On the Use of Computational Fluid Dynamics (CFD) Modelling to Design Improved Dry Powder Inhalers
David F Fletcher, Vishal Chaugule, Larissa Gomes dos Reis, Paul M Young, Daniela Traini, Julio Soria
AAccepted: Pharmaceutical Research (2021)
On the Use of Computational Fluid Dynamics(CFD) Modelling to Design Improved DryPowder Inhalers
David F Fletcher , Vishal Chaugule ,Larissa Gomes dos Reis , Paul M Young ,Daniela Traini , and Julio Soria School of Chemical and Biomolecular Engineering, The University of Sydney, Sydney,Australia Laboratory for Turbulence Research in Aerospace and Combustion (LTRAC),Department of Mechanical and Aerospace Engineering, Monash University, ClaytonCampus, Melbourne, Australia Respiratory Technology, Woolcock Institute of Medical Research and Discipline ofPharmacology, Faculty of Medicine and Health, The University of Sydney, Sydney,Australia * Corresponding author: [email protected]
AbstractPurpose
Computational Fluid Dynamics (CFD) simulations are performed toinvestigate the impact of adding a grid to a two-inlet dry powder in-haler (DPI). The purpose of the paper is to show the importance ofthe correct choice of closure model and modeling approach, as well asto perform validation against particle dispersion data obtained from in-vitro studies and flow velocity data obtained from particle imagevelocimetry (PIV) experiments.
Methods
CFD simulations are performed using the Ansys Fluent 2020R1 soft-ware package. Two RANS turbulence models (realisable k - (cid:15) and k - ω SST) and the Stress Blended Eddy Simulation (SBES) models areconsidered. Lagrangian particle tracking for both carrier and fine par-ticles is also performed.
Results a r X i v : . [ c s . C E ] J a n ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021)
Results a r X i v : . [ c s . C E ] J a n ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021) Excellent comparison with the PIV data is found for the SBES ap-proach and the particle tracking data are consistent with the disper-sion results, given the simplicity of the assumptions made.
Conclusions
This work shows the importance of selecting the correct turbulencemodelling approach and boundary conditions to obtain good agree-ment with PIV data for the flow-field exiting the device. With thisvalidated, the model can be used with much higher confidence to ex-plore the fluid and particle dynamics within the device.
Keywords: dry powder inhaler, CFD, turbulence models, SBES, particletracking
Abbreviations
API - Active Pharmaceutical IngredientsCC - Curvature CorrectionCFD - Computational Fluid DynamicsDPI - Dry Powder InhalerDPM - Discrete Phase ModelFPF - Fine Particle FractionLDV - Laser Doppler VelocimetryLES - Large Eddy SimulationLRN - Low Reynolds NumberNSE - Navier-Stokes EquationsPIV - Particle Image VelocimetryRANS - Reynolds-Averaged Navier-StokesSBES - Stress Blended Eddy SimulationSRS - Scale-Resolving SimulationSST - Shear Stress TransportURANS - Unsteady Reynolds-Averaged Navier-StokesWALE - Wall-Adapting Local Eddy-viscosity2 ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021)
API - Active Pharmaceutical IngredientsCC - Curvature CorrectionCFD - Computational Fluid DynamicsDPI - Dry Powder InhalerDPM - Discrete Phase ModelFPF - Fine Particle FractionLDV - Laser Doppler VelocimetryLES - Large Eddy SimulationLRN - Low Reynolds NumberNSE - Navier-Stokes EquationsPIV - Particle Image VelocimetryRANS - Reynolds-Averaged Navier-StokesSBES - Stress Blended Eddy SimulationSRS - Scale-Resolving SimulationSST - Shear Stress TransportURANS - Unsteady Reynolds-Averaged Navier-StokesWALE - Wall-Adapting Local Eddy-viscosity2 ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021)
Pharmaceutical aerosol generated through a dry powder inhaler (DPI) isa multi-phase flow comprising a continuous phase (air) and a disperse phase(particles), which contains the active pharmaceutical ingredients (API). Dur-ing aerosolization, there is an interaction between the two phases - the airflow contributes to the dispersion and deposition of the particles, and thepresence and motion of particles modulates the air flow-field. The transi-tion of local flow from laminar to turbulent and the high volume fraction ofparticles near the release point, relative to the fluid volume in a DPI, leadsto complex particle-flow interactions. In addition, these particle-flow inter-actions are symbiotic with the device design, the inhalation flow, and theformulation and properties of the drug which further increases its complex-ity. Experimental investigations of these phenomena have significant prac-tical challenges, thus computational modelling of the fluid flow and particledynamics has been performed to study these processes and optimize devicedelivery (1–3).The modelling of the continuous phase of a DPI has been performed us-ing Computational Fluid Dynamics (CFD), which has traditionally involvedsolving the Reynolds-Averaged Navier-Stokes (RANS) equations numerically,with suitable turbulence closure models. These equations are time-averagedforms of the governing continuity and momentum equations (Navier-Stokesequations (NSE)), and the turbulence model serves to close this system ofmean-flow equations. However, time-averaging leads to a loss of informa-tion and some turbulence models have limitations in accurately modellingturbulent swirling flows that are inherent in a DPI (4). These issues canbe mitigated by using Large Eddy Simulation (LES), that solves the filteredNSE and can resolve large-scale turbulence eddies and detailed flow struc-tures, depending on the applied local filter width. LES has been shown toprovide more high-fidelity information of the flow field compared with RANS,but it has not been widely used for DPI modelling because of the higher com-putational requirements, especially if it is applied in boundary layers (5).One of the earliest CFD studies on DPIs was conducted by Coates etal. (6) in which they studied the flow-field and particle trajectories in theAerolizer ® DPI for different design parameters of the inhaler mouthpieceand grid. The flow-field was simulated using the RANS approach with the k - ω Shear Stress Transport (SST) turbulence model (7) and with particlestracked using a Lagrangian approach. Flow field validation was carried outby comparing the simulation results with laser doppler velocimetry (LDV)3 ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021)
Pharmaceutical aerosol generated through a dry powder inhaler (DPI) isa multi-phase flow comprising a continuous phase (air) and a disperse phase(particles), which contains the active pharmaceutical ingredients (API). Dur-ing aerosolization, there is an interaction between the two phases - the airflow contributes to the dispersion and deposition of the particles, and thepresence and motion of particles modulates the air flow-field. The transi-tion of local flow from laminar to turbulent and the high volume fraction ofparticles near the release point, relative to the fluid volume in a DPI, leadsto complex particle-flow interactions. In addition, these particle-flow inter-actions are symbiotic with the device design, the inhalation flow, and theformulation and properties of the drug which further increases its complex-ity. Experimental investigations of these phenomena have significant prac-tical challenges, thus computational modelling of the fluid flow and particledynamics has been performed to study these processes and optimize devicedelivery (1–3).The modelling of the continuous phase of a DPI has been performed us-ing Computational Fluid Dynamics (CFD), which has traditionally involvedsolving the Reynolds-Averaged Navier-Stokes (RANS) equations numerically,with suitable turbulence closure models. These equations are time-averagedforms of the governing continuity and momentum equations (Navier-Stokesequations (NSE)), and the turbulence model serves to close this system ofmean-flow equations. However, time-averaging leads to a loss of informa-tion and some turbulence models have limitations in accurately modellingturbulent swirling flows that are inherent in a DPI (4). These issues canbe mitigated by using Large Eddy Simulation (LES), that solves the filteredNSE and can resolve large-scale turbulence eddies and detailed flow struc-tures, depending on the applied local filter width. LES has been shown toprovide more high-fidelity information of the flow field compared with RANS,but it has not been widely used for DPI modelling because of the higher com-putational requirements, especially if it is applied in boundary layers (5).One of the earliest CFD studies on DPIs was conducted by Coates etal. (6) in which they studied the flow-field and particle trajectories in theAerolizer ® DPI for different design parameters of the inhaler mouthpieceand grid. The flow-field was simulated using the RANS approach with the k - ω Shear Stress Transport (SST) turbulence model (7) and with particlestracked using a Lagrangian approach. Flow field validation was carried outby comparing the simulation results with laser doppler velocimetry (LDV)3 ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021) data at the exit of the device. An increase in the size of the grid openingsreduced the flow straightening effect, and also the turbulence intensity, justdownstream of the grid. Consequently, particle collisions with the grid alsodecreased, but led to an increase in particle-wall collisions in the mouthpiece.This balancing effect, of lower turbulence intensity and particle-grid collisionswith higher particle-wall collisions in the mouthpiece, was found to result insimilar values of fine particle fraction (FPF) for these design changes.In a follow-up study on the effect of flow rates on DPI performance (8),they reported the expected increase of turbulence intensity, integral scalestrain rates and particle-wall collisions with an increase in air flow rates. Thisled to an improvement in powder de-agglomeration and thus its dispersionin the flow, but only up to a flow rate of 65 l/min. A later study by Coateset al. (9) on the effect of tangential inlet size on the inhaler flow-field showedthat a reduction of inlet area size resulted in higher turbulence intensitiesand velocity of particle-wall collisions in