Professional Documents
Culture Documents
A R T I C L E I N F O A B S T R A C T
Keywords: Turbulent reactive flow simulation often requires accounting for turbulence-chemistry interactions and the sub-
Turbulent Reactive Flow Simulation grid phenomena. Their complexity leads to a trade-off between computational efficiency on one hand and
Artificial Neural Network computational accuracy on the other. We attempt to bridge this gap by coupling the power of machine learning
Lagrangian PDF Method
with the turbulent reactive flow simulation, specifically in the form of a deep artificial neural network. The
Turbulence-Chemistry Interaction
Lagrangian Monte Carlo method is chosen as a demonstration case as it is one of the most accurate models for
Sub-Grid Effect
turbulent reactive flow simulation, but also one of the most time-consuming. The workflow consists of training
data generation, deep neural network construction, and implementation in ANSYS-Fluent, followed by an
evaluation of model accuracy and efficiency, which results in an order of magnitude faster simulation without
loss of accuracy thanks to our data-driven deep neural network. This approach can be of universal relevance in
speeding up time-consuming models in the field of reactive flow simulation.
https://doi.org/10.1016/j.cej.2021.132442
Unlike the moment closure, the Probability Density Function closure learning, i.e. an ANN containing a certain number of hidden layers, has
scheme uses distribution functions instead of lumped values to describe shown to be very performant in many diverse fields [26–30]. After
scalars in a computational grid cell, to account for Sub-Grid Effects providing relevant variables as inputs and desired variables as outputs,
(SGE) [16]. Two representative methods can be found: (1) the deep ANN links the two via learning, i.e. credit assignment to weights
presumed-PDF method and (2) the full or transport PDF method. As its and bias when minimizing properly defined loss functions [31]. Deep
name suggested, the former assumes a distribution shape in advance for ANN exhibits powerful capabilities in data regression and pattern
the PDF, such as β-distribution based models [17–19] and δ-function recognition in general and in chemical (engineering) research in
based models [8,9]. Although a presumed distribution form is an particular [32]. Examples include transport scalars prediction [33,34],
effective way to close the scalar covariance term, the shortcoming is the species properties prediction [35,36], synthetic route design [37,38],
presumption itself, which leads to different model performance [20]. optimization framework construction [39], flow pattern classification
The transport PDF method, however, requires no presumption of the [40] to name a few. In addition, it has also been successfully coupled to
distribution shape by introducing the Monte Carlo method to solve the CFD simulations [31,41–47]. Particularly, a deep ANN works quite well
PDF transport equations [21]. The Monte Carlo algorithm defines in approximating numerical solutions for underlying partial differential
notional particles in computational grids to represent the PDF and take equations (PDEs) [48–50]. Since the reaction rate in turbulent reactive
the SGE into account. Since the turbulent characteristic properties, such flow is always highly nonlinear, it is possible to use a deep ANN to
as the turbulent kinetic energy and dissipation rate, are involved when calculate e.g. the chemical source terms present in the PDEs corre
tracking the Lagrangian notional particles, the TCI eventually plays a sponding to the transport PDF method. In this case, a data-driven
crucial role in the closure of the chemical source term. It is worth approach is expected to obtain a fast simulation of turbulent reactive
mentioning that solving the transport PDF could also be under the flow based on the accuracy of the full/transport PDF method, which is
Eulerian framework, examples being the direct quadrature method of also the main motivation of this work. As pointed out in a recent
moments combined with the interaction-by-exchange-with-the-mean research of developing a digital twin of a combustion furnace, the
micromixing model (DQMOM-IEM) and stochastic fields method concept of the reduced-order modelling is essential for CFD results being
[2,22,23]. The Lagrangian Monte Carlo PDF method is by far believed to used in real time prediction owing to the quick response of reduced-
be one of the most accurate methods for turbulent reactive flow simu order models compared to CFD simulations [51]. As such, ANN-based
lation but at the same time one of the most time-consuming ones strategies can be considered as an important approach for developing
[24,25]. In fact, to some extent, the low computational efficiency of the the reduced-order model.
model limits its broad application [3]. In this work we will demonstrate a novel approach to build a
A possible way to tackle this limitation is via machine learning, framework for coupling a deep ANN and turbulent reactive flow simu
especially in the form of an Artificial Neural Network (ANN). Deep lations and investigate the maximum speedup of the simulations by the
2
Y. Ouyang et al. Chemical Engineering Journal 429 (2022) 132442
generated ANN in respect to the accuracy. Specifically, the Lagrangian fine structure can be expressed based on the turbulent Reynolds number
PDF method is chosen as the target model due to its high accuracy Ret as follows:
combined with a high computational cost. Essential for the success are ( )14 ( )14
the input/output data collection from the TCI model, ANN architecture/ ξ=
3CD2 1
(3)
training, and implementation in the solver framework. The target TCI
2
4CD1 Ret
model is thus replaced by the ANN-based model, and the prediction
Between the fine structures and surroundings, mass transfer occurs to
performance as well as a speedup factor are assessed. Strengths and
supply the reactant species to the reaction region, i.e. the fine structure.
weaknesses are discussed.
The time scale for the mass transfer process is estimated by the Kol
mogorov time scale τη as follows:
2. Turbulence-chemistry interaction models
( )12
CD2
Simulating turbulent reactive flow requires a proper closure for the τ= τη (4)
3
chemical source term, which usually translates in the need for a model to
calculate the reaction rate for each computational cell in either an And finally, the net source of species i based on EDCM is expressed as:
explicit or implicit manner. In what follows, a summary of how to ξ2 ( )
calculate the reaction rate, expressed in the form of the net rate of Ri,EDC = Mi ( ) Ci* − Ci (5)
τ 1 − ξ3
species production in various TCI models is given.
