Machine Learning Prediction

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 8

Nuclear Engineering and Design 393 (2022) 111794

Contents lists available at ScienceDirect

Nuclear Engineering and Design


journal homepage: www.elsevier.com/locate/nucengdes

Machine learning based prediction of subcooled bubble condensation


behavior, validation with experimental and numerical results
Vijay Mohan Nagulapati a, 1, SalaiSargunan S Paramanantham a, b, 1, Aleksey Ni a,
Senthil Kumar Raman c, Hankwon Lim a, d, *
a
School of Energy and Chemical Engineering, Ulsan National Institute of Science and Technology, Ulsan 44919, Republic of Korea
b
Carbon Value Co, Ltd, 2802 A-dong, 97, Centum Jungang-ro Haeundae-gu, Busan, Republic of Korea
c
School of Automotive and Mechanical Engineering (SAME), Kalasalingam Academy of Research and Education, Krishnankoil, Tamil Nadu 626126, India
d
Department of Energy Engineering, Ulsan National Institute of Science and Technology, Ulsan 44919, Republic of Korea

A R T I C L E I N F O A B S T R A C T

Keywords: Measuring a full life cycle of condensing subcooled bubbles using either the experimental and/or numerical
Nuclear reactor approaches is a very challenging problem. In present study this problem is solved through Machine Learning
Machine Learning techniques using existing data sets from both experiment and numerical results. Two different machine leaning
Subcooled flow boiling
methods, Linear Regression (LR) and Gaussian Process Regression (GPR) are trained to predict the bubble
Direct contact condensation
Bubble condensation prediction
condensing life-history. The models are trained with 70% of data and validated using 30 % data from the
collected datasets. The predicted results are compared with both numerical and experimental results and model
prediction obtained good agreement. Additionally, the validated machine learning models are used to predict
various bubble diameters ranging between 1 and 6 mm. These predicted results give a much better understanding
of subcooled bubble condensation behavior without the need for extensive experiments and numerical studies.

1. Introduction biochar by Zhu et al. (Zhu et al., 2019), to understand the complex
electrochemical degradation of lithium ion battery and for predicting
Machine Learning (ML) and artificial intelligence (AI) based ap­ state of health and capacity estimations by Nagulapati et al. (Nagulapati,
proaches have revolutionised the engineering sector by facilitating 2021; Nagulapati, 2021), to predict the performance of rotating ma­
mechanistic understanding of complex phenomena. These ML and AI chinery by Han et al. (Han et al., 2018), ML based phase change model
based methods have the ability to process data and correlate between for simulating bubble condensation by Tang et al. (Tang, 2021). The
process variables to predict a particular process characteristic. This has application of ML based approach, which is essentially a black-box
enabled to circumvent the need to build complex first-principle models approach that does not require detailed assumptions or governing
while maintaining predictive accuracies. This is hugely helpful in en­ equations is to overcome the difficulties associated with building a first-
gineering applications where first principle understanding is difficult principle model for these complex engineering processes while still
and elusive to capture using numerical approaches. ML methods have having sufficient predictive accuracy. However, given the fact that these
been implemented in a wide variety of engineering applications like for ML models are predominantly black box models, deriving a physical
example, solving heat transfer equation in manufacturing and engi­ relationship and to produce interpretable solution is not possible. To
neering applications by Zobeiry et al. (Zobeiry and Humfeld, 2021), avoid this drawback, use of regression analysis techniques like Linear
computing vapour-water two-phase flow in tube with uniform and non- Regression (LR) and Gaussian Process Regression (GPR), that are
uniform heating by Shang et al. (Shang, 2005), for understanding and interpretable and allow for understanding the physical relationships are
predicting circulating fluidized bed (CFB) riser flow characteristics by emphasized (Theßeling et al., 2020).
Chew et al. (Chew and Cocco, 2020) and predicting axial solid holdup by Bubble condensation is one of the primary phenomena in the fluid-
Nagulapati et al. (Upadhyay et al., 2022), predicting metal sorption into based heat transfer industry such as nuclear power plants, electronics

* Corresponding author.
E-mail address: hklim@unist.ac.kr (H. Lim).
1
Co-first authors.

https://doi.org/10.1016/j.nucengdes.2022.111794
Received 7 December 2021; Received in revised form 15 March 2022; Accepted 26 April 2022
Available online 4 May 2022
0029-5493/© 2022 Elsevier B.V. All rights reserved.
V.M. Nagulapati et al. Nuclear Engineering and Design 393 (2022) 111794

Fig. 1. Schematic for machine learning prediction of bubble condensation based on experimental and CFD generated data.

