Download as pdf or txt
Download as pdf or txt
You are on page 1of 12

Chemical Engineering Science 62 (2007) 7078 – 7089

www.elsevier.com/locate/ces

Development of support vector regression (SVR)-based correlation for


prediction of overall gas hold-up in bubble column reactors for various
gas–liquid systems
Ankit B. Gandhi a , Jyeshtharaj B. Joshi a , Valadi K. Jayaraman b , Bhaskar D. Kulkarni b,∗
a Institute of Chemical Technology, Matunga, Mumbai 400 019, India
b Chemical Engineering & Process Development Division, National Chemical Laboratory, Pune 411008, India

Received 10 April 2007; received in revised form 23 July 2007; accepted 31 July 2007
Available online 7 August 2007

Abstract
The objective of this study was to develop a unified data-driven correlation for the overall gas hold-up for various gas–liquid systems using
support vector regression (SVR)-based modeling technique. Over the years, researchers have amply quantified the hydrodynamics of bubble
column reactors in terms of the overall gas hold-up. In this work, about 1810 experimental points were collected from 40 open sources
spanning the years 1965–2007. The model for overall gas hold-up was established as a function of several parameters which include superficial
gas velocity, superficial liquid velocity, gas density, molecular weight of gas, sparger type, sparger hole diameter, number of sparger holes,
liquid viscosity, liquid density, liquid surface tension, operating temperature, operating pressure and column diameter of the gas–liquid system.
For understanding the hold-up behavior, the data used for training the model was grouped into various gas–liquid systems viz., air–water,
gas–aqueous viscous liquids, gas–organic liquids, gas–aqueous electrolyte solutions and gas–liquid systems operated over a wide range of
pressure. A generalized model established using SVR was evaluated for its performance for various gas–liquid systems. Statistical analysis
showed that the proposed generalized SVR-based correlation for overall gas hold-up has prediction accuracy of 97% with average absolute
relative error (% AARE) of 12.11%. A comparison of this correlation with the selected system specific correlations in the literature showed
that the developed SVR-based correlation significantly gives enhanced prediction of overall gas hold-up.
䉷 2007 Published by Elsevier Ltd.

Keywords: Bubble column reactor; Overall gas hold-up; Support vector regression

1. Introduction which the reaction occurs. Indirectly, the hold-up profiles gov-
ern the liquid phase flow pattern and hence the rates of mixing,
Bubble columns are widely used in chemical, petrochemi- heat transfer and mass transfer. Thus an adequate knowledge
cal, biochemical and metallurgical industries. The advantages of overall gas hold-up and its profile is needed for modeling,
of bubble columns include good heat and mass transfer char- design, and scale-up of bubble column reactors.
acteristics, no moving parts and thus reduced wear and tear, Over the years, overall gas hold-up has been studied
higher catalyst durability, ease of operation, compactness and extensively with various measurement techniques, ranging
low maintenance. In bubble columns, the gas phase is in the from measuring pressure drop to computed tomography. In the
form of dispersed bubbles in a continuous liquid phase. Over- literature, numerous empirical correlations have been proposed
all gas hold-up, defined as the fraction of the reactor dynamic for overall gas hold-up for various gas–liquid systems. An
volume occupied by the gas is one of the important parame- excellent review on the topic is given by Joshi et al. (1998).
ters for bubble column design and scale-up. The gas hold-up Some of the important correlations for overall gas hold-up for
directly decides the reactor volume depending upon the phase in various gas–liquid systems and considered for this study are
listed in Table 1.
∗ Corresponding author. Tel.: +91 20 5893095; fax: +91 20 5893041. It has now been recognized that the number of variables
E-mail address: bdk@ems.ncl.res.in (B.D. Kulkarni). affecting gas hold-up are large in number and they interact with
0009-2509/$ - see front matter 䉷 2007 Published by Elsevier Ltd.
doi:10.1016/j.ces.2007.07.071
A.B. Gandhi et al. / Chemical Engineering Science 62 (2007) 7078 – 7089 7079

Table 1
Summary of overall gas hold-up correlations for various gas–liquid systems

Researcher Gas–liquid system Correlation

Hikita and Kikukawa (1974) Gas–water/gas–organic G = 0.505VG0.47 (72/L )7/3 (1/L )0.05
Hikita et al. (1980) Gas–water/gas–organic G = 0.672(VG L /L )0.578 (4L g/L 3L )−0.131 (g /L )0.062 (G /L )0.107
Kelkar et al. (1983) Gas–aqueous electrolyte solution G = 0.475VG0.37
Reilly et al. (1986) Gas–water G = (296VG0.44 −0.16
L −0.98
L 0.19
G ) + 0.009
Wilkinson et al. (1992) Gas–water/high pressure VG < Vtrans ; G = VG /Vs,b
VG > Vtrans ; G = (Vtrans /Vs,b ) + [(VG − Vtrans )/Vl,b ]
Vtrans /Vs,b = 0.5 exp(−193−0.61
G 0.5
L L )
0.11

L Vs,b /L = 2.25(3L L /g 4L )−0.273 (L /G )0.077


L Vl,b /L = L Vs,b /L + 2.4{L (VG − Vtrans )/L }0.75
× (3L L /g 4L )−0.07 (L /G )0.077

Reilly et al. (1994) Gas–water/high pressure G = 2.84VG 0.04


G / L
0.12

Zahradnik et al. (1995) Gas–aqueous electrolyte solution G = 0.87VG0.62


  0.15   0.15   0.15   0.15
w
Joshi et al. (1998) Gas–water/organic/aqueous elec- G = 0.62VG0.56 w G w
  a L
trolyte/aqueous viscous solution
Jordan and Schumpe (2001) Gas–organic/high pressure G /(1 − G ) = b1 (g L dB2 /L )0.16 (g 2L dB3 /2L )0.04 (VG /(gd B )0.5 )0.7

×{1 + 27(VG /(gd B )0.5 )0.52 (G /L )0.58 },

where b1 is the sparger dependent constant, dB the bubble dia.(m)