the region just downstream of theinlets.A RANS approach using the k - ω SST turbulence model was used byDonovan et al. (10) to study the flow-field and particle trajectories in theAerolizer ® and Handihaler ® DPI geometries. The particles were modelledusing a Stokesian drag law with non-spherical corrections to account for parti-cle shape effects. The swirling flow in the Aerolizer ® intensified particle-wallcollisions, which lead to an improvement in drug detachment, whereas theabsence of swirling flow in the Handihaler ® lead to fewer particle collisionswith the inhaler wall, and thus lower aerosol performance. It was also shownthat increasing the mean particle diameter increased the number of particle-wall collisions due to the increased Stokes number leading to more ballistictrajectories.The application of RANS with various models for turbulent flow (stan-dard k - (cid:15) , RNG k - (cid:15) and k- ω SST) was used by Milenkovic et al. (5) tomodel the flow in a Turbuhaler ® DPI geometry. They also used LES, butfor only a single parametric case, which was then compared with the RANSsolutions. The LES generated radial and tangential flows within the deviceshowed enhanced presence of eddies and secondary flow structures that weremost similar to those obtained with the k - ω SST model. In a later study,Milenkovic et al. (11) modelled the dynamic flow in the same DPI geome-try instead of a steady flow. This dynamic flow comprised an initial rapidincrease of flow rate that gradually plateaued to a steady flow rate, and wassimulated by imposing dynamic outlet pressures. They showed that the nor-malised dynamic flow-field velocities were similar for peak inspiratory flow4 ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021)
Pharmaceutical aerosol generated through a dry powder inhaler (DPI) isa multi-phase flow comprising a continuous phase (air) and a disperse phase(particles), which contains the active pharmaceutical ingredients (API). Dur-ing aerosolization, there is an interaction between the two phases - the airflow contributes to the dispersion and deposition of the particles, and thepresence and motion of particles modulates the air flow-field. The transi-tion of local flow from laminar to turbulent and the high volume fraction ofparticles near the release point, relative to the fluid volume in a DPI, leadsto complex particle-flow interactions. In addition, these particle-flow inter-actions are symbiotic with the device design, the inhalation flow, and theformulation and properties of the drug which further increases its complex-ity. Experimental investigations of these phenomena have significant prac-tical challenges, thus computational modelling of the fluid flow and particledynamics has been performed to study these processes and optimize devicedelivery (1–3).The modelling of the continuous phase of a DPI has been performed us-ing Computational Fluid Dynamics (CFD), which has traditionally involvedsolving the Reynolds-Averaged Navier-Stokes (RANS) equations numerically,with suitable turbulence closure models. These equations are time-averagedforms of the governing continuity and momentum equations (Navier-Stokesequations (NSE)), and the turbulence model serves to close this system ofmean-flow equations. However, time-averaging leads to a loss of informa-tion and some turbulence models have limitations in accurately modellingturbulent swirling flows that are inherent in a DPI (4). These issues canbe mitigated by using Large Eddy Simulation (LES), that solves the filteredNSE and can resolve large-scale turbulence eddies and detailed flow struc-tures, depending on the applied local filter width. LES has been shown toprovide more high-fidelity information of the flow field compared with RANS,but it has not been widely used for DPI modelling because of the higher com-putational requirements, especially if it is applied in boundary layers (5).One of the earliest CFD studies on DPIs was conducted by Coates etal. (6) in which they studied the flow-field and particle trajectories in theAerolizer ® DPI for different design parameters of the inhaler mouthpieceand grid. The flow-field was simulated using the RANS approach with the k - ω Shear Stress Transport (SST) turbulence model (7) and with particlestracked using a Lagrangian approach. Flow field validation was carried outby comparing the simulation results with laser doppler velocimetry (LDV)3 ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021) data at the exit of the device. An increase in the size of the grid openingsreduced the flow straightening effect, and also the turbulence intensity, justdownstream of the grid. Consequently, particle collisions with the grid alsodecreased, but led to an increase in particle-wall collisions in the mouthpiece.This balancing effect, of lower turbulence intensity and particle-grid collisionswith higher particle-wall collisions in the mouthpiece, was found to result insimilar values of fine particle fraction (FPF) for these design changes.In a follow-up study on the effect of flow rates on DPI performance (8),they reported the expected increase of turbulence intensity, integral scalestrain rates and particle-wall collisions with an increase in air flow rates. Thisled to an improvement in powder de-agglomeration and thus its dispersionin the flow, but only up to a flow rate of 65 l/min. A later study by Coateset al. (9) on the effect of tangential inlet size on the inhaler flow-field showedthat a reduction of inlet area size resulted in higher turbulence intensitiesand velocity of particle-wall collisions in the region just downstream of theinlets.A RANS approach using the k - ω SST turbulence model was used byDonovan et al. (10) to study the flow-field and particle trajectories in theAerolizer ® and Handihaler ® DPI geometries. The particles were modelledusing a Stokesian drag law with non-spherical corrections to account for parti-cle shape effects. The swirling flow in the Aerolizer ® intensified particle-wallcollisions, which lead to an improvement in drug detachment, whereas theabsence of swirling flow in the Handihaler ® lead to fewer particle collisionswith the inhaler wall, and thus lower aerosol performance. It was also shownthat increasing the mean particle diameter increased the number of particle-wall collisions due to the increased Stokes number leading to more ballistictrajectories.The application of RANS with various models for turbulent flow (stan-dard k - (cid:15) , RNG k - (cid:15) and k- ω SST) was used by Milenkovic et al. (5) tomodel the flow in a Turbuhaler ® DPI geometry. They also used LES, butfor only a single parametric case, which was then compared with the RANSsolutions. The LES generated radial and tangential flows within the deviceshowed enhanced presence of eddies and secondary flow structures that weremost similar to those obtained with the k - ω SST model. In a later study,Milenkovic et al. (11) modelled the dynamic flow in the same DPI geome-try instead of a steady flow. This dynamic flow comprised an initial rapidincrease of flow rate that gradually plateaued to a steady flow rate, and wassimulated by imposing dynamic outlet pressures. They showed that the nor-malised dynamic flow-field velocities were similar for peak inspiratory flow4 ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021) rates (PIFR) of 30, 50 and 70 l/min.A Lagrangian approach with one-way coupling was used by Sommerfeldand Schmalfuß (12) to determine the fluid stresses experienced by the car-rier particles along their path through a DPI. The RANS equations with the k - ω SST turbulence model were solved for steady flow through the inhaler.Their results indicated that wall collisions largely prevailed in particle mo-tion, wherein de-agglomeration of drug powder mainly occurred due to wallimpacts in the swirl chamber and with the grid placed just after it. The wall-collision frequency of the particles was found to increase with particle sizedue to their increased inertia, but this reduced their wall-impact velocities.Longest et al. (13) performed CFD simulations using the low Reynoldsnumber (LRN) k - ω turbulence model and employed a Lagrangian particletracking algorithm to predict individual particle trajectories and determineparticle interaction with the mean turbulent flow-field. Six different inhalerdesigns were studied and they explored both turbulence and impaction aspotential particle break-up mechanisms. It was found that turbulence wasthe primary de-aggregation mechanism for carrier-free particles, with highturbulence kinetic energy, long exposure time, and small characteristic eddylength scales. However, in a later study by Longest and Farkas (14), onpowder dispersion in a dose aerosolization and containment unit, they foundan undesirable increase in aerodynamic diameter when flow turbulence wasincreased.It is important to keep in mind that CFD simulations can only be usedwith confidence once they have been validated. It is for this reason that we areemploying three complementary methods in our current investigation of theimpact of inhaler design on performance. CFD can provide information onthe flow field and particle behaviour both inside and outside of the inhaler,however there are many uncertainties pertaining to turbulence modellingand the dynamics and break-up of particle agglomerates. Particle imagevelocimetry (PIV) studies provide high quality data on the flow field outsideof the device. Finally, in-vitro studies provide a means of studying deviceperformance for a powder formulation and the interaction of the inhaledparticle cloud with the respiratory tract. Ultimately, models should reliablydetermine particle deposition inside the device as this in turn affects thedetermination of emitted FPF from simulations. The size, distribution andvelocity of aerosol particles upon exiting the DPI mouthpiece govern theirmotion and deposition in the respiratory tract, which is of utmost importancein assessing the performance of the DPI.In a previous study (15) we presented both PIV data and in-vitro studies5 ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021)