In the above equations, CD1 and CD2 are model parameters. The
turbulence Reynolds number Ret is defined as kνε , where ν is the kinematic
2
2.1. Finite-Rate Model
( )12 2
viscosity. The Kolmogorov time scale τη is calculated as νε . ξτ repre
As discussed above, the FRM does not account for TCI, but it is a good
starting point to discuss the reaction rate calculation. The FRM is based sents the intra-cell mass transfer rate between the fine structures and
on the Arrhenius equation and the general reaction rate expression. For surroundings. ξ3 represents the volume fraction of the fine structures
an irreversible elementary reaction, a general form of the net source of that can be used to calculate the average scalars of a cell. Note that the
species i due to reaction is given by: expressions of the intra-cell mass transfer rate and average scalars of a
cell can take different forms [52–54], but Eq. (5) is the form used in this
∏
N
( )n
Ri,FR = vi Mi kr Cj (1) work. C*i is determined by applying the Arrhenius rate at fine structure
j=1 conditions after reacting over the time scale τ. Ci is the mean concen
tration at the cell level.
where vi is the stoichiometric coefficient of species i, Mi the molecular
weight of species i, kr is the rate coefficient, N represents the number of
2.4. β-PDF model
reactant species, n is the rate exponent of species j, and Cj is the molar
concentration of reactant species j.
The β-PDF model is one of the representative models for the
It’s important to note that the Cj term represents the mean state
presumed-PDF method. The reaction rate can be calculated in the form
concentration if a Reynolds-averaged Navier–Stokes (RANS) equations
of Eq. (1), but the species concentration follows a beta distribution
based turbulence model is applied, hence why it is considered as a first-
instead of a moment closure. However, if multiple variables (e.g. two or
order moment closure. The concentration fluctuation or any turbulent
more species) are involved for chemistry, developing a multi-
characteristic parameter is not accounted for. The mean state concen
dimensional PDF becomes difficult [3]. One solution is to introduce a
tration is also used in the EDM and EDCM model.
new variable–the mixture fraction to replace the reactant composition
state in each cell. For instance, in Baldyga et al.’s work, the net source of
2.2. Eddy-Dissipation Model species i is given by [17,18]:
∫1
The EDM takes the turbulent characteristic properties into account ∏N
( )n
Ri,β− = vi Mi kr 〈 Cj 〉 = vi Mi kr C(f )Φ(f )df (6)
and assumes the reaction is turbulent mixing-limited [13]: PDF
j=1 0
( )
ε
Ri,EDM = vi Mi A min
C1 C2 Cj Cp
, , ⋯, , B (2) In Eq. (6), the notation of 〈⋯〉 denotes the mean value. f is the
k v1 v2 vj vp mixture fraction, a conserved scalar quantity defined by a linear trans
formation of the species concentration, representing a local equivalence
where A and B are empirical constants, ε is the turbulent kinetic energy ratio [2]. Φ(f) is the beta distribution of the mixture fraction. Most
dissipation rate, k is the turbulent kinetic energy, and the subscript p importantly, there are two parameters that determine the shape of the
represents the product species. The limiting factors of reactant and beta distribution, which needs closure. Different closure approaches
product species (which generally means the limiting factor of flame have been reported [15,17], but no matter what the closure is, the tur
spread in combustion simulation) are described by taking the minimum bulent characteristic properties play an important role in determining
values in the equation. According to Eq. (2), the intrinsic kinetics are not the two parameters.
considered in the EDM. The mean reaction rate is only a function of
turbulent characteristic properties and species concentration. Therefore,
a commonly used method is to calculate both Eqs. (1) and (2) and choose 2.5. Lagrangian PDF method
the smaller one as the actual net production rate i.e. chemical source for
the corresponding cell. The transport PDF method does not assume the shape of the PDF a
priori, but derives the PDF transport equation from the Navier-Stokes
equation [55]:
2.3. Eddy-Dissipation Concept Model
( ) ( )
∂ ∂ ∂ ∂ ρμt ∂F ∂ 1 ∂ ∂Jk,i
(ρF) + (ρUk F) + (ρRi F) = + ρ〈 |ψ 〉F
EDCM assumes that a cell contains turbulent fine structures and ∂t ∂xk ∂ψ i ∂xk Sct ∂xk ∂ψ i ρ ∂xk ∂xk
surroundings, and that reactions take place in the region of the fine (7)
structures where the turbulent kinetic energy dissipates [14]. The latter
accounts for the SGE in EDM. The characteristic length fraction of this where Ri is the net source of species i due to reaction, F is the Favre joint
3
Y. Ouyang et al. Chemical Engineering Journal 429 (2022) 132442
PDF of composition, ρ is the mean fluid density, Uk is the Favre mean function of reaction rate coefficient and species concentration. However,
fluid velocity vector, ψ i is the composition space vector for species i, Sct taking TCI into account leads to various approaches, such as direct
is the turbulent Schmidt number and Jk,i is the molecular diffusion flux closures in the EDM or EDCM, or indirect ways where the PDFs are
vector. The notation of 〈P|Q〉 represents the conditional probability of determined first, after which they are coupled with the mean reaction
event P. Particularly, μt is the turbulent viscosity, calculated from a rate calculation. Regardless of which approach is used, the turbulent
2
RANS model, such as ρCμ k∊ , used in this work (Cμ is a turbulence model characteristic properties are very important in determining the chemical
constant which is usually taken to be 0.09 [56]). source term. As the intention of this work is to use ANN to replace and
Solving Eq. (7) could be under the Eulerian framework, one example speed up certain TCI models, understanding the determining factors is
being the DQMOM approach [2,22,23], but a presumption of the PDF essential for selecting the ANN input features.