−0.24 0.22
Anabtawi et al. (2003) Gas–aqueous viscous solution G = 0.60VG0.60 H −0.38 0.02
L L L
−0.18 −0.12 [0.3 exp(−9L )]
Urseanu et al. (2003) Gas–aqueous viscous G = 0.21VG D
0.58 L G
solution/high pressure
  2/3
dS
Mouza et al. (2005) Gas–aqueous viscous G = 0.001 Fr0.5 A0.1
r E0
2.2 ,
D
solution/gas–organic
where [Fr = VG /Dg], [Ar = D 3 2L g/2L ], [E0 = D 2 L g/L ]

each other in a complex way. For instance, superficial gas ve- gas hold-up in bubble column reactors. The SVR-based model’s
locity (VG ) affects G in very different ways: (i) in homogenous application pertaining to the field of chemical engineering has
and heterogeneous regimes; (ii) with a change in L , L , L , been described by Agrawal et al. (2003) and Nandi et al. (2004).
etc.; (iii) temperature and pressure of operation; (iv) sparger SVR is a data driven modeling tool, based rigorously on sta-
design; (v) column diameter, etc. A similar statement can be tistical learning theory. This recently developed machine learn-
made on the effect of any other variable on G . Therefore, ing formalism is gaining popularity due to its many attractive
though very large number of correlations have been developed features and promising empirical performance. The salient fea-
in the past 50 years, its results have provided a very limited tures of SVR includes: (i) The objective function in SVR be-
success. This is mainly because a simple (visibly comprehen- ing of a convex quadratic form possesses a single minimum
sible) correlation will always have limited success when used (Fig. 1A), thus avoiding chances of getting trapped into poor
for strongly interacting multi parameter systems. local minima and thereby aiding in finding the global minima.
Data-driven modeling employing artificial intelligence and This is in direct contrast to usage of ANN and the conven-
machine learning methods are finding increasing relevance and tional non-linear least squares regression methodology for es-
importance in chemically reacting systems. The goal of data- tablishing the empirical correlations, where there is a very high
driven modeling is to build a prototype that can adapt and learn probability of getting trapped into poor local minima (Fig. 1B).
from practical data. Artificial neural networks (ANN) is the (ii) In SVR, the inputs are first nonlinearly mapped into a high-
most commonly used modeling tool under this category. Re- dimensional feature space wherein they are correlated linearly
cently Shaikh and Al-Dahhan (2003) and Behkish et al. (2005) with the output. Such a linear regression in high-dimensional
have developed an ANN-based correlation for predicting the feature space reduces the algorithm complexity enabling high
overall gas hold-up in bubble column reactors and slurry bubble predictive capabilities of both training and unseen test exam-
column reactors, respectively. However, their respective ANN ples. Further, introduction of kernel functions simplifies the
models were limited to a maximum prediction accuracy of 91% computational procedure by enabling calculations in the input
for overall gas hold-up. space itself. (iii) Sparseness property: decision boundary (final
In this study we propose to develop a support vector regres- regression function) is expressed in terms of a limited number
sion (SVR)-based data driven model for correlating the overall of support vectors (subset of a given training data).
7080 A.B. Gandhi et al. / Chemical Engineering Science 62 (2007) 7078 – 7089

Fig. 1. Nature of objective functions (A) SVR, (B) ANN.

Thus the objective of this study is to develop a unified


SVR-based data-driven correlation for the fractional gas
hold-up for various gas–liquid systems based on the overall
gas hold-up data available in the open literature. For under-
standing the hold-up behavior, the gas–liquid systems were
grouped into- (i) air–water, (ii) gas–aqueous viscous liquids,
(iii) gas–organic liquids, (iv) gas–aqueous electrolyte solu-
tions and (v) gas–liquid systems operated over a wide range
of pressures. A generalized unified correlation applicable for
all types of aforementioned gas–liquid systems over a wide
range of operating conditions, physical properties and column
dimensions has been established.

2. SVR-based modeling

The SVR is an adaptation of a recently introduced statis-


tical/machine learning theory based classification paradigm
known as, support vector machines (Vapnik et al., 1996). Con- Fig. 2. A schematic illustration of SVR using -sensitive loss function by
sider a training data set T = {(x1 , y1 ), (x2 , y2 ), (xN , yN )}, such Nandi et al. (2004).
that xi ∈ RN is a vector of input variables and yi ∈ RN the
corresponding scalar output (target) value. The objective over into high-dimensional feature space, F(x → (x)) and sub-
here is to build a -SVR model (Vapnik et al., 1996) to fit a sequently regressing it linearly. To fulfill the stated goal, SVR
regression function, y = f (x), such that it accurately predicts considers the following linear estimation function in the high-
the outputs {yi } corresponding to a new set of input examples, dimensional feature space:
{xi }. With this the linear SVR formulation can be given as
f (x, w) = (wQ(x) + b), (2)
f (x, w) = (wQx + b), (1)
where (x) is the function termed feature and (wQ(x)) the
where f is the regression function to be established, w the dot product in the feature space, F, such that  : x → F, and
weight vector, b the constant (bias). w ∈ F.
However, in most of the practical cases, the input data can- Here in -SVR model,  represents the radius of the tube lo-
not be correlated to the required output linearly. The options cated around the regression function, f (x) (Fig. 2). The region
available in such case are to use non-linear regression paradigm enclosed by the tube is known as ‘-insensitive’ zone. The loss
like non-linear least squares regression or ANN, in order to function assumes a zero value in this zone and as a result it does
deal with the non-linearity associated with the data set. How- not penalize the prediction errors with magnitudes smaller than
ever, as mentioned before, the problem associated with these . However, practically it is not possible to establish a func-
techniques is with the solution being getting trapped into the tion, f with errors less than . In that case some allowance for
poor local minima, thereby again resulting into poor predic- errors has to be considered, which can be represented in
tion accuracy. Using the same SVR formulation (Eq. (1)) it is terms of slack variables i and ∗i . These slack variables have
possible to overcome such a problem by first mapping the data non-zero values outside the [, −] (-insensitive) zone and
A.B. Gandhi et al. / Chemical Engineering Science 62 (2007) 7078 – 7089 7081