Pharmaceutical aerosol generated through a dry powder inhaler (DPI) isa multi-phase flow comprising a continuous phase (air) and a disperse phase(particles), which contains the active pharmaceutical ingredients (API). Dur-ing aerosolization, there is an interaction between the two phases - the airflow contributes to the dispersion and deposition of the particles, and thepresence and motion of particles modulates the air flow-field. The transi-tion of local flow from laminar to turbulent and the high volume fraction ofparticles near the release point, relative to the fluid volume in a DPI, leadsto complex particle-flow interactions. In addition, these particle-flow inter-actions are symbiotic with the device design, the inhalation flow, and theformulation and properties of the drug which further increases its complex-ity. Experimental investigations of these phenomena have significant prac-tical challenges, thus computational modelling of the fluid flow and particledynamics has been performed to study these processes and optimize devicedelivery (1–3).The modelling of the continuous phase of a DPI has been performed us-ing Computational Fluid Dynamics (CFD), which has traditionally involvedsolving the Reynolds-Averaged Navier-Stokes (RANS) equations numerically,with suitable turbulence closure models. These equations are time-averagedforms of the governing continuity and momentum equations (Navier-Stokesequations (NSE)), and the turbulence model serves to close this system ofmean-flow equations. However, time-averaging leads to a loss of informa-tion and some turbulence models have limitations in accurately modellingturbulent swirling flows that are inherent in a DPI (4). These issues canbe mitigated by using Large Eddy Simulation (LES), that solves the filteredNSE and can resolve large-scale turbulence eddies and detailed flow struc-tures, depending on the applied local filter width. LES has been shown toprovide more high-fidelity information of the flow field compared with RANS,but it has not been widely used for DPI modelling because of the higher com-putational requirements, especially if it is applied in boundary layers (5).One of the earliest CFD studies on DPIs was conducted by Coates etal. (6) in which they studied the flow-field and particle trajectories in theAerolizer ® DPI for different design parameters of the inhaler mouthpieceand grid. The flow-field was simulated using the RANS approach with the k - ω Shear Stress Transport (SST) turbulence model (7) and with particlestracked using a Lagrangian approach. Flow field validation was carried outby comparing the simulation results with laser doppler velocimetry (LDV)3 ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021) data at the exit of the device. An increase in the size of the grid openingsreduced the flow straightening effect, and also the turbulence intensity, justdownstream of the grid. Consequently, particle collisions with the grid alsodecreased, but led to an increase in particle-wall collisions in the mouthpiece.This balancing effect, of lower turbulence intensity and particle-grid collisionswith higher particle-wall collisions in the mouthpiece, was found to result insimilar values of fine particle fraction (FPF) for these design changes.In a follow-up study on the effect of flow rates on DPI performance (8),they reported the expected increase of turbulence intensity, integral scalestrain rates and particle-wall collisions with an increase in air flow rates. Thisled to an improvement in powder de-agglomeration and thus its dispersionin the flow, but only up to a flow rate of 65 l/min. A later study by Coateset al. (9) on the effect of tangential inlet size on the inhaler flow-field showedthat a reduction of inlet area size resulted in higher turbulence intensitiesand velocity of particle-wall collisions in the region just downstream of theinlets.A RANS approach using the k - ω SST turbulence model was used byDonovan et al. (10) to study the flow-field and particle trajectories in theAerolizer ® and Handihaler ® DPI geometries. The particles were modelledusing a Stokesian drag law with non-spherical corrections to account for parti-cle shape effects. The swirling flow in the Aerolizer ® intensified particle-wallcollisions, which lead to an improvement in drug detachment, whereas theabsence of swirling flow in the Handihaler ® lead to fewer particle collisionswith the inhaler wall, and thus lower aerosol performance. It was also shownthat increasing the mean particle diameter increased the number of particle-wall collisions due to the increased Stokes number leading to more ballistictrajectories.The application of RANS with various models for turbulent flow (stan-dard k - (cid:15) , RNG k - (cid:15) and k- ω SST) was used by Milenkovic et al. (5) tomodel the flow in a Turbuhaler ® DPI geometry. They also used LES, butfor only a single parametric case, which was then compared with the RANSsolutions. The LES generated radial and tangential flows within the deviceshowed enhanced presence of eddies and secondary flow structures that weremost similar to those obtained with the k - ω SST model. In a later study,Milenkovic et al. (11) modelled the dynamic flow in the same DPI geome-try instead of a steady flow. This dynamic flow comprised an initial rapidincrease of flow rate that gradually plateaued to a steady flow rate, and wassimulated by imposing dynamic outlet pressures. They showed that the nor-malised dynamic flow-field velocities were similar for peak inspiratory flow4 ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021) rates (PIFR) of 30, 50 and 70 l/min.A Lagrangian approach with one-way coupling was used by Sommerfeldand Schmalfuß (12) to determine the fluid stresses experienced by the car-rier particles along their path through a DPI. The RANS equations with the k - ω SST turbulence model were solved for steady flow through the inhaler.Their results indicated that wall collisions largely prevailed in particle mo-tion, wherein de-agglomeration of drug powder mainly occurred due to wallimpacts in the swirl chamber and with the grid placed just after it. The wall-collision frequency of the particles was found to increase with particle sizedue to their increased inertia, but this reduced their wall-impact velocities.Longest et al. (13) performed CFD simulations using the low Reynoldsnumber (LRN) k - ω turbulence model and employed a Lagrangian particletracking algorithm to predict individual particle trajectories and determineparticle interaction with the mean turbulent flow-field. Six different inhalerdesigns were studied and they explored both turbulence and impaction aspotential particle break-up mechanisms. It was found that turbulence wasthe primary de-aggregation mechanism for carrier-free particles, with highturbulence kinetic energy, long exposure time, and small characteristic eddylength scales. However, in a later study by Longest and Farkas (14), onpowder dispersion in a dose aerosolization and containment unit, they foundan undesirable increase in aerodynamic diameter when flow turbulence wasincreased.It is important to keep in mind that CFD simulations can only be usedwith confidence once they have been validated. It is for this reason that we areemploying three complementary methods in our current investigation of theimpact of inhaler design on performance. CFD can provide information onthe flow field and particle behaviour both inside and outside of the inhaler,however there are many uncertainties pertaining to turbulence modellingand the dynamics and break-up of particle agglomerates. Particle imagevelocimetry (PIV) studies provide high quality data on the flow field outsideof the device. Finally, in-vitro studies provide a means of studying deviceperformance for a powder formulation and the interaction of the inhaledparticle cloud with the respiratory tract. Ultimately, models should reliablydetermine particle deposition inside the device as this in turn affects thedetermination of emitted FPF from simulations. The size, distribution andvelocity of aerosol particles upon exiting the DPI mouthpiece govern theirmotion and deposition in the respiratory tract, which is of utmost importancein assessing the performance of the DPI.In a previous study (15) we presented both PIV data and in-vitro studies5 ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021) for four different inhalers having two tangential inlets, six tangential inlets,two inlets with an inlet grid and two inlets with an exit grid. Given that thetwo and six inlet cases showed very similar results, in this paper we presenta CFD study of the two inlet cases and compare our results with both the in-vitro and PIV data. The inhaler geometries studied here are shown inFigure 1. (a) no-grid (b) entry-grid (c) exit-grid Figure 1:
DPI device models examined in this study
The PIV experimental setup, which the CFD model geometry replicates,is shown in Fig. 2. The DPI device models used in the PIV experiments weregeometrically scaled-up three times to that of the original models shown inFig. 1. Each model was placed in a tank with a closed-loop water flow sys-tem, wherein a steady water flow-rate was maintained through the model toattain a Reynolds number of ≈ ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021)
The PIV experimental setup, which the CFD model geometry replicates,is shown in Fig. 2. The DPI device models used in the PIV experiments weregeometrically scaled-up three times to that of the original models shown inFig. 1. Each model was placed in a tank with a closed-loop water flow sys-tem, wherein a steady water flow-rate was maintained through the model toattain a Reynolds number of ≈ ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021) Figure 2: PIV experimental setup
In all cases time-dependent simulations were performed as, not unex-pectedly, convergence to a steady flow could not be achieved. Therefore,simulations were started in steady mode to establish an initial flow field andthen time-dependent simulations were performed. Once these had estab-lished realistic flow fields, transient statistics were evaluated to enable themean flow velocities and the Reynolds stresses to be obtained for compari-son with the PIV experimental data. All simulations were performed usingAnsys ® Fluent 2020R1 (16) and were run in double precision to eliminaterounding error.