cooling systems, boilers, chemical engineering applications and so on. a microsecond. Therefore, numerical simulation is another preferred
The bubble condensing behaviour is usually studied as a direct contact method for study the condensing bubble dynamics (S Paramanantham
condensation through which heat and mass transfer is observed (Warrier et al., 2018).
et al., 2002). Understanding the bubble condensing behavior is chal­ Numerical simulation is a flexible tool for studying the static and
lenging even with current technology. The condensing behavior occurs dynamics characteristics of bubble condensation and allows to carry out
depending on the bubble column condition, heated wall material, and research that goes far beyond what is possible experimentally. In nu­
application. Even though the life of the condensing bubble is in micro merical modelling, it is necessary to apply various schemes and ap­
level time step, during this time bubble changes its shape, and size. proaches in order for results to be conformed to the experiment.
Furthermore, the complexity of bubble dynamics has a great impact on Following this path, the researchers have implemented different
bubble lifetime, pressure drop and flow instability in the bubble column. methods and approaches to obtain accurate results, through numerical
Our main target is to develop the better understanding of nuclear method for the problem of bubble dynamics. The following methods,
reactor safety monitor systems. In the rose, we implement the machine schemes and approaches were used in modeling, namely: semi-implicit
learning algorithm (Sobes et al., 2021) for prediction of bubble moving particles, moving boundary, interfacial particles of two-phase
condensation behavior. In the process of reactor core cooling, the interface, coupling between momentum and energy equation, coales­
nucleate bubble is generated from the core wall, and it gets condensed in cence of stagnant liquid and various types of heat and mass transfer
subcooled liquid. The process of the nucleated bubble getting detached model, based on mesh approach such as adaptive mesh, and also fixed
from wall and condensing within the subcooled liquid is influenced by grid method and moving grid (Tian et al., 2010; Chen et al., 2010; Chen
many factors such as subcooled temperature, pressure, velocity, bubble et al., 2011; Zuo et al., 2013). Furthermore, various types of bubble
size, and shape. In the nucleated boiling, bubble is one of the sources for conditions such as interfacial heat transfer for water temperature, ve­
heat and mass transfer and the bubble condensing time is based on the locity and bubble condensing diameter, bubble vertex generation
size of the nucleate bubble (Paramanantham, 2020). So, prediction of mechanisms, bubble nucleation, generated and detached, film boiling,
bubble condensation with respect to size of the bubble helps to under­ flow boiling, and condensation, single and multiple bubble dynamics (S
stand the design of cooling system. Utilization of machine learning Paramanantham et al., 2018; Bahreini et al., 2015; Zhenyu et al., 2015;
method to predict the bubble condensation behavior will facilitate Xiaohang et al., 2017; Paramanantham et al., 2020; Paramanantham
better understanding of bubble condensing phenomenon that will help et al., 2020; Liu, 2020; Bahreini, 2021). Even though the numerical
to develop nuclear reactor safety monitor systems. method can freely implement various approaches to bubble condensa­
Generally, studying of condensing bubble behavior is based on tion when using a stationary grid with a moving boundary for bubble
theoretic investigation, experimental assessment, and numerical condensation, the volume fraction of the occupied bubble level de­
models. Using the analytical study, behavior of static and dynamics creases during condensation, and after a certain amount of bubble
bubble condensing behavior was investigated (Zuber, 1960; Okhotsim­ condensation, the loss of the bubble grid volume cannot be tracked. Due
skii, 1988), but there was an insufficient degree of accuracy and also the to loss of mesh volume, it is impossible to capture the full bubble
influence of the shape and size of bubbles was not taken into account. On condensation history. For example, 4.9 mm size bubble condensation
the other hand, experimental methods are applied for studying the history was captured until the bubble diameter reached 1.5 mm and the
behavior of condensing bubble, which makes it possible to take into rest of the remaining bubble history did not capture. To capture the
account/detect almost any parameters, to a certain degree, but also re­ remaining bubble condensation history, but without conducting addi­
quires a significant investment of time and money and not always tional simulation and/or experiment the machine learning (ML)
possible to perform. For the last few decades, researchers studied methods were used on existing data to predict the bubble condensation
various aspects of subcooled bubble condensing behavior such as history.
interfacial heat transfer, changing dynamics of bubble shape, and size, Machine learning methods have proved their capability in different
influence of fluid numbers. Furthermore, condensing behavior of bubble applications related to process industry (Ge et al., 2017; Yan, 2022).
life, bubble velocity, collapse behavior, and low-and high-pressure ef­ Machine learning methods have received increased attention in the field
fects were studied through various experiments (Kim and Park, 2011; of fluid mechanics in recent times (Babanezhad et al., 2020; Brunton
Lucasa and Prasserb, 2007; Harada et al., 2010; Brucker and Sparrow, et al., 2020; Mesgarpour et al., 2022). Bubble columns are widely used in
1977; Yang, 2021; Qiu, 2020). There are other drawbacks in experi­ chemical industry in various exothermic reactions where heat and mass
mental approach. For example, due to the many factors studied in ex­ transfer are crucial. By learning from data collected from previous ex­
periments, it is sometimes difficult to accurately trace the entire high- periments and numerical studies, a machine learning model is trained to
speed dynamics of a bubble due to condensation occurring in less than predict the bubble condensation by correlating various influencing

2
V.M. Nagulapati et al. Nuclear Engineering and Design 393 (2022) 111794

covariances between the input variables based on which the outputs are
calculated. LR and GPR based ML models are considered given their
ability to map the dependencies and covariances between the input/
independent and output/dependent variables.
In this paper, machine learning techniques were implemented to
capture the complete subcooled condensing bubble life history. The
bubble condensing history is one of the important factors required for
understanding the heat and mass transfer from the heated surface to the
subcooled fluid flow channel. This process is difficult to measure and/or
capture, both experimentally and numerically. To solve this problem, a
machine learning method is used, utilizing existing datasets created
from experiments and numerical results. This prediction uses LR and
GPR techniques to train the data to observe bubble life history to better
design the subcooled flow path for various purposes.