measure the deviation (yi − f (xi )) from the boundaries of the Lagrange multipliers  and ∗ give the value of weight vector,
-insensitive zone (penalty). At the same time function should w followed by the expanded form of the SVR,
be as flat as possible. In order to achieve such flat function
the weight vector ‘w’, has to be minimized which can be 
N
w= (i − ∗i )(xi ), (5)
achieved by minimizing the Euclidean norm, w2  (complex-
i=1
ity term). Mathematically expressing it in the form of convex
optimization problem we get 
Nsv
f (x, i , ∗i ) = (i − ∗i )((xi )Q(xj )) + b. (6)
1  N
i + ∗i
(∗) i=1
Minimize f (w, i ) = w2  + C
2
i=1 However, for the above-mentioned optimization problem
Subject to (Eq. (4)) with the increase in the input dimensions the dimen-
((wQ(x)) + b − yi )  + ∗i , sions in the high-dimensional feature space further increases
(yi − (wQ(x)) − b)  + ∗i , i , ∗i 0. (3) by many folds and thus becomes a computationally intractable
problem. Such a problem can be overcome by defining appro-
This penalty is regulated with the aid of cost function C em-
priate Kernel functions in place of the dot product of the input
ployed to obtain a trade-off between the flatness of the regres-
vectors in high-dimensional feature space.
sion function and the amount to which deviations larger than
 can be tolerated. Thus establishing the f (x) for the training K(xi , xj ) = ((xi )Q(xj )). (7)
data, the SVR minimizes the training set error by minimizing
not just only i and ∗i , but also w2  in order to increase the The advantage of a kernel function is that the dot product in the
flatness of the function. This serves to avoid an under-fitting and feature space can now be computed without actually mapping
also over-fitting of the training data. Fig. 2 depicts the situation the input vectors, xi into high-dimensional feature space. Thus,
graphically, where the points outside the tube contribute to the when using a kernel function all the necessary computations
cost insofar, as the deviations are penalized in the linear fashion. can be performed implicitly in the input space instead of in the
The above-mentioned convex optimization problem (Eq. (3)), feature space.
can be solved more easily in its dual form rather than solving With this the dual optimization problem (Eq. (4)) gets revised
in primal form. Thus after algebraic transformation to dual to the following form:
objective function (Gunn, 1998; Smola and Scholkopf, 2004;

N 
N
yi (i − ∗i ) −  (i + ∗i )
Vapnik et al., 1996), Eq. (3) takes the following form: (∗)
Maximize L (i,j ) =

N 
N i=1 i=1
yi (i − ∗i ) −  (i + ∗i )
(∗)
Maximize L(i,j ) = 1 
N N

i=1 i=1 − (i − ∗i )(j − ∗j )K(xi , xj )


2
i=1 j =1
1 
N N
− (i − ∗i )(j − j ∗) Subject to constraints
2
i=1 j =1 
N
× ((xi )Q(xj )) C i , ∗i 0 and (i − ∗i )yi = 0. (8)
Subject to constraints i=1


N Having obtained the Lagrange multipliers the basic SVR for-
C i , ∗i 0 and (i − ∗i )yi = 0. (4) mulation takes following the form:
i=1

Nsv
The coefficients  and ∗
mentioned in Eq. (4) are obtained f (x, , ∗ ) = (i − ∗i )K(x, xi ) + b. (9)
by solving this convex quadratic programming (QP). Owing i=1
to the specific character of the above-described quadratic pro-
gramming problem, only some of the coefficients, (i − ∗i ), Also the bias parameter, b, can be computed by applying
are non-zero and the corresponding input vectors, xi , are called Karush–Kuhn–Tucker (KKT) conditions, which states that at
support vectors (SVs). These SVs are known to be the most the optimal solution the product between dual variables and
informative data points that compress the information content constraints has to vanish. Thus giving,
of the training set, thereby representing the entire SVR func- b = {yi − (i − ∗i )K(x, xi ) − } for i ∈ 0, C,
tion. The coefficients i and ∗i have an intuitive interpretation
b = {yi − (i − ∗i )K(x, xi ) + } for ∗i ∈ 0, C, (10)
as forces pushing and pulling the regression estimate f (xi ) to-
wards the measurements, yi . It can be seen in Fig. 2, that SVs where xi and yi , respectively, denote the ith SV and the corre-
are depicted as points lying on the surface of the tube and the sponding target output.
regression function can be fully characterized by these support For using these kernel functions, it has to satisfy the Mercer’s
vectors. Owing to this characteristic the final regression model condition (Smola and Scholkopf, 2004), which states that any
can be defined with the help of relatively small numbers of positive semi-definite, symmetric kernel function, K, can be
input vectors. These SVs, xi and the corresponding non-zero expressed as a dot product in the high-dimensional space.
7082 A.B. Gandhi et al. / Chemical Engineering Science 62 (2007) 7078 – 7089