form is still made for in this case. A more accurate way to solve the PDF
transport equation is by employing the Monte Carlo algorithm and 3. Methods
defining a certain number of notational particles in each cell, which is
known as the Lagrangian PDF method. The particles have mass, 3.1. CFD model
depending on the cell mass in which the particles are located. Generally,
the principal steps in the Lagrangian PDF method are summarized as The focus of this work is to develop a methodology of using a data-
follows [2]: (1) advancing the Finite-volume code, i.e. solving the mo driven ANN to replace and speed up certain TCI models, specifically
mentum transport equation and turbulence model (such as a RANS the Lagrangian PDF method. Therefore, a simple example case is built to
model) to get the velocity and turbulent scalars (such as turbulent vis demonstrate the workflow and prove its viability. Fig. 1 shows the ge
cosity in Eq. (7)) at each cell; (2) interpolating the scalars to the intra- ometry of the example case. Inlet a and inlet b, both with a width of 2
cell particle based on the particle’s location; (3) advancing the PDF mm, introduce species A and B into the reaction domain. Due to the
code based on the Monte Carlo algorithm. For the step 3, the particles regular geometry, a structured mesh can be easily created.
move through the computational domain due to convection, mixing and The model reaction used in this work is given below:
reaction. Several models were proposed to adjust the particle composi
(12)
k
A+B→C
tion to account for particle mixing, such as the Modified Curl Model [57]
and the Euclidean Minimum Spanning Tree Model [58]. The latter is After the species A and B are injected, the mixing of species A and B
used in this work. The Arrhenius rate is applied at the particle level. The takes place in the computational domain, and the reaction occurs once
PDF is thus described by the intra-cell particle information, similar to the two species contact, generating species C. The rate equation is given
constructing a histogram to estimate the PDF. As a result, the Lagrangian by r = kCA CB , where k = 1000m3 /(kmol∙s).
PDF method simulates the reaction in an indirect way, i.e. through re The FRM, EDCM and the Lagrangian PDF method as well as the
action within each of the notional particles. The mean reaction rate at subsequent established ANN-based model are implemented to simulate
the cell level can then be calculated by the mass-weighted average the above case. In addition, other numerical models have been kept
scalars of the intra-cell particles. The Lagrangian PDF method is theo identical for different simulations. For instance, given the high Reynolds
retically by far the most accurate model to take into account the SGE and number in the computational domain (2.0 × 104 –8.0 × 104), the
TCI, but the intra-cell particle tracking makes the model extremely time- standard k-ε turbulence model is adopted for all simulations. For a
consuming. straightforward comparison, the direct integration chemistry solver is
For a proper numerical efficiency two important choices need to be applied for all cases. All the species concentrations are monitored during
made in the Lagrangian PDF method: the particle number per cell and the calculation and the iteration does not stop until the species profiles
the particle local time step [59]. Note that increasing the particle remain steady. The convergence criteria for all equation residuals are set
number per cell or decreasing the particle local time step will reduce to 1.0 × 10-6.
statistical errors but will always increase computational cost. To obtain a
reliable prediction, one should increase the particle number per cell to 3.2. Methodology
make sure the simulation results are independent of the particle number
[2]. In general, the particle number is roughly in the range of 20–200 Fig. 2 shows the proposed methodology of using a deep ANN to
[24,59–61]. In this work, the number of particles per cell is set to 50. As replace and speed up the Lagrangian PDF method under the ANSYS-
for the local time step, it is determined by: Fluent framework. The procedure briefly consists of (1) performing
( ) simulations of the training cases using the Lagrangian PDF method, (2)
Δt = min Δtconv , Δtdiff , Δtmix (8)
exporting the cell data and particle data of the simulation results and
further linking the cell data, e.g. turbulent characteristic properties and
where:
species concentration, to the particle data of reaction rate, (3) con
L structing and training the deep ANN based on the input layer of the cell
Δtconv = Cconv (9)
U data and the output layer of the mean reaction rate, (4) compiling the
ANN, i.e. the weights, bias and activation functions in Fluent using user-
L2 defined functions and (5) performing simulations using the ANN-based
Δtdiff = Cdiff (10)
Dt model to calculate the reaction rate of each cell. The detailed descrip
tion is as follows:
k
Δtmix = Cmix (11)
∊
In Eqs. (9)–(11), L is the characteristic length of a cell, U is the mean
velocity, and Dt is the turbulent diffusivity. Without the constants Cconv ,
Cdiff and Cmix , Eqs. (9)–(11) provide the characteristic time scale of flow
convection, turbulent diffusivity and turbulent mixing. The 3 constants
usually take values of less than 1 to determine the local time step to
advance the intra-cell particle. They are all set as 0.5 in this work, in line
with previous research [59].
To sum up, if no TCI is considered, the mean reaction rate is only a
Fig. 1. 2D geometry of the simulation case.
4
Y. Ouyang et al. Chemical Engineering Journal 429 (2022) 132442
Fig. 2. Schematic of the deep ANN methodology for modeling turbulence chemistry interaction.