2. Mathematical formulation

2.1. In-house code details

Predicting the bubble condensation history using Machine Learning


(ML) techniques need the data set for further investigation of bubble
condensing history. To derive bubble condensation datasets for the
prediction of bubble condensation, author’s previous studies were taken
into account. A previous study used in-house code developed for a two-
dimensional, two-phase homogeneous mixture method based on the
fully compressible Reynolds-averaged Navier-Stokes (RANS) method. In
addition, the interface between the vapour and liquid phases was
defined utilizing the continuous surface force approach (Ha and Park,
Fig. 2. Computational domain for single bubble condensation. 2016; Ha et al., 2017; Hashemi et al., 2020). To solve the governing
equations the implicit dual-time techniques and equations were imple­
parameters. Knowledge about bubble condensation behavior is crucial mented. The stepwise formation, numerical schemes, bubble conditions
in certain applications such as in terms of reactor performance as the and procedures, and other simulation related details can be found in the
bubble size is directly proportional to the specific area and thus to heat authors previous works (S Paramanantham et al., 2018; Para­
and mass transfer rates. In this study, Linear Regression (LR) (Kunz, manantham et al., 2020; Paramanantham et al., 2020).
2021) and Gaussian Process Regression (GPR) (Byun, 2021) based ma­ The numerical simulation was based on a multiphase homogenous
chine learning models are trained based on a database where data points mixture flow. A mixture density with compressible, thermal effects, and
are collected from the previous experiments and numerical simulations an axisymmetric two-dimensional model were applied. The system of
as shown in the schematic in Fig. 1. Interpretable solutions are desired generalised curvilinear coordinates are given as follows:
for wider applicability of ML methods in process industry. Regression ̂ ̂ ∂( E
̂− E v
̂ ) ∂( F
̂− F v
̂ )
(
̂ ξ ∂Ĥη
)
∂Q ∂Q ∂H
analysis helps in understanding the dependencies between the depen­ Γe +Γ + + =̂
S + σk + (1)
∂t ∂τ ∂ξ ∂η ∂ξ ∂η
dent and independent variables and to identify statistically significant
predictors. Gaussian Processes are commonly applied in regression. where τ is the pseudo-time; J denotes the Jacobian matrix; ξ and η
Regression through gaussian processes is performed by defining denote the curvilinear coordinates; Q
̂ denote the flow vector, ̂
E and ̂
F

Fig. 3. Comparison of equivalent bubble diameter with numerical and experimental results. a) 1.008 mm (Kamei and Hirata, 1990), and b) 4.9 mm (Kim and
Park, 2011).

3
V.M. Nagulapati et al. Nuclear Engineering and Design 393 (2022) 111794

Fig. 4. Comparative plots for LR, and GPR with experimental and simulation results for predicting a) 1.008 mm, and b) 4.9 mm condensing bubble diameters.

( ) ( )
Table 1 ̃
k = − ∇.
∇α
n= −
= − ∇.̂
∂̂n x ∂̂n y
+ (7)
RMSE and MAPE values for the validation plots of LR and GPR models. |∇α| ∂y ∂y
Diameter Model 1.008 mm 4.9 mm where ̂
n is the unit normal vector to the interface:
RMSE MAPE RMSE MAPE
∇αL
LR 0.000489 3.48% 0.026869 1.72%
n=
̂ (8)
|∇αL |
GPR 0.007125 12.31% 0.012908 1.34%
The interface advection equation for the vapor phase is.
( )
v v ∂αv ∂αv ∂αv αv 1 − αv
denote the flux terms; ̂
E and ̂ F denote the viscous terms; and ̂
S denotes +U +V = − + ṁ (9)
the source term: ∂t ∂ξ ∂η ρL ρv
⎛ ⎞ ⎛ ⎞ ⎛ ⎞ The surface tension term is given as.
p YL ρV YL ρU
⎜u⎟ ⎜ ρuV + ηx p ⎟ ⎜ ρuU + ξx p ⎟ ( )
⎜ ⎟ ⎜ ⎟ ⎜ ⎟ ∂̂n x ∂̂n x ∂̂n y ∂̂n y
̂ = 1 ⎜ v ⎟F ̂ 1⎜ ⎟̂ 1 ⎜ ⎟ (2) Pσ = σ̃
k = − σ ξx + ηx + ξy + ηy (10)
Q
J⎜ ⎟ = J ⎜ ρvV + ηy p ⎟ E = J ⎜ ρvU + ξy p ⎟, ∂ξ ∂η ∂ξ ∂η
⎝T⎠ ⎝ ρht V ⎠ ⎝ ρht U ⎠
Yv Yv ρV Yv ρU
where the phase interface equation is given as.
⎛ ⎞ ⎛ ⎞
0 0 ξx ∂α∂ξv + ηx ∂α v