There exist several choices of kernel function K like lin- 3.2. Procedure for estimating regression function
ear, polynomial and Gaussian radial basis function. The most
commonly used kernel function is the Gaussian radial basis In the present study, a SVR-implementation known as
function (RBF) defined below in Eq. (11) (Gunn, 1998; Vapnik “-SVR” in the LIBSVM software library (Chang and Lin,
et al., 1996), 2001) has been used to develop the SVR-based model for
  overall gas hold-up. The LIBSVM package utilizes a fast and
1 efficient method known as sequential minimal optimization for
K(xi , xj ) = ((xi )Q(xj )) = exp − 2 xi − xj  2
2 solving large quadratic programming problems. Mentioned be-
i, j = 1, . . . , N, (11) low procedure was followed for estimating regression function
using LIBSVM software:
where  denotes the width of the RBF.
A brief algorithm on the establishment of the afore- (i) Training data set—Choose an appropriate training set of
mentioned regression function has been explained as input and output features.
Appendix A. (ii) Scaling—Conduct linear scaling of each data attribute to
the range [0, 1] in order to avoid data in greater numeric
3. Establishment of SVR-based correlation ranges to dominate those in smaller numeric ranges.
(iii) Kernel selection—In general RBF kernel is a reasonable
3.1. Collection of data set first choice.
(iv) Computing kernel function—For the selected value of ,
As mentioned earlier, over the years researchers have amply kernel function can be computed with the help of training
quantified the hydrodynamics of bubble column reactors based data set.
on the overall gas holdup. In this work, about 1810 experi- (v) Use the best parameters for establishment of regression
mental points have been collected from 40 sources spanning function—Having known the best estimated values of C, 
the years 1965–2007. This wide range of database (Table 2) and the kernel function (best values of these parameters are
includes experimental information for different gas–liquid obtained by using k-fold cross validation procedure that
systems. The value of gas hold-up is believed to be depen- minimizes the average absolute relative error (AARE) and
dent upon several parameters which include superficial gas enhances the values of correlation coefficient CCs towards
velocity, superficial liquid velocity, gas density, molecular unity, which is explained in the subsequent section), the
weight of gas, sparger type, sparger hole diameter, num- regression model can be established with the entire train
ber of sparger holes, liquid viscosity, liquid density, liq- data set. This is done by implementing QP (Eq. (8)) to
uid surface tension, operating temperature, operating pres- generate non-zero Lagrange multipliers ( − ∗ ) and the
sure and the column diameter of the gas–liquid system. corresponding support vectors. Knowing these also the bias
term ‘b’ is estimated, thereby defining the final regression
Table 2 function (Eq. (9)).
Range of input and output parameters of SVR-based correlation for overall
gas hold-up
3.3. Procedure for predicting the gas hold-up
Overall gas hold-up 0.0006–0.65

Operating conditions The final SVR function/model thus established is represented


Pressure, kPa 100–15600 by the support vectors, name of the kernel used, model pa-
Temperature, K 284–538 rameters used (C, and ), non-zero Lagrange multipliers
Superficial gas velocity, m/s 0.00030–0.72
( − ∗ ) corresponding to the support vectors and the value
Superficial liquid velocity, m/s 0–0.32
of the bias term ‘b’. All these parameters are considered for
Gas–liquid properties calculating/predicting the new output based on any new input
Gas
features. However due to large number of support vectors ob-
Molecular weight 2–44
Density, kg/m3 0.083–171 tained (around 869) for the SVR-based model for the predic-
tion of overall gas hold-up, they have been made available on
Liquid
the web. Such a provision has been made by preparing a web-
Viscosity, Pa s 0.0003–0.55
Density, kg/m3 655–1462 based SVR-simulation tool for the prediction of the overall gas
Surface tension, N/m 0.018–0.076 hold-up for bubble column reactors.
This tool is named as ‘SVR_GH_BC’, and can be down-
Reactor geometry
Column diameter, m 0.045–5.5 loaded from following web link http://www.esnips.com/web/
Sparger type Ring, Single hole, SVR. This tool is in the form of a Microsoft Excel file. Within
Multiple hole, the file, exist an excel sheet by name ‘SVR model’, is the sheet
Bubble cap, of interest and has provision for inputting various design and
Sintered plate,
operating conditions (13 inputs: (1) superficial gas velocity, (2)
Ejector and Injector type
Sparger hole diameter, m 0.00001–0.4 superficial liquid velocity, (3) gas molecular weight, (4) gas
density (5) sparger type, (6) sparger hole diameter, (7) number
A.B. Gandhi et al. / Chemical Engineering Science 62 (2007) 7078 – 7089 7083

of sparger holes, (8) liquid viscosity, (9) liquid density, (10) 4. Results and discussion
liquid surface tension, (11) system temperature, (12) system
pressure, and (13) column diameter) for the prediction of the As mentioned earlier, in the present study, an SVR-
overall gas hold-up for the defined conditions. Cell by name implementation known as “-SVR” in the LIBSVM software
‘predicted gas hold-up’ gives the predicted value of the overall library has been used to develop the SVR-based model for
gas hold-up. As stated earlier these calculations for the predic- overall gas hold-up in bubble column reactors. RBF kernel
tion of the overall gas hold-up are carried out mainly with the resulted in the minimum % AARE values and maximum CC
help of list of support vectors, and are listed in sheet by name values for the training data sets. In this study, standard k-fold
‘Support vectors’. These SVs exist in form of their scaled for- cross-validation was carried out for the estimation of the best
mat. Further calculation of the kernel elements (Eq. (11)) re- model parameters viz, C, and . In this approach, for a given
quired for the prediction, is carried out in sheet by name ‘Ker- combination of width of RBF kernel (), cost coefficient (C)
nel elements’. The information possessed within these sheets is and loss function parameter (), the training set is first ran-
vital and thus are locked and cannot be altered, however they domly divided into k equal sized subsets. Next, k number of
do help one to understand and supports the afore-mentioned models are constructed by leaving out a different subset each
calculations for the predictive output. The summation of the time, with the remaining (k − 1) subsets collectively repre-
product of the so obtained kernel elements and the respective senting the training set. An average of the error corresponding
non-zero Langrange multipliers for each support vectors, with
further addition of the bias term to it gives the single prediction Table 6
output (Eq. (9)). This particular prediction is shown within the Performance indicators for previous literature correlations for various
cell by name ‘predicted gas hold-up’ of sheet ‘SVR model’. gas–liquid systems
Gas–liquid system Correlation CC AARE (%)
Table 3
Parameter selection for SVR based G model Gas–water Hikita and Kikukawa (1973) 0.60 50.28
Reilly et al. (1986) 0.63 47.34
Model C = 1
22
 Number of Number of SVR (this work) 0.98 9.90
support training
vectors data points Gas–organic Hikita et al. (1980) 0.66 40.36
liquids
Overall gas 14.7 41.5 0.013 869 1810 Mouza et al. (2005) 0.60 48.34
hold-up SVR (this work) 0.98 6.32

