational capacities over the last decades led to an improved prediction of wind pressures on buildings by CFD (e.g., Nozawa and Tamura, 2002; Ono , 2008; Montazeri and Blocken, 2013; Ricci , 2018; Guichard, 2019; Lamberti , 2020; Xu 2023; Xing , 2023; and Ricci, 2024). However, the reliability of CFD (as an alternative to WT) as source of external wind pressure (P_e) for VNM to predict indoor pressure (P_i) is not known yet. Despite the great efforts of the computational wind engineering community in recent decades to improve the prediction of external wind pressures by CFD, the accurate prediction of pressure fluctuation and, in particular, peak pressures is still challenging (e.g., Potsis , 2023).

This paper combines external wind pressure (P_e) time series on an internally depressurized building with a carefully designed ventilation network to analyze indoor pressure (P_i) and containment breach duration. The P_e time series and statistical parameters are obtained by reduced-scale WT tests performed on an idealized isolated building (Jayakumari , 2023), on the one hand, and by CFD simulations of large-eddy simulation (LES) and scale-adaptive simulation (SAS) (Menter and Egorov, 2010) on the other hand. SAS is a hybrid RANS-LES approach recently gaining widespread applicability in various wind engineering applications (e.g., García , 2015; Jadidi , 2018; Rezaeiha , 2019; van Druenen and Blocken, 2023; Žužul , 2023; and Qin , 2024). The SAS approach integrates aspects from both LES and RANS to perform transient simulations, by resolving large-scale turbulent structures of the flow (similar to LES) rather than modeling the impact of these on the mean flow like the RANS approach. Its advantage lies in its ability to perform transient simulations with a lower resolution of grid and time step size near the walls (compared to LES), while still capturing a large part of the energy turbulence spectrum. Additionally, unlike other hybrid models such as DES, SAS can dynamically switch from LES-like to RANS-like approach when temporal or spatial resolution is inadequate for achieving LES-like solutions. For this study, SAS is particularly useful as it enables the generation of P_e time series required for the ventilation network model SYLVIA (IRSN, 2022). The P_e time series from WT, LES, and SAS are used as input to the ventilation network model SYLVIA (IRSN, 2022). The depressurization value used is -40 Pa, and the reference wind speed (U_ref) at building height is 12.65 m/s. The goal of the paper is threefold: (1) comparing the P_e data as obtained by WT and CFD in terms of time series and statistics; (2) assessing the impact of uncertainties in the P_e as generated by CFD on the resulting P_i and on the breach duration percentage; (3) estimating the P_i statistics and breach duration percentage for an initial case study with the indoor depressurization of −40 Pa and a fixed U_ref of 12.65 m/s at building height. The objective (2) aims to determine whether, despite the inherent uncertainties in CFD results, the combinations LES-SYLVIA (LES-S) or SAS-SYLVIA (SAS-S) could be used as alternatives to the WT-SYLVIA (WT-S) combination in future studies.

The manuscript is structured as follows. Section II describes the methodology and the case study and introduces the experimental and numerical techniques used in the study. Section III discusses the results from the experiments and numerical simulations. Section IV closes the paper with summary, limitations, and conclusions.

The methodology used in this study is illustrated in the workflow of Fig. 1. In the first stage, WT tests and CFD (i.e., SAS and LES) simulations are performed on an idealized isolated cubical building model (see Sec. II B) to generate the time series of external pressures (P_e). The P_e results from SAS and LES simulations are validated by means of the WT results. In the second stage, the corresponding P_e results from the WT tests and CFD (i.e., SAS and LES) simulations are used as input for SYLVIA to predict the internal pressure (P_i) variation over time within the asbestos containment zone. In the third stage, the P_i results from SAS-S and LES-S are compared to those obtained from WT-S, to assess the effect of different methods for obtaining P_e on the prediction of P_i for a mechanically ventilated asbestos containment zone. The results are analyzed and presented in terms of external and internal pressure coefficients (C_pe and C_pi). The percentage containment breach predicted by SYLVIA is also compared among WT-S, SAS-S, and LES-S at multiple locations on the building facades.