(11)
∂η
⎜ ξx τxx + ξy τxy ⎟ ⎜ ηx τxx + ηy τxy ⎟ η x = √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
̂ ( ) ( )
⎜ ⎟ ⎜ ⎟
̂ = 1 ⎜ ξx τyx + ξy τyy ⎟ F 1⎜
v v 2 2
E ̂
⎟ = J ⎜ ηx τyx + ηy τyy ⎟
⎟ (3) ξx ∂α∂ξv + ηx ∂α v
+ ξy ∂α∂ξv + ηy ∂α v
J⎜⎝ ξx bx + ξy by ⎠ ⎝ η bx + η by ⎠
∂η ∂η
x y
0 0
ξy ∂α∂ξv + ηy ∂α∂ηv
⎛ ⎞ ⎛ ⎞ η y = √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
̂ ( ) ( ) (12)
0 0 2 2
⎜ ξx αv ⎟ ⎜ ηx αv ⎟ ξx ∂α∂ξv + ηx ∂α∂ηv + ξy ∂α∂ξv + ηy ∂α∂ηv
⎜ ⎟ ⎜ ⎟
̂ ξ = 1 ⎜ ξy αv ⎟ H
H ̂ η = 1 ⎜ ηy αv ⎟ ̂
J⎝⎜ ⎟ J⎜ ⎟S The details of the numerical implementation of the dual-time pre­
0 ⎠ ⎝ 0 ⎠
0 0 conditioned system are found in the studies of Ha and Park (Ha et al.,
⎛ ⎞ 2017) and Ha et al. (Ha and Park, 2016). The discretization form of Eq.
ṁ (− YL ρv/y )/
⎜ ρgx − ( ρuv − τxy)/ y ⎟ (1) is given as.
1⎜ ⎟
= ⎜ ⎜ ρgy − ρv2 − τyy y )/ ⎟,
⎟ (4) [ ( ) ]n+1,m
J⎝ ( ̃+ + ∇ξ ΓA
∇ξ ΓA ̃− n+1,m
(13)
⎠ n+1,m
ṁhlv + uτxy + vτyy − ρht v − qy y I + Γ−s 1 +
̂
ΔQ ̂
= − Γ−s 1 R
− ṁ − Yv ρv/y ̃
+∇η ΓB + ∇η ΓB ̃−

where I denote the identity matrix,Γs = Γ/Δτ + 1.5t; Δ and ∇ denote


bxi = uj τxi xj − qxi
the first-order forward and backward differencing respectively; super­
qx = − λ(ξx ∂ξ T + ηx ∂η T) (5)
qy = − λ(ξy ∂ξ T + ηy ∂η T) scripts n + 1, m and m + 1 denotes the physical-time level, actual and
new pseudo-time level, respectively; Δt and Δτ are the physical- and
U = ξx u + ξy vV = ηx u + ηy v (6) pseudo- time steps, respectively. A
̃ ± and B
̃± are the preconditioned flux
Jacobians. R
̂ is the residual given as.
The interfacial tension force was applied as a continuous surface
force (CSF) in this simulation according to Brackbill et al. (Brackbill
et al., 1992). The curvature is defined as.

4
V.M. Nagulapati et al. Nuclear Engineering and Design 393 (2022) 111794

Fig. 5. Machine learning predictions for condensing bubble diameter using LR model for a) 1.008 mm b) 3 mm, c) 4 mm, d) 4.9 mm, and e) 6 mm bubble diameters.

5
V.M. Nagulapati et al. Nuclear Engineering and Design 393 (2022) 111794

Fig. 6. Interpolation Remaining bubble condensing history using machine learning predicting of various bubble diameter. a) 3 mm, and b) 6 mm.

̂m ̂ n ̂ n− 1 ( )m sets of variables, where the dependent variable or response can be a


̂ m = − Γe Q − 4 Q + Q + E
R ̂ i+1/2 − Ê i− 1/2 + F ̂ j+1/2 − F
̂ j− function of the independent variables or predictor given by.
(14)
1/2
( v 2Δt )m
(19)
v v v m
− Ê ̂ ̂ ̂
i+1/2 − E i− 1/2 + F j+1/2 − F j− 1/2 − ̂S Y = β 0 + β 1 X1 + β 2 X2 + ⋯ + β n Xn