Gas–aqueous Kelkar et al. (1983) 0.14 65.91


electrolyte
solutions
Table 4 Zahradnik et al. (1995) 0.11 70.66
Variation of SVR-based performance parameters with respect to number of SVR (this work) 0.95 7.81
support vectors Gas–aqueous Godbole et al. (1982) 0.76 35.38
Run no. Number % Support Train data Test data viscous solu-
of support vectors set correlation set correlation tions
vectors coefficient coefficient Urseanu et al. (2003) 0.65 44.24
SVR (this work) 0.99 8.01
1 1687 93.2 0.980 0.840
High pressure Wilkinson et al. (1992) 0.10 74.35
2 1223 67.5 0.979 0.860
Jordan and Schumpe (2001) 0.50 55.92
3 1013 56.0 0.977 0.865
SVR (this work) 0.98 8.43
4 869 48.0 0.976 0.870
5 719 39.7 0.966 0.860 Generalized Reilly et al. (1986) 0.41 58.13
6 497 27.4 0.960 0.850 Joshi et al. (1998) 0.38 59.21
7 375 20.7 0.940 0.790 Jordan and Schumpe (2001) 0.21 62.45
8 246 13.6 0.900 0.770 SVR (this work) 0.97 12.11

Table 5
Sample of scaled support vectors for the SVR-based model for overall gas hold-up for bubble column reactors

Non-zero VG VL Gas G Kd d0 N0 L L L T P D
Langrange (m/s) (m/s) molecular (Kg/m3 ) (dimensionless) (m) (Pa s) (Kg/m3 ) (N/m) (K) (kPa) (m)
multipliers weight
(  − ∗ )

−3.63 0.13797 0 0.639048 0.006414 0.70 3.8E−6 0 0.00189 0.40892 0.7597 0.0275 0 0.010
14.75 0.07525 0 0.639048 0.018467 0.58 1.4E−7 4.E−6 1.29E−3 0.427509 0.9522 0.0551 0.0180 0.0192
0.03 0.02263 0.00025 0.639048 0.006063 0.70 2.9E−6 0 9.53E−4 0.164684 0.2170 0.1141 0 0.0100
−0.34 0.01526 0.00064 0.639048 0.006063 0.97 1.1E−6 0 2.24E−3 0.525403 0.9623 0.0551 0 0.0174
7084 A.B. Gandhi et al. / Chemical Engineering Science 62 (2007) 7078 – 7089

0.8 0.7

0.6

OVERALL GAS HOLD-UP, εG (-)


PREDICTED GAS HOLD-UP, εG (-)

0.6
0.5

0.4
0.4

0.3

0.2
0.2

0.1

0 0
0 0.2 0.4 0.6 0.8 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
EXPERIMENTAL GAS HOLD-UP, εG (-) SUPERFICIAL GAS VELOCITY, VG (m/s)

0.8 0.7

0.6
PREDICTED GAS HOLD-UP, εG (-)

PREDICTED GAS HOLD-UP, εG (-)

0.6
0.5

0.4
0.4

0.3

0.2 0.2

0.1

0
0
0 0.2 0.4 0.6 0.8
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
EXPERIMENTAL GAS HOLD-UP, εG (-)
EXPERIMENTAL GAS HOLD-UP, εG (-)

Fig. 3. Parity plots based on entire data for overall gas hold-up, (A) 
Fig. 4. Gas hold-up values for, ◦ Letzel et al. (1999), × Camarasa et al.
SVR-based correlation, (B)  Reilly et al. (1986), × Jordan and Schumpe
(1999),  Han and Al-Dahhan (2007). (A) Variation of gas hold-up with VG
(2001).
for gas–water system. (B) Parity plot for gas hold-up at VG = 0.15 m/s.

to the left-out subsets, known as “cross-validation error” gives case, increasing ‘eps’, i.e., the width of the tube surrounded
an estimate of the model performance if a large-sized data set around the regression function, decreased the number of support
was available for building the model. After evaluation of the vectors and vice versa. However in that case the model starts
model for the wide range of model parameters (grid search becoming more and more generalized and leads to increase
methodology), gives the optimal values of them. So obtained in the value of AARE and drop in the value of CC. Table 4
optimal values of the three model parameters, corresponding shows the effect of number of support vectors on the model
to overall gas hold-up model are listed in Table 3. Optimally performance. As seen, for the SVR-based model represented
selecting these model parameters, automatically decides over by 1687 numbers of support vectors, shows excellent predic-
the optimal number of support vectors (869), which plays a tion accuracy for the train data set; however, it is not enough
vital role in the performance of the SVR-based model. In our for a model to show good prediction accuracy over train data
A.B. Gandhi et al. / Chemical Engineering Science 62 (2007) 7078 – 7089 7085

Table 7
Simulation conditions and results for various gas–liquid systems

Gas–liquid system Gas–water Gas–organic Gas–aqueous Gas–aqueous High pressure 1 High pressure 2
(air–water) liquids (air–25% electrolyte solutions viscous solutions (air–paratherm) (nitrogen–heptane)
v/v methanol) (air–0.2 M NaCl) (air–engine oil)

Author Akita and Akita and Akita and Anabtawi Lau et al. (2004) Tarmy et al. (1984)
Yoshida (1973) Yoshida (1973) Yoshida (1973) et al. (2003)
Superficial gas 0.016–0.25 0.016–0.20 0.05–0.25 0.05–0.25 0.05–0.25 0.05–0.20
velocity, m/s
Superficial liquid 0 0 0 0 0.0017 0.06–0.15
velocity, m/s
Gas molecular 28.84 28.84 28.84 28.84 28.84 28
weight
Gas density, kg/m3 1.18 1.18 1.18 1.18 1.18/9.0/16.0 1.18/3.95
Sparger type Single nozzle Single nozzle Single nozzle Single nozzle Perforated plate Perforated
plate/bubble cap
Sparger hole 0.005 0.005 0.005 0.01 0.0015 0.001/0.0005
diameter, m
No. of sparger holes 1 1 1 1 120 50/1000
Liquid viscosity, Pa s 0.001 0.007 0.00122 0.194 0.0317 0.0004
Liquid density, kg/m3 998 970 1018 917 870 685
Liquid surface tension, N/m 0.0728 0.045 0.0722 0.028 0.0295 0.02
Temperature, K 295 295 295 293 298 298
Pressure, kPa 100 100 100 100 100/790/1480 100/345
Column diameter, m 0.152 0.152 0.152 0.074 0.101 0.15/0.61
CC 0.99 0.95 0.98 0.99 0.98 0.97
% AARE 18.90 17.10 17.87 4.40 11.67 10.67

