This document discusses using statistical learning techniques and direct numerical simulation (DNS) data to develop closure relationships for a simplified two-fluid model of bubbly two-phase flow in a vertical channel. DNS is used to simulate bubbly upflow in the channel, and the data is averaged and used to train a neural network model to relate unknown closure terms in the two-fluid model equations to resolved variables. The model predictions are then tested against additional DNS data. The presence of walls adds complexity compared to previous work on periodic domains by introducing new closure terms related to surface tension effects near walls.
This document summarizes a numerical study on free-surface flow conducted using a computational fluid dynamics (CFD) solver. The study examines the wave profile generated by a submerged hydrofoil through several test cases varying parameters like the turbulence model, grid resolution, and hydrofoil depth. The document provides background on the governing equations solved by the CFD solver and the interface capturing technique used to model the free surface. Five test cases are described that investigate grid convergence, the impact of laminar vs turbulent models, the relationship between hydrofoil depth and wave height, and the effect of discretization schemes.
This document summarizes research on using direct numerical simulation (DNS) to study bubbly flows in vertical channels and develop closure terms for two-fluid models of multiphase flows. It describes DNS of small laminar systems with several spherical bubbles that show a non-monotonic transient evolution as bubbles initially move toward walls then slowly return to the channel center. Larger turbulent simulations are also discussed. The goal is to use DNS data to provide values for unresolved terms in simplified averaged models through statistical learning, with prospects for filtering interface structures in large-eddy simulations discussed.
This document summarizes direct numerical simulations (DNS) of multiphase flows performed by Grétar Tryggvason and colleagues. It discusses DNS of bubbly flows in vertical channels, including the effects of bubble deformability and size on turbulent upflow. Machine learning methods are applied to DNS data to derive closure relationships for modeling averaged multiphase flows. More complex gas-liquid flows involving many bubbles of different sizes in turbulent channel flow are also examined.
Changes in dam break hydrodynamic modelling practice - Suter et alStephen Flood
Abstract: Today, many organisations rely on hydrodynamic modelling to assess the consequences of dam break failure on downstream populations and infrastructure. The availability of finite volume shock-capturing schemes and flexible mesh schematisations in widely used software platforms imply that dam break modelling projects will be carried out differently in the future: Finite volume based platforms allow widespread application of shock-capturing methods and flexible mesh platforms can represent features in the study area more realistically and are more flexible thanks to varying mesh resolutions. Furthermore, the recent adoption of Graphics Processing Unit (GPU) technology in mainstream scientific and engineering computing will also significantly decrease computation times at relatively low cost.
This paper examines the application of finite volume, flexible mesh and GPU technologies to dam break modelling. One-dimensional (1D) modelling results are compared to those from two-dimensional (2D) finite difference and finite volume approaches. The results demonstrate that there are differences between modelling approaches and that the computational speeds of 2D simulations can be significantly reduced by the use of GPU processors.
The document discusses using computational fluid dynamics to model hydrodynamic forces on coastal highway bridges during hurricanes. It summarizes that:
1) Investigations found bridge failures were often due to uplift forces exceeding the capacity of bent cap connections, allowing superstructure elements to be displaced.
2) The study uses Navier-Stokes equations to simulate wave-bridge interactions and calculate hydrodynamic forces with reasonable accuracy.
3) Simulations show that air vents cored into bridge decks and diaphragms can effectively reduce hydrodynamic forces and mitigate damage, validating their potential as a retrofitting option.
This document describes a Kriging component for spatial interpolation of climatological variables in the OMS modeling framework. Kriging is a geostatistical technique that interpolates values based on measured data and the spatial autocorrelation between data points. The component implements ordinary and detrended Kriging algorithms using 10 semivariogram models. It can interpolate both raster and point data and outputs the interpolated climatological variable values. Links are provided for downloading the component code, data, and OMS project files needed to run the interpolation.
Modification of the casagrandes equation of phreatic lineIAEME Publication
The document presents an experimental study to modify Casagrande's equation for determining the phreatic line in an earthen dam. A physical model of an earth-fill dam was constructed using sandy silt soil. Phreatic lines were traced for the model with and without a horizontal filter. The experimental results were compared to Casagrande's analytical solution. Regression modeling was then used to derive a new, more accurate equation to determine the phreatic line position based on the experimental data for dams with sandy silt soils. The modified equation was found to estimate the phreatic line location more precisely than Casagrande's original equation.
Aerodynamic Analysis of Low Speed Turbulent Flow Over A Delta WingIJRES Journal
Delta wing has been a subject of intense research since decades due to decades due to inherent characteristics of generating increased nonlinear lift due to vortex dominated flows. Lot of work has been carried out in order to understand the vortex dominated flows on the delta wing. The delta wing is a wing platform in the form of a triangle. Aerodynamics of wings with moderate sweep angle is recognized by the aerospace community as a challenging problem. In spite of its potential application in military aircraft, the understanding of the aerodynamics of such wings is far from complete. In order to address this situation, the present work is initiated to compute the 3D turbulent flow field over sharp edged finite wings with a diamond shaped plan forms and moderate sweep angle. The detailed flow pattern and surface pressure distribution may further indicate the appropriate kind of flow control during flight operation of such wings. The flow field is computed using an in-house developed CFD code RANS3D.
This document summarizes a numerical study on free-surface flow conducted using a computational fluid dynamics (CFD) solver. The study examines the wave profile generated by a submerged hydrofoil through several test cases varying parameters like the turbulence model, grid resolution, and hydrofoil depth. The document provides background on the governing equations solved by the CFD solver and the interface capturing technique used to model the free surface. Five test cases are described that investigate grid convergence, the impact of laminar vs turbulent models, the relationship between hydrofoil depth and wave height, and the effect of discretization schemes.
This document summarizes research on using direct numerical simulation (DNS) to study bubbly flows in vertical channels and develop closure terms for two-fluid models of multiphase flows. It describes DNS of small laminar systems with several spherical bubbles that show a non-monotonic transient evolution as bubbles initially move toward walls then slowly return to the channel center. Larger turbulent simulations are also discussed. The goal is to use DNS data to provide values for unresolved terms in simplified averaged models through statistical learning, with prospects for filtering interface structures in large-eddy simulations discussed.
This document summarizes direct numerical simulations (DNS) of multiphase flows performed by Grétar Tryggvason and colleagues. It discusses DNS of bubbly flows in vertical channels, including the effects of bubble deformability and size on turbulent upflow. Machine learning methods are applied to DNS data to derive closure relationships for modeling averaged multiphase flows. More complex gas-liquid flows involving many bubbles of different sizes in turbulent channel flow are also examined.
Changes in dam break hydrodynamic modelling practice - Suter et alStephen Flood
Abstract: Today, many organisations rely on hydrodynamic modelling to assess the consequences of dam break failure on downstream populations and infrastructure. The availability of finite volume shock-capturing schemes and flexible mesh schematisations in widely used software platforms imply that dam break modelling projects will be carried out differently in the future: Finite volume based platforms allow widespread application of shock-capturing methods and flexible mesh platforms can represent features in the study area more realistically and are more flexible thanks to varying mesh resolutions. Furthermore, the recent adoption of Graphics Processing Unit (GPU) technology in mainstream scientific and engineering computing will also significantly decrease computation times at relatively low cost.
This paper examines the application of finite volume, flexible mesh and GPU technologies to dam break modelling. One-dimensional (1D) modelling results are compared to those from two-dimensional (2D) finite difference and finite volume approaches. The results demonstrate that there are differences between modelling approaches and that the computational speeds of 2D simulations can be significantly reduced by the use of GPU processors.
The document discusses using computational fluid dynamics to model hydrodynamic forces on coastal highway bridges during hurricanes. It summarizes that:
1) Investigations found bridge failures were often due to uplift forces exceeding the capacity of bent cap connections, allowing superstructure elements to be displaced.
2) The study uses Navier-Stokes equations to simulate wave-bridge interactions and calculate hydrodynamic forces with reasonable accuracy.
3) Simulations show that air vents cored into bridge decks and diaphragms can effectively reduce hydrodynamic forces and mitigate damage, validating their potential as a retrofitting option.
This document describes a Kriging component for spatial interpolation of climatological variables in the OMS modeling framework. Kriging is a geostatistical technique that interpolates values based on measured data and the spatial autocorrelation between data points. The component implements ordinary and detrended Kriging algorithms using 10 semivariogram models. It can interpolate both raster and point data and outputs the interpolated climatological variable values. Links are provided for downloading the component code, data, and OMS project files needed to run the interpolation.
Modification of the casagrandes equation of phreatic lineIAEME Publication
The document presents an experimental study to modify Casagrande's equation for determining the phreatic line in an earthen dam. A physical model of an earth-fill dam was constructed using sandy silt soil. Phreatic lines were traced for the model with and without a horizontal filter. The experimental results were compared to Casagrande's analytical solution. Regression modeling was then used to derive a new, more accurate equation to determine the phreatic line position based on the experimental data for dams with sandy silt soils. The modified equation was found to estimate the phreatic line location more precisely than Casagrande's original equation.
Aerodynamic Analysis of Low Speed Turbulent Flow Over A Delta WingIJRES Journal
Delta wing has been a subject of intense research since decades due to decades due to inherent characteristics of generating increased nonlinear lift due to vortex dominated flows. Lot of work has been carried out in order to understand the vortex dominated flows on the delta wing. The delta wing is a wing platform in the form of a triangle. Aerodynamics of wings with moderate sweep angle is recognized by the aerospace community as a challenging problem. In spite of its potential application in military aircraft, the understanding of the aerodynamics of such wings is far from complete. In order to address this situation, the present work is initiated to compute the 3D turbulent flow field over sharp edged finite wings with a diamond shaped plan forms and moderate sweep angle. The detailed flow pattern and surface pressure distribution may further indicate the appropriate kind of flow control during flight operation of such wings. The flow field is computed using an in-house developed CFD code RANS3D.