Based on the above literature review it was decided to investigate threedifferent turbulence modelling approaches. The realisable k - (cid:15) (17) and the k - ω SST (7) models were chosen as being representative of the unsteady-RANS (URANS) modelling approaches. It is clear that the k - ω SST modelis the most widely used, however a k - (cid:15) model was also included as this ap-proach is widely used in internal flow simulations. It is well-known thatthese two-equation models do not capture swirling flow correctly, so bothwere solved with a Curvature Correction (CC) term included (18), as it hasbeen shown to correctly capture the swirl profile in cyclones (19). WhilstReynolds stress models can in theory provide good solutions for swirling7 ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021)
Based on the above literature review it was decided to investigate threedifferent turbulence modelling approaches. The realisable k - (cid:15) (17) and the k - ω SST (7) models were chosen as being representative of the unsteady-RANS (URANS) modelling approaches. It is clear that the k - ω SST modelis the most widely used, however a k - (cid:15) model was also included as this ap-proach is widely used in internal flow simulations. It is well-known thatthese two-equation models do not capture swirling flow correctly, so bothwere solved with a Curvature Correction (CC) term included (18), as it hasbeen shown to correctly capture the swirl profile in cyclones (19). WhilstReynolds stress models can in theory provide good solutions for swirling7 ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021) flows they are renowned for being numerically stiff and hard to solve, so theywere not investigated in this study.In order to investigate the impact of using a Scale-Resolving Simulation(SRS) approach, simulations were made using the Stress Blended Eddy Sim-ulation (SBES) approach (20) as this takes advantage of the best aspectsof the RANS and LES approaches. In the near wall region, where the flowis attached and LES simulations are prohibitively expensive, the k - ω SSTmodel provides the eddy viscosity. Away from the wall, in regions where themesh is sufficiently fine, the model blends the eddy viscosity with that froman LES modelling approach. The subgrid-scale closure of the Wall-AdaptingLocal Eddy-viscosity (WALE) model (21) was used.In all cases the computational mesh was constructed so that there weresufficient inflation layers adjacent to the inhaler walls that the y + values werelow enough for the flow to be resolved up to the wall in the k - ω models. Carewas taken to ensure that the transition to SRS occurred where expected andthat in this case the unresolved turbulence led to an eddy viscosity consistentwith the LES approach. A recent study that highlights the best practicesand checks to be performed can be consulted for more detail (22). Once the flow was established, the Discrete Phase Model (DPM) wasused to perform time-dependent particle tracking in the time-dependent flowfor the SBES simulations, assuming a drag model appropriate for smoothspheres. The simulations were performed for a low particle loading using one-way coupling as the current work compares the flow field with PIV data inwhich the drug particles are absent. As the large scale turbulence structuresare captured in these simulations, no additional turbulent dispersion wasadded. At the walls, particles were assumed to reflect with coefficients ofrestitution of 0.9 in the tangential direction and 0.7 in the normal direction,based on values determined for typical drug formulations (23). User-definedfunctions were used to capture the number of impacts and the impact kineticenergy of the particles.Two different sets of particle tracking were performed. Firstly, 280 µ mdiameter particles were released from the spherical end cap of the inhaler(dosing cup) to represent the carrier particles, and their impact behaviourwith the wall and grid (if present) was studied. Particle de-agglomerationoccurs when carrier particles impact the wall or each other, knocking activedrug particle off the carrier particle. Here we investigated the importance of8 ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021)
Based on the above literature review it was decided to investigate threedifferent turbulence modelling approaches. The realisable k - (cid:15) (17) and the k - ω SST (7) models were chosen as being representative of the unsteady-RANS (URANS) modelling approaches. It is clear that the k - ω SST modelis the most widely used, however a k - (cid:15) model was also included as this ap-proach is widely used in internal flow simulations. It is well-known thatthese two-equation models do not capture swirling flow correctly, so bothwere solved with a Curvature Correction (CC) term included (18), as it hasbeen shown to correctly capture the swirl profile in cyclones (19). WhilstReynolds stress models can in theory provide good solutions for swirling7 ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021) flows they are renowned for being numerically stiff and hard to solve, so theywere not investigated in this study.In order to investigate the impact of using a Scale-Resolving Simulation(SRS) approach, simulations were made using the Stress Blended Eddy Sim-ulation (SBES) approach (20) as this takes advantage of the best aspectsof the RANS and LES approaches. In the near wall region, where the flowis attached and LES simulations are prohibitively expensive, the k - ω SSTmodel provides the eddy viscosity. Away from the wall, in regions where themesh is sufficiently fine, the model blends the eddy viscosity with that froman LES modelling approach. The subgrid-scale closure of the Wall-AdaptingLocal Eddy-viscosity (WALE) model (21) was used.In all cases the computational mesh was constructed so that there weresufficient inflation layers adjacent to the inhaler walls that the y + values werelow enough for the flow to be resolved up to the wall in the k - ω models. Carewas taken to ensure that the transition to SRS occurred where expected andthat in this case the unresolved turbulence led to an eddy viscosity consistentwith the LES approach. A recent study that highlights the best practicesand checks to be performed can be consulted for more detail (22). Once the flow was established, the Discrete Phase Model (DPM) wasused to perform time-dependent particle tracking in the time-dependent flowfor the SBES simulations, assuming a drag model appropriate for smoothspheres. The simulations were performed for a low particle loading using one-way coupling as the current work compares the flow field with PIV data inwhich the drug particles are absent. As the large scale turbulence structuresare captured in these simulations, no additional turbulent dispersion wasadded. At the walls, particles were assumed to reflect with coefficients ofrestitution of 0.9 in the tangential direction and 0.7 in the normal direction,based on values determined for typical drug formulations (23). User-definedfunctions were used to capture the number of impacts and the impact kineticenergy of the particles.Two different sets of particle tracking were performed. Firstly, 280 µ mdiameter particles were released from the spherical end cap of the inhaler(dosing cup) to represent the carrier particles, and their impact behaviourwith the wall and grid (if present) was studied. Particle de-agglomerationoccurs when carrier particles impact the wall or each other, knocking activedrug particle off the carrier particle. Here we investigated the importance of8 ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021) wall impact by recording both the average number of wall impacts and theaverage impact kinetic energy of the particles. Secondly, 1 . µ m diameterparticles were released from an annulus one nozzle diameter upstream ofthe mouthpiece exit, occupying the outer 20% of the device mouthpiece torepresent the fine particles. This simulation was made to investigate thesubsequent dispersion of these particles assuming they had been releasedfrom wall impaction and had subsequently travelled along the wall region.In both cases a particle density of 1540 kg m − was used, based on that forlactose (24). The model geometry was created to mirror that of the PIV experiment,briefly described in Section 2.1, but for an incompressible fluid of air, atambient conditions. The Reynolds number based on the jet diameter D a was 8400, as used experimentally. The geometry used, showing the externalsurface mesh, is given in Figure 3(a). A spherical region of ambient air ismodelled around the inlet region, as it was found that applying boundaryconditions at the inlets of the inhaler led to an over-constrained flow in thatregion. The air exiting the device enters a box, just as was used in the PIVexperiments, in order provide the same downstream flow domain to allowdirect comparison of the jet behaviour with the experimental data. Figure3(b) shows a section through the computational mesh for the case with a gridat the exit, showing the poly-hexcore structure used, with hexahedral meshin the important central regions, connected to inflation mesh at the wallsby a layer of polyhedra. Local mesh controls were applied to ensure goodresolution where needed. Based on mesh studies, the final mesh comprised ∼ ∼ y + values. For the SST model y + < y + <
