You are on page 1of 11

Journal of Quantitative Spectroscopy & Radiative Transfer 226 (2019) 127–137

Contents lists available at ScienceDirect

Journal of Quantitative Spectroscopy & Radiative Transfer


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

Inverse estimation of number and location of discrete heaters in


radiant furnaces using artificial neural networks and genetic algorithm
Rahul Yadav a, C. Balaji a,∗, S.P. Venkateshan b
a
Department of Mechanical Engineering, Indian Institue of Technology Madras, Chennai, India
b
Department of Mechanical Engineering, Indian Institue of Information Technology, Design and Manufacturing, Kancheepuram, India

a r t i c l e i n f o a b s t r a c t

Article history: The inverse problem of the determination of optimum number and location of discrete heaters in a three
Received 26 June 2018 dimensional radiant heating furnace is solved by developing an artificial neural network based model to
Revised 21 October 2018
replace the radiative transfer equation solver and coupling it with genetic algorithm. The furnace medium
Accepted 23 December 2018
is considered to be participating with a mixture of H2 O and CO2 . For the purpose of estimation of the
Available online 27 December 2018
best configuration of the heaters to achieve a desired uniform heat flux at the design surface, the inverse
Keywords: problem has been treated as an optimization problem. A generalized radiative transfer code based on
Discrete heaters Finite Volume Method and spectral line based weighted sum of gray gases method has been developed
ANN and validated with the available benchmarks. This code serves as the basis to develop the neural network
GA based model for the flux estimation at design surface. The ANN function has been tested upon 5 random
Radiant furnaces test cases. The ANN based replacement of RTE for this problem gives an error less than 4% for the local
heat flux values at the order of 103 times less computation time. A design case is then considered where
the desired flux levels should be uniform on the design surface in a very tight tolerance. Results obtained
show that this novel method of ANN coupled with GA proves to be a fast, robust and efficient tool for
solving such types of inverse problems.
© 2018 Elsevier Ltd. All rights reserved.

1. Introduction the desired variables may produce similar effects. Implementation


of inverse methodology in solving the radiation heat transfer prob-
In many practical applications like in the baking of food items, lems has gained considerable attention of the researchers in the
heat treatment of the semiconductor wafers, reheating of billets last two decades. The inverse solution methodology was discussed
and curing of paints and dyes, radiant heating furnaces are em- in detail by Howell et al. [1] as a more convenient alternative to
ployed. These have discrete heaters mounted over the top surface the traditional trial and error method in designing efficient radi-
that dissipate heat into the medium. This heat, in turn, is trans- ant systems. This was also applied to enclosures with participating
ferred to the specimens lying at the bottom of the furnace. There is medium [2–4], irregular geometry [3,5], specular surfaces [6], tran-
often a requirement to impart uniform heating over an area of the sient systems [7,8] and combined convection and radiation scenar-
specimen in order to achieve the desired surface properties. This ios [9–11]. Primarily, the solution to such inverse problems can be
area of the specimen which is under special consideration is called approached by regularization or optimization techniques. Regular-
the design surface. The task of achieving uniform heating is ac- ization methods reduce the ill-posedness of the problem by using
complished by controlling the power of the heater array and by an special numerical treatments such as singular value decomposition
intelligent arrangement and placement of individual heaters. Due and conjugate gradient regularization to name a few. Optimization
to high operating temperatures of the furnaces, radiation is the methods, on the other hand, try to approach the solution itera-
dominant mode of heat transfer. Assuming that all the other nec- tively where, in general, an initial guess is made for the solution
essary information related to furnace and the specimen is known, and which is then improved over the iterations to reach the opti-
the question before us is the best heater configuration that uni- mal solution. A fair comparison of regularization and optimization
formly heats the specimen surface. This is a classic inverse prob- methods to solve inverse radiative transfer problems can be found
lem. Such problems are ill-conditioned as several combinations of in Daun et al. [12]
With regard to the optimization strategies to solve the inverse

radiation heat transfer problems, methods that have been adopted
Corresponding author.
to optimize the objective function include gradient based methods
E-mail address: balaji@iitm.ac.in (C. Balaji).

https://doi.org/10.1016/j.jqsrt.2018.12.031
0022-4073/© 2018 Elsevier Ltd. All rights reserved.
128 R. Yadav, C. Balaji and S.P. Venkateshan / Journal of Quantitative Spectroscopy & Radiative Transfer 226 (2019) 127–137

One possible way to solve this problem is to pose this as an op-


Nomenclature timization problem. The objective function for the optimization
problem is formulated in such a way that the output of the op-
Ai Area of the control volume faces, m2 timization algorithm gives the best configuration (or placement)
aI Coefficient for discretization at the nodal point of the heaters. Once the objective function is defined, the task
aj Weight corresponding to the jth gray gas then is to determine the solution. Researchers have shown keen
Cj Absorption cross section corresponding to the jth interest in applying heuristic based search algorithms lately par-
gray gas m2 /mol ticularly for this problem, due to aforementioned reasons. For in-
Dci Directional weights stance, Sarvari [19] used genetic algorithm for obtaining the opti-
Fa Absorption line blackbody distribution function mum placement of heaters in a 2D radiant furnace. Ashouri et al.
F Objective function [20] considered a similar problem but the furnace was three di-
Iλ Spectral intensity, W/m2 sr μm mensional. Safavinejad et al. [15] used the micro genetic algorithm
Ibλ Blackbody intensity, W/m2 sr μm to optimize heater power and location in a 2D enclosure with dif-
L Mean beam length, m fuse and specular surfaces. Amiri et al. [16] determined the op-
M, N Total number of directions along θ and φ timal number and location of heaters in a two dimensional enclo-
MSE Mean squared error sure with discrete heaters at the adjacent boundaries for two cases
ni Normal to the wall location namely (i) radiative equilibrium and (ii) specified temperature of
NGG Number of gray gases in SLW model the medium. Lemos et al. [21] used generalized extreme optimiza-
Qh Total heater power, W tion algorithm to optimize the geometric configuration of filament
q Radiative heat flux, kW/m2 heaters. In many similar cases, where the heating has to be per-
R Goodness of fit formed using gas fired burners, the medium inside the radiant fur-
SP Source term in the discretized radiative transfer nace contains participating gases such as H2 O and CO2 , which is
equation likely to affect the radiation heating of the specimens [2,3]. An as-
SSE Sum of Squared Errors sumption of transparent medium may lead to inaccurate results in
SSME Sum of Squared Mean Errors these cases. To address this, Amiri et al. [22] considered a two di-
Tref Reference temperature, K mensional radiant heating furnace with a gray absorbing-emitting
Y Mole fraction of the gas species medium and isotropic scattering and concluded that the changes
in medium and wall properties change the optimal solutions con-
Greek symbols siderably.
V Control volume, m3 From the preceding review of the literature, it is clear that the
 Emissivity focus of the researchers has been towards the implementation of
κj Absorption coefficient of jth gray gas, m-1 efficient optimization algorithms to solve inverse problems of find-
κs Absorption coefficient of soot, m-1 ing power and placement of heaters in radiant furnaces. How-
μ, ζ , η Direction cosines ever, it is clear that a major fraction of the computational load