An idealized cubical building of dimensions 18 × 18 × 18 m³ in full scale is considered for the case study. The asbestos containment zone constitutes the upper half of the building, corresponding to a volume of 18 × 18 × 9 m³. The ventilation system for asbestos abatement consists of air inlets, negative pressure units (NPU) with HEPA filters, an airlock for materials, an airlock for people, and air inlets for tuning. The ventilation system for the considered containment volume is designed following a well-established methodology (Dubernet , 2018) to define an internal depressurization of −40 Pa for the reduced-scale building. This ventilation system comprises 40 air inlets (A1–A40), 4 NPUs (N1–N4), an airlock for materials, an airlock for people, and an air inlet for tuning. A schematic diagram of the reduced-scale building with ventilation components and photographs of the ventilation components from a real worksite are reported in Fig. 2. In addition to these ventilation components, four leakages (L1–L4) positioned close to the edges are considered. The WT tests and CFD simulations are performed on a 1:40 scaled model of the case study building to generate the P_e time series at the locations of each ventilation component (including the leakages). Two wind directions, θ = 0° and θ = 45°, are investigated. The exact positioning of the ventilation components (i.e., where P_e is generated) on the building facades and their labels for reference are shown in Fig. 3.

The computational domain for the 1:40 scaled building is created following the best practice guidelines for CFD simulations in wind engineering and urban and building aerodynamics (e.g., Franke , 2007; Tominaga , 2008) with an upstream fetch of 5 h, a downstream fetch of 15 h, and a height of 6 h, where h is the building height in reduced scale equal to 0.45 m (Fig. 5). The blockage ratio is 0.8%, which is well below the maximum of 3% (Franke , 2007; Tominaga , 2008; and Blocken, 2015). The Gambit 2.4.6 software is used for grid generation (Fluent Inc., 2005). A grid-sensitivity analysis is carried out on three grids by means of a factor of √2 in each of the three Cartesian directions (x, y, z): the coarse grid (4 193 128 cells), the basic grid (17 357 985 cells), and the fine grid (32 208 108 cells) (Fig. 6). For the basic grid, the cells close to the building facades for a distance of h around the building are cubical with a size of 0.0125 m, equivalent to 0.5 m at full scale. Farther away from the building, cells are stretched with a growth rate of 1.02 to limit the overall number of cells while ensuring a close-to-cubical cell shape suitable for LES simulations. The coarse and fine grids are derived from the basic grid by applying the above-mentioned factor for coarsening or refining, respectively. All simulations for the grid-sensitivity analysis are performed with the LES approach.

The boundary conditions for the SAS and LES simulations are reported in Fig. 5, for the wind directions θ = 0° and θ = 45°.

For the LES simulations, similar boundary conditions as for SAS are used. The U(z), k(z), and ε(z) profiles at the inlet are obtained from the WT data using Eqs. (2), (4), and (5), respectively. The Wall-Adapting Local Eddy-Viscosity (WALE) sub-grid scale model proposed by Nicoud and Ducros (1999) is adopted for modeling the sub-grid scale stresses. As in SAS, the spectral synthesizer method is adopted to perturbate the flow field at the inlet.

Both LES and SAS simulations are performed with ANSYS Fluent 2021 R2 (Ansys Inc., 2021), with the SAS approach coupled to the shear-stress transport (SST) k–ω turbulence model (Menter , 2003). The fractional step method is used for pressure–velocity coupling with non-iterative time advancement (NITA). For the spatial discretization, the least squares cell-based method is used for gradient computation, the second-order scheme is used for pressure, a bounded-central-differencing scheme is used for momentum, and the second-order upwind scheme is used for k and ω. For the temporal discretization, the bounded second-order implicit scheme is used. A time step of 6.25 × 10⁻⁴ s is used to keep the Courant–Friedrichs–Lewy number (CFL) smaller than 1. Ten flow-through-time is executed to initialize the flow field and 35 flow-through times for the sampling of the flow statistics. Thus, the simulations are performed for a total duration of 34.4 s (at reduced scale) generating 54998 data samples. All simulations are carried out with the Dutch national supercomputer Snellius (SURFsara), with 2 nodes (128 cores each) and 256 GB RAM. The results of the external pressures from the SAS and LES simulations are presented and discussed in terms of C_pe in Sec. III.

The backward differential formula well suited for stiff equations (Asher and Petzold, 1998) is used for time discretization. Each time step is resolved by an iterative Newton–Raphson algorithm. The procedure described as follows is adopted for building the network and performing the simulations with SYLVIA:

• Step 1: The containment zone is created as a room.