This document summarizes a study that uses neural networks and direct numerical simulations to develop closure models for two-fluid equations describing multiphase bubbly flows. Direct numerical simulations were performed for a system of bubbles rising in liquid and these results were used to train a neural network model. The neural network model was able to reasonably predict the evolution of different initial bubble distributions and velocities compared to direct numerical simulations. This approach shows promise for using computational simulations to develop reduced-order models that can simulate multiphase flows without resolving all length and time scales.
Flow Modeling Based Wall Element TechniqueCSCJournals
Two types of flow where examined, pressure and combination of pressure and Coquette flow of confined turbulent flow with a one equation model used to depict the turbulent viscosity of confined flow in a smooth straight channel when a finite element technique based on a zone close to a solid wall has been adopted for predicting the distribution of the pertinent variables in this zone and examined even with case when the near wall zone was extended away from the wall. The validation of imposed technique has been tested and well compared with other techniques.
Large Eddy Simulation of Turbulence Modeling for wind Flow past Wall Mounted ...IJERA Editor
This paper will present the large eddy simulation of turbulence modeling for wind flow over a wall mounted 3D cubical model. The LES Smagorinsky scheme is employed for the numerical simulation. The domain for this study is of the size of 60 cm x 30 cm x 30 cm. The 3D cube model is taken of the size of 6 cm x 6 cm x 4 cm. The Reynolds number for the flow in respect of the height of the cube i.e, 4 cm is 5.3x104. The hexahedral grids are used for the meshing of the flow domain. The results are discussed in terms of various parameters such as velocity profile around the cube and the computational domain, the pressure distribution over the cube, near wall velocity profile and the shear stress distribution and also the result of drag coefficient is verified by neural network time series analysis using MATLAB. In this present study we have used the OpenFoam platform for the computational and numerical analysis. The numerical scheme employed is the combination of the steady state incompressible Newtonian flow model using SIMPLE algorithm followed by the transient model of incompressible Newtonian flow using PISO algorithm. We have observed that there is a constant positive drag coefficient in case of steady state simulation where as there is a negative lift coefficient in the initial run and a very low lift coefficient at the end of the steady state simulation.
This document summarizes a large eddy simulation of flow around a sharp-edged surface-mounted cube. The simulation was performed using the Petsc-Fem code developed at CIMEC. The flow conditions matched published benchmarks, with a Reynolds number of 40,000. An upstream channel flow was first simulated to provide turbulent inflow conditions. The simulation results are analyzed to validate the LES implementation and identify areas for improving turbulence modeling.
Review and Assessment of Turbulence Transition ModelsIJERDJOURNAL
ABSTRACT: Turbulent boundary layer transition can greatly affect flow characteristics such as skin friction, heat transfer, pressure loads, and boundary layer separation point. Accurate computation of such effect is vital to the design of components and vehicles subjected to turbulent transition flows. In this paper we report our review of the existing boundary layer transition models, selection of the boundary layer transition models most appropriate for the existing Reynolds-averaged Navier-Stokes flow solvers, and implementation and evaluation of the selected transition models with some benchmark test cases, ranging from subsonic to hypersonic flows. The objective of this assessment study is not intended to pick the best transition model (in fact, there is no transition model accurately predict the boundary layer transition for all cases tested here). Rather it is to demonstrate that (1) the flow physics of turbulence transition is very complicated and not yet well understood, (2) the applicability of empirical transition models is limited and their use should be cautious due to different transition characteristics for different flow regimes/environments, and (3) further research on prediction of turbulence transition is warranted to improve the accuracy, efficiency and range of applicability.
Upscaling Improvement for Heterogeneous Fractured Reservoir Using a Geostatis...erissano
Statistical percolation threshold formulation indicator to estimate the fracture network connectivity. Can be used for upscaling of permeabilities.
Available in FracaFlow sofware
CFD and Artificial Neural Networks Analysis of Plane Sudden Expansion FlowsCSCJournals
It has been clearly established that the reattachment length for laminar flow depends on two non-dimensional parameters, the Reynolds number and the expansion ratio, therefore in this work, an ANN model that predict reattachment positions for the expansion ratios of 2, 3 and 5 based on the above two parameters has been developed. The R2 values of the testing set output Xr1, Xr2, Xr3, and Xr4 were 0.9383, 0.8577, 0.997 and 0.999 respectively. These results indicate that the network model produced reattachment positions that were in close agreement with the actual values. When considering the reattachment length of plane sudden-expansions the judicious combination of CFD calculated solutions with ANN will result in a considerable saving in computing and turnaround time. Thus CFD can be used in the first instance to obtain reattachment lengths for a limited choice of Reynolds numbers and ANN will be used subsequently to predict the reattachment lengths for other intermediate Reynolds number values. The CFD calculations concern unsteady laminar flow through a plane sudden expansion and are performed using a commercial CFD code STAR-CD while the training process of the corresponding ANN model was performed using the NeuroShellTM simulator.
This document summarizes a study that used computational fluid dynamics (CFD) to analyze vortex induced vibration on an offshore structure through fluid-structure interaction (FSI) modeling. The study performed 2D and 3D CFD analyses to understand flow patterns and validate results. Preliminary one-way FSI analysis was then conducted by coupling structural and fluid solvers to observe the dynamic response of the structure to periodically varying vortex loads. The goal was to better understand vortex induced loads on offshore structures through numerical simulation.
An Immersed Boundary Method To Solve Flow And Heat Transfer Problems Involvin...Courtney Esco
This document presents an immersed boundary method to solve fluid-structure interaction problems involving a moving rigid solid object. The method treats the solid object as fluid governed by the Navier-Stokes equations, and adds virtual force and heat source terms to model the solid object's prescribed velocity and temperature. The method was validated on benchmark forced and natural convection problems, and used to study a mixed convection problem of a heated cylinder moving inside an enclosure. The immersed boundary method avoids dynamic grid generation and provides efficient computation of thermal and flow fields caused by fluid-structure interaction.
In this paper, the unsteady motion of a spherical particle rolling down an inclined tube in a
Newtonian fluid for a range of Reynolds numbers was solved using a simulation method called
the Differential Transformation Method (DTM). The concept of differential transformation is
briefly introduced, and then we employed it to derive solution of nonlinear equation. The
obtained results for displacement, velocity and acceleration of the motion from DTM are
compared with those from numerical solution to verify the accuracy of the proposed method.
The effects of particle diameter (size), continues phase viscosity and inclination angles was
studied. As an important result it was found that the inclination angle does not affect the
acceleration duration. The results reveal that the Differential Transformation Method can achieve suitable results in predicting the solution of such problems.
This project aims at simulating lid driven cavity flow problem using package MATLAB. Steady Incompressible Navier-Stokes equation with continuity equation will be studied at various Reynolds number. The main aim is to obtain the velocity field in steady state using the finite difference formulation on momentum equations and continuity equation. Reynold number is the pertinent parameter of the present study. Taylor’s series expansion has been used to convert the governing equations in the algebraic form using finite difference schemes.
This document provides an overview of a thesis that used OpenFOAM to simulate unsteady aerodynamic flow over a sphere. The study aimed to analyze flow instabilities and quantify uncertainty in simulations. It used an IDDES turbulence model on refined meshes. Frequency analysis revealed peaks from vortex shedding and shear layer instabilities. Mean flow parameters matched experimental data. Future work could include hybrid RANS-LES models and longer simulations for improved flow statistics.
International Journal of Computational Engineering Research(IJCER)ijceronline
International Journal of Computational Engineering Research(IJCER) is an intentional online Journal in English monthly publishing journal. This Journal publish original research work that contributes significantly to further the scientific knowledge in engineering and Technology
Thin Film Pressure Estimation of Argon and Water using LAMMPSCSCJournals
In this work, we investigate the pressure and density characteristics of water film when simulated using the emerging technique called many body dissipative particle dynamics method. This work also layout the methodology of estimating local pressure from LAMMPS simulation using Harasima scheme. Using the triangular shaped cloud interpolation function, pressure and density are estimated at local bins and compared with the experimental database. Our results show good agreement for the molecular dynamics results of the argon system, while the many body dissipative particle model fails to simulate the water properties at room temperature. In its current form, the many body dissipative particle method cannot be used for accurate liquid vapor interfacial simulations and heat transfer studies.
Cross-checking of numerical-experimental procedures for interface characteriz...Diego Scarpa
This document summarizes the cross-checking of experimental and numerical procedures for characterizing interfaces in quasi 2-D transitional flows. The study uses laser light scattering and PIV techniques experimentally, along with 2D DNS simulations, to obtain interface location, velocity fields, strain rates, and interface density for a reference test case. Comparisons show good agreement between experiments and simulations for interface contours, mean velocity profiles, and interface density profiles along the streamwise direction, but some discrepancies in strain rate profiles. The cross-checking helps validate the experimental results and identifies areas for improving numerical models.
Turbulence structure behind the shock in canonical shockvortical turbulence ...Jovan Belite Squad
This document discusses a study using direct numerical simulation (DNS) to examine the interaction between isotropic turbulence and a normal shock wave. High-resolution simulations were performed across a range of shock Mach numbers (1.1-2.2) and turbulent Reynolds numbers (10-45). The results showed that turbulence quantities from DNS converge to predictions of linear interaction analysis (LIA) as the turbulent Mach number decreases, demonstrating the importance of LIA. LIA was also extended to compute full post-shock flow fields, revealing how the shock significantly changes turbulence structures through symmetrization and increased strain-rotation correlation.
This document discusses the built-in filters of two eddy viscosity subgrid scale models used in large-eddy simulation (LES): the Smagorinsky model and the Schumann model. It presents the theoretical basis of LES filtering and equations. The study identifies the forms of the built-in filters associated with each model by varying a characteristic length scale in the models and analyzing the effective filter rendered by the simulations. The behavior of an equivalent Kolmogorov constant is also compared to theory. The results provide insight into how the modeling terms influence the filtering process and flow dynamics in LES without using explicit filtering.