Solid angle, sr is taken up by the forward estimation (i.e. the RTE solution from
first principles) rather than the inverse one. This necessitates the
Subscripts
need to improve the computational efficiency of the forward RTE
d Design surface properties
solver. One possibility is to utilize the concepts of neural network
E, W, N, S East, West, North and South nodal points with ref-
modelling. Artificial neural networks have gained attention of re-
erence to P
searchers in solving wide variety of engineering problems. In so far
e, w, n, s East, West, North and South control faces
as the field of radiation heat transfer is considered, initial imple-
g Corresponding to gas properties
mentations for the control of a transient heating system was per-
h Heater surface properties
formed by several researchers [23,24]. Some other notable works
j jth gray band
include the use of ANN in obtaining a quick estimation of the
λ Spectral quantities
emissivity of the gas volume [25], which employed the high res-
P Nodal point of the control volume
olution molecular spectroscopic database. The application of such
s Corresponding to soot properties
a model was then realized in scenarios such as fire resistance test
furnaces [26]. Although, the use of ANN in the radiation heat trans-
and heuristic based search methods. Out of these, heuristic based fer area has started over a decade ago, its full strength has not
methods have appeared as vital choice of the researchers lately due been completely realized yet. Specifically, efforts towards expe-
to their robustness and versatility. Genetic algorithm (GA), in par- diting the estimation of radiation heat flux by the use of neural
ticular, has been used to solve a variety of inverse problems in heat network concepts in applications such as discrete heating has not
transfer [13]. These methods follow the evolution strategy, where been adequately addressed in literature.
the first set of solutions is considered as initial population and In view of this, in the present study, a novel strategy of solv-
based on the fitness of the individuals akin to fitness in genet- ing the inverse problem of estimation of discrete heaters number
ics the new population (or set of solutions) is generated by per- and locations is presented, where the actual RTE solver is replaced
forming mating of the potential individuals. In most cases, the GA with an artificial neural network based function for the forward es-
has been used in its traditional form. However, some of the recent timation of the radiative fluxes in the furnace. The fast estimation
applications to the cases of interests have used its other versions of the radiative fluxes using ANN would result in saving significant
namely micro-genetic algorithm [14–16] and hybrid genetic algo- computation time in such problems, where, invariably the forward
rithm [17,18], wherein the variation in choice of mating population model has to be run many times. However, the accuracy of ANN
and strategies is involve in the exchange of genetic information. based function depends upon the accuracy of the forward model.
As far as the present problem of obtaining the best heater For this purpose an efficient forward RTE model is first generated
power and location in a radiant heating furnace is considered, using Finite Volume Method (FVM) as RTE solver and Spectral line
the major complexity is the mathematical formulation for invert- based weighted sum of gray gases (SLW) method as the full spec-
ing the non-linear problem of radiation in a participating medium. trum band model. The output of this formulation is used as the
R. Yadav, C. Balaji and S.P. Venkateshan / Journal of Quantitative Spectroscopy & Radiative Transfer 226 (2019) 127–137 129

basis for generation of the ANN based function. The optimization The prescribed range of absorption cross sections is divided into n
problem is solved using the genetic algorithm, which requires re- discrete absorption cross section values or gray gases. The ALBDF
peated solutions to the forward model, which is facilitated by the for each of these can be estimated from above equation. The rep-
ANN model. resentative absorption coefficient can be evaluated as

2. Mathematical formulation κg, j = N C˜ j−1C˜ j (9)

2.1. RTE Solution: Finite volume method where N is gas molar density. For the mixture of gases, the individ-
ual ALBDFs must be evaluated for each participating gas and then
The directional spectral radiative transfer equation for an should be modified for the mixture. Several methods are listed by
absorbing-emitting medium in Cartesian coordinates, with a mix- Solovjov et al. [29] for generating the ALBDF for a mixture from
ture of gases as participating medium is given as: the ALBDF of the individual gases. In this work, a multiplication
approach has been used along with the newly updated correlations
∂ Iλ ∂ I ∂ I
μ + η λ + ζ λ = −κg,λ Iλ + κg,λ Ib,g,λ (1) provided by Pearson et al. [30]. The ALBDF generated for the mix-
∂x ∂y ∂z ture is then used to estimate the gray gas weights ag,j appearing in
In Eq. (1), μ, η and ζ are the direction cosines in the x, y and Eqs. (6) and (7) as
z directions respectively. Eq. (1) is an integro-differential equation
ag, j = Fa (C˜ j ) − Fa (C˜ j−1 ) (10)
in first order and hence requires only one boundary condition for
the solution which comes from the intensities emanating from the
2.3. SLW-Gray Approximation
wall

1 − w
Iλ (s ) = w Ib + Iλ |nw .
 |d
 (2) In many engineering applications, it is customary to define a
π n w
 < 0 single absorption coefficient for the whole spectral range. Such
where s represents position at the wall. Eq. (1) is integrated over models are called gray approximations. Their accuracy depends
the control volume and the control angle following the finite vol- upon the method of evaluation of effective absorption coefficient.
ume method and is discretized to be represented in terms of face Following the conventional formulation, the total emissivity for a
intensities and areas. A detailed description of the discretization is gas with partial pressure path length L at any given temperature
well documented in literature [27] and hence is not repeated here. based on the weighted sum of individual gray gas emissivities is
For a discretized set of Mθ × Nφ control angles, the directional in- given by
tensity in a direction mn, for the jth wavelength band at the node