set. For any unknown test data set as well, the model should Further the parametric sensitivity analysis of the proposed
show reasonably good prediction accuracy, which fails for this model was carried out by checking the effect of the promi-
particular instance. Such a model established after satisfying nent input parameters on gas hold-up. Fig. 4A shows the vari-
almost 93% of the total data points, falls under the category of ation of the overall gas hold-up with respect to superficial gas
an over fitted model. Also selecting relatively very low numbers velocity for gas–water system. Looking carefully at the plot,
of support vectors (13.6% of the total data points), the model it can be seen that for a particular gas velocity, multiple val-
behaved in highly generalized manner, just like an under fitted ues of gas hold-up exist. Such a variation exists mainly due
model, showing relatively poor train as well as test prediction to variation in the design and operating conditions at the same
accuracy. Based on this comprehensive analysis over the selec- gas velocity. Thus in order to check for the sensitivity and
tion of the number of support vectors, the final SVR-based re- accuracy of the SVR-based model for reliably showing the
gression function for overall gas hold-up was established within effect of these parameters on the values of the gas hold-up,
all 869 numbers of support vectors, showing the maximum test simulations at constant gas velocity of 0.15 m/s were carried
prediction accuracy of 87%. Thus for any number of SVs lesser out. Referring Fig. 4A, it can be seen that at constant VG of
or more than the optimal numbers, suppresses the model per- 0.15 m/s, the G lies in the range of 0.16–0.58. A careful analy-
formance. Some sample numbers of support vectors defining sis reveals that, for the data by Letzel et al. (1999), gas hold-up
this model are listed in Table 5 in their scaled format. was found to increase with pressure, depicted by four different
In order to check the applicability and performance of this points for four different pressures in increasing order. In case
unified model against the literature correlations for various of data by Camarasa et al. (1999) at atmospheric conditions,
gas–liquid systems, data pertaining to the individual gas–liquid gas hold-up by sintered plate sparger was found to be more
system was subjected for test to the model. The simulation than that by perforated plate, which in turn was higher than
results are shown in Table 6 for individual and generalized that by single nozzle sparger. While for the data by Han and
gas–liquid system. It can be seen that for various gas–liquid Al-Dahhan (2007), shows the effect of sparger as well the pres-
system under consideration, SVR-based model performs much sure on gas hold-up. At atmospheric conditions it can be seen
better than the system specific empirical correlations proposed that the gas hold-up for perforated plate sparger was found to
in the literature by various authors. Fig. 3 shows one such com- be more than that for ring sparger. Similarly for perforated plate
parison between the SVR-based generalized correlation and the sparger, gas hold-up at the higher pressure values was found
literature coefficients (Reilly et al., 1986; Jordan and Schumpe, be much higher than that at the atmospheric pressure. Some
2001) for overall gas hold-up. It is clearly seen from Fig. 3 and of these effects of pressure and sparger were well captured by
the values of Table 6 that the SVR-based correlation predicts the SVR-based model and correspondingly predicted the gas
overall gas hold-up far better than the selected correlations for hold-up Parity plot for the same (Fig. 4B), shows an excellent
the available dataset. agreement between the experimental and the predicted values,
7086 A.B. Gandhi et al. / Chemical Engineering Science 62 (2007) 7078 – 7089

Fig. 5. Comparison of SVR prediction with the experimental data of various authors for various gas–liquid systems. (A) Gas–water system,  experimental data
Akita and Yoshida (1973), —SVR. (B) Gas–organic liquid system,  experimental data Akita and Yoshida (1973), (B1) 1—SVR, (B2) 2—SVR (heterogeneous
regime) (C) Gas–aqueous electrolyte solutions system,  experimental data Akita and Yoshida (1973). (C1) 1—SVR, (C2) 2—SVR (heterogeneous regime).
(D) Gas–aqueous viscous solutions system,  experimental data Anabtawi et al. (2003), —SVR. (E) High pressure system 1, experimental data Lau et al.
(2004),  100 kPa,  790 kPa,  1480 kPa, 1—SVR (100 kPa), 2—SVR (790 kPa), 3—SVR (1480 kPa). (F) High pressure system 2, experimental data Tarmy
et al. (1984),  100 kPa (D = 0.15 m),  100 kPa (D = 0.61 m),  345 kPa (D = 0.61 m), 1—SVR (100 kPa, D = 0.15 m), 2—SVR (100 kPa, D = 0.61 m),
3—SVR (345 kPa, D = 0.61 m).

showing the responsiveness of the model against the prominent gas–liquid systems from open literature. Simulation conditions
input parameters. for different gas–liquid systems have been detailed in Table 7,
As a part of this work, we also checked as to how well the with results shown as Fig. 5. For gas–water, gas–organic and
unified SVR-based correlation predicts the overall gas hold-up gas–aqueous electrolyte solutions systems, simulations were
for the data set which was not included in the databank for carried out at the experimental conditions of Akita and Yoshida
training the model i.e., test data set. The simulation conditions (1973). While that for gas–aqueous viscous solution and high
were selected in such a manner so as to see the effect of various pressure systems experimental conditions of Anabtawi et al.
gas–liquid system and pressure on the gas hold-up. For this SVR (2003) and Lau et al. (2004), Tarmy et al. (1984) were consid-
simulations were carried out using the test dataset for various ered respectively.
A.B. Gandhi et al. / Chemical Engineering Science 62 (2007) 7078 – 7089 7087