1. Prior to an ANN architecture, a training dataset is needed. Hence, the 4. Results and discussion
first step is to perform simulations of several turbulent reactive flow
cases using the Lagrangian PDF method and export the data both at 4.1. Training and test set generation
the cell level and at the particle level.
2. The exported cell data contains the velocity fields, turbulent char Table 1 gives an overview of the cases that are simulated for
acteristic properties, species concentration, etc. at locations of all the generating the training and test set. Various combinations of inlet ve
computational cells. The exported particle data is similar. Based on locity, mass fraction and geometry are explored. The specified inlet
the spatial location of the cells and particles, the cell data can be velocities in Table 1 correspond to a Reynold number range of 2.0 × 104
linked to intra-cell particle data by employing the ball tree algorithm –8.0 × 104, which ensures the flow is turbulent. In addition, considering
[62,63] and the Chebyshev distance [64]. Most importantly, the the rate constant and the specified inlet reactant concentrations, the
mean reaction rate of each cell can be then calculated based on the reaction characteristic time scale is at least one order smaller than the
corresponding intra-cell particles. turbulence time scale for all the cases in Table 1, which ensures a fast
3. The training data is then fed to an ANN, the architecture of which is chemistry [2] and the TCI is the dominating factor for the reaction
shown in Fig. 2. The turbulent kinetic energy, dissipation rate, and process. The Lagrangian PDF method has been implemented for all the
the mass fractions of reactant A & B at the cell level are chosen as the cases, but only the exported data from Cases 1–5 is used for the ANN
input layer data. The output layer consists of the mean reaction rate training data. In other words, the conditions of Cases 6–11 are our test
as discussed in step 2. To obtain a proper prediction accuracy, the set, which will be used to evaluate the performance of the ANN-based
Keras Tuner [65] is applied to search for an optimal set of hyper model. In particular, Cases 5–9 are specified to evaluate the prediction
parameters of the ANN architecture, i.e. a proper combination of performance for different operating conditions while Cases 10–11
various activation functions, the number of hidden layers and neu evaluates different geometries. More information on specifying the
rons. As the number of combinations is huge, the hyperband tuning training and test set is provided in the Supplementary Information. The
algorithm in the Keras Tuner is invoked to search for high-performing inlet velocity, mass fraction and geometry will have a direct impact on
candidate models, which is done by using an adaptive resource the turbulence flow and reaction process, and eventually on the con
allocation strategy. For example, in the initial stage, only a few centration profile of the product species C.
epochs are specified to select the best half candidate models for the
next round, similar to a sports championship style bracket. After a
thorough evaluation 3 hidden layers with respectively 32, 64, 128 4.2. Grid independence test
neurons are employed. The rectified linear unit (ReLU) and sigmoid
are chosen as the activation functions, which introduces a degree of Fig. 3 shows a grid independence test based on the mass fraction
nonlinearity into the network. A MinMaxScaler [66] is used to scale distribution of the product species C along the horizontal axis. The
the input and output data to a range of 0 and 1. The python deep
learning library Keras [67] with the backend of Tensorflow [68] is Table 1
used to construct the deep ANN. Dataset construction: Training and test set.
4. After a deep ANN with acceptable accuracy is obtained, the weights, Case # Dataset Inlet velocity Mass Mass
bias and activation functions of the deep ANN are translated into type [m/s] fraction of A fraction of B
user-defined functions (UDFs), which is further compiled in Fluent. Case-1 Training 10 0.1 0.1
To be more specific, the UDFs invoke the cell level data of the tur Case-2 Training 10 0.2 0.2
Case-3 Training 10 0.05 0.05
bulent kinetic energy, dissipation rate, and the mass fractions of
Case-4 Training 5 0.1 0.1
reactant A & B and return the reaction rate to the corresponding cell. Case-5 Training 20 0.1 0.1
5. When performing a simulation based on the above UDFs, the mean Case-6 Test 10 0.075 0.075
reaction rate of each computational cell is determined by a well- Case-7 Test 10 0.15 0.15
trained deep ANN. Case-8 Test 15 0.075 0.075
Case-9 Test 10 0.15 0.075
Case-10 (different Test 15 0.075 0.075
geometry)
Case-11 (different Test 15 0.075 0.075
geometry)
5
Y. Ouyang et al. Chemical Engineering Journal 429 (2022) 132442
results are obtained based on the FRM. As the total cell number in
creases, the distributions tend to converge, especially in the region
where the mixing and reaction initially occur. To quantify the result of
the independence test, the grid convergence index (GCI) is applied [69].
Grid-2/Grid-4 and Grid-4/Grid-5 are chosen as two sets of fine/coarse
grids as both sets of grid spacing ratios are equal to 2. Based on the GCI
approach and the average mass fraction values, the GCI2/4 and GCI4/5
are calculated to be 0.05008% and 0.24427%, respectively, indicating
the solutions are in the asymptotic range of convergence [69]. There
fore, the grid with 14,364 cells is adopted in this work (Grid-2 in Fig. 3).
Fig. 4. Parity plot for the reaction rate prediction: (a) training dataset; (b) split dataset for validation.
6
Y. Ouyang et al. Chemical Engineering Journal 429 (2022) 132442
Fig. 6. Comparison of the FRM, EDCM, Lagrangian PDF method and ANN-based model: (a) Case-1; (b) Case-2; (c) Case-3; (d) Case-4; (e) Case-5.