Effect of Geometry on Variation of Heat Flux and Drag for Launch Vehicle -- Z...Abhishek Jain
Above Research Paper can be downloaded from www.zeusnumerix.com
The research paper aims at studying the variation of the geometry of the launch vehicle nose and its effect on heat flux. CFDExpert software is first validated on NASA's hyperballistic model and then used on proposed geometries. Various nose radius and blending shapes are studied for effect on drag and heat flux. Cone ogive shape is found to decrease heat flux with an insignificant increase in drag. Authors Abhishek Jain (Zeus Numerix), Rohan Kedar and Prof V Kalamkar (SPCOE).
An Adaptive Moving Mesh Method For Two-Dimensional Relativistic Magnetohydrod...Scott Faria
This document describes an adaptive moving mesh method for solving two-dimensional thin film flow equations that include surface tension. The method uses moving mesh partial differential equations (MMPDEs) and mesh density functions to adaptively move the computational mesh based on solution characteristics like curvature. Numerical results show this r-adaptive moving mesh technique accurately captures the moving contact line and associated fingering instability with reduced computational effort compared to a fixed uniform mesh.
Multi-task learning using non-linear autoregressive models and recurrent neur...IJECEIAES
Tide level forecasting plays an important role in environmental management and development. Current tide level forecasting methods are usually implemented for solving single task problems, that is, a model built based on the tide level data at an individual location is only used to forecast tide level of the same location but is not used for tide forecasting at another location. This study proposes a new method for tide level prediction at multiple locations simultaneously. The method combines nonlinear autoregressive moving average with exogenous inputs (NARMAX) model and recurrent neural networks (RNNs), and incorporates them into a multi-task learning (MTL) framework. Experiments are designed and performed to compare single task learning (STL) and MTL with and without using non-linear autoregressive models. Three different RNN variants, namely, long short- term memory (LSTM), gated recurrent unit (GRU) and bidirectional LSTM (BiLSTM) are employed together with non-linear autoregressive models. A case study on tide level forecasting at many different geographical locations (5 to 11 locations) is conducted. Experimental results demonstrate that the proposed architectures outperform the classical single-task prediction methods.
This document summarizes a study that uses neural networks and direct numerical simulations to develop closure models for two-fluid equations describing multiphase bubbly flows. Direct numerical simulations were performed for a system of bubbles rising in liquid and these results were used to train a neural network model. The neural network model was able to reasonably predict the evolution of different initial bubble distributions and velocities compared to direct numerical simulations. This approach shows promise for using computational simulations to develop reduced-order models that can simulate multiphase flows without resolving all length and time scales.
Flow Modeling Based Wall Element TechniqueCSCJournals
Two types of flow where examined, pressure and combination of pressure and Coquette flow of confined turbulent flow with a one equation model used to depict the turbulent viscosity of confined flow in a smooth straight channel when a finite element technique based on a zone close to a solid wall has been adopted for predicting the distribution of the pertinent variables in this zone and examined even with case when the near wall zone was extended away from the wall. The validation of imposed technique has been tested and well compared with other techniques.
Large Eddy Simulation of Turbulence Modeling for wind Flow past Wall Mounted ...IJERA Editor
This paper will present the large eddy simulation of turbulence modeling for wind flow over a wall mounted 3D cubical model. The LES Smagorinsky scheme is employed for the numerical simulation. The domain for this study is of the size of 60 cm x 30 cm x 30 cm. The 3D cube model is taken of the size of 6 cm x 6 cm x 4 cm. The Reynolds number for the flow in respect of the height of the cube i.e, 4 cm is 5.3x104. The hexahedral grids are used for the meshing of the flow domain. The results are discussed in terms of various parameters such as velocity profile around the cube and the computational domain, the pressure distribution over the cube, near wall velocity profile and the shear stress distribution and also the result of drag coefficient is verified by neural network time series analysis using MATLAB. In this present study we have used the OpenFoam platform for the computational and numerical analysis. The numerical scheme employed is the combination of the steady state incompressible Newtonian flow model using SIMPLE algorithm followed by the transient model of incompressible Newtonian flow using PISO algorithm. We have observed that there is a constant positive drag coefficient in case of steady state simulation where as there is a negative lift coefficient in the initial run and a very low lift coefficient at the end of the steady state simulation.
This document summarizes a large eddy simulation of flow around a sharp-edged surface-mounted cube. The simulation was performed using the Petsc-Fem code developed at CIMEC. The flow conditions matched published benchmarks, with a Reynolds number of 40,000. An upstream channel flow was first simulated to provide turbulent inflow conditions. The simulation results are analyzed to validate the LES implementation and identify areas for improving turbulence modeling.
Review and Assessment of Turbulence Transition ModelsIJERDJOURNAL
ABSTRACT: Turbulent boundary layer transition can greatly affect flow characteristics such as skin friction, heat transfer, pressure loads, and boundary layer separation point. Accurate computation of such effect is vital to the design of components and vehicles subjected to turbulent transition flows. In this paper we report our review of the existing boundary layer transition models, selection of the boundary layer transition models most appropriate for the existing Reynolds-averaged Navier-Stokes flow solvers, and implementation and evaluation of the selected transition models with some benchmark test cases, ranging from subsonic to hypersonic flows. The objective of this assessment study is not intended to pick the best transition model (in fact, there is no transition model accurately predict the boundary layer transition for all cases tested here). Rather it is to demonstrate that (1) the flow physics of turbulence transition is very complicated and not yet well understood, (2) the applicability of empirical transition models is limited and their use should be cautious due to different transition characteristics for different flow regimes/environments, and (3) further research on prediction of turbulence transition is warranted to improve the accuracy, efficiency and range of applicability.
Upscaling Improvement for Heterogeneous Fractured Reservoir Using a Geostatis...erissano
Statistical percolation threshold formulation indicator to estimate the fracture network connectivity. Can be used for upscaling of permeabilities.
Available in FracaFlow sofware
CFD and Artificial Neural Networks Analysis of Plane Sudden Expansion FlowsCSCJournals
It has been clearly established that the reattachment length for laminar flow depends on two non-dimensional parameters, the Reynolds number and the expansion ratio, therefore in this work, an ANN model that predict reattachment positions for the expansion ratios of 2, 3 and 5 based on the above two parameters has been developed. The R2 values of the testing set output Xr1, Xr2, Xr3, and Xr4 were 0.9383, 0.8577, 0.997 and 0.999 respectively. These results indicate that the network model produced reattachment positions that were in close agreement with the actual values. When considering the reattachment length of plane sudden-expansions the judicious combination of CFD calculated solutions with ANN will result in a considerable saving in computing and turnaround time. Thus CFD can be used in the first instance to obtain reattachment lengths for a limited choice of Reynolds numbers and ANN will be used subsequently to predict the reattachment lengths for other intermediate Reynolds number values. The CFD calculations concern unsteady laminar flow through a plane sudden expansion and are performed using a commercial CFD code STAR-CD while the training process of the corresponding ANN model was performed using the NeuroShellTM simulator.
This document summarizes a study that used computational fluid dynamics (CFD) to analyze vortex induced vibration on an offshore structure through fluid-structure interaction (FSI) modeling. The study performed 2D and 3D CFD analyses to understand flow patterns and validate results. Preliminary one-way FSI analysis was then conducted by coupling structural and fluid solvers to observe the dynamic response of the structure to periodically varying vortex loads. The goal was to better understand vortex induced loads on offshore structures through numerical simulation.
An Immersed Boundary Method To Solve Flow And Heat Transfer Problems Involvin...Courtney Esco
This document presents an immersed boundary method to solve fluid-structure interaction problems involving a moving rigid solid object. The method treats the solid object as fluid governed by the Navier-Stokes equations, and adds virtual force and heat source terms to model the solid object's prescribed velocity and temperature. The method was validated on benchmark forced and natural convection problems, and used to study a mixed convection problem of a heated cylinder moving inside an enclosure. The immersed boundary method avoids dynamic grid generation and provides efficient computation of thermal and flow fields caused by fluid-structure interaction.
In this paper, the unsteady motion of a spherical particle rolling down an inclined tube in a
Newtonian fluid for a range of Reynolds numbers was solved using a simulation method called
the Differential Transformation Method (DTM). The concept of differential transformation is
briefly introduced, and then we employed it to derive solution of nonlinear equation. The
obtained results for displacement, velocity and acceleration of the motion from DTM are
compared with those from numerical solution to verify the accuracy of the proposed method.
The effects of particle diameter (size), continues phase viscosity and inclination angles was
studied. As an important result it was found that the inclination angle does not affect the
acceleration duration. The results reveal that the Differential Transformation Method can achieve suitable results in predicting the solution of such problems.
This project aims at simulating lid driven cavity flow problem using package MATLAB. Steady Incompressible Navier-Stokes equation with continuity equation will be studied at various Reynolds number. The main aim is to obtain the velocity field in steady state using the finite difference formulation on momentum equations and continuity equation. Reynold number is the pertinent parameter of the present study. Taylor’s series expansion has been used to convert the governing equations in the algebraic form using finite difference schemes.
This document provides an overview of a thesis that used OpenFOAM to simulate unsteady aerodynamic flow over a sphere. The study aimed to analyze flow instabilities and quantify uncertainty in simulations. It used an IDDES turbulence model on refined meshes. Frequency analysis revealed peaks from vortex shedding and shear layer instabilities. Mean flow parameters matched experimental data. Future work could include hybrid RANS-LES models and longer simulations for improved flow statistics.
International Journal of Computational Engineering Research(IJCER)ijceronline
International Journal of Computational Engineering Research(IJCER) is an intentional online Journal in English monthly publishing journal. This Journal publish original research work that contributes significantly to further the scientific knowledge in engineering and Technology
Thin Film Pressure Estimation of Argon and Water using LAMMPSCSCJournals
In this work, we investigate the pressure and density characteristics of water film when simulated using the emerging technique called many body dissipative particle dynamics method. This work also layout the methodology of estimating local pressure from LAMMPS simulation using Harasima scheme. Using the triangular shaped cloud interpolation function, pressure and density are estimated at local bins and compared with the experimental database. Our results show good agreement for the molecular dynamics results of the argon system, while the many body dissipative particle model fails to simulate the water properties at room temperature. In its current form, the many body dissipative particle method cannot be used for accurate liquid vapor interfacial simulations and heat transfer studies.