The prediction results for gas–water system as shown in d0 sparger hole diameter
Fig. 5A, is in excellent agreement with the experimental data. dS sparger diameter
However, the simulated values for the gas–organic, Fig. 5B (line D bubble column diameter
1) show little over prediction for the entire range of gas ve- f (x) regression function
locity, while that for gas–aqueous electrolyte solutions system F high-dimensional feature space
(Fig. 5C, line 1) the model under predicts mainly in the hetero- H clear liquid height
geneous regime. This observation suggests that grouping the K(xi , xj ) kernel function
data for different flow regimes and having a correlation for Kd sparger distribution coefficient
each flow regimes would benefit the accuracy of the predic- L Lagrangian function (dual form)
tion. Therefore it was thought desirable to develop two differ- N0 number of sparger holes
ent SVR-based models for each of the regimes i.e., homoge- Nsv number of support vectors
nous and heterogeneous regimes were established. Fig. 5B and P system pressure
C also shows the simulated results for SVR-based model ex- R input space
clusively for heterogeneous regime (line 2), with the improved T system temperature
prediction of the gas hold-up values. Moreover, it can be seen VG superficial gas velocity
that for similar design and operating conditions, gas hold-up VL superficial liquid velocity
for gas–organic and gas–aqueous electrolyte solutions systems Vtrans transition gas velocity
is seen to be higher than that by plain air–water system. These w weight vector
simulated results decently follow the trend of obtaining in- xi ith inputvector
creased gas hold-up for non-coalescing gas–organic and elec- yi target output corresponding to the ith vector
trolyte systems in comparison with coalescing gas–water sys-
tem. Further for gas–viscous liquid (Fig. 5D) and high pres- Greek letters
sure systems (Fig. 5E and F), simulation results are in good
(∗)
agreement with the experimental conditions. For high pressure ij Lagrange multipliers
 
system 1 (Fig. 5E) and high pressure system 2 (Fig. 5F), the gamma 21 2
effect of pressure on gas hold-up follows an excellent trend of
 loss function
increase in hold-up values with the experimental values. The
G overall gas hold-up, dimensionless
overall simulations performed using the SVR-based correlation
G gas phase viscosity
predicts the effect of different parameters and gas–liquid sys-
L liquid phase viscosity
tems on overall gas holdup as per the trend reported in the lit-
erature. This proves its utility as a design estimation tool for i
(∗)
slack variables
bubble column reactors.
G gas phase density
L liquid phase density
5. Conclusion  width of radial basis function (RBF) kernel
L liquid phase surface tension
SVR-based correlation shows noticeable improvement in the
(xi ) mapping function to high dimensional feature space
prediction of overall gas hold-up in comparison to the avail-
for input vector x
able literature correlations. The generalized SVR-based corre-
lation for the prediction of over gas hold-up yields % AARE Subscript
of 12.11% and % prediction accuracy of 97%, which is far bet-
ter than those obtained for the selected literature correlations. G overall gas phase
Also, the SVR-based correlation gave enhanced and more ac-
Superscript
curate predictions for a variety of gas–liquid systems over a
wide range of operating pressures, and various column diam- N number of training data points
eters. Hence the developed SVR-based correlation should be
useful in the scale-up of bubble column reactors.

Notation


ypredicted −yexperimental


AARE average absolute relative error (1/N ) N
1
yexperimental

b bias term
C cost function
CC correlation
coefficient
N
(y −y )(y −y )
i=1 experimental(i) experimental(mean) predicted(i) predicted(mean)
N
(y −y ) 2 N (y −y 2
1 experimental(i) experimental(mean) 1 predicted(i) predicted(mean) )
7088 A.B. Gandhi et al. / Chemical Engineering Science 62 (2007) 7078 – 7089

Acknowledgments Akita, K., Yoshida, F., 1973. Gas holdup and volumetric mass transfer
coefficient in bubble columns. Industrial and Engineering Chemistry
AG thanks Council of Scientific and Industrial Research Research 12 (1), 76–80.
(CSIR), the Govt. of India, New Delhi, for a Senior Research Anabtawi, M.Z.A., Abu-Eishah, S.I., Hilal, N., Nabhan, M.B.W., 2003.
Hydrodynamic studies in both bi-dimensional and three-dimensional
Fellowship (SRF). VKJ acknowledges for the financial support bubble columns with a single sparger. Chemical Engineering and
received from Department of Science and Technology, New Processing 42, 403–408.
Delhi, India. Behkish, A., Lemoine, R., Sehabiague, L., Oukaci, R., Morsi, B.I., 2005.
Prediction of the gas holdup in industrial-scale bubble columns and
Appendix A. Algorithm for the establishment of the regres- slurry bubble column reactors using back-propagation neural networks.
International Journal of Chemical Reactor Engineering 3, 1–35.
sion function Camarasa, E., Vial, C., Poncin, S., Wild, G., Midoux, N., Bouillard, J., 1999.
Influence of coalescence behaviour of the liquid and of gas sparging on
Given a training data set of N (1810) training examples, with hydrodynamics and bubble characteristics in a bubble column. Chemical
each consisting of a vector, x having 13 input parameters and Engineering and Processing 38, 329–344.
a corresponding scalar output value, y (overall gas hold-up), Chang, C.-C., Lin, C.-J., 2001. LIBSVM: a library for support vector machi-
the steps involved in establishing the regression function are as nes. Software available at http://www.csie.ntu.edu.tw/∼cjlin/libsvm.
Godbole, S.P., Honath, M.F., Shah, Y.T., 1982. Holdup structure in highly
follows:
viscous Newtonian and non-Newtonian liquids in bubble columns.
(1) Derivation of a -SVR model in the high dimensional fea- Chemical Engineering Communications 16, 119–134.
ture space, Gunn, S.R., 1998. Support vector machines for classification and regression.
Technical Report, department of electronics and computer science.