NGG
of the control volume is written as = a j (T )(1 − e−κ j L ) (11)
amn mn
p I j,p = amn mn mn mn mn mn
E I j,E + aW I j,W + aN I j,N
j=0

+ amn mn mn mn mn mn
S I j,S + aT I j,T + aB I j,B + Smn
j,p (3) where aj is the weight of the jth gray gas. The term within the
brackets on right hand side gives the individual gray gas emissiv-
where the coefficients appearing in above equations are defined as
ities. The term j = 0 corresponds to the window in the spectrum
of the gas and is considered as a clear gas with κ = 0. Knowing
amn
I = max(−Ai Dmn
ci , 0 ) (4)
the gray gas weights and individual gray gas absorption coefficients
 through the SLW model, the total emissivity can be calculated from
amn
p = max(Ai Dci , 0 ) + (κg, j ) p V 
mn (5) the above equation with the help of knowledge of the mean beam
i=e,w,n,s length of the enclosure. The formulation for the representative ab-
In Eq. (4), amn is the coefficient of discretization at the faces of the sorption coefficient for a single gray gas is then given by
I
control volume with nodal point P. The source term appearing in 1
Eq. (3) can be written for gaseous emission as κ = − ln(1 −  ) (12)
L
j,p = ag, j (Tg )κg, j Ib,g
Smn (6) where L is the mean beam length of the enclosure, sometimes also
defined based on the path length of the corresponding control vol-
where ag, j is the weighting factor for the gas corresponding to its
ume. Though, there is no physical basis for this formulation [31],
temperature and jth gray band. In the present work, the evaluation
it certainly helps in determining a first cut approximation of the
of ag, j is performed using the spectral line based weighted sum
radiative fluxes at a substantially lower computational cost. Fur-
of gray gases (SLW) method which is explained in the next sub-
thermore, under the conditions of radiative equilibrium, this model
section.
performs better than the other gray models such as Gray-Planck
model and thus can be considered where computational load is
2.2. SLW Band model
one of the primary concerns [32].
Consider a non gray spectral RTE with gas emission-absorption
and no scattering 2.4. Artificial neural networks (ANN)

dI j Artificial neural networks have gained popularity over the last


= −κg, j I j + ag, j κg, j Ib (7)
ds few decades due to their ability to handle highly non linear prob-
where κ g, j and ag, j are the absorption coefficient and the weights lems with acceptable accuracy. A typical architecture of the net-
for the jth gray gas respectively. In the SLW model, an absorption work is illustrated in Fig. 1, where the inputs and outputs are
line blackbody distribution function (ALBDF) is generated for the separated by a hidden layer. There are connections between the
complete spectrum which gives the fractional blackbody emission input and the hidden layer and the hidden layer and the out-
from a set of wavenumbers which fall under a predefined absorp- put. Each connection carries a weight and a bias which are ad-
tion cross section range given as [28] justed at every iteration during the training phase until the solu-
 tion reaches acceptable accuracy with respect to the target values
1
Fa (C ) = E ( T )d η (8) or outputs, when known outputs are matched against the known
Eb (Tb ) η:Cη <C bη b
130 R. Yadav, C. Balaji and S.P. Venkateshan / Journal of Quantitative Spectroscopy & Radiative Transfer 226 (2019) 127–137

some time, under the conditions of equilibrium, the gas tempera-


ture adjusts itself to allow all the heat available from the heaters
to pass on to the design surface. The time taken to reach equilib-
rium depends upon the composition of the gas and design surface
properties. Following this, from a simple energy balance, the total
power supplied to the equally powered heaters in order to achieve
the total flux qd at the design surface can be determined as

i qd,i Ad,i
qh =  (13)
j Ah, j

where Ad,i is the area of the ith element on the design surface
and Ah,j is the area of the jth element on the heater surface.
Upon knowing the total power and design surface temperature, for
a particular configuration of heaters all the input variables for the
model are known and the estimation of flux values at the discrete
location at the design surface can be performed. However, it is to
be noted that different configurations of heaters would produce
different heat fluxes at the design surface. The challenge then is
to determine the best possible configuration of the heaters in or-
der to achieve a nearly uniform heat fluxes at all locations on the
Fig. 1. A typical architecture of an artificial neural network. design surface with very tight tolerance. In other words, the error
between the desired heat flux and the estimated heat flux for a
particular configuration should be minimum. Based on this, if the
inputs. These weights and bias are called the network parame- desired uniform heat flux at a location i on the design surface is
ters. Each neuron performs a mapping or transfer operation to get qd,i and the heat flux estimated by the solver is qe,i , the objective
the output from the input. Typical mapping functions include tan- function may be defined as
sigmoid, log-sigmoid and pure linear functions. The mapping starts  
with the initial guess of the weights and the bias and depend- 
F = min (qd,i − qe,i )2 (14)
ing upon the error in the estimation, the values of weights and
i
bias are readjusted to provide the best fitting of the input and
output data. The value of the weight of a connection determines Eq. (14) can be solved by an appropriate optimization algorithm
the contribution of the specific input in the mapping function, It like simulated annealing, genetic algorithm or conventional search
can be positive, negative or zero. Upon completion of the map- algorithms and since a closed form expression for qd does not ex-
ping, the accuracy of the network is checked and the values of ist, the RTE equation cannot be inverted directly to obtain the de-
weights and bias are readjusted. This is called network training sired heat flux on all the individual heaters. In the present study,
or learning and there exist several training algorithms to perform the standard genetic algorithm has been utilized. Genetic algo-
this function. Among these, back-propagation algorithms are very rithm is a heuristic based search method which works on the prin-
common and they consider the input, output and the participating ciple of evolution. An initial population of n individuals is chosen
neurons as distinct layers, which can feed their output to the adja- which is improved systematically over the generations using the
cent layers. The errors are back propagated to adjust the weights. principles of genetics and evolution. As the generations progress,
Some of the widely used feed forward back propagation algorithms the procedure results in better and better individuals. GA is a very
are ’Levenberg-Marquardt’ [33] and ’Bayesian-Regularization’ [34]. powerful tool for the estimation of optimum in some complex sit-
Although the Levenberg-Marquardt algorithm is fast, one primary uations. In the present study, an in-house code for the genetic al-
advantage of using Bayesian Regularization is that it choses only gorithm has been developed and has been coupled with the ANN
the effective number of parameters participating in the network no model.
matter how large the number of total parameters in the network
would be. In simple words, it allows one to train the network for a 3.1. ANN As a surrogate to RTE solver and its coupling with GA
large number of neurons and then selects only the significant pa-
rameters amongst them. This is very useful in situations where the For the purpose of developing a neural network function, a set
data set is limited. of input variables and corresponding outputs should be defined.
In the present problem, the outputs are the flux values at the de-
3. Description of the test problem: 3D radiant furnace sign surface. In order to monitor these flux values, five flux sensors
(or monitoring nodes) are considered at different locations on the
The problem under consideration is a three dimensional rect- design surface. A pictorial representation of these sensor locations
angular enclosure with a participating medium, similar to a typical along with the heater nomenclature is given in Fig. 3.
radiant heating furnace. The geometry of the enclosure is shown As far as the input variables are concerned, there are a total
in Fig. 2. The cubical shaped radiant heating furnace has a heater of 6 variables in the problem namely total heater power (Qh ), de-
surface and a design surface, which is to be heat treated. The di- sign surface temperature (Tdes ), design surface emissivity ( d ), gas
mensions of the furnace are 1 × 1 × 1 (all in m). There are a total compositions (YH2O and YCO2 ) and the configuration of the heaters,
16 heaters of equal sizes (0.25m × 0.25m) that are installed on the as shown in Table 1. To generate the different configurations of
heater surface. The left, right, back and the front walls are adia- the heaters, a binary bit 0 or 1 is used to define the state of the
batic. The design surface (at the bottom), which can be considered heaters, where 0 means heater is OFF and 1 means that the heater
as the part of the specimen undergoing heat treatment is 0.4m is ON, starting from the first heater as shown in Fig. 3. Following
long and 0.4m wide. The participating gas contains a mixture of this, based on individual state of the heaters, a 16 bit binary string
H2 O and CO2 in proportions that are variable. As the heaters dissi- is generated and the decimal equivalent of this binary string is
pate the heat into the furnace, the gas temperature changes. After used to represent the placement and configuration of the heaters.
R. Yadav, C. Balaji and S.P. Venkateshan / Journal of Quantitative Spectroscopy & Radiative Transfer 226 (2019) 127–137 131