3, meaning that the model was resolving the flow to the wall.For the realisable k - (cid:15) , 11 < y + < ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021)
3, meaning that the model was resolving the flow to the wall.For the realisable k - (cid:15) , 11 < y + < ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021) (a) (b)Figure 3: CFD model geometry: (a) Geometry ; (b) Mesh method to achieve high accuracy. A second order differencing scheme wasused for the pressure, a bounded central differencing scheme for momentum,a second order upwind scheme for the turbulence quantities and a boundedsecond order implicit scheme for the transient terms. The solution requiredthe use of small time steps ( ∼ µ s) and typically 5 - 8 iterations per timestep. Initially we investigated the effect of the choice of the turbulence mod-elling approach. Figure 4 shows a comparison of the time-averaged axial U and radial V velocity components predicted by the CFD modelling withthe PIV data. The axial and radial coordinates are represented by x and y ,respectively. The velocity components have been normalised by the jet-exitmean velocity U a , and the spatial coordinates by the jet-exit diameter D a .Comparisons are presented at two representative downstream lines, locatedjust after the exit from the device and two diameters further downstream. Itis evident that in all cases the SBES predictions are closer to the experimental10 ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021)
3, meaning that the model was resolving the flow to the wall.For the realisable k - (cid:15) , 11 < y + < ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021) (a) (b)Figure 3: CFD model geometry: (a) Geometry ; (b) Mesh method to achieve high accuracy. A second order differencing scheme wasused for the pressure, a bounded central differencing scheme for momentum,a second order upwind scheme for the turbulence quantities and a boundedsecond order implicit scheme for the transient terms. The solution requiredthe use of small time steps ( ∼ µ s) and typically 5 - 8 iterations per timestep. Initially we investigated the effect of the choice of the turbulence mod-elling approach. Figure 4 shows a comparison of the time-averaged axial U and radial V velocity components predicted by the CFD modelling withthe PIV data. The axial and radial coordinates are represented by x and y ,respectively. The velocity components have been normalised by the jet-exitmean velocity U a , and the spatial coordinates by the jet-exit diameter D a .Comparisons are presented at two representative downstream lines, locatedjust after the exit from the device and two diameters further downstream. Itis evident that in all cases the SBES predictions are closer to the experimental10 ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021) data. In particular the realisable k - (cid:15) URANS models tends to over-predictthe back-flow at the device outlet and the radial velocities distributions aremuch closer to the measured data for the SBES model. Given the importanceof the prediction of the jet spreading rate, the use of the URANS models wasdiscontinued.
The impact of the grid on the flow field is shown in Figure 5, whichpresents the axial and swirl velocity components on a centre-plane. FromFigure 5(a) it is apparent that the case with no-grid shows a large vortexbreakdown region at the exit of the device which leads to back-flow in thecentral region and as a consequence the wide dispersion of the axial flow.The entry-grid case shows much reduced jet spreading and the exit-grid caseshows focusing of the high velocity jet generated by the grid towards thecentral axis. Both flow fields for devices with grids are potentially beneficialin that they are likely to focus particles along the centre of the jet.The swirl velocities, given in Figure 5(b), show the strong swirling flowexiting the device in the absence of a grid and that it is significantly reducedby the presence of the grid. In the entry-grid case the region of strong swirlis small and this may have an effect on particle de-agglomeration, whereasthe exit-grid model shows strong swirl within the device being suppressed atthe exit. 11 ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021)
The impact of the grid on the flow field is shown in Figure 5, whichpresents the axial and swirl velocity components on a centre-plane. FromFigure 5(a) it is apparent that the case with no-grid shows a large vortexbreakdown region at the exit of the device which leads to back-flow in thecentral region and as a consequence the wide dispersion of the axial flow.The entry-grid case shows much reduced jet spreading and the exit-grid caseshows focusing of the high velocity jet generated by the grid towards thecentral axis. Both flow fields for devices with grids are potentially beneficialin that they are likely to focus particles along the centre of the jet.The swirl velocities, given in Figure 5(b), show the strong swirling flowexiting the device in the absence of a grid and that it is significantly reducedby the presence of the grid. In the entry-grid case the region of strong swirlis small and this may have an effect on particle de-agglomeration, whereasthe exit-grid model shows strong swirl within the device being suppressed atthe exit. 11 ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021)
Figure 4: Impact of the turbulence model on the comparison with the PIV data. Mean velocities for the no-grid model : mean axial velocity at (a) x/D a = 0; (b) x/D a = 2; mean radial velocity at (c) x/D a =0; (d) x/D a = 2; RKE; SST; SBES; PIV. Validation of the above flow fields was performed via comparison withdetailed PIV data. Figure 6 shows a comparison of the mean axial and radialvelocity components with the PIV data. It is evident that in all cases there isgood agreement between simulations and experiment. Mean axial velocitiesare well predicted with the worse agreement being a slight under-predictionof the central values at x/D a = 3 for the entry-grid case. There are alsosome differences in the radial velocity in this case but the velocities are smalland much less important in determining the flow field.12 ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021)
Figure 4: Impact of the turbulence model on the comparison with the PIV data. Mean velocities for the no-grid model : mean axial velocity at (a) x/D a = 0; (b) x/D a = 2; mean radial velocity at (c) x/D a =0; (d) x/D a = 2; RKE; SST; SBES; PIV. Validation of the above flow fields was performed via comparison withdetailed PIV data. Figure 6 shows a comparison of the mean axial and radialvelocity components with the PIV data. It is evident that in all cases there isgood agreement between simulations and experiment. Mean axial velocitiesare well predicted with the worse agreement being a slight under-predictionof the central values at x/D a = 3 for the entry-grid case. There are alsosome differences in the radial velocity in this case but the velocities are smalland much less important in determining the flow field.12 ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021) (a)(b)Figure 5: Flow-field contour plots: (a) mean axial velocity ; (b) mean swirl velocity Figures 7 and 8 show the axial and radial velocity fluctuations andReynolds stress comparisons with the PIV data. The best agreement isobserved for the no-grid case. However, whilst there are some deviations inthe cases where grids are present, these are relatively small and are mostpronounced close to the device in the exit-grid case. In this case, small de-viations of measuring locations and fabrication tolerances would have themost pronounced effect. What is clear is that the CFD results correctlycapture the magnitude and trends of these quantities in all cases, providing13 ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021)