For the time discretization, the physical- and pseudo-time derivatives where Y represents the dependent variable or response, β0 to βn are
were approximated using the second- and first-order backward differ­ equation parameters and X1 to Xn are independent variables. Therefore,
ences, respectively. For the spatial discretization, the inviscid terms the mean of Y is a linear function of X, but the change in Y value is not
were approximated using a MUSCL upwind scheme, while the viscous dependent on the X value. Given the uncorrelated nature of the errors
and surface tension terms were discretized using the second-order cen­ the responses are also uncorrelated.
tral differencing scheme.
The discretization of the interface advection Eq. (9) is given as. 2.2.2. Gaussian process Regression(GPR)
⎡ ( ) ⎤ GPR can be described as a non parametric model based on Bayesian
4α n
− αn− 1
2 U n+1 α*v,i+1/2 − α*v,i− 1/2 ( )
αv 1 − αv n n theory, where a limited number of arbitrary variables following a joint
⎢ ⎥
αn+1 = v Gaussian distribution defined by their mean and covariance (Li et al.,
v
v − Δt⎣ ( ) ⎦− + ṁ
3 3 +V n+1 α*v,j+1/2 − α*v,j− 1/2 ρL ρv
2020; Goebel et al., 2008). Assuming f(x), m(x) and k (x, x’) as real
(15) process, mean function and the co-variance function, respectively, the
mean is given by.
where
m(x) = E[f (x)] (20)
{
αnv,i±1/2,LS
α*v,i±1/2 = If U > 0 (16) and co-variance function is given by.
αnv,i±1/2,RS
k(x, x’ ) = E[(m(x) − f (x) )(m(x’ ) − f (x’ ) )] (21)
The void fraction at the left (LS) and right (RS) of cell boundary were
considered in the same the way as in Ha et al. (Ha et al., 2015). GPR is used to estimate the output f(x), where the input vector is x.
The output function is f(x) GP[m(x), k(x, x’ ) ]. The co-variance function
α n n
= α + 0.25[(1 + χ )rLS φ(1/rLS ) + (1 − χ )φ(rLS ) ]δi− (10) ( )
v,i+1/2,LS v,i 1/2
k xi , xj is the measurement between xi and xj inputs and is given by.
( ) ( ) ( )
αnv,i+1/2,RS = αnv,i − 0.25[(1 + χ )rRS φ(1/rRS ) + (1 − χ )φ(rRS ) ]δi+3/2 (17) k xi , xj = kf xi , xj + kn xi , xj (22)

with where kf and kn are the functions describing the unknown part of the
system and the noise of the system, respectively. The approach of
ϕ(r) = max(0, min(rβ, 1.0), min(r, β) )
gaussian process is the assumption of a simple form of f(x), instead of
rLS =
δi+1/2
, rRS =
δi+1/2
, δi+1/2 = αi+1 − αi (18) going for a parametric function f (x, θ) and then estimating θ.
δi− 1/2 δi+3/2 Squared Exponential (SE) is a covariance function given by.
β = 64, χ = 1/3 ( )
1
(23)
′ ′
κSE (x, x ) = θ2f exp − 2 ‖x − x ‖2
θl
2.2. Machine learning details
The SE kernel maps the correlation between the given points as a
2.2.1. Linear regression function of the difference between their inputs x − x’ while assuming the
In Linear regression (LR) the model tries to build a relationship by function as smooth. This assumption is a harsh condition for batteries
adopting a linear equation to the observed data and the variables. One given the nature of their degradation phenomena. As an alternative, the
set of variables are explanatory, and the other set are dependent vari­ matern covariance is considered which is given by.
ables. The linear equation is the relationship function between the two

6
V.M. Nagulapati et al. Nuclear Engineering and Design 393 (2022) 111794

( ′ )v ( )
21− v √̅̅̅̅̅ (x − x ) √̅̅̅̅̅ (x − x′ ) 3.2. Analysis of Machine learning predictions.
(24)