Table 2
Sample configuration numbers and their binary equivalents signifying the heaters’
state numbered in Fig. 2.

Configuration number Binary equivalent Heaters’ state

3855 1111000011110000 1,2,3,4 - ON


5,6,7,8 - OFF
9,10,11,12- ON
13,14,15,16- OFF
61680 0000111100001111 1,2,3,4 - OFF
5,6,7,8 - ON
9,10,11,12 - OFF
13,14,15,16 - ON
52428 0011001100110011 1,2,5,6 - OFF
3,4,7,8 - ON
9,10,13,14 - OFF
11,12,15,16 - ON
13107 1100110011001100 1,2,5,6 - ON
3,4,7,8 - OFF
Fig. 2. Sample problem of discrete heating in a cubical enclosure.
9,10,13,14 - ON
11,12,15,16 - OFF

is that it should be sufficient enough to provide efficient train-


ing to the network and at the same time avoid over-fitting of the
network. The LHS technique divides the complete input variable
range into discrete intervals and assigns equal probability to each
of them. Upon doing so, one sample is selected from each inter-
val of all the variables to form the hypercube combination. In con-
trast with other sampling techniques, LHS provides more uniform
spread in the sample values over the complete range. After the
generation of these 200 sets of input variables, the forward model
Fig. 3. The cubical enclosure with discrete heaters at the top. RTE solver is executed and the net radiative heat flux at the sen-
sor locations is obtained and stored. This input-output dataset for
Table 1
200 cases is utilized to train and generate the artificial neural net-
Lower and upper bounds of the inputs for the ANN model.
work for the possible replacement of the RTE solver. In the ensuing
Variables Lower bound Upper bound Unit section, a detailed validation of the RTE solver with the available
Qh 10 0 0 50 0 0 W benchmark cases in literature is presented to demonstrate the ac-
Tdes 300 500 K curacy of the present formulation in terms of predicting radiative
d 0 1 - fluxes.
YH2O 0 0.25 -
YCO2 0 0.25 -
Configuration 1 65535 -

4. Verification of the accuracy of RTE solver

For the present case under consideration, there could be 216 pos- For the purpose of verification of the RTE solver developed, a
sible configurations of the heaters. Thus, on analysing, the lower case of non-gray radiation in a three dimensional rectangular en-
bound of the configuration variable would take a value of 1 where closure with a mixture of 20% H2 O and 10% CO2 on molar basis
only the first heater is ON and the upper bound would take up a has been considered. This case corresponds to the case solved by
value of 65535 in which case, all the heaters are ON. For a bet- Liu [37] with the use of the statistical narrow band (SNB) model
ter understanding, some sample configuration values and their bi- and Monte Carlo (MC) method. The results of this method are con-
nary equivalents are shown in the Table 2. From the Table, it can sidered to be the most accurate and hence are used as a bench-
be seen that a configuration value of 3855 would signify a heater mark by many researchers to validate their codes. The geometry
state such that first row heaters are ON, second row heaters are of the case is shown in Fig. 4. The dimensions of the furnace
OFF, third row heaters are ON and fourth row heaters are OFF. Sim- are 2 × 2 × 4 (all in m). The walls of the furnaces are all black at
ilarly, a configuration value of 61680 would take the heater states 300 K. The temperature of the gas varies in the z direction as,
as alternate OFF and ON rows. T = (Tc − Te ) f (r ) + Te , where f (r ) = 1 − 3(r/R )2 + 2(r/R )3 . Here, r
For reducing the computational time substantially, with a view is the distance of the point from the center axis and R is the ra-
to solving the problem of determination of the best heater con- dius of a hypothetical circular region around the center axis (with
figuration, which is essentially an inverse problem, a MATLAB R=1). Tc is the centerline temperature, which increases linearly
[35] based ANN implementation has been utilized with Bayesian from 400 K at z = 0 to 1800 K at z = 0.375, and then decreases
Regularization as the training algorithm. A two layer (one hidden to Tc = Te = 800K at z = 4. The temperature outside of this circu-
layer and one output layer) feed forward network is used with tan- lar region is taken to be constant at Te . The radiative heating to
sigmoid and pure linear as the transfer functions in hidden and out- the wall of the furnace at the location (2m, 1m, z) has been esti-
put layer respectively. For getting sufficient amount of data to train mated by the FVM-SLW method (using 7 and 10 gray gases) with
the network, a total of 200 cases have been generated by varying two grid sizes namely 12 × 12 × 24 and 24 × 24 × 48. Angular dis-
the input variables within the prescribed range (see Table 1) us- cretization is taken as 12 × 20. The values of the incident radiative
ing the Latin Hypercube Sampling (LHS) [36]. The choice of num- heat flux against the wall location are shown in Fig. 4. From the
ber of training points is random. Even so, the basic requirement figure, it is clear that the results of the present FVM-SLW formula-
132 R. Yadav, C. Balaji and S.P. Venkateshan / Journal of Quantitative Spectroscopy & Radiative Transfer 226 (2019) 127–137