• Step 2: All the ventilation components (air inlets, airlocks, leaks, and NPUs with filter) are modeled connecting the containment zone with a node representing the boundary with the exterior environment.

• Step 3: The properties of the fluid (air in this case) and the characteristic properties of the containment zone (the initial depressurization and the volume) are set up. An initial depressurization value of −40 Pa is imposed in the containment zone.

• Step 4: The geometrical properties (L and S) and the aeraulic characteristics (Q vs ΔP) of each ventilation component as well as the fan operating curve are input in SYLVIA.

• Step 5: Volumetric flow into and out of the containment zone is checked to assure there is flow balance satisfying the continuity equation.

• Step 6: Q and ΔP are used to establish the corresponding airflow resistances (R) and the flow exponent (n) for each component within SYLVIA using the following equation:

  $ΔP=RQnρ.$

  (10)

• Step 7: Initial conditions, in the absence of external wind effects, are established based on the inputs from steps 3 and 4, with the time (t) equal to zero in Eqs. (8) and (9).

• Step 8: The P_e time series at the locations of the different ventilation components are used as a boundary condition and the resulting corresponding P_i time series solution is obtained.

A grid-sensitivity analysis is performed with the LES approach. The results in terms of mean (C_pe,mean), root mean square (C_pe,rms), peak positive and negative (C_pe,max and C_pe,min) external pressure coefficients obtained from the coarse, basic and fine grids, for θ = 0° and 28 pressure locations (at a height of 0.7 h) are presented in Fig. 9. Then, the average absolute difference between (δ_c) coarse and basic and (δ_f) fine and basic is calculated for comparison and reported in Table I. It can be observed that C_pe,mean and C_pe,rms predicted by all the three grids are in good agreement with each other as evident from δ_c and δ_f of 0.02. A relatively good agreement in terms of C_pe,max between the basic and the fine grids (δ_f = 0.05) is also found, while there is an increased average absolute difference of δ_c = 0.07 between the coarse and the basic grids. However, notable deviations are observed for C_pe,min with increased differences of δ_c = 0.14 and δ_f = 0.12. Although the difference in C_pe,min between the fine grid and the basic grid is not negligible, the basic grid can be considered a good compromise between computational accuracy and computational costs. Therefore, the basic grid is used for further SAS and LES simulations.

A visual comparison of the coherent turbulent structures in the flow by LES and SAS is made by means of the Q-criterion (Jeong and Hussain, 1995). The Q-criterion is defined as the second invariant of velocity gradient tensor, i.e., $Q=12Ω2−S2$⁠, where Ω² represents the rotation rate and S² represents the strain rate. It implies that a positive Q iso-surface is a representation of the areas where the strength of rotation overcomes the strain and can, thus, be considered as vortex envelopes (Dubief and Delcayre, 2000). Figure 11 shows the Q iso-surfaces (colored by instantaneous C_pe) corresponding to a Q value of 1.0 × 10⁴ s⁻² for both wind directions (θ = 0° and θ = 45°) for LES and SAS, for the time instance t = 34.4 s. It is evident that the small-scale turbulent structures of the flow field are not as well resolved in SAS as in LES, with major turbulence structures developing only in the proximity of the building. The possible effect of this on the prediction of C_pe is discussed in Sec. III C.

This subsection discusses the validation of the LES and SAS results of the external pressure statistical parameters with the WT data. The comparison in terms of C_pe time series for both wind directions is presented in Fig. 12. Only a duration of 15 s is displayed for clarity. For θ = 0°, Figs. 12(a) and 12(c) show the results at positions S1 and S3 in a positive pressure region (windward facade) for θ = 0° and θ = 45°, respectively. Figures 12(b) and 12(d) show the results at S2 and S4 positioned in a negative pressure region, i.e., side facade for θ = 0° and leeward facade for θ = 45°. The CFD and WT time series show some similarities but also some clear discrepancies. The fluctuations at the windward facades (S1 and S3) are not well reproduced by LES and SAS for both θ = 0° and 45° [Figs. 12(a) and 12(c)]. The C_pe magnitude is significantly underpredicted by SAS when compared to WT, with a similar (though less pronounced) discrepancies observed in LES, particularly for θ = 45°. As observed in Fig. 11, this may be due to the missing turbulent structures in the incident flow and the small-scale turbulent structures not being resolved by SAS. The agreement between LES and SAS with WT is considerably better for S2 and S4 positioned in negative pressure zones [Figs. 12(b) and 12(d)]. However, an underprediction of the C_pe with SAS can still be observed for S4 from the upward shift of SAS time series in Fig. 12(d).