κMa (x − x ) = σ 2 2v Rv 2v
Γ(v) ρ ρ
In this work, machine learning based LR, and GPR models are
where the hyper parameter related to smoothness is denoted by v. Larger developed based on CFD simulation data for bubble condensation. The
v corresponds to smoother function. Commonly given by v = 32 andv = 52. CFD model is validated based on experimental data as shown in Fig. 1
The log-likelihood function is maximised to optimize the unknown schematic. LR and GPR models are trained using the CFD generated data
hyper parameters in the covariance function (Richardson et al., 2019; and the thus trained models are used to predict bubble condensation.
Datong, 2012). Different bubble diameter data were used to train the models. 1.008
mm, 3 mm, 4 mm, 4.9 mm, and 6 mm bubble diameters were considered
3. Results and discussion for training and prediction and the results compared. Both LR and GPR
models are compared for prediction accuracy with Root Mean Square
3.1. Bubble condensation and its data sets Error (RMSE) and Mean Absolute Percentage Error (MAPE) as the per­
formance metrics. Fig. 4 shows the validation plots for LR and GPR
To predict bubble condensation behavior using machine learning plotted against experimental and numerical data for bubble diameter
methods, available numerical simulation data were used. The developed 1.008 mm in Fig. 4 (a) and for bubble diameter 4.9 mm in Fig. 4 (b). The
in-house code to were used to simulate the condensing bubble behavior results show the predictive capabilities of machine learning models.
under various aspects such as single and multiple bubble condensing Both LR and GPR models show good agreement with the experimental
behavior, bubble diameter, flow, and physical conditions. Fig. 2 shows and numerical data. However, it can be seen that the predictions deviate
initial and boundary condition of computational domain. Initially the slightly from the actual values over time. This is assumed to be due to the
bubble was injected into subcooled flow fluid from the bottom of the fact that the training datasets lack sufficient data from the later stages of
flow channel. Based on the initial and boundary conditions, the behavior bubble condensation as the data was split sequentially for training:
of the bubble was observed for the studied conditions. The temperature testing datasets. This leads to an increase in prediction error over time as
of the entire domain was considered as the saturation temperature, the the prediction moves further away from the last known training point. In
initial bubble diameter was 1 to 6 mm, and the pressure was 0.1 MPa. spite of these shortcomings the models exhibited acceptable level of
Fig. 3 illustrates a suitable value of the condensation constant defines prediction accuracy with low RMSE and MAPE values as can be seen in
the amount of the vapor phase converted into the liquid phase at the Table 1. From the RMSE and MAPE values in Table 1 it can be assumed
interface. The result shows that the bubble life history derived from the that the LR model exhibits better prediction accuracy compared to GPR
numerical results is highly sensitive to the condensation coefficient, and model. Given the superior performance of LR model, further predictions
the Lee phase transfer model is that the condensation constant is a of bubble condensation behavior with various bubble diameters are
function of pressure and temperature. The validation of in-house code- performed with LR based machine learning model. Fig. 5 (a)-(e) display
based simulation against the experimental results. The present numeri­ the actual vs prediction plots with LR model for 1.008 mm, 3 mm, 4 mm,
cal result shows good agreement with the experiment (Kim and Park, 4.9 mm and 6 mm bubble diameter, respectively. Table 1 displays the
2011; Kamei and Hirata, 1990). During the simulation, pressure, ve­ corresponding RMSE values. It can be observed that the LR model pre­
locity, temperature, and volume fraction were measured for investi­ dictions are well in agreement with the actual values with low RMSE
gating the condensing bubble behavior. This data is used to form a values demonstrating the predictive capabilities of the model. Finally,
machine learning prediction dataset to match the diameter of the the well trained LR model is used to predict the remaining bubble
condensing bubbles and the condensation history of the remaining condensation phenomenon for 3 mm and 6 mm bubble diameters as
bubbles. For validation of condensing bubble prediction, the experiment shown in Fig. 6. The predictions allow us to understand the complete
data for 1.008 mm and 4.9 mm size bubble diameter and their corre­ bubble condensation behaviour over time. It can be examined for 3 mm
sponding error percentage with the experimental results are 4.87 and bubble diameter prediction in Fig. 6 (a), the prediction ends with a
2.18%, respectively (S Paramanantham et al., 2018; Paramanantham bubble diameter of 0.4 mm after a time 49 ms. The prediction for 6 mm
et al., 2020). bubble diameter in Fig. 6 (b), the prediction reaches 0 mm in diameter
We are using bubble diameter ranging between 1 and 6 mm. E.g., the after a time of 90 ms. We have predicted for 3 mm and 6 mm bubble
condensing time for 1.008 mm bubble condensation took until 3.3 mi­ diameters to show the interpolation and extrapolation abilities of the
croseconds with corresponding bubble size 0.18 mm. The data were machine learning LR model.
extracted during the simulation of bubble condensing behavior
amounting to nearly 100 data points with different time intervals. 4. Conclusion
However, interval between two-time steps is not equal as the extracted
data is based on the iteration steps. The CFL used in the simulation will In this study, the behavior of bubble condensation life history pre­
be accelerating based on the iteration convergence criteria. So, for the diction using machine learning techniques from the experiments and
data extracted with constant iteration steps, time interval would be numerical datasets was investigated. The datasets are composed of
different. Numerical simulation data for various bubble diameters of bubble condensation history and time. Experimental and numerical
1.008 mm, 3 mm, 4 mm, 4.9 mm and 6 mm is used for training and datasets for 1.008 mm and 4.9 mm were used for prediction, of which
validation of the machine learning models. Models trained in this way 30% of the data for testing and 70% of the actual sample data for
are used to predict bubble condensation behavior. The dataset consists training the models. The actual and predicted data were compared with
of pressure, velocity, temperature, and void fraction collected at random both experiment and numerical results. A validated machine learning
time intervals as inputs and with the corresponding bubble diameter model was used to predict 3 mm, and 6 mm of subcooled bubble
data as the output. The data available for training the model is divided condensing life-history. The proposed approach gives deep insights into
by a ratio of 70:30, i.e., 70% of the data is used for training machine bubble condensation phenomenon without the need for extensive
learning models, and 30% of the data is used for testing. The data was experimental and numerical studies. This data-driven machine learning
split sequentially in the ratio of 70:30 for training: testing, respectively. approach coupled with CFD-derived data is an effective approach for
Splitting the data sequentially translates to the data from the origin/ predicting bubble condensation. Using CFD-derived data increases the
formation of bubble to 70% of its condensation life is used for training prediction accuracy of the machine learning model by providing extra
and the remaining 30% i.e., till the bubble is completely condensed, is data points for the model to train on.
used for testing.