Cross-checking of numerical-experimental procedures for interface characteriz...Diego Scarpa
This document summarizes the cross-checking of experimental and numerical procedures for characterizing interfaces in quasi 2-D transitional flows. The study uses laser light scattering and PIV techniques experimentally, along with 2D DNS simulations, to obtain interface location, velocity fields, strain rates, and interface density for a reference test case. Comparisons show good agreement between experiments and simulations for interface contours, mean velocity profiles, and interface density profiles along the streamwise direction, but some discrepancies in strain rate profiles. The cross-checking helps validate the experimental results and identifies areas for improving numerical models.
Turbulence structure behind the shock in canonical shockvortical turbulence ...Jovan Belite Squad
This document discusses a study using direct numerical simulation (DNS) to examine the interaction between isotropic turbulence and a normal shock wave. High-resolution simulations were performed across a range of shock Mach numbers (1.1-2.2) and turbulent Reynolds numbers (10-45). The results showed that turbulence quantities from DNS converge to predictions of linear interaction analysis (LIA) as the turbulent Mach number decreases, demonstrating the importance of LIA. LIA was also extended to compute full post-shock flow fields, revealing how the shock significantly changes turbulence structures through symmetrization and increased strain-rotation correlation.
This document discusses the built-in filters of two eddy viscosity subgrid scale models used in large-eddy simulation (LES): the Smagorinsky model and the Schumann model. It presents the theoretical basis of LES filtering and equations. The study identifies the forms of the built-in filters associated with each model by varying a characteristic length scale in the models and analyzing the effective filter rendered by the simulations. The behavior of an equivalent Kolmogorov constant is also compared to theory. The results provide insight into how the modeling terms influence the filtering process and flow dynamics in LES without using explicit filtering.
Effect of Geometry on Variation of Heat Flux and Drag for Launch Vehicle -- Z...Abhishek Jain
Above Research Paper can be downloaded from www.zeusnumerix.com
The research paper aims at studying the variation of the geometry of the launch vehicle nose and its effect on heat flux. CFDExpert software is first validated on NASA's hyperballistic model and then used on proposed geometries. Various nose radius and blending shapes are studied for effect on drag and heat flux. Cone ogive shape is found to decrease heat flux with an insignificant increase in drag. Authors Abhishek Jain (Zeus Numerix), Rohan Kedar and Prof V Kalamkar (SPCOE).
An Adaptive Moving Mesh Method For Two-Dimensional Relativistic Magnetohydrod...Scott Faria
This document describes an adaptive moving mesh method for solving two-dimensional thin film flow equations that include surface tension. The method uses moving mesh partial differential equations (MMPDEs) and mesh density functions to adaptively move the computational mesh based on solution characteristics like curvature. Numerical results show this r-adaptive moving mesh technique accurately captures the moving contact line and associated fingering instability with reduced computational effort compared to a fixed uniform mesh.
Multi-task learning using non-linear autoregressive models and recurrent neur...IJECEIAES
Tide level forecasting plays an important role in environmental management and development. Current tide level forecasting methods are usually implemented for solving single task problems, that is, a model built based on the tide level data at an individual location is only used to forecast tide level of the same location but is not used for tide forecasting at another location. This study proposes a new method for tide level prediction at multiple locations simultaneously. The method combines nonlinear autoregressive moving average with exogenous inputs (NARMAX) model and recurrent neural networks (RNNs), and incorporates them into a multi-task learning (MTL) framework. Experiments are designed and performed to compare single task learning (STL) and MTL with and without using non-linear autoregressive models. Three different RNN variants, namely, long short- term memory (LSTM), gated recurrent unit (GRU) and bidirectional LSTM (BiLSTM) are employed together with non-linear autoregressive models. A case study on tide level forecasting at many different geographical locations (5 to 11 locations) is conducted. Experimental results demonstrate that the proposed architectures outperform the classical single-task prediction methods.
Multi-task learning using non-linear autoregressive models and recurrent neur...
Manuscript
1. Using statistical learning to close two-fluid multiphase
flow equations for bubbly flows in vertical channels
Ming Ma, Jiacai Lu, Gretar Tryggvason∗
aDepartment of Aerospace and Mechanical Engineering, University of Notre Dame, Notre
Dame, IN 46556, USA
Abstract
Data generated by direct numerical simulations (DNS) of bubbly up-flow in a
periodic vertical channel is used to generate closure relationships for a simplified
two-fluid model for the average flow. Nearly spherical bubbles, initially placed
in a fully developed parabolic flow, are driven relatively quickly to the walls,
where they increase drag and slowly reduce the flow rate. Once the flow rate has
been decreased enough, some of the bubbles move back into the channel interior
and the void fraction there approaches the value needed to balance the weight
of the mixture and the imposed pressure gradient. A database is generated
by averaging the DNS results over planes parallel to the walls, and a Model
Averaging Neural Network (MANN) is used to find the relationships between
unknown closure terms in a simple model equations for the average flow and
the resolved variables. The closure relations are then tested, by following the
evolution of different initial conditions, and it is found that the model predictions
are in reasonably good agreement with DNS results.
Keywords: Bubble dynamics, Gas/liquid flow, DNS, Two fluid models,
Closure terms, Model averaging neural networks, Statistical learning
∗Corresponding author
Email address: gtryggva@nd.edu (Gretar Tryggvason)
Preprint submitted to International Journal of Multiphaseflow December 23, 2015
2. 1. Introduction
For many multiphase flows the physics is well understood in the sense that
the dynamics is believed to be accurately described by the Navier-Stokes equa-
tions. The large range of temporal and spatial scales in industrial size systems
does, however, make solutions of the equations impractical, even on modern5
supercomputers. For many purposes a low order description of the system, such
as the average motion, is all that is needed. A reduced order description can be
derived by averaging or filtering the Navier-Stokes equations, but the process
introduces unknown closure terms that account for the effect of the unresolved
scales on the resolved ones. Those have to be modelled. The fundamental10
assumption behind modelling of multiphase flows is that the closure terms ap-
pearing in the equations for the averaged or filtered quantities depend on the
averaged flow, such as the deformation tensor and the void fraction, and some-
times on some integral measures of the unresolved motion, such as the turbulent
kinetic energy, dissipation rate, and surface area density. The latter quantities15
evolve by their own equations with their own closure terms that need to be
related to other quantities, in the same way as for the equations for the average
flow. The averaged and filtered equations can be written in several different
forms and depending on the exact form we need different closure relations. For
reviews of modelling of two phase flows see, for example, [1, 2, 3, 4] and [5].20
Here we examine an infinitely long vertical channel with flow driven by an
imposed pressure gradient and gravity. The average flow is homogeneous in the
streamwise and spanwise direction and the averaged quantities depend only on
the wall-normal coordinate, as well as time. The averaged equations were exam-
ined analytically for laminar bubbly flow by [6] and later by [7]. For turbulent25
flow [8] solved the equations for the average flow using asymptotic analysis, and
numerical solutions have been reported by a number of investigators, including
[9, 10, 11, 12]. The results, both for laminar and turbulent flows, generally agree
with the flow structure seen experimentally, once the model coefficients have
been tuned appropriately. For nearly spherical bubbles this includes uniform30
2
3. velocity and void fraction in the core, wall peaking for upflow and a wall-layer
without bubbles for downflow.
In [13] we used neural networks to obtain closure relationships for laminar
bubbly upflows in fully periodic domains from DNS results and found that the
relationships allowed us to predict the evolution of different initial conditions35
(velocity and void fraction profiles) using a simple two-fluid model for the av-
erage flow. Here we examine the evolution of laminar bubbly flows in vertical
channels, again using DNS results to find closure relationships for the aver-
age two-fluid model by model averaging neural networks. Although the flow is
similar to the one examined in our earlier paper, the presence of walls adds con-40
siderable complexity. First of all, and even though the governing parameters are
chosen so that the bubbles remain nearly spherical, surface tension effects are
important near the wall and must be included in the average model, resulting
in new terms that must be closed. Secondly, finding the neural network coeffi-
cients for the relationships between the closure terms and the resolved average45
variables is much more difficult and we have had to adopt new strategies to
train the model. Since the bubbles are driven to the wall relatively quickly at
the early stage, but the flow then changes relatively slowly, new data sampling
strategies have been used. Those strategies are discussed in the method section.
Applications of various statistical techniques have a long history in studies50
of complex flows ([14], for example). As the use of such techniques has become
more widespread for a large range of other problems, ranging from predict-
ing consumer preferences to identifying security threats, methods for statistical
learning have advanced rapidly. For turbulence flows, applications include the
development of an adaptive controller for drag reduction by [15], and various55
reconstructions of the flow. [16] used neural networks to reconstruct near-wall
turbulent flows and [17] used neural network to interpolate DNS data for a back-
ward facing step. Other applications include the use of statistical learning to
quantify the uncertainty in model predictions. [18] used Bayesian probabilistic
approach to quantify the uncertainty in RANS model predictions for turbulent60
channel flow, [19] examined the uncertainty for turbulent boundary layers, and
3
4. [20] used machine learning to compare RANS results with DNS and LES predic-
tions to identify regions of high uncertainty, presumably where the assumptions
behind the RANS model no longer hold. For our purpose, the use of statisti-
cal learning to automate, at least partially, the development and evaluation of65
reduced order models, is most relevant. For turbulent flows, examples includes
the development of a subgrid model for LES by [21], the optimal estimation
of subgrid models for LES by [22], the parametrization of surface features in
coarse LES by [23], and the use of neural networks to optimize the model con-
stants of the k − ε turbulence model applied to simulations of data centres by70
[24]. Most recently, [25] have used statistical learning to determine the func-
tional dependency of the closure terms for data generated by Spalart-Allmaras
turbulence model, rather than full DNS, and [26] and [27] have used inverse
modelling to obtain spatially distributed functional terms to aid closure mod-
elling, instead of inferring model parameters directly. For multiphase flows, [28]75
computed the response of fully resolved particles to shocks and used a neural
network to develop closure laws for macroscopic simulations of the gas-particle
mixture. [29] discuss various techniques to bridge the scales between detailed
microscopic simulations and macroscopic models, focusing on the convergence
rate for various model problems.80
2. Problem Setup
We examine bubbly flows in a rectangular vertical channel, between two ver-
tical plates, with periodic boundaries in the streamwise and spanwise direction,
as sketched in figure 1. At steady state the flow of nearly spherical bubbles has
a simple structure, consisting of a nearly homogeneous core region and a thin85
bubble-rich wall-layer ([30]). The structure is determined by the lift forces on
the bubbles, which drive them toward the side where the relative liquid velocity
is largest, in a frame of reference moving with the bubble. Thus, in upflow the
bubbles are driven toward the walls. Studies of the magnitude of the lift force
and how it depends on the shear and other properties of the flow include [31, 32]90
4
5. g
W
L
H
y
x
z
Figure 1: A sketch of the computational domain. The domain is bounded by vertical walls
and is periodic in the other directions.
and [33]. The lift force depends strongly on the deformability of the bubble and
larger bubbles (and thus more deformable) tend to experience either little lift
or a lift in the opposite direction from the nearly spherical ones ([34]; [35]).