Table 3
Network parameters for the training data.

No of neurons MSE R

2 2.3534E-03 0.9786
4 7.7770E-06 0.9999
6 5.8774E-05 0.9995
8 2.9802E-06 0.9999
10 2.0478E-06 0.9999
12 3.3779E-06 0.9999
14 1.6602E-05 0.9998

Table 4
Network parameters for the testing data.

No of neurons MSE R

2 4.4126E-03 0.9351
4 2.5245E-03 0.9761
6 3.6612E-04 0.9967
8 3.6700E-03 0.9752
10 8.8174E-04 0.9913
12 7.0091E-04 0.9956
14 3.0616E-03 0.9772

where, SSE and SSME are sum of squared errors and sum of
squared mean errors respectively. They are calculated as


N
SSE = (qi − qt,i )2
Fig. 4. Variation of incident radiative heat flux at the wall (2, 1, z), with respect to
i=1
location- Comparison of the present method with SNB-MC ([37]).

N
SSME = (q̄i − qt,i )2 (16)
i=1
tion are in close agreement with those obtained using the SNB-MC
method. Here N is the total number of input output dataset. qi is the
actual output value and qt,i is the output value obtained by the
trained network.
The results obtained during the training and testing of the net-
5. Results and discussion works with different number of neurons are presented in Tables 3
and 4 respectively. A single hidden layer has been chosen, and
For the problem under consideration, the output in terms of calculations with multiple hidden layers showed no significant
radiative fluxes at the design furnace for 200 different combina- improvement in the results. Based on the results presented in
tions of the input variables (generated using LHS) is determined Tables 3 and 4, it can be seen that the network with 6 and 12
using the FVM-SLW code developed in this work. The participat- number of neurons provides the best fit and the lowest errors.
ing medium is considered to be in radiative equilibrium with the However, when tested on the external 10% dataset which were not
relevant properties being calculated from the SLW-Gray approxi- introduced to the network, the network with 6 and 12 number of
mation. A spatial grid of 20 × 20 × 20 and angular grid of 8 × 12 neurons provided a mean square error value of 2.2106 × 10−3 and
has been considered. Further refinement in the grid size did not 6.5206 × 10−4 respectively. Thus, the network with 12 number of
change the results and the difference in the prediction of mean neurons has been chosen to solve the inverse problem.
heat flux was less than 0.2% in both the cases. Before introducing
the database to the ANN, the 200 sets of input-output data that 5.1. Comparison of ANN prediction with RTE solution
are stored, are divided in to two parts (i) A set with 90% of the
data to be input to the neural network algorithm in MATLAB after For the purpose of comparison of the ANN predictions with the
normalizing and (ii) the remaining data (10%) for performing the RTE solution, 5 additional test cases have been generated which
additional validation of the network. were not a part of the training or validation but lie in the specified
To make the data consistent for an efficient fitting of the net- input range (Table 1). A description of the test cases is provided in
work the heat input and design temperature are normalized as the Table 5. For these cases, the flux values at the sensor locations
Qh = Qh /Qh,max , Tdes
 = T /T
des des,max , while the local output flux is have been obtained by both the actual RTE solver and the ANN.
normalized as qd = qd /qd,max , where qmax = Qh,max /Ad , Ad is The differences in the estimation of local flux values by both these
the area of design surface. The effectiveness of the network after methods in terms of relative error are plotted in Fig. 5. From the
the training is tested with the following performance parameters figure, it can be seen that the estimation of the local heat fluxes
(i) Mean Squared Error (MSE) and (ii) goodness of fit (R). These by the ANN and the RTE show a good agreement with maximum
are given by: errors falling under 4%. The efficiency of the network in terms of
the computational time is assessed for the same five cases. The
1
N computational time taken by these cases when solved by the RTE
MSE = (qi − qt,i )2 and ANN on a Dell T1650 workstation with 3.4 GHz processor and
N
i=1 10 GB memory is plotted in the Fig. 6. The figure shows that the
2 SSE present ANN model is 60 0 0 times faster (average of the 5 cases)
R =1− (15)
SSME than the actual RTE solver and hence can be employed for prob-
R. Yadav, C. Balaji and S.P. Venkateshan / Journal of Quantitative Spectroscopy & Radiative Transfer 226 (2019) 127–137 133

Fig. 5. Comparison of the results obtained by actual forward model (RTE) and the Fig. 6. Comparison of the computational time taken by the actual forward model
Neural network based model (ANN) for five different cases mentioned in Table 5. (RTE) and the Neural network based model (ANN) for the solution five for different
cases mentioned in Table 5.
Table 5
Test cases to assess the accuracy of the ANN model in 3D rectangular geome- Table 6
try. Design case considered for performing optimization of the heater configuration
in the three dimensional rectangular geometry.
Case Qh (W) Tdes (K) d YH2O YCO2 Configuration
Qh (W) Tdes (K) d YH2O YCO2 Configuration
Test case 1 3800 320 0.1 0.025 0.025 32768
Test case 2 30 0 0 360 0.9 0.225 0.225 45875 Design case 20 0 0 400 0.4 0.2 0.1 to be optimized
Test case 3 2200 480 0.3 0.175 0.075 58982
Test case 4 4600 440 0.7 0.125 0.175 19661
Test case 5 1400 400 0.5 0.075 0.125 6554 Table 7
Description of various cases obtained by varying the
individual parameters one at a time while keeping
the values of the other parameters equal to design
lems involving multiple estimations, under uniform boundary con- case values. Note: Case 2 resembles the design case.
ditions such as the problem of optimization of discrete heaters. Case Qh (W) Tdes (K) d YH2O YCO2