N University of Southampton. http://trevinca.ei.uvigo.es/∼cernadas/tc03/mc/
yi = f (x, i , ∗i ) = (i − ∗i )((xi )Q(xj )) + b, SVM.pdf.
i=1
Han, L., Al-Dahhan, M.H., 2007. Gas–liquid mass transfer in a high
pressure bubble column reactor with different sparger designs. Chemical
where i = 1 → N (total number of vectors/training data Engineering Science 62, 131–139.
set). Hikita, H., Kikukawa, H., 1974. Liquid-phase mixing in bubble columns:
effect of liquid properties. The Chemical Engineering Journal 8, 191–197.
(2) Employment of a suitable kernel for implicitly performing
Hikita, H., Asai, S., Tanigawa, K., Segawa, K., Kitao, M., 1980. Gas hold-up
necessary computations in the input space instead of in the in bubble columns. The Chemical Engineering Journal 20, 59–67.
high-dimensional feature space, Jordan, U., Schumpe, A., 2001. The gas density effect on mass transfer in
bubble columns with organic liquids. Chemical Engineering Science 56,
K(xi , xj ) = ((xi )Q(xj )). 6267–6272.
Joshi, J.B., Veera, U.P., Prasad, C.V., Phanilumar, D.V., Deshphande, N.S.,
Radial basis function (RBF) kernel is the ideal choice for Thakre, S.S., Thorat, B.N., 1998. Gas hold-up structures in bubble column
it. reactors. PINSA 64 (A), 441–567.
(3) Solve the QP problem (Eq. (8)) to obtain a subset of the Kelkar, B.G., Phulgaonkar, S.R., Shah, Y.T., 1983. The effect of electrolyte
total training vectors, N , having non-zero Lagrange mul- solutions on hydrodynamics and back mixing characteristics in bubble
tipliers, which is subsequently employed for the establish- columns. The Chemical Engineering Journal 27, 125–133.
Lau, R., Peng, W., Velazquez-Vargas, L.G., Yang, G.Q., Fan, L.S., 2004.
ment of the above mentioned regression function. Thus, Gas–liquid mass transfer in high-pressure bubble columns. Industrial and
i = 1 → Nsv (number of support vectors, 869). Thus, the Engineering Chemistry Research 43, 1302–1311.
final SVR formulation takes following the form: Letzel, H.M., Schouten, J.C., Krishna, R., van den Bleek, C.M., 1999. Gas
holdup and mass transfer in bubble column reactors operated at elevated

Nsv pressure. Chemical Engineering Science 54, 2237–2246.
f (x, , ∗ ) = (i − ∗i )K(x, xi ) + b. Mouza, A.A., Dalakoglou, G.K., Paras, S.V., 2005. Effect of liquid properties
i=1 on the performance of bubble column reactors with fine porous spargers.
Chemical Engineering Science 60, 1465–1475.
(4) Compute the bias parameter, b, (Eq. (10)) by applying Nandi, S., Badhe, Y., Lonari, J., Sridevi, U., Rao, B.S., Tambe, S.S.,
Karush–Kuhn–Tucker (KKT) conditions. Kulkarni, B.D., 2004. Hybrid process modeling and optimization
(5) Knowing the values of (i − ∗i ) for each of the support strategies integrating neural networks/support vector regression and genetic
algorithms: study of benzene isopropylation on hbeta catalyst. Chemical
vector and the bias parameter (global parameter), for any
Engineering Journal 97, 115–129.
single new input vector, the corresponding target output Reilly, I.G., Scott, D.S., De Bruijin, T.J.W., Piskorz, J., 1986. A correlation
is predicted. This is done first by constructing the kernel for gas holdup in turbulent coalescing bubble columns. The Canadian
matrix (Eq. (11)) from a single new input vector, ‘x’ and Journal of Chemical Engineering 64, 705–716.
the set of support vectors ‘xi ’, where i = 1 → number of Reilly, I.G., Scott, D.S., De Bruijin, T.J.W., MacIntyre, D., 1994. The role of
support vectors, Nsv . Finally after summation of product gas phase momentum in determining gas holdup and hydrodynamic flow
regimes in bubble column operations. The Canadian Journal of Chemical
of non-zero Lagrange multipliers and kernel elements, bias Engineering 72, 3–12.
term ‘b’ is added to get the predicted value. Shaikh, A., Al-Dahhan, M., 2003. Development of an artificial neural network
correlation for prediction of overall gas holdup in bubble column reactors.
References Chemical Engineering Processing 42, 599–610.
Smola, A.J., Scholkopf, B., 2004. A tutorial on support vector regression.
Agrawal, M., Jade, A.M., Jayaraman, V.K., Kulkarni, B.D., 2003. Support Statistics and Computing 14, 199–222.
vector machines: a useful tool for process engineering applications. Tarmy, B.L., Chang, M., Coulaloglou, C.A., Ponzi, P.R., 1984. The three
Chemical Engineering Progress 98 (1), 57–62. phase hydrodynamic characteristics of the EDS coal liquefaction reactors:
A.B. Gandhi et al. / Chemical Engineering Science 62 (2007) 7078 – 7089 7089

their development and use in reactor scaleup. Institution of Chemical Wilkinson, P.M., Spek, A.P., van Dierendonck, L.L., 1992. Design parameters
Engineers, Symposium series 87, 303–317. estimation for scale-up of high pressure bubble columns. A.I.Ch.E. Journal
Urseanu, M.I., Guit, R.P.M., Stankiewicz, A.G., Lommen van, K.J.H.G.M., 38 (4), 544–554.
2003. Influence of operating pressure on the gas hold-up in bubble columns Zahradnik, J., Fialova, M., Kastanek, F., Green, K.D., Thomas, N.H., 1995.
for high viscous media. Chemical Engineering Science 60, 1465–1475. The effect of electrolyte on bubble columns coalescence and gas holdup in
Vapnik, V., Golowich, S., Smola, A.J., 1996. Support vector method for bubble column reactors. Transactions of Institution of Chemical Engineers
function approximation, regression estimation and signal processing. 73 (A), 341–346.
Advanced Neural Information Processing System 9, 281–287.

You might also like