(accuracy here means how close the result of the ANN-based model is to unseen data. Results show that the ANN-based model remains in good
the result of the Lagrangian PDF method, not validation based on agreement with the target model, indicating the ANN-based model is
experimental data), and more importantly, this level of accuracy is not able to simulate new conditions after training with a limited number of
obtained by the other considered TCI models such as the EDC model. cases. From a practical point of view, one can use a similar approach to
employ deep ANN(s) to replace the Lagrangian PDF method or even
other time-consuming TCI models for specific application cases.
4.5. ANN-based model to predict testing dataset As long as suitable training data is available, an ANN-based model is
essentially independent of the cases, indicating this methodology can be
A key question that remains is if the model can simulate new con applied to other practical simulations. While the methodology itself can
ditions next to replicate data that it was trained for. Fig. 7 shows the be of universal relevance, the generated ANN in this work is not, as the
results of the ANN-based model simulating Cases 6–9, which are selection of training dataset, input parameters and architecture of the
different from the operating conditions of the training cases. Fig. 8 ANN can be case dependent. For instance, in a case with different
shows the results of simulating Cases 10–11, which have different ge reactant species/species number and reactor/burner geometries, the
ometries. To evaluate the model performance, Cases 6–11 are simulated ANN construction and the training process should be re-run to fit the
by the Lagrangian PDF method as well. Meanwhile, Fig. 9 summarizes specific case. However, the procedure to build this framework, i.e.
and compares all the data points of the results when using the generating a training dataset, specifying input parameters, constructing
Lagrangian PDF method and the ANN-based model, i.e. the data points the deep ANN and compiling the ANN in ANSYS-Fluent, should be the
of the two models in Figs. 6, 7 and 8 for all the training and test cases. same as presented in this work. The first step would be choosing a set of
Both training cases and testing cases are within ± 10% deviation, proper training data cases, preferably the cases with representative
meaning overfitting is avoided and the ANN-based model can predict
7
Y. Ouyang et al. Chemical Engineering Journal 429 (2022) 132442
Fig. 7. ANN-based model to predict turbulent reactive flow under various operating conditions: (a) Case-6; (b) Case-7; (c) Case-8; (d) Case-9.
Fig. 8. ANN-based model to predict turbulent reactive flow for different geometries: (a) Case-10; (b) Case-11.
conditions in the full map of operating conditions. In practical uses of work [7], in which the TCI is essential for an accurate simulation.
the methodology in this study, the method of experimental design can be
used for the training cases, e.g., orthogonal experimental design. The 5. Speedup factor assessment
training cases that cover the operating condition range will improve the
prediction performance of the ANN-based method as the necessary in To further quantify the performance of the ANN-based model it is
formation can be learned by the ANN. Next, after driving a well-trained necessary to evaluate the computational efficiency. The Lagrangian PDF
ANN model, the fundamental model can be replaced to some extent method is considered as our reference. The procedure for calculating the
while maintaining accuracy. For example this will become very effective speedup factor start with simulating Cases 6–11 via the Lagrangian PDF
when it comes to parametric analysis and reactor/burner geometry method and the ANN-based model. Identical convergence criteria are set
optimization, as both of them require a large number of similar simu for both models. The calculations are carried out on a serial processor of
lations [70]. Actually, the original motivation of this work is to increase Intel core i7-8650U. The calculation time is recorded from the initiali
the computational efficiency of the Lagrangian PDF method when per zation of the simulation to achieving a steady-state result. The speedup
forming parametric analysis in a liquid–liquid micromixing simulation factors for Cases 6–11 are then calculated by normalizing the efficiency
8
Y. Ouyang et al. Chemical Engineering Journal 429 (2022) 132442
Fig. 9. Comparison of the simulation results based on the Lagrangian PDF method and the ANN-based model: (a) training Cases 1–5; (b) test Cases 6–11.
9
Y. Ouyang et al. Chemical Engineering Journal 429 (2022) 132442
[3] T. Poinsot, D. Veynante, Theoretical and Numerical Combustion, RT Edwards, Inc., [30] P.P. Plehiers, S.H. Symoens, I. Amghizar, G.B. Marin, C.V. Stevens, K.M. Van Geem,
2005. Artificial Intelligence in Steam Cracking Modeling: A Deep Learning Algorithm for
[4] B. Merci, D. Roekaerts, B. Naud, S. Pope, Comparative study of micromixing Detailed Effluent Prediction, Engineering. 5 (2019) 1027–1040, https://doi.org/
models in transported scalar PDF simulations of turbulent nonpremixed bluff body 10.1016/j.eng.2019.02.013.
flames, Combust. Flame. 146 (1-2) (2006) 109–130, https://doi.org/10.1016/j. [31] K.O. Lye, S. Mishra, D. Ray, Deep learning observables in computational fluid
combustflame.2006.04.010. dynamics, J. Comput. Phys. 410 (2020), 109339, https://doi.org/10.1016/j.
[5] Y. Pei, E.R. Hawkes, S. Kook, A Comprehensive Study of Effects of Mixing and jcp.2020.109339.
Chemical Kinetic Models on Predictions of n-heptane Jet Ignitions with the PDF [32] V.G. Maltarollo, K.M. Honório, A.B.F. da Silva, Applications of artificial neural
Method, Flow Turbul. Combust. 91 (2) (2013) 249–280, https://doi.org/10.1007/ networks in chemical engineering, in: Artif. Neural Networks-Architectures Appl.,
s10494-013-9454-z. 2013: pp. 203–223.