7
V.M. Nagulapati et al. Nuclear Engineering and Design 393 (2022) 111794

CRediT authorship contribution statement Qiu, B., et al., 2020. Experimental investigation on the interface shape of bubble
condensation for vertical upward steam jet at low mass flux. Int. J. Heat Mass Transf.
157, 119909.
Vijay Mohan Nagulapati: Conceptualization, Methodology, Vali­ Paramanantham, S., Ha, C.-T., Park, W.-G., 2018. Numerical investigation of single and
dation, Formal analysis, Investigation, Writing – original draft, Writing – multiple bubble condensing behaviors in subcooled flow boiling based on
review & editing. SalaiSargunan S. Paramanantham: Conceptualiza­ homogeneous mixture model. Int. J. Mech. Sci. 136, 220–233.
Tian, W., Ishiwatari, Y., Ikejiri, S., Yamakawa, M., Oka, Y., 2010. Numerical computation
tion, Methodology, Validation, Formal analysis, Investigation, Writing – of thermally controlled steam bubble condensation using Moving Particle Semi-
original draft, Writing – review & editing. Aleksey Ni: Writing – review implicit (MPS) method. Ann. Nucl. Energy 37 (1), 5–15.
& editing. Senthil Kumar Raman: Writing – review & editing. Hank­ Chen, R., Tian, W., Su, G.H., Qiu, S., Ishiwatari, Y., Oka, Y., 2010. Numerical
investigation on bubble dynamics during flow boiling using moving particle semi-
won Lim: Writing – original draft, Supervision, Funding acquisition. implicit method. Nucl. Eng. Des. 240 (11), 3830–3840.
Chen, R.H., Tian, W.X., Su, G.H., Qiu, S.Z., Ishiwatari, Y., Oka, Y., 2011. Numerical
Declaration of Competing Interest investigation on coalescence of bubble pairs rising in a stagnant liquid. Chem. Eng.
Sci. 66 (21), 5055–5063.
Zuo, J., Tian, W., Chen, R., Qiu, S., Su, G., 2013. Two-dimensional numerical simulation
The authors declare that they have no known competing financial of single bubble rising behavior in liquid metal using moving particle semi-implicit
interests or personal relationships that could have appeared to influence method. Prog. Nucl. Energy 64, 31–40.
Bahreini, M., Ramiar, A., Ranjbar, A.A., 2015. Numerical simulation of bubble behavior
the work reported in this paper. in subcooled flow boiling under velocity and temperature gradient. Nucl. Eng. Des.
293 (1), 238–248.
Acknowledgements Zhenyu, L., Bengt, S., Huiying, W., 2015. Numerical modeling of multiple bubbles
condensation in subcooled flow boiling. J. Therm. Sci. Eng. Appl. 7 (3).
Xiaohang, Q., Shripad, T.R., Maocheng, T., 2017. Numerical simulation of bubble
This work was supported by the National Research Foundation of formation and condensation of steam air mixture injected in subcooled pool. Nucl.
Korea (NRF) grant funded by the Korea government (NRF- Eng. Des. 320, 123–132.
2019M1A2A2065614) and by the Hydrogen Energy Innovation Tech­ Paramanantham, S.S., Phan, T.-H., Park, W.-G., 2020. Numerical analysis of bubble
condensation behavior under high-pressure flow conditions. Proc. Inst. Mech. Eng.,
nology Development Program of the National Research Foundation of Part C: J. Mech. Eng. Sci. 234 (18), 3725–3741.
Korea (NRF) funded by the Korean government (Ministry of Science and Paramanantham, S.S., Kim, D.-H., Park, W.-G., 2020. Numerical study of bubble behavior
ICT (MSIT)) (NRF-2019M3E6A1064290). under gradient flows during subcooled flow boiling in vertical flow channel.
Symmetry 12 (4), 611.
Liu, H., et al., 2020. An assessment and analysis of phase change models for the
References simulation of vapor bubble condensation. Int. J. Heat Mass Transf. 157, 119924.
Bahreini, M., et al., 2021. Numerical study on multiple bubbles condensation in
Zobeiry, N., Humfeld, K.D., 2021. A physics-informed machine learning approach for subcooled boiling flow based on CLSVOF method. Int. J. Therm. Sci. 170, 107121.
solving heat transfer equation in advanced manufacturing and engineering Ge, Z., Song, Z., Ding, S.X., Huang, B., 2017. Data mining and analytics in the process
applications. Eng. Appl. Artif. Intell. 101, 104232. industry: The role of machine learning. IEEE Access 5, 20590–20616.
Shang, Z., 2005. Application of artificial intelligence CFD based on neural network in Yan, Y., et al., 2022. Machine learning and in-silico screening of metal–organic
vapor–water two-phase flow. Eng. Appl. Artif. Intell. 18 (6), 663–671. frameworks for O2/N2 dynamic adsorption and separation. Chem. Eng. J. 427,
Chew, J.W., Cocco, R.A., 2020. Application of machine learning methods to understand 131604.
and predict circulating fluidized bed riser flow characteristics. Chem. Eng. Sci. 217, Babanezhad, M., Pishnamazi, M., Marjani, A., Shirazian, S., 2020. Bubbly flow prediction
115503. with randomized neural cells artificial learning and fuzzy systems based on k–ε
Upadhyay, M., Nagulapati, V.M., Lim, H., 2022. Hybrid CFD-neural networks technique turbulence and Eulerian model data set. Sci. Rep. 10 (1).
to predict circulating fluidized bed reactor riser hydrodynamics. J. Cleaner Prod. Brunton, S.L., Noack, B.R., Koumoutsakos, P., 2020. Machine learning for fluid
337, 130490. mechanics. Annu. Rev. Fluid Mech. 52 (1), 477–508.