Figures 13 and 14 allow a quantitative comparison between CFD and WT, in terms of mean (C_pe,mean), root mean square (C_pe,rms), positive peak (C_pe,max), and negative peak (C_pe,min) pressure coefficients, for θ = 0° and θ = 45°, respectively. The C_pe,mean and C_pe,rms [Figs. 13(a), 13(b), 14(a), and 14(b)] predicted by CFD show a good agreement with WT, with the majority of locations within a deviation of 15% from WT. On the contrary, the C_pe,max and C_pe,min [Figs. 13(c), 13(d), 14(c), and 14(d)] predicted by CFD show higher deviations from the WT results, at certain locations even larger than 30%. Such deviations in peak pressures were also reported by other researchers (e.g., Kataoka , 2020; Nozawa and Tamura, 2002; Selvam, 1997; and Wang and Chen, 2022), indicating that accurate prediction of peak C_pe is still a quite challenging task even when using more sophisticated CFD approaches as LES. One key reason for these discrepancies is the sensitivity of peak pressures to the turbulence characteristics of the CFD incident wind flow, including turbulence intensities, turbulent length scales, and power spectra of turbulence. Discrepancies in these characteristics between CFD and WT can significantly impact peak pressure predictions, as also observed by Lamberti and Gorlé (2020). These discrepancies arise primarily due to the inflow generation method used. The recent past has seen increased research aimed at improving inflow generation methods (e.g., Yan and Li, 2015; Lamberti , 2018; Vasaturo , 2018; Guichard, 2019; Patruno and De Miranda, 2020; Melaku and Bitsuamlak, 2021; and Potsis , 2024) to better replicate incident flow conditions at the building location. While the precursor method has demonstrated better representation of incident wind flow characteristics (Vasaturo , 2018), its high computational cost limits its practicality for industrial applications. Additionally, while the vortex method predicts the mean pressures relatively well, it can lead to unrealistic pressure fluctuations. Therefore, in this research study, the in-built spectral synthesizer inflow generation method in ANSYS Fluent 2021 R2 (Ansys Inc., 2021) has been used without any additional code manipulation, in order to investigate its performance for the intended industrial application. The subsequent section discusses the significance (or potential insignificance) of accurate peak P_e predictions for input in SYLVIA in determining the resulting P_i.

Overall, it can be noted that LES outperforms SAS in the prediction of C_pe statistics. Furthermore, the SAS approach proves to be less efficient in this case requiring more computational resources than the corresponding LES simulations, despite using similar computational settings (see Sec. II D 2) and the same grid (i.e., the basic), as also observed by Jadidi (2018). Specifically, the SAS simulation took 42 h to complete compared to the 29 h required for the LES simulation using the same computational resources already mentioned in Sec. II D.

This subsection compares the temporal variation in P_i within the containment volume as predicted by SYLVIA using the P_e time series from WT vs CFD as input. Figure 15 shows the C_pi time series for θ = 0° [Figs. 15(a) and 15(b)] and θ = 45° [Figs. 15(c) and 15(d)]. The comparison for θ = 0° shows that LES-S and SAS-S are in fairly good agreement with WT-S, despite the discrepancies observed in terms of C_pe (see Sec. III C). However, for θ = 45° the C_pi is underpredicted by both CFD-S cases, with the SAS-S showing a larger underprediction than LES-S. This trend is consistent with the C_pe results commented in Sec. III C. This is evident from Figs. 15(c) and 15(d) in the upward shift of the LES-S and especially the SAS-S time series.