[6] J. Bałdyga, R. Pohorecki, Turbulent micromixing in chemical reactors—a review, [33] R. Laubscher, P. Rousseau, Application of generative deep learning to predict
Chem. Eng. J. Biochem. Eng. J. 58 (2) (1995) 183–195, https://doi.org/10.1016/ temperature, flow and species distributions using simulation data of a methane
0923-0467(95)02982-6. combustor, Int. J. Heat Mass Transf. 163 (2020), 120417, https://doi.org/
[7] Y. Ouyang, Y. Xiang, H. Zou, G. Chu, J. Chen, Flow characteristics and micromixing 10.1016/j.ijheatmasstransfer.2020.120417.
modeling in a microporous tube-in-tube microchannel reactor by CFD, Chem. Eng. [34] L. Zhu, J. Tang, Z. Luo, Machine learning to assist filtered two-fluid model
J. 321 (2017) 533–545, https://doi.org/10.1016/j.cej.2017.03.151. development for dense gas–particle flows, AIChE J. 66 (2020), https://doi.org/
[8] D.L. Marchisio, A.A. Barresi, CFD simulation of mixing and reaction: the relevance 10.1002/aic.16973.
of the micro-mixing model, Chem. Eng. Sci. 58 (2003) 3579–3587. [35] Y. Yan, T. Mattisson, P. Moldenhauer, E.J. Anthony, P.T. Clough, Applying
[9] D.L. Marchisio, R.O. Fox, A.A. Barresi, G. Baldi, On the comparison between machine learning algorithms in estimating the performance of heterogeneous,
presumed and full PDF methods for turbulent precipitation, Ind. Eng. Chem. Res., multi-component materials as oxygen carriers for chemical-looping processes,
American Chemical Society 40 (23) (2001) 5132–5139, https://doi.org/10.1021/ Chem. Eng. J. 387 (2020), 124072, https://doi.org/10.1016/j.cej.2020.124072.
ie0010262. [36] G. Banin, R. Pioli Vieira, L. Maria Ferrareso Lona, Artificial neural networks
[10] M.T. Lewandowski, J. Pozorski, Assessment of turbulence-chemistry interaction towards average properties targets in styrene ARGET-ATRP, Chem. Eng. J. (2020)
models in the computation of turbulent non-premixed flames, in, J. Phys. Conf. 126999, https://doi.org/10.1016/j.cej.2020.126999.
Ser., Institute of Physics Publishing 760 (2016) 012015, https://doi.org/10.1088/ [37] C.W. Coley, D.A. Thomas, J.A.M. Lummiss, J.N. Jaworski, C.P. Breen, V. Schultz,
1742-6596/760/1/012015. T. Hart, J.S. Fishman, L. Rogers, H. Gao, R.W. Hicklin, P.P. Plehiers, J. Byington, J.
[11] A. De, A. Dongre, Assessment of turbulence-chemistry interaction models in MILD S. Piotti, W.H. Green, A. John Hart, T.F. Jamison, K.F. Jensen, A robotic platform
combustion regime, Flow Turbul. Combust. 94 (2) (2015) 439–478, https://doi. for flow synthesis of organic compounds informed by AI planning, Science (80)
org/10.1007/s10494-014-9587-8. (2019) 365, https://doi.org/10.1126/science.aax1566.
[12] J. Cheng, X. Feng, D. Cheng, C. Yang, Retrospect and perspective of micro-mixing [38] P.P. Plehiers, C.W. Coley, H. Gao, F.H. Vermeire, M.R. Dobbelaere, C.V. Stevens, K.
studies in stirred tanks, Chinese J Chem. Eng. 20 (1) (2012) 178–190, https://doi. M. Van Geem, W.H. Green, Artificial intelligence for computer-aided synthesis in
org/10.1016/S1004-9541(12)60378-4. flow: analysis and selection of reaction components, Front. Chem. Eng. 2 (2020) 5,
[13] B.F. Magnussen, B.H. Hjertager, On mathematical modeling of turbulent https://doi.org/10.3389/fceng.2020.00005.
combustion with special emphasis on soot formation and combustion, Symp. [39] Y. Su, T. Ng, Z. Li, J.H. Davidson, Sparse scattered high performance computing
Combust. 16 (1) (1977) 719–729, https://doi.org/10.1016/S0082-0784(77) data driven artificial neural networks for multi-dimensional optimization of
80366-4. buoyancy driven heat and mass transfer in porous structures, Chem. Eng. J. 397
[14] B.F. Magnussen, On the structure of turbulence and a generalized eddy dissipation (2020), 125257, https://doi.org/10.1016/j.cej.2020.125257.
concept for chemical reaction in turbulent flow, American Institute of Aeronautics [40] S. Godfrey Nnabuife, B. Kuang, J.F. Whidborne, Z. Rana, Non-intrusive
and Astronautics (AIAA) (1981), https://doi.org/10.2514/6.1981-42. classification of gas-liquid flow regimes in an S-shaped pipeline riser using a
[15] X. Duan, X. Feng, C. Yang, Z.S. Mao, Numerical simulation of micro-mixing in Doppler ultrasonic sensor and deep neural networks, Chem. Eng. J. 403 (2021),
stirred reactors using the engulfment model coupled with CFD, Chem. Eng. Sci. 140 126401, https://doi.org/10.1016/j.cej.2020.126401.
(2016) 179–188, https://doi.org/10.1016/j.ces.2015.10.017. [41] H. Bonakdari, S. Baghalian, F. Nazari, M. Fazti, Numerical analysis and prediction
[16] A.W. Cook, J.J. Riley, Subgrid-scale modeling for turbulent reacting flows, of the velocity field in curved open channel using artificial neural network and
Combust. Flame 112 (1998) 593–606, https://doi.org/10.1016/S0010-2180(97) genetic Algorithm, Eng. Appl. Comput. Fluid Mech. 5 (2011) 384–396, https://doi.