Figure 4: Impact of the turbulence model on the comparison with the PIV data. Mean velocities for the no-grid model : mean axial velocity at (a) x/D a = 0; (b) x/D a = 2; mean radial velocity at (c) x/D a =0; (d) x/D a = 2; RKE; SST; SBES; PIV. Validation of the above flow fields was performed via comparison withdetailed PIV data. Figure 6 shows a comparison of the mean axial and radialvelocity components with the PIV data. It is evident that in all cases there isgood agreement between simulations and experiment. Mean axial velocitiesare well predicted with the worse agreement being a slight under-predictionof the central values at x/D a = 3 for the entry-grid case. There are alsosome differences in the radial velocity in this case but the velocities are smalland much less important in determining the flow field.12 ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021) (a)(b)Figure 5: Flow-field contour plots: (a) mean axial velocity ; (b) mean swirl velocity Figures 7 and 8 show the axial and radial velocity fluctuations andReynolds stress comparisons with the PIV data. The best agreement isobserved for the no-grid case. However, whilst there are some deviations inthe cases where grids are present, these are relatively small and are mostpronounced close to the device in the exit-grid case. In this case, small de-viations of measuring locations and fabrication tolerances would have themost pronounced effect. What is clear is that the CFD results correctlycapture the magnitude and trends of these quantities in all cases, providing13 ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021)
Figure 6: Mean axial and radial velocities for: (a) and (d) no-grid ; (b) and (e) entry-grid ; (c) and (f) exit-grid models; SBES: x/D a = 0, x/D a = 1, x/D a = 3; PIV: x/D a = 0, x/D a = 1, x/D a = 3. confidence for it to be used to investigate the entire flow field. The measured pressure drop data are compared with the mean valuesobtained from the simulation in Figure 9 for an air flow rate of 60 l min − .In the absence of a grid the values are very close, while the trend is correctlypredicted, the value is under-predicted by about 35%, for the two cases witha grid. The reason for this is unclear but is most likely related to smalldifferences between the CAD geometry used to construct the CFD modeland the 3D-printed physical device model, and the surface roughness of thephysical model. 14 ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021)
Figure 6: Mean axial and radial velocities for: (a) and (d) no-grid ; (b) and (e) entry-grid ; (c) and (f) exit-grid models; SBES: x/D a = 0, x/D a = 1, x/D a = 3; PIV: x/D a = 0, x/D a = 1, x/D a = 3. confidence for it to be used to investigate the entire flow field. The measured pressure drop data are compared with the mean valuesobtained from the simulation in Figure 9 for an air flow rate of 60 l min − .In the absence of a grid the values are very close, while the trend is correctlypredicted, the value is under-predicted by about 35%, for the two cases witha grid. The reason for this is unclear but is most likely related to smalldifferences between the CAD geometry used to construct the CFD modeland the 3D-printed physical device model, and the surface roughness of thephysical model. 14 ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021) Figure 7: RMS axial and radial fluctuating velocities for: (a) and (d) no-grid ; (b) and (e) entry-grid ; (c)and (f) exit-grid models; SBES: x/D a = 0, x/D a = 1, x/D a = 3; PIV: x/D a = 0, x/D a = 1, x/D a = 3.Figure 8: Reynolds shear-stress for: (a) no-grid ; (b) entry-grid ; (c) exit-grid models; SBES: x/D a = 0, x/D a = 1, x/D a = 3; PIV: x/D a = 0, x/D a = 1, x/D a = 3. As discussed in Section 2.2.2, carrier particles were released in the dosingcup of the device and their paths were tracked to collect data on spreading15 ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021)
Figure 6: Mean axial and radial velocities for: (a) and (d) no-grid ; (b) and (e) entry-grid ; (c) and (f) exit-grid models; SBES: x/D a = 0, x/D a = 1, x/D a = 3; PIV: x/D a = 0, x/D a = 1, x/D a = 3. confidence for it to be used to investigate the entire flow field. The measured pressure drop data are compared with the mean valuesobtained from the simulation in Figure 9 for an air flow rate of 60 l min − .In the absence of a grid the values are very close, while the trend is correctlypredicted, the value is under-predicted by about 35%, for the two cases witha grid. The reason for this is unclear but is most likely related to smalldifferences between the CAD geometry used to construct the CFD modeland the 3D-printed physical device model, and the surface roughness of thephysical model. 14 ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021) Figure 7: RMS axial and radial fluctuating velocities for: (a) and (d) no-grid ; (b) and (e) entry-grid ; (c)and (f) exit-grid models; SBES: x/D a = 0, x/D a = 1, x/D a = 3; PIV: x/D a = 0, x/D a = 1, x/D a = 3.Figure 8: Reynolds shear-stress for: (a) no-grid ; (b) entry-grid ; (c) exit-grid models; SBES: x/D a = 0, x/D a = 1, x/D a = 3; PIV: x/D a = 0, x/D a = 1, x/D a = 3. As discussed in Section 2.2.2, carrier particles were released in the dosingcup of the device and their paths were tracked to collect data on spreading15 ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021)
Figure 9: Pressure drop across the device models: Measured; CFD. and wall impacts. Figure 10(a) shows the radial distribution of particlesacross the device at the exit and one jet-exit diameter downstream. For allthree cases the exit distribution is very similar with particles clustered aroundthe device wall. Even in the case with an entry-grid there is sufficient swirlto keep the particles at the wall. However, once they exit the device there isa very clear difference in behaviour. The particles in the no-grid case haveall moved in the radial direction by one jet-exit diameter and continue tomove along that trajectory (data not shown). In the entry-grid case thereis a small amount of outward spreading and in the exit-grid case there isspreading both inwards and outwards. The Stokes number for the particlesis in the intermediate range ( ∼ entry-grid case, followed by the exit-grid and the worse is the case with no-grid , with the median number of impacts in these cases being 16, 11 and8, respectively. Clearly, the presence of a grid promotes particle-wall impactsbut it is interesting that the entry-grid case has the best performance in thissense. The same trend is present in the data for the mean particle impactenergy in Figure 10(c), with the median value for the entry-grid case beingabout twice that of the other two cases.16 ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021)
Figure 9: Pressure drop across the device models: Measured; CFD. and wall impacts. Figure 10(a) shows the radial distribution of particlesacross the device at the exit and one jet-exit diameter downstream. For allthree cases the exit distribution is very similar with particles clustered aroundthe device wall. Even in the case with an entry-grid there is sufficient swirlto keep the particles at the wall. However, once they exit the device there isa very clear difference in behaviour. The particles in the no-grid case haveall moved in the radial direction by one jet-exit diameter and continue tomove along that trajectory (data not shown). In the entry-grid case thereis a small amount of outward spreading and in the exit-grid case there isspreading both inwards and outwards. The Stokes number for the particlesis in the intermediate range ( ∼ entry-grid case, followed by the exit-grid and the worse is the case with no-grid , with the median number of impacts in these cases being 16, 11 and8, respectively. Clearly, the presence of a grid promotes particle-wall impactsbut it is interesting that the entry-grid case has the best performance in thissense. The same trend is present in the data for the mean particle impactenergy in Figure 10(c), with the median value for the entry-grid case beingabout twice that of the other two cases.16 ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021) (a) (b)(c)Figure 10: Cumulative distributions of particle variables: (a) Particle radial location, no-grid , entry-grid , exit-grid at x/D a = 0, no-grid , entry-grid , exit-grid at x/D a = 1 ; (b) Number of particle-wall impacts ; (c) Average particle-impact kinetic energy; no-grid , entry-grid , exit-grid . Based on the above data, it is clear that a grid is important to reducethe particle spread and that the presence of a grid increases both the numberof particle wall impacts and their energy. According to these predictions the entry-grid device should perform best.