Zhu, X., Wang, X., Ok, Y.S., 2019. The application of machine learning methods for Mesgarpour, M., Abad, J.M.N., Alizadeh, R., Wongwises, S., Doranehgard, M.H.,
prediction of metal sorption onto biochars. J. Hazard. Mater. 378, 120727. Jowkar, S., Karimi, N., 2022. Predicting the effects of environmental parameters on
Nagulapati, V.M., et al., 2021. A novel combined multi-battery dataset based approach the spatio-temporal distribution of the droplets carrying coronavirus in public
for enhanced prediction accuracy of data driven prognostic models in capacity transport- A machine learning approach. Chem. Eng. J. 430, 132761.
estimation of lithium ion batteries. Energy and AI 5, 100089. Kunz, M.R., et al., 2021. Data driven reaction mechanism estimation via transient
Nagulapati, V.M., et al., 2021. Capacity estimation of batteries: Influence of training kinetics and machine learning. Chem. Eng. J. 420, 129610.
dataset size and diversity on data driven prognostic models. Reliab. Eng. Syst. Saf. Byun, M., et al., 2021. Machine learning based predictive model for methanol steam
216, 108048. reforming with technical, environmental, and economic perspectives. Chem. Eng. J.
Han, T.e., Jiang, D., Zhao, Q.i., Wang, L., Yin, K., 2018. Comparison of random forest, 426, 131639.
artificial neural networks and support vector machine for intelligent diagnosis of Ha, C.T., Park, W.G., 2016. Evaluation of a new scaling term in preconditioning schemes
rotating machinery. Trans. Inst. Meas. Control 40 (8), 2681–2693. for computations of compressible cavitating and ventilated flows. Ocean Eng. 126,
Tang, J., et al., 2021. A machine-learning based phase change model for simulation of 432–466.
bubble condensation. Int. J. Heat Mass Transf. 178, 121620. Ha, C.-T., Kim, D.-H., Park, W.-G., Jung, C.-M., 2017. A compressive interface-capturing
Theßeling, C., Grünewald, M., Biessey, P., 2020. Determination of bubble sizes in bubble scheme for computation of compressible multi-fluid flows. Comput. Fluids 152,
column reactors with machine learning regression methods. Chem. Eng. Res. Des. 164–181.
163, 47–57. Hashemi, M.R., Ryzhakov, P.B., Rossi, R., 2020. An enriched finite element/level-set
Warrier, G.R., Basu, N., Dhir, V.K., 2002. Interfacial heat transfer during subcooled flow method for simulating two-phase incompressible fluid flows with surface tension.
boiling. Int. J. Heat Mass Transf. 45 (19), 3947–3959. Comput. Methods Appl. Mech. Eng. 370, 113277.
Sobes, V., Hiscox, B., Popov, E., Archibald, R., Hauck, C., Betzler, B., Terrani, K., 2021. Brackbill, J.U., Kothe, D.B., Zemach, C., 1992. A continuum method for modeling surface
AI-based design of a nuclear reactor core. Sci. Rep. 11 (1). tension. J. Comput. Phys. 100 (2), 335–354.
Paramanantham, S.S., Numerical modeling and analysis of multiphase flows for bubble Ha, C.T., Park, W.G., Jung, C.M., 2015. Numerical simulations of compressible flows
condensation phenomena, in School of Mechanical Engineering. 2020, Pusan using multi-fluid models. Int. J. Multiph. Flow 74 (Supplement C), 5–18.
National University. Li, X., Yuan, C., Li, X., Wang, Z., 2020. State of health estimation for Li-Ion battery using
Zuber, N., 1960. The dynamics of vapor bubbles in nonuniform temperature fields. Int. J. incremental capacity analysis and Gaussian process regression. Energy 190, 116467.
Heat Mass Transf. 2 (1–2), 83–98. Goebel, K., Saha, B., Saxena, A., A comparison of three data-driven techniques for
Okhotsimskii, A.D., 1988. The thermal regime of vapour bubble collapse at different prognostics. in 62nd meeting of the society for machinery failure prevention
Jacob numbers. Int. J. Heat Mass Transf. 31 (8), 1569–1576. technology (mfpt). 2008.
Kim, S.-J., Park, G.-C., 2011. Interfacial heat transfer of condensing bubble in subcooled Richardson, R.R., Osborne, M.A., Howey, D.A., 2019. Battery health prediction under
boiling flow at low pressure. Int. J. Heat Mass Transf. 54 (13-14), 2962–2974. generalized conditions using a Gaussian process transition model. J. Storage Mater.
Lucasa, D., Prasserb, H.M., 2007. Steam bubble condensation in sub-cooled water in case 23, 320–328.
of co-current vertical pipe flow. Nucl. Eng. Des. 237 (5), 497–508. Datong, L., et al., 2012. Data-driven prognostics for lithium-ion battery based on
Harada, T., Nagakura, H., Okawa, T., 2010. Dependence of bubble behavior in subcooled Gaussian Process Regression. In: Proceedings of the IEEE 2012 Prognostics and
boiling on surface wettability. Nucl. Eng. Des. 240 (12), 3949–3955. System Health Management Conference (PHM-2012 Beijing), pp. 1–5.
Brucker, G.G., Sparrow, E.M., 1977. Direct contact condensation of steam bubbles in Kamei, S., Hirata, M., 1990. Condensing phenomena of a single vapor bubble into
water at high pressure. Int. J. Heat Mass Transf. 20 (4), 371–381. subcooled water. Exp. Heat Transfer 3 (2), 173–182.
Yang, Q., et al., 2021. Experimental study on the influence of buoyancy on steam bubble
condensation at low steam mass flux. Exp. Therm Fluid Sci. 129, 110467.

You might also like