00173-9. org/10.1080/19942060.2011.11015380.
[17] J. Baldyga, M. Henczka, L. Makowski, J. Bałdyga, M. Henczka, Makowski, Effects [42] P. Plehiers, L.A. Vandewalle, G.B. Marin, C.V. Stevens, K.M. Van Geem,
of mixing on parallel chemical reactions in a continuous-flow stirred-tank reactor, Accelerating reactive CFD simulations with detailed pyrolysis chemistry using
Chem. Eng. Res. Des. 8 (2001) 895–900, https://doi.org/10.1205/ artificial neural networks, in: 2019 AIChE Annu. Meet., Orlando, USA, 2019. http
02638760152721109. s://biblio.ugent.be/publication/8639401.
[18] J. Bałdyga, Ł. Makowski, CFD modelling of mixing effects on the course of parallel [43] J. Tompson, K. Schlachter, P. Sprechmann, K. Perlin, Accelerating eulerian fluid
chemical reactions carried out in a stirred tank, Chem. Eng. Technol. 27 (3) (2004) simulation with convolutional networks, in: 34th Int. Conf. Mach. Learn. ICML
225–231, https://doi.org/10.1002/ceat.200401992. 2017. 7 (2017) 5258–5267.
[19] L. Vicum, S. Ottiger, M. Mazzotti, Ł. Makowski, J. Bałdyga, Multi-scale modeling of [44] R. Laubscher, Utilization of artificial neural networks to resolve chemical kinetics
a reactive mixing process in a semibatch stirred tank, Chem. Eng. Sci. 59 (8-9) in turbulent fine structures of an advanced CFD combustion model, Stellenbosch
(2004) 1767–1781, https://doi.org/10.1016/j.ces.2004.01.032. University, Stellenbosch, 2017 https://scholar.sun.ac.za (accessed November 10,
[20] A.A. Öncül, G. Janiga, D. Thévenin, A.A. Öncül, G. Janiga, D. Thévenin, 2020).
Comparison of various micromixing approaches for computational fluid dynamics [45] H. Jung, S. Yoon, Y. Kim, J.H. Lee, H. Park, D. Kim, J. Kim, S. Kang, Development
simulation of barium sulfate precipitation in tubular reactors, Ind. Eng. Chem. Res. and evaluation of data-driven modeling for bubble size in turbulent air-water
48 (2008) 999–1007, https://doi.org/10.1021/ie800364k. bubbly flows using artificial multi-layer neural networks, Chem. Eng. Sci. 213
[21] R. Cao, S.B. Pope, Numerical integration of stochastic differential equations: Weak (2020) 115357, https://doi.org/10.1016/j.ces.2019.115357.
second-order mid-point scheme for application in the composition PDF method, [46] A. Marcato, G. Boccardo, D. Marchisio, A computational workflow to study particle
J. Comput. Phys. 185 (2003) 194–212, https://doi.org/10.1016/S0021-9991(02) transport and filtration in porous media: Coupling CFD and deep learning, Chem.
00054-2. Eng. J. 417 (2021) 128936, https://doi.org/10.1016/j.cej.2021.128936.
[22] D.L. Marchisio, R.O. Fox, Solution of population balance equations using the direct [47] L. Zhu, B. Ouyang, H. Lei, Z. Luo, Conventional and data-driven modeling of
quadrature method of moments, J. Aerosol Sci. 36 (1) (2005) 43–73, https://doi. filtered drag, heat transfer, and reaction rate in gas–particle flows, AIChE J.
org/10.1016/j.jaerosci.2004.07.009. (2021), e17299, https://doi.org/10.1002/aic.17299.
[23] J. Akroyd, A.J. Smith, L.R. McGlashan, M. Kraft, Comparison of the stochastic [48] I.E. Lagaris, A. Likas, D.I. Fotiadis, Artificial neural networks for solving ordinary
fields method and DQMoM-IEM as turbulent reaction closures, Chem. Eng. Sci. 65 and partial differential equations, IEEE Trans. Neural Networks. 9 (1998)
(20) (2010) 5429–5441, https://doi.org/10.1016/j.ces.2010.06.039. 987–1000, https://doi.org/10.1109/72.712178.
[24] H. Möbus, P. Gerlinger, D. Brüggemann, Comparison of Eulerian and Lagrangian [49] M. Raissi, A. Yazdani, G.E. Karniadakis, Hidden Fluid Mechanics: A Navier-Stokes
Monte Carlo PDF methods for turbulent diffusion flames, Combust. Flame 124 (3) Informed Deep Learning Framework for Assimilating Flow Visualization Data,
(2001) 519–534, https://doi.org/10.1016/S0010-2180(00)00207-8. (2018). http://arxiv.org/abs/1808.04327 (accessed November 10, 2020).
[25] D.C. Haworth, Progress in probability density function methods for turbulent [50] J. Han, A. Jentzen, Deep Learning-Based Numerical Methods for High-Dimensional
reacting flows, Prog. Energy Combust. Sci. 36 (2) (2010) 168–259, https://doi.org/ Parabolic Partial Differential Equations and Backward Stochastic Differential
10.1016/j.pecs.2009.09.003. Equations, Commun. Math Stat. 5 (4) (2017) 349–380, https://doi.org/10.1007/
[26] Y. LeCun, Y. Bengio, G. Hinton, Deep learning, Nature 521 (7553) (2015) 436–444, s40304-017-0117-6.
https://doi.org/10.1038/nature14539. [51] G. Aversano, M. Ferrarotti, A. Parente, Digital twin of a combustion furnace
[27] J. Schmidhuber, Deep Learning in neural networks: An overview, Neural Networks. operating in flameless conditions: Reduced-order model development from CFD
61 (2015) 85–117, https://doi.org/10.1016/j.neunet.2014.09.003. simulations, Proc. Combust. Inst. (2020), https://doi.org/10.1016/j.