Very little is known about the lift on many bubbles in high void fraction flows,
although preliminary studies can be found in [36]. The horizontal migration of95
bubbles from the core to the wall increases the average density of the mixture
in the core and if there are enough bubbles the weight of the mixture eventu-
ally balances the imposed pressure gradient and as the shear becomes zero the
migration of bubbles stops. The void fraction in the core and the wall layer are
therefore easily predicted. One of the main conclusions of these studies is that100
the steady state flow does not depend on the magnitude of the lift coefficient—
only its sign. This suggests that if we intend to use the DNS results to evaluate
and help construct models of the average flow, it is necessary to examine the
transient motion. Initial efforts to do so can be found in [37], who studied the
transient motion of bubbles in a vertical Couette flow.105
The problem specification is similar to [13]. The width of the domain, in
the wall normal direction is H = 2, in the stream wise direction the length is
L = 1.5 and the span wise width is W = 1.0. To make the computations as
easy as possible, the density and viscosity are taken to be ρl/ρg = 2.5/0.125;
5
6. and µl/µg = 0.01/0.005, where the subscript l denotes the liquid and g stands110
for the gas. Gravity acceleration is gy = 1.0 and total pressure gradient in
the streamwise direction is β = −0.04, where β = dpo/dy + ρavggy. The flow
contains N = 49 bubbles of diameter de = 0.18, giving an average void fraction
of αavg = 0.049. The surface tension coefficient is σ = 0.4. These parameters
result in a Morton number of M = gyµ4
l /ρlσ3
= 2.56 × 10−8
and an E¨otv¨os115
number of Eo = ρlgyd2
/σ = 0.2025. The initial velocity is a parabolic profile
U(y) =
H2
β
2µl
y
H
2
−
y
H
, (1)
where the wall shear balances the pressure gradient and the weight of the mix-
ture. The maximum velocity is Umax = 3.00, and the bulk velocity of the
channel is Ub = 2.0. The channel Reynolds number based on the bulk velocity
and width of the channel is 1000. The domain is resolved by 192 × 192 × 128120
grid points in the stream wise, wall-normal and span wise direction. The grid
is stretched in the wall-normal direction to increase the resolution at the wall.
For further details, see [30]. At steady state, the average wall shear stress τw is
related to the pressure gradient and the weight of the bubble/liquid mixture by
a streamwise momentum balance:125
τw = −(dP0/dz + ρavggy)h = −βH/2, (2)
which is found by taking a simple force balance for the whole domain. Since the
liquid and the bubbles are incompressible, ρavg is constant, and β is therefore
constant, since we assume that the imposed pressure gradient does not change.
The direction of the flow depends directly on the sign of β and for upflow we
must have β < 0.130
The Navier-Stokes equations are solved by an explicit projection method
that is second order in time, using a fixed, staggered grid where the advection
terms are resolved by a QUICK scheme and the viscous terms by a second order
centred approximation. In order to keep the boundary between the different
fluids sharp, to advect the density and the viscosity fields, and to accurately135
compute the surface tension, the fluid interface is tracked by connected marker
6
7. points (the “front”). The front points, which are connected to form an un-
structured surface grid, are advected by the fluid velocity, interpolated from
the fixed grid. As the front deforms, surface markers are dynamically added
and deleted. The surface tension is represented by a distribution of singularities140
(delta-functions) located at the front. The gradients of the density and viscos-
ity become delta functions when the change is abrupt across the boundary. To
transfer the front singularities to the fixed grid, the delta functions are approx-
imated by smoother functions, with a compact support, on the fixed grid. At
each time step, after the front has been advected, the density and the viscosity145
fields are reconstructed by integration of the smooth grid-delta function. The
surface tension is then added to the nodal values of the discrete Navier-Stokes
equations. Finally, an elliptic pressure equation is solved by a multigrid method
to impose a divergence-free velocity field. The method has been applied to sev-
eral multifluid problems and tested and validated in a number of ways. Those150
tests include comparisons with analytical solutions for simple problems, other
numerical computations, and experiments. The actual resolution requirement
varies with the governing parameters of the problem and high Reynolds numbers
flows, for example, generally require finer resolution than low Reynolds number
flows, as in other numerical calculations. Examples of validations and a detailed155
description of the method can be found in [38].
The DNS results contain a detailed description of the flow so that the closure
terms in the averaged equations can be calculated at every spatial and temporal
location, as well as the average velocity, void fraction, and any other quantity of
interest. The fundamental assumption behind modeling the average flow is that160
the closure terms appearing in the model equations depend on the average flow
and possibly some integral measures of the unresolved motion. For turbulent
flows it is essential to include integral measures of the unresolved motion but
here we assume that such terms are not needed, since all the unsteadiness is due
to the motion of the bubbles. To derive equations for the average void fraction165
and the average vertical velocity as functions of the wall-normal coordinate, we
integrate the mass and momentum equations over planes parallel to the walls,
7
8. taking the density and the viscosity of the gas to be zero. The averaging is
essentially the same as is worked out in detail in [13], except here we do not
assume that the bubbles remain completely spherical. We define the average170
liquid volume fraction and the average phasic vertical liquid velocity by
αl =
1
Azy
χda, < v >l=
1
αlAzy
χvda. (3)
where χ is an indicator function such that χ = 1 in the liquid phase and χ = 0
in the gas. Azy is the area of a plane parallel to the walls, as sketched in figure
1. The results of the averaging are two equations for the liquid volume fraction
and the average vertical velocity of the liquid. The equation for the average175
liquid volume fraction is:
∂αl
∂t
+
∂Fl
∂x
= 0. (4)
The horizontal flux of liquid is given by
Fl =
1
αlAzy
χuda = αl < u >l . (5)
The gas volume fraction, αg, and average horizontal velocity, < u >g, are found
in the same way. Adding the equations for the gas and the liquid, using that the
volume fractions must add up to unity, αl + αg = 1, and that the u velocities180
at the walls are zero, shows that
Fg + Fl = αg < u >g +αl < u >l= 0. (6)
Thus, gas flow in one horizontal direction is always balanced by liquid flow in
the opposite direction.
The averaged equation for the vertical liquid velocity < v >l is:
∂
∂t
αl < v >l +
∂
∂x
αl < u >l< v >l=
−
1
ρl
dp
dy
−
1
ρl
αlρlgy +
1
ρl
∂
∂x
αlµl
∂ < v >l
∂x
−
∂
∂x
αl < u v >l + (fσ)y. (7)
where the last term is the average surface tension. Notice that the momentum185
equation contains the liquid fluxes since the average horizontal liquid velocity
is given by αl < ul >= Fl. Equations (4) and (7) give the evolution of αl
8
9. and < v >l, but require closure relations for Fl, < u v >l and (fσ)y. The
averaged equations are solved by a simple first order upwind method using a
large number of grid points and very small time steps to ensure that the results190
have converged. The code has been written in R, to simplify the incorporation
of the functional relationships for the closure terms (see below).
The data generated by the DNS is averaged over planes parallel to the walls,
resulting in a database containing the average gas flux, liquid streaming stresses,
and surface tension, along with void fraction, void fraction gradient, gradient of195
the average vertical liquid velocity, and the distance to the nearest wall, at the
horizontal grid point, at all time intervals. We now assume that each of the first
three quantities depends uniquely on the latter four, at least in some average
sense (since we expect considerably scatter):
Fg = f(x), < u v >l= g(x), (fσ)y = s(x), (8)
where200
x =
∂ < vl >
∂x
, αg,
∂αg
∂x
, dw . (9)
The functions f, g, and s are determined using the DNS database. While these
functional relationships can be found in many ways, here we use Model Averag-
ing Neural Networks (MANN). The learning strategy for the Neural Networks
part is similar to [13], since we expect the relationships to be nonlinear. The
unknown coefficients are initially set to small values near zero and and then205
adjusted by an iterative process. Since the iterations are stopped before they
have reached a final converged value, to avoid over-fitting, different initial values
can lead to different fits. To reduce the variance between training sessions with
different initial values two techniques are combined here: in the first method we
create several NN models using different randomly selected starting coefficients,210
and aggregate all the NN models to generate the averaged NN fit. This pro-
cedure has been shown to result in more stable predictions ([39, 40, 41]). The
other technique is to use resampling and bagging the data during the training of
each NN, where we resample the training data with replacement for each NN’s
training dataset before aggregating all the models, and this has been shown215
9
10. to result in improvements for unstable procedures like NN and decision tree
method ([42, 43]). Both MANN and data resampling techniques are available
in the Caret package, implemented in R ([44]).