Figure 16 shows the comparison in terms of C_pi statistics—i.e., mean, root mean square (rms), peak maximum (max), and peak minimum (min)—for the WT and CFD cases coupled with SYLVIA. The experimental uncertainty of the WT data are also reported in the figure. This uncertainty is estimated based on C_pe, by comparing the magnitudes of C_pe,mean, C_pe,rms, C_pe,max, and C_pe,min at symmetric locations on the cube. The 90th percentile of all absolute differences in magnitude is then selected as the uncertainty. In general, the C_pi statistics from LES-S and SAS-S show a satisfactory agreement with WT-S data for both θ = 0° [Fig. 16(a)] and θ = 45° [Fig. 16(b)]. The mean and rms of C_pi by LES-S and SAS-S show a difference as small as 0.01 from the WT-S data. The C_pi values from CFD-S show a larger deviation in terms of the max and min with respect to the WT-S. However, the absolute differences in terms of C_pi mean, rms, and peak between WT-S and CFD-S cases fall within the uncertainties associated with WT data (see Fig. 16) in spite of some clear deficiencies in reproducing the WT C_pe data by LES and especially SAS, as discussed in Subsection III C. This indicates that the prediction of the internal pressure statistics is not always very sensitive to the aforementioned deviations in the C_pe data used as input. This is because the C_pi value at any time step results from the combined effect of all C_pe inputs and the instantaneous differences (i.e., the differences in fluctuations and individual peaks) tend to be compensated rather than accumulate. Moreover, as discussed in Sec. III C, significant discrepancies in C_pe are observed only at positions within the positive pressure region (i.e., the windward facade), while in all other regions, the results remain comparable.

The primary objective of using SYLVIA for a pollutant containment zone is to predict the possible breaches in containment due to wind effects. This breach occurs when the local external pressure is lower than the internal pressure. In this subsection, the containment breach is quantified for all the locations of the ventilation components as the percentage duration for which C_pi exceeds C_pe with respect to the total duration of the SYLVIA simulations. Figure 17 compares the duration of breach by WT-S, LES-S, and SAS-S for θ = 0° and θ = 45°. The locations of the breaches (in red) on the building are also shown in Fig. 17. The breaches occur in regions of flow separation characterized by pronounced negative external pressure, as on facades 1–2 and 3–0 for θ = 0° and on facades 2-3 and 3-0 for θ = 45° (Jayakumari , 2023). This observation is consistent for all three cases (WT-S, SAS-S, and LES-S), with breach occurrence at all locations along facades 1-2 and 3-0 for θ = 0° and on facades 2-3 and 3-0 for θ = 45° as marked in Fig. 17. In general, the percentage duration of the breaches predicted by LES-S and SAS-S is comparable to WT-S with an average difference in breach duration of 4% and 5%, respectively. The largest deviation in percentage breach duration between LES-S and WT-S is 10% for θ = 0° (at A36) and 7% for θ = 45° (at A35 and A37), while for the SAS-S, this is about 22% for θ = 0° (at L3) and 14% for θ = 45° (at A35). For θ = 45°, CFD-S is found to consistently underpredict the breach percentage, with LES-S predicting values closer to WT-S than SAS-S. This is a reflection of the underprediction of C_pi as discussed in Subsection III D. Thus, LES-S outperforms SAS-S for the prediction of percentage breach duration.

Overall, the agreement between CFD-S and WT-S is acceptable for industrial applications when considering the possible uncertainties of measurements as well as simulations.

Hazardous pollutant containment zones should be maintained at a pressure lower than the outdoor atmospheric pressure to prevent pollutants from escaping to the outdoor environment. However, atmospheric wind conditions can cause breaching of the containment zone that is established through mechanical ventilation. This paper combines external wind pressure (P_e) time series on an internally depressurized building with a carefully designed ventilation network to analyze indoor pressure (P_i) and containment breach duration. The P_e time series and statistical parameters are obtained by reduced-scale wind tunnel (WT) tests performed on an idealized isolated building on the one hand, and by computational fluid dynamics (CFD simulations) by large-eddy simulation (LES) and scale-adaptive simulation (SAS) on the other hand. These P_e time series are used as input to the ventilation network model SYLVIA. The depressurization value used is −40 Pa, and the reference wind speed (U_ref) at building height is 12.65 m/s. The goal of the paper is threefold: (1) comparing the P_e data as obtained by WT and CFD in terms of time series and statistics; (2) assessing the impact of uncertainties in the P_e as generated by CFD on the resulting P_i and on the breach duration percentage; and (3) estimating the P_i statistics and breach duration percentage for an initial case study with the indoor depressurization of −40 Pa and a fixed U_ref of 12.65 m/s at building height. The objective (2) aims to determine whether, despite the inherent uncertainties in CFD results, the combinations LES-SYLVIA (LES-S) or SAS-SYLVIA (SAS-S) could be used as alternatives to the WT-SYLVIA (WT-S) combination in future studies. The results are discussed in terms of dimensionless pressure coefficients (C_pe and C_pi). It is worth noting that the study has the following limitations:

1. Only two wind directions (i.e., θ = 0° and θ = 45°) have been investigated. While these are two representative wind directions for the present building, for more geometrically complex buildings with less symmetry, additional wind directions should be considered.