1 20 0 0 300 0.4 0.2 0.1


5.2. Optimization of heater state and locations using genetic 2 20 0 0 400 0.4 0.2 0.1
algorithm (GA) 3 20 0 0 500 0.4 0.2 0.1
4 20 0 0 400 0.2 0.2 0.1
As shown in Fig. 3, a total of 16 heaters mounted at the top 5 20 0 0 400 0.6 0.2 0.1
6 20 0 0 400 0.8 0.2 0.1
surface are numbered 1 to 16 starting from the right side wall. 7 20 0 0 400 1 0.2 0.1
The design surface of the specimen lying at the bottom is to be 8 20 0 0 400 0.4 0.1 0.1
heat treated. The objective of the problem now is to determine the 9 20 0 0 400 0.4 0.2 0.05
best possible placement and configuration of the heaters in order 10 20 0 0 400 0.4 0.1 0.05
to achieve a uniform desired value of heat flux at all five locations
on the design surface. The problem is relevant in many industries
since prior knowledge of correct number, placement and configu- sensor locations. The objective function of the problem has already
ration of the heaters helps to improve the cost and efficiency of the been derived in Eq. (14).
process. To test the applicability of the ANN based model coupled The optimization problem is solved using the standard Genetic
with GA in these problems, the problem is set up as mentioned Algorithm. Initially, a set of 80 individuals are chosen for the first
below. A uniform heat flux of 12.5 kW/m2 is desired at all the generation. Any further increase in number of individuals did not
locations on the design surface. Since the medium is in radiative show any impact on the solution. The choice of individuals is made
equilibrium and all other walls are adiabatic, a simple energy bal- on the basis of uniformly equi-spaced intervals in the feasible do-
ance, as given by Eq. (13), can be used to determine the effective main for all the variables. However, a comparison with individual
power to be supplied. However, the uniformity in the flux values choices using the random selection has also been made and with
would depend upon the configuration of the heaters. This config- respect to the two strategies, the change in the value of mean
uration in turn may change with respect to change in other vari- fitness of the population was always found to be less than 0.1%.
ables of the problem i.e. design surface temperature, design surface Upon generation of the first set of individuals, the ANN is used
emissivity and gas compositions. To assess this, the analysis can be to predict the fluxes qe, i at the sensor locations. The solution is
started with a sample design case, the details of which are men- compared with the desired heat flux and the value of the objec-
tioned in Table 6. For this case and with an assumed configuration tive function is estimated. The individuals are sorted in the order
of heaters, all the input variables of the problem are defined and of best fitness and are allowed to perform cross-over, mutation to
the ANN prediction can be used to estimate the heat fluxes at the produce the new generation. A certain fraction of these individu-
134 R. Yadav, C. Balaji and S.P. Venkateshan / Journal of Quantitative Spectroscopy & Radiative Transfer 226 (2019) 127–137

Fig. 8. Optimized configuration for the design case, showing the values of the input
parameters and a depiction of the heater configuration.

variables in the problem (see Table 1) has been varied one at a


time and the corresponding optimum configurations for each case
have been generated. These cases are mentioned as cases 1 to 10
in the order of variation of design surface temperature (from 300
to 500 K, Case 1 to Case 3), emissivity of the design surface (from
0.2 to 1, Case 4 to Case 7) and the gas compositions (from 0 to
0.25, Case 8 to Case 10). A depiction of these cases is presented
in Table 7. Table 8 shows the results of the ANN-GA algorithm in
terms of the optimized configurations for the above mentioned 10
Fig. 7. Sample flow chart for the evaluation of optimum through Genetic Algorithm.
cases and corresponding values of objective function along with
als (0.05) which give best solutions are chosen as elite population values of heat flux to the heaters. Figs. 9–11 show the values of
and are directly moved to the next generation. A fraction of pop- the local heat fluxes with the optimum configuration for the above
ulation (0.7) undergoes uniform bit cross over and new individu- mentioned cases, as against the desired heat flux. The actual de-
als are generated. A very small fraction of them (0.04) undergoes sired flux should be a constant of 12.5 kW/m2 . The values obtained
mutation. These fractions have been varied in their relevant range are remarkably close to the design value with the maximum devi-
and the results obtained did not change with respect to the choice ation being 0.06 kW/m2 , on a baseline of 12.5 kW/m2 , showing a
of cross-over, mutation and elitism fraction. The new generation very good match with the desired heat flux for all the cases.
is then evaluated and tested based on the value of the objective
function it provides. The procedure is repeated till the change in 5.4. Indirect validation of optima with forward model calculations
mean fitness value of the population in the subsequent generation
is less than the tolerance (10−8 ). The best individual in the final The optimum configurations have been input in the forward
population gives the minimum value of the objective function and model and the relative error between the estimation of radiative
the corresponding optimal configuration. It is instructive to men- fluxes through ANN and RTE solution is calculated for all the above
tion that all individuals in the final population show a low value mentioned cases. The maximum relative error for each of the cases
of the objective function close to the global minimum as it should has been given in Table 9. From the table, it is seen that the pre-
be intuitively apparent with a method that tries to mimic evolu- dictions of the heat flux value through ANN are very close to the
tion. The complete solution algorithm of ANN coupled with GA for RTE, with the maximum errors observed to be less than 1% for all
the problem under consideration is shown in Fig. 7. cases.

5.3. Optimal configuration for design case 5.5. Sensitivity study of the optimum configurations

The above methodology in implemented on the problem under Despite its potential to yield accurate solutions, Genetic algo-
consideration and the optimal heater configurations for the design rithm is essentially a search technique which means the true op-
case (Table 6) are generated using genetic algorithm. The optimum timum can never be reached but only be approached. In order to
configuration of heaters (with red arrow marks showing that those gain confidence over the solutions obtained through genetic algo-
heaters should be ON) and the corresponding value of the objec- rithm procedure, an exhaustive search is performed only for one
tive function are shown in Fig. 8. Following this, each one of the case namely the design case (Table 6). As mentioned, a total 216
R. Yadav, C. Balaji and S.P. Venkateshan / Journal of Quantitative Spectroscopy & Radiative Transfer 226 (2019) 127–137 135

Table 8
Optimized configurations, heat input to the heaters and value of objective function obtained using ANN-GA for various cases given in Table 7.