Figure 11 shows the spreading of the fine particles once they exit thedevice. At one jet-exit diameter downstream, the particles in the no-grid case17 ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021)
Figure 11 shows the spreading of the fine particles once they exit thedevice. At one jet-exit diameter downstream, the particles in the no-grid case17 ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021)
Figure 11: Cumulative distribution of radial location for the fine particles: no-grid , entry-grid , exit-grid at x/D a = 1, no-grid , entry-grid , exit-grid at x/D a = 5. have already spread out around 1.5 jet diameters from the axis, whereas forthe entry-grid there is almost no spreading and there is a slight focusing effectin the exit-grid case. At 5 jet-exit diameters downstream, the fine particlesare spread over 5 jet diameters in the no-grid case, whereas the spreading isonly 1.5 and 2 diameters for the exit-grid and entry-grid , respectively. Thusif reduced dispersion, and consequently less mouth-cavity deposition of theactive ingredient is the aim, the exit-grid device is to be preferred based onthese results. The objective of this paper was to perform CFD studies for a numberof inhaler designs and to confirm the results with experimental data in orderto determine the utility of appropriate CFD simulations. Of course, this canonly be done if the simulations results are of high quality and the modelsused are correctly applied. It takes experience and significant knowledge todo this correctly, so we have tried to outline the important questions to askwhen setting up models and checking the results. For example, it was foundthat the common practice of applying boundary conditions at the device in-lets leads to a non-physical influence on the flow in a very important part ofthe device. Similarly, the impact of using the correct turbulence modellingapproach is highlighted. Whilst it is no surprise that these strongly swirling18 ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021)
Figure 11: Cumulative distribution of radial location for the fine particles: no-grid , entry-grid , exit-grid at x/D a = 1, no-grid , entry-grid , exit-grid at x/D a = 5. have already spread out around 1.5 jet diameters from the axis, whereas forthe entry-grid there is almost no spreading and there is a slight focusing effectin the exit-grid case. At 5 jet-exit diameters downstream, the fine particlesare spread over 5 jet diameters in the no-grid case, whereas the spreading isonly 1.5 and 2 diameters for the exit-grid and entry-grid , respectively. Thusif reduced dispersion, and consequently less mouth-cavity deposition of theactive ingredient is the aim, the exit-grid device is to be preferred based onthese results. The objective of this paper was to perform CFD studies for a numberof inhaler designs and to confirm the results with experimental data in orderto determine the utility of appropriate CFD simulations. Of course, this canonly be done if the simulations results are of high quality and the modelsused are correctly applied. It takes experience and significant knowledge todo this correctly, so we have tried to outline the important questions to askwhen setting up models and checking the results. For example, it was foundthat the common practice of applying boundary conditions at the device in-lets leads to a non-physical influence on the flow in a very important part ofthe device. Similarly, the impact of using the correct turbulence modellingapproach is highlighted. Whilst it is no surprise that these strongly swirling18 ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021) flows are time-dependent, it is clear that simply switching on transient flow,to change a RANS simulation to a URANS simulation, is not the correct ap-proach. Doing this does indeed allow a transient simulation to made and thehigh residuals associated with an unconverged steady-state to be reducedbut the URANS approach does not provide a physical description of theturbulence structure (20). What is now evident is that the earliest studiesused relatively coarse computational meshes, lower order numerics and sim-pler turbulence models (without, for example, curvature correction terms) sothat the flows often appeared much steadier than they do now because theswirl was artificially dissipated. The approach advocated here gives muchmore realistic turbulent flow fields meaning that both the swirl behaviourand the impact of the flow on particle transport are captured much moreaccurately.Validation against detailed PIV data has allowed the models to be as-sessed and it is clear that the SBES is a good approach, especially given thenature of the flow where there are significant regions of the flow domain occu-pied by attached boundary layers, which are known to be captured well usingthe SST model. The comparisons with PIV data presented herein providevery good validation of the modelling approach. This is important as CFDcan then be used with confidence to explore the flow behaviour within thedevice itself, a region very difficult to access experimentally, and to screenideas for new device designs.The impact of the grid on mouth-cavity deposition is well captured inthe simulations as the results conform with the in-vitro results showing thatthere was a significant difference between the devices, with most depositionin the no-grid case and least in exit-grid case. The in-vitro studies showedthat more drug remained in the device for the exit-grid case, a parameter thatwas not assessed in this model. Moreover, the fine particle fraction (FPF)in the in-vitro study was similar amongst the devices, with values of 52.83% ± ± ± no-grid , entry-grid and exit-grid , respectively. From the CFD results presented here, the presence ofthe grid led to a higher mean number of impacts and increased impact kineticenergy of the particles, which is expected to translate into greater drug de-tachment from the carrier particles. Although there was a numerical increasein FPF, the increased number of particle-wall impacts observed in the CFDdid not lead to a significant increase in FPF, as shown in a previous study(15). During aerosolization, drug detachment from the carrier is thoughtto derive from both particle-wall and particle-particle collisions. From CFDresults, the entry-grid case was predicted to have a better performance dueto its greater de-agglomeration potential resulting from the higher number19 ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021)
Figure 11: Cumulative distribution of radial location for the fine particles: no-grid , entry-grid , exit-grid at x/D a = 1, no-grid , entry-grid , exit-grid at x/D a = 5. have already spread out around 1.5 jet diameters from the axis, whereas forthe entry-grid there is almost no spreading and there is a slight focusing effectin the exit-grid case. At 5 jet-exit diameters downstream, the fine particlesare spread over 5 jet diameters in the no-grid case, whereas the spreading isonly 1.5 and 2 diameters for the exit-grid and entry-grid , respectively. Thusif reduced dispersion, and consequently less mouth-cavity deposition of theactive ingredient is the aim, the exit-grid device is to be preferred based onthese results. The objective of this paper was to perform CFD studies for a numberof inhaler designs and to confirm the results with experimental data in orderto determine the utility of appropriate CFD simulations. Of course, this canonly be done if the simulations results are of high quality and the modelsused are correctly applied. It takes experience and significant knowledge todo this correctly, so we have tried to outline the important questions to askwhen setting up models and checking the results. For example, it was foundthat the common practice of applying boundary conditions at the device in-lets leads to a non-physical influence on the flow in a very important part ofthe device. Similarly, the impact of using the correct turbulence modellingapproach is highlighted. Whilst it is no surprise that these strongly swirling18 ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021) flows are time-dependent, it is clear that simply switching on transient flow,to change a RANS simulation to a URANS simulation, is not the correct ap-proach. Doing this does indeed allow a transient simulation to made and thehigh residuals associated with an unconverged steady-state to be reducedbut the URANS approach does not provide a physical description of theturbulence structure (20). What is now evident is that the earliest studiesused relatively coarse computational meshes, lower order numerics and sim-pler turbulence models (without, for example, curvature correction terms) sothat the flows often appeared much steadier than they do now because theswirl was artificially dissipated. The approach advocated here gives muchmore realistic turbulent flow fields meaning that both the swirl behaviourand the impact of the flow on particle transport are captured much moreaccurately.Validation against detailed PIV data has allowed the models to be as-sessed and it is clear that the SBES is a good approach, especially given thenature of the flow where there are significant regions of the flow domain occu-pied by attached boundary layers, which are known to be captured well usingthe SST model. The comparisons with PIV data presented herein providevery good validation of the modelling approach. This is important as CFDcan then be used with confidence to explore the flow behaviour within thedevice itself, a region very difficult to access experimentally, and to screenideas for new device designs.The impact of the grid on mouth-cavity deposition is well captured inthe simulations as the results conform with the in-vitro results showing thatthere was a significant difference between the devices, with most depositionin the no-grid case and least in exit-grid case. The in-vitro studies showedthat more drug remained in the device for the exit-grid case, a parameter thatwas not assessed in this model. Moreover, the fine particle fraction (FPF)in the in-vitro study was similar amongst the devices, with values of 52.83% ± ± ± no-grid , entry-grid and exit-grid , respectively. From the CFD results presented here, the presence ofthe grid led to a higher mean number of impacts and increased impact kineticenergy of the particles, which is expected to translate into greater drug de-tachment from the carrier particles. Although there was a numerical increasein FPF, the increased number of particle-wall impacts observed in the CFDdid not lead to a significant increase in FPF, as shown in a previous study(15). During aerosolization, drug detachment from the carrier is thoughtto derive from both particle-wall and particle-particle collisions. From CFDresults, the entry-grid case was predicted to have a better performance dueto its greater de-agglomeration potential resulting from the higher number19 ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021) of particle-wall impacts. However, particle-particle collisions were not mod-elled in this study, which is the likely explanation for the differences observedbetween CFD and in-vitro results. This paper has shown that provided the correct modelling choices aremade and the simulations are executed with the appropriate care and knowl-edge, CFD can provide significant insights into DPI performance. Simula-tions using the Stress Blended Eddy Simulation (SBES) approach are wellsuited for this task, which is supported by the very good agreement withthe PIV data. This turbulence modelling choice is important as it allowsthe transient nature of the flow and the significant turbulence generationby highly swirling flows to be captured. This has a follow-on effect on thedispersion of the fine particles that have low Stokes numbers and follow theturbulent eddies. This work shows that it is possible to improve upon theuse of RANS or URANS significantly without going to a full LES simulation.In particular, the proposed approach uses the optimal turbulence modellingapproach in each zone: RANS in attached boundary layers at the walls andLES in the regions of separated flow and wakes. Use of pure LES is notpractical as it requires locally refined meshes in all three dimensions at thewall if the boundary layer is to be captured correctly.The simulations capture important experimental observations of the re-duction in radial spreading of the flow and fine particles due to the presence ofa grid, with the exit-grid geometry performing best, in line with the reducedmouth-cavity deposition observed in the experiments. Keeping in mind thatthe experiments did not use a throat geometry and the simulations did notmodel all aspects of the particle behaviour, specifically particle-particle in-teractions and particle detachment the adopted CFD approach captured thedispersion data quite well.