During the training, we pool all the spatial and temporal data together for
the upflow channel case, and, unlike for the fully periodic case ([13]) where all220
the data is sampled uniformly at the early time, here we select the frequency
of the training data following a exponential distribution based on time, since
for the upflow case considered here, the bubbles are initially driven relatively
quickly to the wall, but at the late stage the flow evolves very slowly. The final
NN is averaged by 5 NNs, each with a 3 layer structure: 4 inputs, 12 neurons225
and 1 output for each unknown closures. All the unknown coefficients for each
NN start randomly below 0.1 and the maximum number of iteration for each
NN’s training is 250. We have tested the “tuning” parameters such as the range
of small starting random numbers, number of neurons, iteration steps, and the
number of NNs needed before aggregation, and selected the numbers given here230
for the best performance.
Neural networks are, of course, not the only statistical learning technique
that can be used, and we have also experimented with linear regressions. Unlike
MANN, where the functional form of the closure terms are unknown, here we
postulate a functional form for the relationship with unknown coefficients and235
determine the unknown coefficients from the DNS data. We assume that the
formula for gas flux is:
Fg = C(−sign(
d < v >l
dx
))αm
g (
d < v >l
dx
)n
d(x)p
. (10)
The direction of the gas flux is modelled by adding the −sign(d < v >l /dx)
term, since the sign the of the gas flux is determined by the direction of the
lift force. Considering the gas flux to be correlated with αg, d < v >l /dx240
and d(x), the distance to the nearest wall, we propose a nonlinear closure with
unknown powers. We find the unknown parameters from the same DNS data
used for neural networks’ training and use the bootstrap method to reduce the
uncertainty of the training results. We note that we have not included the
10
11. gradient of αg, since at locations where dαg/dx = 0, the gas flux Fg is not245
necessarily zero.
3. Results
Figure 2 shows the bubbles and the velocity field in the mid-plane through
the domain at six times. The first frame is at time zero where the flow is
parabolic and the bubbles are spread nearly homogeneously across the channel.250
In the next frame, most of the bubbles have moved toward the wall. There is a
layer of bubbles at both walls, and only a couple of bubbles in the middle. The
bubble motion is fairly unsteady and the wall-layer is often broken up by the
motion of the bubbles just outside of it. This development continues in the third
frame, where the middle of the channel is mostly free of bubbles. In the fourth255
frame, all the bubbles are near the walls, and in frame five, the bubbles stay
close to the wall, but the average liquid velocity is lower. In the sixth frame,
at a much later time, some of the bubbles have moved back into the channel
interior and the void fraction there approaches the value needed to balance the
weight of the mixture and the imposed pressure gradient.260
3.1. Training of the model
The results shown in figure 2 provide the training data for our model. Before
we plug in the models for the closure terms into the averaged equations, we
need a reliable training procedure for the neural networks. We collect various
quantities averaged over planes parallel to the walls, together with the unknown265
closure terms, at different times and use neural networks for the fittings. Here
we assume that the unknown closure terms in equations (6) and (7) depend only
on the averaged quantities and not on any average description of the unresolved
field. Obviously there cannot be any dependency on < v >l directly, so to the
lowest order we assume a dependency on the velocity gradient. The closure270
terms can, however, depend on the void fraction and we include the possibility
that there is also a dependency on the gradient of αg. We have done a test of
11
12. Figure 2: The bubble distribution at several different times, starting with the initial conditions
in the top left corner.
12
13. the model variable selection, and in [13] these three variables were found to be
enough. For wall bounded flow we must also allow for a possible dependency on
the distance to the wall dw. Figure 3 shows the quality of the fitting. In the left275
column we show a histogram of the training error distribution for the gas fluxes,
streaming stress and surface tension. Most of the training errors are clustered
around zero, and the number of instance falls off rapidly as the error becomes
larger. The distribution is nearly Gaussian, though there is a slight skewness.
In the right column we show a scatter plot of the predicted average gas fluxes280
versus the average fluxes from the DNS for the test data. Although there is
some spread, the points are mostly clustered around the forty five degree line.
The output from the Caret package is a function with the relationship and we
use these functions for each closure term in an R code to solve equations (6)
and (7). The predictions of the model are shown in figure 4 where we plot the285
average liquid velocity at five times in the left column and the void fraction
at the same times in the right column, along with the averaged DNS results
and from the model where the closure terms are found by linear regression.
The results again show that when nearly spherical bubbles are injected into
parabolic flow the evolution toward steady state is highly non-monotonic. First,290
all the bubbles migrate towards the walls, leaving the centre region nearly free
of bubbles. Then the presence of the bubbles near the wall increases the friction
there and reduces the flow rate. As the flow rate is reduced sufficiently some
of the bubbles migrate back into the core region until the mixture there is in
hydrostatic equilibrium. Thus, the initially parabolic velocity decreases rapidly295
in the middle of the channel and is nearly uniform across almost all of the
channel at the last time shown, going to nearly zero only very near the walls.
The void fraction profile is more complex but a careful inspection shows that the
bubbles rapidly accumulate near the walls, with the void fraction going to zero
in the middle of the channel and then returning to a profile consisting of nearly300
uniform value across most of the channel with peaks near the walls in a layer
of thickness of about one bubble diameter. The initial migration of the bubbles
to the wall takes place relatively fast, but the slowing down of the flow and the
13
14. −0.01 −0.005 0 0.005 0.01
0
500
1000
1500
2000
2500
3000
3500
4000
4500
5000
F
g
Fitted Error Histgram
Instances
−0.01 −0.005 0 0.005 0.01
−0.01
−0.005
0
0.005
0.01
DNS
Model
R=0.93602, F
g
Regression Plot
Data
Fit
Y = T
−2 −1 0 1 2
x 10
−3
−2
−1.5
−1
−0.5
0
0.5
1
1.5
2
x 10
−3
DNS
Model
R=0.8666, <u’v’> Regresion Plot
Data
Fit
Y = T
−0.002 −0.001 0 0.001 0.002
0
1000
2000
3000
4000
5000
6000
7000
<u’v’> Fitted Error Histgram
Errors
Instances
−0.2 0 0.2 0.4 0.6 0.8
−0.2
0
0.2
0.4
0.6
0.8
DNS
Model
R=0.99416, f
σ
Regression Plot
Data
Fit
Y = T
−0.05 −0.025 0 0.025 0.05
0
1000
2000
3000
4000
5000
6000
7000
8000
9000
f
σ
Fitted Error Histgram
Error
Instances
Figure 3: The training error histograms (left column) and the regression plots (right column)
for the three closure terms: Fg, < u v >, and fσ.
14
15. coefficients Mean Value Standard Deviation
C 2.84e-2 2.24e-3
m 6.54e-1 6.88e-2
n 7.88e-1 6.68e-2
p 3.77 5.57e-1
Table 1: The estimated regression coefficients and corresponding standard deviations, using
bootstrap.
migration of the bubbles back into the core is a much slower process. Here we
do not allow the bubbles to coalesce, but if they could then the migration to305
the wall might promote the formation of either larger bubbles that would move
away from the wall relatively quickly or possibly a gas film at the wall. The
linear regression is more sensitive to the data sampling techniques and data size
than the MANN. Here a bootstrapping method is used for selecting the data
with replacement for 2000 times (we gradually increase the data sample size310
from 500, and end at 2000), generating 2000 sub-dataset. In each subset we do
a linear regression and then take the mean value of the unknown coefficients and
estimate its standard deviation, in order to reduce the uncertainty and get more
reliable results. Using bootstrap we can also estimate the standard deviation
for each coefficient, as given in Table 1315
In addition to directly checking the average void fraction and liquid velocity,
we have also examined the unsteady evolution of the flow as shown in figure 5,
where we plot the average wall shear (top) and the average flow rate (bottom)
as obtained from DNS data and from the model. Here the liquid flow rate Q
and average wall shear rate γ are defined as:320
Q =
H
0
< v >l dx, γ =
d < v >l
dx
|x=0 (11)
Initially the shear stress is in balance with the steady state flow of the liquid
but as bubbles accumulate near the walls, the wall shear increases and the flow
starts to slow down. The wall shear quickly reaches maximum, when most of
15
16. 0 0.5 1 1.5 2
0
0.5
1
1.5
2
Location
T=80.0s
0 0.5 1 1.5 2
0
0.1
0.2
0.3
Void Fraction
0 0.5 1 1.5 2
0
0.1
0.2
0.3
0 0.5 1 1.5 2
0
0.1
0.2
0.3
0 0.5 1 1.5 2
0
0.1
0.2
0.3
0 0.5 1 1.5 2
0
0.1
0.2
0.3
Location
0 0.5 1 1.5 2
0
0.5
1
1.5
2
Vertical Velocity
T=5.0s
DNS
Neural Networks
Linear Regression
0 0.5 1 1.5 2
0
0.5
1
1.5
2
T=10.0s
0 0.5 1 1.5 2
0
0.5
1
1.5
2
T=20.0s
0 0.5 1 1.5 2
0
0.5
1
1.5
2
T=40.0s
Figure 4: A comparison of the prediction of the model with averaged DNS data. The average
liquid velocity is shown on the left, at several times, and the average void fraction on the
right.
16
17. 0 10 20 30 40 50 60 70 80
Time
1.5
2.5
3.5
4.5
5.5
6.5
7.5
8.5
d<v>l/dxx=0
Shear Rate at Wall
DNS
ANN Model
Regression Model
0 10 20 30 40 50 60 70 80
Time
0
0.5
1
1.5
2
2.5
3
H
0<v>ldx
Liquid Flow Rate
DNS
Regression Model
ANN Model
Figure 5: The average wall shear (top) and the flow rate (bottom) versus time, for the simu-
lation shown in figure 4.
17
18. the bubbles are at the wall, and then drops slowly as the flow rate continues to
slow down.325
3.2. Validation of the Model: applications to different initial conditions
Ideally the relationship derived in the last section by MANN is sufficiently
general so that it can be used for flows that are different from the “training” case.
To examine how well it does, while still remaining within the rather restrictive
setup in which it was developed, we have examined a few different conditions.330
Those include: up flow cases with different shear rate; more complex initial
velocity shaped as sin function; down flow cases with different shear rates; and
a vertical Couette flow.