Case Optimized configuration (index of heaters that are switched ON) Value of objective function (F) Heat input to the heaters (qh in kW/m2 )

1 1,4,5,6,7,8,9,10,11,12,14,15 0.0125 2.67


2 2,3,4,5,7,10,11,16 0.0013 4.00
3 2,8,10,11 0.0 0 01 8.00
4 1,4,5,7,10,11,13,15,16 0.0 0 03 3.56
5 11,13,14 0.0010 10.67
6 2,3,5,8,9,12,13,14 0.0016 4.00
7 8,10,11,13,14 0.0025 6.40
8 1,3,4,5,6,8,11,12,13,14,15,16 0.0011 2.67
9 5,8,9,12,13,16 0.0018 5.33
10 1,2,3,4,6,9,10,15 0.0014 4.00

Fig. 11. Value of the net radiative heat fluxes at the design surface for the opti-
Fig. 9. Value of the net radiative heat fluxes at the design surface for the optimized mized configuration for various gas compositions (Baseline values are x = YH2O =
configuration with varying design surface temperatures. 0.2 and y = YCO2 = 0.1.).

Table 9
Maximum relative error in the predictions of
local radiative heat flux between the ANN and
the RTE for the optimized configurations.

Maximum relative error


Case (| RT E−ANN
RT E
× 100|)

1 0.4363
2 0.2194
3 0.4044
4 0.4536
5 0.8457
6 0.1848
7 0.8148
8 0.1215
9 0.1681
10 0.2364

possible configurations of the heaters exist for this problem. On


performing an exhaustive search, it turns out that due to the na-
ture of the problem, around 1% of all the possible configurations
are workable solutions with the similar objective function values.
One of these workable solutions is what GA returns and this con-
firms the credibility of GA. Although the configurations obtained
are not symmetric about a vertical plane passing through the cen-
Fig. 10. Value of the net radiative heat fluxes at the design surface for the opti- ter of the furnace, the solutions are still workable since they give
mized configuration for different design surface emissivities. the minimum value of the objective function possible with the
136 R. Yadav, C. Balaji and S.P. Venkateshan / Journal of Quantitative Spectroscopy & Radiative Transfer 226 (2019) 127–137

Table 10
Results with normal random perturbation of inputs of the design case.

Case Qh (W) Tdes (K) d YH2O YCO2 config F

Design case 20 0 0 400 0.4 0.2 0.1 34398 0.00126


1 2007.75 395.82 0.3991 0.2046 0.1054 34398 0.01442
2 20 0 0.20 403.61 0.4107 0.1967 0.0898 34398 0.00134
3 2045.25 403.28 0.3939 0.1866 0.1111 34398 0.41081
4 1990.15 396.24 0.4082 0.2128 0.0973 34398 0.01824
5 2067.00 401.30 0.4081 0.1867 0.0927 34398 0.89045