Acknowledgments
The research was supported by the Australian Research Council. Theauthors acknowledge the University of Sydney for providing High Perfor-mance Computing resources that have greatly contributed to the researchresults reported here (http://sydney.edu.au/research support). The research20 ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021)
The research was supported by the Australian Research Council. Theauthors acknowledge the University of Sydney for providing High Perfor-mance Computing resources that have greatly contributed to the researchresults reported here (http://sydney.edu.au/research support). The research20 ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021) was also benefited from computational resources provided through the NC-MAS, supported by the Australian Government. The computational facilitiessupporting this project included the Multi-modal Australian ScienceS Imag-ing and Visualisation Environment (MASSIVE) at Monash.21 ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021)
The research was supported by the Australian Research Council. Theauthors acknowledge the University of Sydney for providing High Perfor-mance Computing resources that have greatly contributed to the researchresults reported here (http://sydney.edu.au/research support). The research20 ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021) was also benefited from computational resources provided through the NC-MAS, supported by the Australian Government. The computational facilitiessupporting this project included the Multi-modal Australian ScienceS Imag-ing and Visualisation Environment (MASSIVE) at Monash.21 ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021)
References [1] Wong W, Fletcher DF, Traini D, Chan HK, Young PM. The use of com-putational approaches in inhaler development. Advanced Drug DeliveryReviews. 2012;64(4):312–322.[2] Islam N, Cleary MJ. Developing an efficient and reliable dry powderinhaler for pulmonary drug delivery - A review for multidisciplinaryresearchers. Medical Engineering and Physics. 2012;34(4):409–427.[3] Sommerfeld M, Cui Y, Schmalfuß S. Potential and constraints forthe application of CFD combined with Lagrangian particle tracking todry powder inhalers. European Journal of Pharmaceutical Sciences.2019;128:299–324.[4] Yang Y, Knudsen Kaer S. Comparison of Reynolds averaged Navier-Stokes based simulation and large-eddy simulation for one isothermalswirling flow. Journal of Thermal Science. 2012;21(2):154–161.[5] Milenkovic J, Alexopoulos AH, Kiparissides C. Flow and particle depo-sition in the Turbuhaler: A CFD simulation. International Journal ofPharmaceutics. 2013;448(1):205–213.[6] Coates MS, Fletcher DF, Chan HK, Raper JA. Effect of design on theperformance of a dry powder inhaler using computational fluid dynam-ics. Part 1: Grid structure and mouthpiece length. Journal of Pharma-ceutical Sciences. 2004;93(11):2863–2876.[7] Menter FR. Two-equation eddy-viscosity turbulence models for engi-neering applications. AIAA Journal. 1994;32(8):1598–1605.[8] Coates MS, Chan HK, Fletcher DF, Raper JA. Influence of air flowon the performance of a dry powder inhaler using computational andexperimental analyses. Pharmaceutical Research. 2005;22(9):1445–1453.[9] Coates MS, Chan HK, Fletcher DF, Raper JA. Effect of design onthe performance of a dry powder inhaler using computational fluid dy-namics. Part 2: Air inlet size. Journal of Pharmaceutical Sciences.2006;95(6):1382–1392.[10] Donovan MJ, Kim SH, Raman V, Smyth HD. Dry powder inhaler deviceinfluence on carrier particle performance. Journal of PharmaceuticalSciences. 2012;101(3):1097–1107.22 ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021)
References [1] Wong W, Fletcher DF, Traini D, Chan HK, Young PM. The use of com-putational approaches in inhaler development. Advanced Drug DeliveryReviews. 2012;64(4):312–322.[2] Islam N, Cleary MJ. Developing an efficient and reliable dry powderinhaler for pulmonary drug delivery - A review for multidisciplinaryresearchers. Medical Engineering and Physics. 2012;34(4):409–427.[3] Sommerfeld M, Cui Y, Schmalfuß S. Potential and constraints forthe application of CFD combined with Lagrangian particle tracking todry powder inhalers. European Journal of Pharmaceutical Sciences.2019;128:299–324.[4] Yang Y, Knudsen Kaer S. Comparison of Reynolds averaged Navier-Stokes based simulation and large-eddy simulation for one isothermalswirling flow. Journal of Thermal Science. 2012;21(2):154–161.[5] Milenkovic J, Alexopoulos AH, Kiparissides C. Flow and particle depo-sition in the Turbuhaler: A CFD simulation. International Journal ofPharmaceutics. 2013;448(1):205–213.[6] Coates MS, Fletcher DF, Chan HK, Raper JA. Effect of design on theperformance of a dry powder inhaler using computational fluid dynam-ics. Part 1: Grid structure and mouthpiece length. Journal of Pharma-ceutical Sciences. 2004;93(11):2863–2876.[7] Menter FR. Two-equation eddy-viscosity turbulence models for engi-neering applications. AIAA Journal. 1994;32(8):1598–1605.[8] Coates MS, Chan HK, Fletcher DF, Raper JA. Influence of air flowon the performance of a dry powder inhaler using computational andexperimental analyses. Pharmaceutical Research. 2005;22(9):1445–1453.[9] Coates MS, Chan HK, Fletcher DF, Raper JA. Effect of design onthe performance of a dry powder inhaler using computational fluid dy-namics. Part 2: Air inlet size. Journal of Pharmaceutical Sciences.2006;95(6):1382–1392.[10] Donovan MJ, Kim SH, Raman V, Smyth HD. Dry powder inhaler deviceinfluence on carrier particle performance. Journal of PharmaceuticalSciences. 2012;101(3):1097–1107.22 ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021) [11] Milenkovic J, Alexopoulos AH, Kiparissides C. Deposition and fine par-ticle production during dynamic flow in a dry powder inhaler: A CFDapproach. International Journal of Pharmaceutics. 2014;461(1-2):129–136.[12] Sommerfeld M, Schmalfuß S. Numerical analysis of carrier particle mo-tion in a dry powder inhaler. Journal of Fluids Engineering, Transactionsof the ASME. 2016;138(4). 41308.[13] Longest PW, Son YJ, Holbrook L, Hindle M. Aerodynamic factors re-sponsible for the deaggregation of carrier-free drug powders to formmicrometer and submicrometer aerosols. Pharmaceutical Research.2013;30(6):1608–1627.[14] Longest W, Farkas D. Development of a new inhaler for high-efficiencydispersion of spray-dried powders using computational fluid dynamics(CFD) modeling. AAPS Journal. 2019;21(2):1–15.[15] Gomes dos Reis L, Chaugule V, Fletcher DF, Young PM, Traini D,Soria J. In-vitro and particle image velocimetry studies of dry powderinhalers. International Journal of Pharmaceutics. 2021, in-press;.[16] Ansys ® Fluent 2020R1;. Available from: .[17] Shih TH, Liou WW, Shabbir A, Yang Z, Zhu J. A new k- (cid:15) eddy viscositymodel for high reynolds number turbulent flows. Computers and Fluids.1995;24(3):227–238.[18] Smirnov PE, Menter FR. Sensitization of the SST turbulence model torotation and curvature by applying the Spalart-Shur correction term.Journal of Turbomachinery. 2009;131(4):1–8.[19] Alahmadi YH, Nowakowski AF. Modified shear stress transport modelwith curvature correction for the prediction of swirling flow in a cycloneseparator. Chemical Engineering Science. 2016;147:150–165.[20] Menter F. Stress-blended eddy simulation (SBES)—A new paradigm inhybrid RANS-LES modeling. Notes on Numerical Fluid Mechanics andMultidisciplinary Design. 2018;137:27–37.[21] Nicoud F, Ducros F. Subgrid-scale stress modelling based on the squareof the velocity gradient tensor. Flow, Turbulence and Combustion.1999;62(3):183–200. 23 ccepted: Pharmaceutical Research (2021)ccepted: Pharmaceutical Research (2021)