[28] E.D. Koronaki, A.M. Nikas, A.G. Boudouvis, A data-driven reduced-order model of proci.2020.06.045.
nonlinear processes based on diffusion maps and artificial neural networks, Chem. [52] I.R. Gran, B.F. Magnussen, A numerical study of a bluff-body stabilized diffusion
Eng. J. 397 (2020), 125475, https://doi.org/10.1016/j.cej.2020.125475. flame. Part 2. Influence of combustion modeling and finite-rate chemistry,
[29] D. Serrano, D. Castelló, Tar prediction in bubbling fluidized bed gasification Combust. Sci. Technol. 119 (1-6) (1996) 191–217, https://doi.org/10.1080/
through artificial neural networks, Chem. Eng. J. 402 (2020), 126229, https://doi. 00102209608951999.
org/10.1016/j.cej.2020.126229.
10
Y. Ouyang et al. Chemical Engineering Journal 429 (2022) 132442
[53] B.F. Magnussen, Modeling of NOx and soot formation by the eddy dissipation [62] G. Varoquaux, L. Buitinck, G. Louppe, O. Grisel, F. Pedregosa, A. Mueller, Scikit-
concept, in: Int. Flame Res. Found. First Top. Oriented Tech. Meet., 1989: pp. learn, GetMobile Mob, Comput. Commun. 19 (1) (2015) 29–33, https://doi.org/
17–19. 10.1145/2786984.2786995.
[54] B.F. Magnussen, The eddy dissipation concept: A bridge between science and [63] S.M. Omohundro, Five balltree construction algorithms, International Computer
technology, in: ECCOMAS Themat. Conf. Comput. Combust., Libson, Portugal, Science Institute, 1989.
2005: p. 24. [64] C.D. Cantrell, Modern mathematical methods for physicists and engineers,
[55] S.B. Pope, PDF methods for turbulent reactive flows, Prog. Energy Combust. Sci. 11 Cambridge University Press, 2000.
(2) (1985) 119–192. [65] T. O’Malley, E. Bursztein, J. Long, F. Chollet, H. Jin, L. Invernizzi, A. Others., Keras
[56] B.E. Launder, D.B. Spalding, Lectures in mathematical models of turbulence, Tuner, (2019). https://github.com/keras-team/keras-tuner (accessed November
Academic press, 1972. 15, 2020).
[57] J. Janicka, W. Kolbe, W. Kollmann, Closure of the transport equation for the [66] F. Pedregosa Gael Varoquaux Alexandre Gramfort Vincent Michel Bertrand
probability density function of turbulent scalar fields, J. Non-Equilibrium Thirion, O. Grisel, Blondel Mathieu, P. Prettenhofer, R. Weiss, V. Dubourg, J.
Thermodyn. 4 (1979) 47–66, https://doi.org/10.1515/jnet.1979.4.1.47. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot Edouard Duchesnay,
[58] S. Subramaniam, S.B. Pope, A mixing model for turbulent reactive flows based on Scikit-learn: Machine Learning in Python, J. Mach. Learn. Res. 12 (2011)
Euclidean minimum spanning trees, Combust. Flame. 115 (1998) 487–514, 2825–2830. doi: 10.5555/1953048.2078195.
https://doi.org/10.1016/S0010-2180(98)00023-6. [67] F. Chollet, A. Others., Keras: The python deep learning library, (2020). https://
[59] M. Muradoglu, S.B. Pope, Local time-stepping algorithm for solving probability keras.io/.
density function turbulence model equations, AIAA J. 40 (9) (2002) 1755–1763, [68] M. Abadi, P. Barham, J. Chen, Z. Chen, A. Davis, J. Dean, M. Devin, S. Ghemawat,
https://doi.org/10.2514/2.1880. G. Irving, M. Isard, Tensorflow: A system for large-scale machine learning, in: 12th
[60] A. Kaufmann, M. Moreau, O. Simonin, J. Helie, Comparison between Lagrangian {USENIX} Symp. Oper. Syst. Des. Implement. ({OSDI} 16), 2016: pp. 265–283.
and mesoscopic Eulerian modelling approaches for inertial particles suspended in [69] P.J. Roache, Verification and Validation in Computational Science and
decaying isotropic turbulence, J. Comput. Phys. 227 (13) (2008) 6448–6472, Engineering, 1998. http://www.hermosa-pub.com/hermosa (accessed November
https://doi.org/10.1016/j.jcp.2008.03.004. 16, 2020).
[61] S. James, M.S. Anand, S.B. Pope, The lagrangian PDF transport method for [70] J.N. Dedeyne, M. Geerts, P.A. Reyniers, F. Wéry, K.M. Van Geem, G.B. Marin,
simulations of gas turbine combustor flows, in: 38th AIAA/ASME/SAE/ASEE Jt. Computational fluid dynamics-based optimization of dimpled steam cracking
Propuls. Conf. Exhib. (2002) 1–8. 10.2514/6.2002-4017. reactors for reduced CO2 emissions, AIChE J. 66 (2020), https://doi.org/10.1002/
aic.16255.
11