present set of input data. Furthermore, to ensure that the GA has [8] Daun KJ, Howell JR, Morton D. Optimization of transient heater settings to pro-
reached the local minimum of the solution space, the input vari- vide spatially uniform heating in manufacturing processes involving radiant
heating. Numer Heat Transfer, Part A 2004;46(7):651–67.
ables of the design case (Table 6) are perturbed and the corre- [9] França FH, Ezekoye OA, Howell JR. Inverse boundary design combining radia-
sponding value of objective function has been obtained by the tion and convection heat transfer. J Heat Transfer 2001;123(5):884–91.
ANN. Table 10 shows the results of 1% random perturbation as- [10] Mossi A, Vielmo H, França F, Howell JR. Inverse design involving combined
radiative and turbulent convective heat transfer. Int J Heat Mass Transf
suming a normal distribution around the actual input values for 2008;51(11–12):3217–26.
the design case. From the Table, it is seen that even a small change [11] Moghadassian B, Kowsary F. Inverse boundary design problem of natural con-
in the input variable causes large changes in the value of objective vection–radiation in a square enclosure. Int J Therm Sci 2014;75:116–26.
[12] Daun KJ, Howell JR. Inverse design methods for radiative transfer systems. J
function, and hence it can be concluded that optimized configura- Quant Spectrosc Radiat Transfer 2005;93(1–3):43–60.
tions are sensitive to the changes in the input variables and for the [13] Gosselin L, Tye-Gingras M, Mathieu-Potvin F. Review of utilization of
defined set of input variables, the optimized configuration obtained genetic algorithms in heat transfer problems. Int J Heat Mass Transf
2009;52(9–10):2169–88.
through GA gives the minimum value of the objective function and
[14] Safavinejad A, Maruyama S, Mansouri SH, Sakurai A. Optimal boundary de-
hence a nearly uniform heat flux at the design surface. sign of radiant enclosures using micro-genetic algorithm. J Therm Sci Technol
2008;3(2):179–94.
[15] Safavinejad A, Mansouri S, Sakurai A, Maruyama S. Optimal number and loca-
6. Conclusions tion of heaters in 2-D radiant enclosures composed of specular and diffuse sur-
faces using micro-genetic algorithm. Appl Therm Eng 2009;29(5–6):1075–85.
The problem of inverse estimation of the best configuration of [16] Amiri H, Mansouri SH, Safavinejad A, Coelho PJ. The optimal number and
location of discrete radiant heaters in enclosures with the participating
discrete heaters in a radiant furnace was solved by posing it as
media using the micro genetic algorithm. Numer Heat Transfer, Part A
an optimization problem. In view of the large computational re- 2011;60(5):461–83.
sources required to run the full RTE model to drive the optimiza- [17] Kim KW, Baek SW, Kim MY, Ryou HS. Estimation of emissivities in a two-di-
mensional irregular geometry by inverse radiation analysis using hybrid ge-
tion problem, an ANN based surrogate of the RTE solver was devel-
netic algorithm. J Quant Spectrosc Radiat Transfer 2004;87(1):1–14.
oped using the feed forward back propagation method. This model [18] Kim KW, Baek SW. Inverse surface radiation analysis in an axisymmetric cylin-
was used as a surrogate in the optimization algorithm in the place drical enclosure using a hybrid genetic algorithm. Numer Heat Transfer, Part A
of actual RTE solver for the prediction of heat fluxes at the design 2004;46(4):367–81.
[19] Sarvari SH. Optimum placement of heaters in a radiant furnace using the ge-
surface. The optimized configurations for a test problem were gen- netic algorithm. In: International Heat Transfer Conference 13. Begel House
erated using the genetic algorithm and the change in the optimum Inc.; 2006.
configurations with respect to change in design variables was also [20] Ashouri J, Sarvari SH, Farahat S. Optimum arrangement of heaters in a three-
-dimensional radiant furnace using the genetic algorithm. In: 7th International
assessed. With the ANN based model, the computational time was Conference on Heat Transfer, Fluid Mechanics and Thermodynamics; 2010.
found to reduce by the order of 103 s. The relative error between [21] Lemos LD, Brittes R, França FH. Application of inverse analysis to determine the
ANN and RTE predictions of heat flux was found to be less than geometric configuration of filament heaters for uniform heating. Int J Therm
Sci 2016;105:1–12.
4%. The validation of the optimum configurations for the design [22] Amiri H, Coelho PJ. A parametric study of optimal number and location of ra-
case was also performed by making 1% random normal perturba- diant heaters in enclosures with participating media. Scientia Iranica Trans B,
tion and the objective function values for such perturbations were Mech Eng 2016;23(1):206–17.
[23] Ertürk H, Ezekoye OA, Howell JR. Inverse radiation heat transfer within enclo-
found to be more than the value at the design case for the same
sures with nonisothermal participating media. In: International Heat Transfer
configuration, thereby confirming the adequacy of the optimization Conference, 8; 1998. p. 729–34.
strategy proposed in this study. It was observed that the ANN cou- [24] Howell JR, Daun KJ, Erturk H, Gamba M, Sarvari MH. The use of inverse meth-
ods for the design and control of radiant sources. JSME Int J Ser B Fluids Ther
pled with GA is able to solve the inverse problem of design of the
Eng 2003;46(4):470–8.
radiant heaters with high accuracy and robustness. [25] Yuen WW. RAD-NNET, A neural network based correlation developed for a re-
alistic simulation of the non-gray radiative heat transfer effect in three-dimen-
References sional gas-particle mixtures. Int J Heat Mass Transf 2009;52(13–14):3159–68.
[26] Yuen WW, Tam W, Chow W. Assessment of radiative heat transfer char-
acteristics of a combustion mixture in a three-dimensional enclosure using
[1] Howell JR, Ezekoye O, Morales J. Inverse design model for radiative heat trans-
RAD-NETT (with application to a fire resistance test furnace). Int J Heat Mass
fer. J Heat Transfer 20 0 0;122(3):492–502.
Transf 2014;68:383–90.
[2] Franca F, Morales JC, Oguma M, Howell JR. Inverse radiation heat transfer
[27] Chai JC, Lee HS, Patankar SV. Finite volume method for radiation heat transfer.
within enclosures with nonisothermal participating media. In: International
J Thermophys Heat Transfer 1994;8(3):419–25.
Heat Transfer Conference, 7; 1998. p. 433–8.
[28] Denison MK, Webb BW. An absorption-line blackbody distribution function for
[3] Sarvari SH, Mansouri S, Howell JR. Inverse boundary design radiation prob-
efficient calculation of total gas radiative transfer. J Quant Spectrosc Radiat
lem in absorbing-emitting media with irregular geometry. Numer Heat Trans-
Transfer 1993;50(5):499–510.
fer Part A 2003;43(6):565–84.
[29] Solovjov VP, Webb BW. SLW Modeling of radiative transfer in multicomponent
[4] Sarvari SH, Mansouri S, Howell JR. Inverse design of three-dimensional enclo-
gas mixtures. J Quant Spectrosc Radiat Transfer 20 0 0;65(4):655–72.
sures with transparent and absorbing-emitting medial using an optimization
[30] Pearson JT, Webb BW, Solovjov VP, Ma J. Efficient representation of the ab-
technique. Int Commun Heat Mass Transfer 2003;30(2):149–62.
sorption line blackbody distribution function for H2 O, CO2 , and CO at variable
[5] Sarvari SH. Inverse determination of heat source distribution in conductive–ra-
temperature, mole fraction, and total pressure. J Quant Spectrosc Radiat Trans-
diative media with irregular geometry. J Quant Spectrosc Radiat Transfer
fer 2014;138:82–96.
2005;93(1–3):383–95.
[31] Coelho PJ. Numerical simulation of radiative heat transfer from non-gray
[6] Daun KJ, Morton D, Howell JR. Geometric optimization of radiant enclosures
gases in three-dimensional enclosures. J Quant Spectrosc Radiat Transfer
containing specular surfaces. J Heat Transfer 2003;125(5):845–51.
2002;74(3):307–28.
[7] Ertürk H, Ezekoye OA, Howell JR. The application of an inverse formulation in
[32] Amiri H, Lari K. Comparison of global radiative models in two-dimensional en-
the design of boundary conditions for transient radiating enclosures. J Heat
closures at radiative equilibrium. Int J Therm Sci 2016;104:423–36.
Transfer 2002;124(6):1095–102.
R. Yadav, C. Balaji and S.P. Venkateshan / Journal of Quantitative Spectroscopy & Radiative Transfer 226 (2019) 127–137 137

[33] Levenberg K. A method for the solution of certain non-linear problems in least [36] McKay MD, Beckman RJ, Conover WJ. Comparison of three methods for select-
squares. Q top Q Appl Math 1944;2(2):164–8. ing values of input variables in the analysis of output from a computer code.
[34] MacKay DJ. Bayesian interpolation. Neural Comput 1992;4(3):415–47. Technometrics 1979;21(2):239–45.
[35] MATLAB. Version 8.3.0 (R2014a). Natick, Massachusetts: The MathWorks Inc.; [37] Liu F. Numerical solutions of three-dimensional non-grey gas radiative transfer
2014. using the statistical narrow-band model. J Heat Transfer 1999;121(1):200–3.

You might also like