In figure 6 we show a comparison of the model predictions and averaged
DNS data for downflow, at five times, using the closure relationships derived335
for the upflow presented earlier. The average liquid velocity is shown in the
left column and the void fraction at the same times in the right column, along
with the averaged DNS results. The changes in the flow are far less dramatic
than for the upflow. The bubbles move slightly closer to the channel centre and
increase the buoyancy there, flattening the velocity profile, resulting in a nearly340
uniformly void fraction in the middle of channel and well-defined bubble free
layers near the walls.
Similar to the upflow case, figure 7 shows a comparisons of the wall shear
rate and the liquid flow rate, between DNS data and model predictions, for
the downflow case. Overall the agreement between the model and the DNS is345
good and the minor differences observed are mainly during the early times. In
general, if we train and test the model on the same case, we would expect a
better prediction performance than testing the learned model on different cases.
However, it is not always the case since we also need to consider the complexity
of the physics. From figures 6 and 7, we can see that even though this is a350
testing case, the closure results in predictions that are somewhat better than
for the original training case.
The agreement between the model predictions and the averaged DNS date
18
19. 0 0.5 1 1.5 2
-3
-2.5
-2
-1.5
-1
-0.5
0
T=4.0s
Vertical Velocity
DNS
Model
0 0.5 1 1.5 2
-3
-2.5
-2
-1.5
-1
-0.5
0
T=8.0s
0 0.5 1 1.5 2
-3
-2.5
-2
-1.5
-1
-0.5
0
T=16.0s
0 0.5 1 1.5 2
-3
-2.5
-2
-1.5
-1
-0.5
0
T=24.0s
0 0.5 1 1.5 2
Location
-3
-2.5
-2
-1.5
-1
-0.5
0
T=28.0s
0 0.5 1 1.5 2
0
0.1
0.2
0.3
Void Fraction
0 0.5 1 1.5 2
0
0.1
0.2
0.3
0 0.5 1 1.5 2
0
0.1
0.2
0.3
0 0.5 1 1.5 2
Location
0
0.1
0.2
0.3
0 0.5 1 1.5 2
0
0.1
0.2
0.3
Figure 6: A comparison of model predictions and averaged DNS results for down flow. The
average liquid velocity is shown on the left, at several times, and the average void fraction on
the right.
19
20. 0 5 10 15 20 25
1.5
3
4.5
6
7.5
9
d < v >l /dx|x=0
Time
DNS
Model
0 5 10 15 20 25
Time
3
3.5
4
4.5
H
0 | < v >l |dx
DNS
Model
Figure 7: The average wall shear (top) and the flow rate (bottom) versus time for the simu-
lation shown in figure 6.
20
21. Case velocity Initial void fraction MSD1 MSD2
Train 2(−x2
+ 2x) uniform αg =4.99% 0.05661 0.002536
Test 1 3(−x2
+ 2x) uniform αg =4.99% 0.08361 0.006262
Test 2 2(x2
− 2x) uniform αg =4.99% 0.01632 0.001637
Test 3 3(x2
− 2x) uniform αg =4.99% 0.03955 0.002374
Test 4 sin πx uniform αg =4.99% 0.04801 0.006202
Test 5 −x + 1 uniform αg =4.99% 0.07569 0.004565
Table 2: The various initial conditions used to train and test the model for the closure terms.
The last two columns are defined by equation (12) and (13).
for other conditions, using the same closure relations, are similar. To summarize
the results, we show in figure 8 the difference between the model and the DNS355
data versus time, for six cases, listed in Table 2. The results are condensed even
further in the two rightmost columns in Table 2 where the mean square different
of the error, defined by
MSD1 =
1
T
(γDNS − γmodel
2
dt, (12)
MSD2 =
1
T
(QDNS − Qmodel
2
dt, (13)
In all cases the differences between the model predictions and the DNS data
are of comparable order of magnitude, unlike the situation in [13], where the360
agreement is generally worse than for the training data. For wall bounded
bubbly flow, the dynamics of the training case itself is more complex than for
some of the testing cases, like the downflow ones.
21
22. 0 5 10 15 20 25
0
0.25
0.5
0.75
1
Time
|QDNS − QM odel|/(QDNS)max
Train
Test 1
Test 2
Test 3
Test 4
Test 5
0 5 10 15 20 25
0
0.25
0.5
0.75
1
Time
|γDNS − γM odel|/(γDNS)max
Train
Test 1
Test 2
Test 3
Test 4
Test 5
Figure 8: The difference between the liquid flow rate (top frame) and wall shear rate (bottom
frame) predicted by the model and found from the DNS data versus time for all the initial
conditions in Table 2. The differences have been normalized by the maximum value of liquid
flow rate and wall shear rate in the DNS results.
22
23. 3.3. Closure Terms Analysis
The results of the fitting, using the neural network, are transferred to the365
numerical implementation of the model as functions that can be treated as a
“black box.” A visual inspection does, however, help us evaluate the quality of
the fit and understand what physical processed are captured. In figure 9 and
10 we show how the gas fluxes, the streaming stresses and the average surface
tension depend on the average liquid shear rate. In the top frame of figure 9370
we set the void fraction equal to αg = 0.1 and show the relationships as the
distance to the wall gradually increases. In the bottom frame of figure 9 the
distance to the wall is set to two bubble diameters and we show the flux for
a few different void fractions. The wall distance function is linear and reaches
its maximum value 1 at the centre of the domain. We have also experienced375
with other types of wall distance function and not found much difference since
the NN automatically adjusts the power of the wall distance function. For zero
velocity gradient the gas fluxes are zero and for gradient near zero the gas fluxes
are a decreasing function of the liquid shear rate, as we would expect for nearly
spherical bubbles. The flux is lowest for bubbles close to the wall and the effects380
of the wall become smaller as dw increases. Eventually we would expect the wall
to have no effect and the dashed line in figure 9 is a fit from simulations with
no walls. Obviously the current results away from the walls agree well for low
shear rates. As the shear rate is increased, however, the flux reaches a maximum
and then decreases. This can, at least in principle, be due to two effects: The385
increased shear deforms the bubbles more and thus the lift changes sign, or this
can be a result of the highest shear being found close to the walls, where the
flux is reduced due to wall effects and not deformability. We suspect that it is
the latter, as discussed below.
23
24. −20 −15 −10 −5 0 5 10 15 20
−0.015
−0.01
−0.005
0
0.005
0.01
0.015
d<v>
l
/dx
F
g
Plot of Fg, αg = 0.1
d=0.5D
d=1.0D
d=2.0D
d=3.0D
d=4.0D
d=5.0D
−20 −15 −10 −5 0 5 10 15 20
−0.015
−0.01
−0.005
0
0.005
0.01
0.015
d<v>
l
/dx
F
g
Plot of Fg, d = 2D
a=0.05
a=0.10
a=0.15
a=0.20
a=0.25
a=0.30
Figure 9: The closure term Fg as a function of the liquid shear rate for different distances
to the wall at a fixed value of void fraction αg = 0.1 (top frame), and with different void
fractions at a distance to the wall of two bubble diameters (bottom frame).
24
25. −12 −8 −4 0 4 8 12
−0.04
−0.02
0
0.02
0.04
d<v>
l
/dx
−<u’v’>
Plot of < u′
v′
>, α = 0.1
d=0.5D
d=1.0D
d=2.0D
d=3.0D
d=4.0D
d=5.0D
−12 −8 −4 0 4 8 12
0
0.1
0.2
0.3
0.4
d<v>
l
/dx
-fσ
Plot of fσ, d = 2D
a=0.05
a=0.10
a=0.15
a=0.20
a=0.25
a=0.30
Figure 10: The closure term < u v >l as functions of the liquid shear rate with different
distances to the wall at a fixed value of void fraction αg = 0.1 (top frame), and the closure
term fσ with different void fractions at a fixed distance to the wall equal to two bubble
diameters (bottom frame).
25
26. -20 -15 -10 -5 0 5 10 15 20
d < v >l /dx
-15
-10
-5
0
5
10
15
Fb
Comparison of Learned Closures
ANN α=0.1
Linear regression, α=0.1
ANN α=0.2
Linear regression, α=0.2
Figure 11: Comparisons of fitted closure term Fg with MANN model (solid line) and linear
regression model (dash line) at αg = 0.1 (red line) and αg = 0.2 (black line).
The fit for the streaming stress shows a similar trend. For zero shear the390
stress is zero and for low shear rate it is a decreasing functions of the liquid
shear rate. It is lowest close to the wall and its depending on the distance to
the wall is reduced as we move further away. As the top frame in figure 10
shows, the slop of the streaming stress is converging to a constant value in the
low shear rate region, as the distance to the wall increases, and approaches the395
slope fitted by the periodic case without walls, shown by the dash line. The
bottom frame of Figure 10 shows the average surface tension term versus the
fluid shear for a distance to the wall equal to two bubble diameter. As the void
fraction increases, more bubbles are near the wall, so the average surface tension
increases. Looking at the fluxes directly can give us insights into the closure400
relationships and allow us to check if they are aligned with our intuition. Such
an inspection can help us identify problems. In our case, for example, the gas
fluxes go to zero for high shear rate for any distance to the wall. While it is
true that high shear rates deform the bubbles and reduce lift and thus gas flux,
the reason is much more likely to be that we do not have any data points for405
high shear away from the walls. We suspect that the similar behaviour for the
streaming stress is for similar reasons.
26
27. We have also examined the fluxed found using the linear regression and they
are generally in good agreement with the fluxes found by the MANN, as shown
in figure 11, where we plot the gas fluxes versus the shear rate in the liquid, for410
two values of the void fraction and a distance to the wall equal to two bubble
diameters, as obtained by both approaches. Notice however that the linear
regression neglect any dependency on the void fraction gradient, so we would
not expect a perfect agreement.