2. The SAS simulations have been performed with the same settings as LES. However, a fine grid near the walls (as for LES) might not be needed for SAS, since RANS is applied in these regions. Therefore, the observations that SAS requires a longer computational time than LES should not be generalized. Further simulations could be performed in order to optimize the SAS grid by decreasing the computational time and improving the results.

3. This study has been performed for an idealized building, for a fixed depressurization value and for a fixed reference wind speed of 12.65 m/s.

Despite the above listed limitations, the study has reached its objectives based on which future studies for more complex building geometries, different depressurization values, different reference wind speeds, and wind directions can be considered. The following conclusions can be drawn from the study.

1. Mean and rms C_pe predicted by LES and SAS simulations are comparable to those predicted by WT tests; however, substantial deviations are observed for the maximum and minimum peak C_pe by LES and especially by SAS compared to the WT results.

2. In spite of the deviations in C_pe, the C_pi time series predicted by LES-S and SAS-S matches fairly well with that by WT-S for θ = 0°. Larger discrepancies, however, are identified for SAS-S for θ = 45°, consistent with those detected for C_pe.

3. Nevertheless, statistics of the C_pi predicted by LES-S and SAS-S show a rather good agreement with that predicted by WT-S. The absolute differences in terms of C_pi mean, rms, and peak between WT-S and CFD-S cases fall within the uncertainties associated with WT measurements. The largest observed absolute difference in C_pi statistics is 0.1, corresponding to the peak C_pi prediction when θ = 45°. A main source of uncertainty in the LES and SAS simulations is the use of the spectral synthesizer as inflow generation method that has been used as is, i.e., without any additional code manipulation. The rather good agreement in terms of C_pi indicates that its prediction is not very sensitive to the aforementioned deviations in the C_pe data used as input.

4. The percentage duration of a possible containment breach predicted by LES-S and SAS-S shows an overall fair agreement with that predicted by WT-S, with an average difference of 4% and 5%, respectively, from WT-S. LES-S is found to slightly outperform SAS-S in the prediction of containment breach probability, as higher differences between SAS-S and WT-S are observed at localized regions for θ = 0° and at all locations for θ = 45° compared to that between LES-S and WT-S. This suggests that in future studies, if needed and if more convenient, LES-S can be used instead of WT-S.

5. For this initial case study with the −40 Pa depressurization and U_ref = 12.65 m/s at the building height, the breach duration percentage can go up to 80%, which is a reason for concern and which stresses the importance of this type of studies.

The project was undertaken in collaboration with Institut National de Recherche et de Sécurité (INRS), France, and their support throughout this project is greatly appreciated. The authors acknowledge the partnership with ANSYS CFD. They also acknowledge the support received from the Institut de radioprotection et de sûreté nucléaire (IRSN) in relation to SYLVIA. The CFD simulations were carried out on the Dutch national e-infrastructure with the support of SURF Cooperative (Grant Nos. EINF-816 and NWO-2021.027/L1), and their support is greatly acknowledged.

The authors have no conflicts to disclose.

A. K. R. Jayakumari: Data curation (lead); Formal analysis (lead); Visualization (lead); Writing – original draft (lead). A. Ricci: Conceptualization (equal); Investigation (equal); Methodology (equal); Supervision (equal); Writing – review & editing (equal). R. Guichard: Conceptualization (equal); Funding acquisition (equal); Investigation (equal); Methodology (equal); Supervision (equal); Writing – review & editing (equal). S. Gillmeier: Conceptualization (equal); Investigation (equal); Methodology (equal); Supervision (equal); Writing – review & editing (equal). B. Blocken: Conceptualization (equal); Funding acquisition (equal); Investigation (equal); Methodology (equal); Supervision (equal); Writing – review & editing (equal).

The data that support the findings of this study are available from the corresponding author upon reasonable request.