4. Conclusions415
Here we have examined how DNS results can be used to provide closure
relations for a simple model of the average flow. We have done this in the
context of a fairly simple setup, following the transient evolution of bubbly
flow in a vertical channel. The bubbles are all of the same size, the initial
liquid velocity is taken to be laminar and parabolic, and the Reynolds number420
is sufficiently low so that velocity disturbances are likely to be only due to the
bubble motion. We use a neural network that is trained on one DNS data set
and then applied to other situations where the initial velocity, void fraction and
boundary conditions are different, but for which we also have DNS results to
compare with. Overall, the average equations with the neural network closure425
reproduce the main aspects of the DNS results, including the generally rapid
change initially, unsteadiness remaining after the initial adjustment and the final
steadiness state.
We have also found the closure terms by postulating a relationship with un-
known coefficients that are found using linear regression. Both approaches give430
comparable results, but given the slightly better accuracy and their generality,
we believe that the neural networks are the preferred approach for more complex
flows.
We emphasize that the system considered here, as well as the two-equation
model, are relatively simple. Specifically, no average descriptions of unresolved435
quantities appear to be necessary (which might be governed by their own evolu-
27
28. tion equations with unknown closure terms, like the turbulent kinetic energy).
Furthermore, the gas flux appears to be a function of the local state and no
evolution equation for the bubble velocity needs to be solved. We have also
examined a model where a new equation governing the evolution of gas flux440
term is added, and the averaged horizontal pressure gradient appears as a new
closure terms instead of gas flux term. The three-equation model does not dra-
matically increase the prediction accuracy, at least not for the current parameter
range. We close by noting that while most multiphase flows are considerably
more complex than the system considered here, we believe that there is nothing445
that intrinsically limits using statistical learning for the closure terms to simple
systems only.
Acknowledgement
This work is supported by the National Science Foundation Grant CBET-
1335913.450
References
[1] M. Ishii, Thermo-Fluid Dynamic Theory of Two-Phase Flow, Eyrolles,
1975.
[2] I. Kataoka, A. Serizawa, Basic equations of turbulence in gas liquid two-
phase flow, Int. J. Multiphase Flow 12 (1989) 745–758.455
[3] D. Z. Zhang, A. Prosperetti, Ensemble phase-averaged equations for bubbly
flows, Phys. Fluids 6 (1994) 2956–2970.
[4] D. A. Drew, S. L. Passman, Theory of multicomponent fluids, Springer,
1999.
[5] A. Prosperetti, G. Tryggvason, Computational Methods for Multiphase460
Flow, Cambridge University Press, 2007.
28
29. [6] S. P. Antal, R. T. Lahey, J. E. Flaherty, Analysis of phase distribution in
fully developed laminar bubbly two-phase flows, Int. J. Multiphase Flow
15 (1991) 635–652.
[7] O. E. Azpitarte, G. C. Buscaglia, Analytical and numerical evaluation of465
two-fluid model solutions for laminar fully developed bubbly two-phase
flows, Chem. Eng. Sci. 58 (2003) 3765–3776.
[8] D. A. Drew, R. T. Lahey, Jr., Analytical modeling of multiphase flow, In:
Particulate Two-phase Flow (1993) 509–566.
[9] M. Lopez De Bertodano, R. L. Jr., O. Jones, Development of a k– model470
for bubbly two-phase flow, J. Fluids Eng. 13 (1987) 327–343.
[10] M. Lopez De Bertodano, R. L. Jr., O. Jones, Phase distribution in bubbly
two-phase flow in vertical ducts., Int. J. Multiphase Flow 20 (1994) 805–
818.
[11] T. C. Kuo, C. Pan, C. C. Chieng, Eulerian–Lagrangian computations on475
phase distribution of two-phase bubbly flows, Int. J. Numer. Meth. Fluids
24 (1997) 579–593.
[12] S. Guet, G. Ooms, R. Oliemans, Aiche j., Simplified two-fluid model for
gas-lift efficiency predictions 51 (2005) 1885–1896.
[13] M. Ma, J. Lu, G. Tryggvason, Using statistical learning to close two-fluid480
multiphase flow equations for a simple bubbly system, Physics of Fluids
(1994-present) 9 (2015) 092101.
[14] G. Huffman, C. Jones, R. Brodkey, Optimization of turbulence models by
means of a logical search algorithm, Applied Scientific Research 27 (1)
(1973) 321–334.485
[15] C. Lee, J. Kim, D. Babcock, R. Goodman, Application of neural networks
to turbulence control for drag reduction, Physics of Fluids 9 (6) (1997)
1740–1747.
29
30. [16] M. Milano, P. Koumoutsakos, Neural network modeling for near wall tur-
bulent flow, Journal of Computational Physics 182 (2002) 1–26.490
[17] E. Rajabi, M. R. Kavianpour, Intelligent prediction of turbulent flow over
backward-facing step using direct numerical simulation data, Engineering
Applications of Computational Fluid Mechanics 6 (4) (2012) 490–503.
[18] T. A. Oliver, R. D. Moser, Bayesian uncertainty quantification applied to
rans turbulence models, in: Journal of Physics: Conference Series, Vol.495
318, IOP Publishing, 2011, p. 042032.
[19] S. H. Cheung, T. A. Oliver, E. E. Prudencio, S. Prudhomme, R. D. Moser,
Bayesian uncertainty analysis with applications to turbulence modeling,
Reliability Engineering & System Safety 96 (9) (2011) 1137–1149.
[20] J. Ling, J. Templeton, Evaluation of machine learning algorithms for predic-500
tion of regions of high reynolds averaged navier stokes uncertainty, Physics
of Fluids (1994-present) 27 (8) (2015) 085103.
[21] F. Sarghini, G. de Felice, S. Santini, Neural networks based subgrid scale
modeling in large eddy simulations, Computers & Fluids (2003) 97–108.
[22] A. Moreau, O. Teytaud, J.-P. Bertoglio, Optimal estimation for large-eddy505
simulation of turbulence and application to the analysis of subgrid models,
Physics of Fluids 18 (10) (2006) 105101.
[23] I. Esau, G. Rieber, On application of artificial neural network methods
in large-eddy simulations with unresolved urban surfaces, Modern Applied
Science 4 (2010) 3–11.510
[24] S. Yarlanki, B. Rajendran, H. Hamann, Estimation of turbulence closure
coefficients for data centers using machine learning algorithms, in: Thermal
and Thermomechanical Phenomena in Electronic Systems (ITherm), 2012
13th IEEE Intersociety Conference on, IEEE, 2012, pp. 38–42.
30
31. [25] B. Tracey, K. Duraisamy, J. J. Alonso, A machine learning strategy to515
assist turbulence model development, in: AIAA Paper 2015-1287, 2015.
[26] K. Duraisamy, Z. J. Zhang, A. P. Singh, New approaches in turbulence and
transition modeling using data-driven techniques, in: AIAA Paper 2015-
1283, 2015.
[27] E. J. Parish, K. Duraisamy, A paradigm for data-driven predictive model-520
ing using field inversion and machine learning, Journal of Computational
Physics 305 (2016) 758–774.
[28] C. Lu, S. Sambasivan, A. Kapahi, H. Udaykumar, Multi-scale modeling of
shock interaction with a cloud of particles using an artificial neural network
for model representation, Procedia IUTAM 3 (2012) 25–52.525
[29] O. Sen, S. Davis, G. Jacobs, H. Udaykumar, Evaluation of convergence be-
havior of metamodeling techniques for bridging scales in multi-scale multi-
material simulation, Journal of Computational Physics 294 (2015) 585–604.
[30] J. Lu, S. Biswas, G. Tryggvason, A DNS study of laminar bubbly flows in
a vertical channel, Int. J. Multiphase Flow 32 (2006) 643–660.530
[31] A. Rust, M. Manga, Orientation and deformation of bubbles in shear flows,
Journal of Colloid and Interface Science 249 (2002) 476–480.
[32] R. Zenit, Y. H. Tsang, D. L. Koch, A. S. Sangani, Shear flow of a suspension
of bubbles rising in an inclined channel, Journal of Fluid Mechanics 515
(2004) 261–292.535
[33] A. Tomiyama, H. Tamai, I. Zun, S. Hosokawa, Transverse migration of
single bubbles in simple shear flows, Chem. Eng. Sci. 57 (2002) 1849–1858.
[34] A. Kariyasaki, Behavior of a single gas bubble in a liquid flow with a
linear velocity profile, in: Proceedings of the 1987 ASME-JSME Thermal
Engineering Joint Conference, Vol. 5, 1987, pp. 261–267.540
31
32. [35] E. Ervin, G. Tryggvason, The rise of bubbles in a vertical shear flow, ASME
J. Fluid Engineering 119 (1997) 443–449.
[36] K. Sankaranarayanan, X. Shan, I. G. Kevrekidis, S. Sundaresan, Analysis
of drag and virtual mass forces in bubbly suspensions using an implicit
formulation of the lattice Boltzmann method, J. Fluid Mech. 452 (2002)545
61–96.
[37] J. Palacios, G. Tryggvason, The transient motion of buoyant bubbles in a
vertical couette flow, AMD Contemporary Mathematics Series 466 (2008)
135–146.
[38] G. Tryggvason, R. Scardovelli, Zaleski, Direct Numerical Simulations of550
Gas-Liquid Multiphase Flow, Cambridge University Press, 2011.
[39] M. P. Perrone, L. N. Cooper, When networks disagree: Ensemble methods
for hybrid neural networks, in: R. J. Mammone (Ed.), Neural Networks for
Speech and Image Processing,, Chapman and Hall, 1993.
[40] B. D. Ripley, Pattern Recognition and Neural Networks, Cambridge Uni-555
versity Press, 1996.
[41] K. Tumer, J. Ghosh, Analysis of decision boundaries in linearly combined
neural classifiers, Pattern Recognition 29 (2) (1996) 341–348.
[42] L. Breiman, Bagging predictors, Machine Learning 24 (1996) 123–140.
[43] C. M. Bishop, Neural networks for pattern recognition, Claridon Press,560
Oxford, 1995.
[44] M. Kuhn, K. Johnson, Applied predictive modeling, Springer,, 2013.
32