Artificial Neural Networks (ANNs) have a wide application in the prediction
of the rate of adsorption of an adsorbent in solid-liquid system. The ANN is
an advanced mathematical tool which is used to determine the network output
based on the available experimental information. The neural networks are made
up of simple elements i.e., input layer, hidden layer and output layer, operating
in parallel which are inspired by biological nervous systems. A typical neural
network can be trained by adjusting the values of weights and bias between elements
(Bhattacharya et al., 2008; Sato
et al., 1999). The main purpose of the training was to make the experimental
output reach as close as to the desired target output for a particular input.
Till now various researchers used the ANN for predicting the performance of
various adsorption systems successfully (Du et al.,
2007). Among all these adsorption process the adsorption of dye stuffs from
different industrial effluents has become a major environmental concern (Dekhil
et al., 2011; Rajeswari et al., 2011).
Most of the dyes are non biodegradable in nature and produce toxic amines over
incomplete degradation which has an adverse effect on the aquatic life (Rajendran
et al., 2011). These colored substances also retard the rate of photosynthesis
by preventing the sunlight penetration into the water bodies (Wanchanthuek
and Thapol, 2011; Arunachalam and Annadurai, 2011).
Therefore, removal of these substances from the industrial wastewater is utmost
important (Wanchanthuek and Nunrung, 2011; Dutta
and Basu, 2012). In most of the cases, the adsorption of dyes is carried
out by using commercial activated carbon (Eckenfelder, 1989;
Dutta and Basu, 2011). However, its higher cost limits
its widespread uses. So far limited applications has been found on the removal
of direct dyes from industrial waste water by using microwave assisted activated
Therefore, in the present investigation the adsorption capacity of microwave assisted activated carbon had been tested against a typical textile dye direct blue-86 (DB 86) through artificial neural network. For this purpose, a multilayer feed-forward neural network model was constructed which was found to predict the decolorization efficiency of DB 86 successfully through a non linear regression analysis.
MATERIALS AND METHODS
Materials: The scrap wood of Acacia auriculiformis was collected from local saw mill of Kharagpur. Direct blue-86 (DB-86), hydrochloric acid, sulphuric acid and ammonia were procured from Merck Specialities Private Limited, Mumbai, India.
Preparation of the activated carbon: The wood scraps were first cut into small pieces of 2 mm width and 40 mm of length, cleaned with distilled water and was sun dried for 24 h prior to the carbonization process. The wood pieces were kept on a ceramic boat which was placed at the center of a 40 mm i.d tubular furnace. The material was then heated from ambient temperature to the carbonization temperature of 750°C at the rate of 4°C min-1 in a continuous flow of N2 (300 mL min-1) and then it was kept at this temperature for 1 h for subsequent activation. The product was then allowed to cool to ambient temperature in presence of N2 flow and sieved to obtain the desired size fractions. It was further stored in a desiccator over silica gel. The char prepared in this way is termed as C750N.
The C750N was further activated in a domestic microwave oven (IF20PG3S) for 5 min at a constant input power of 800 W and a frequency of 2450 MHz. After treating in the microwave oven C750N was termed as AC750NMW5.
Characterization of activated carbon: The surface area and the total pore volume of the prepared char and the activated carbons were determined by using N2 adsorption-desorption method by using Brunauer Emmett Teller (BET) apparatus (Autosorb-1, Quantacrome). The surface properties and the microporous structure were investigated by using Scanning Electron Microscope (SEM) (Hitachi, model SU-70).
Batch adsorption study: The effect of contact time and initial concentration on the amount of dye adsorbed were investigated at an initial dye concentration of 40-100 mg L-1 whereas, the contact time was varied from 10-140 min. The pH effect was studies over a pH range of 1.5-9.4. The adsorbent dosage was varied from 1-3 g L-1 in order to investigate the effect of adsorbent dosage on the instantaneous uptake of DB 86. The kinetic study was carried out by analyzing adsorptive uptake of the dye from aqueous solution at different time intervals. The samples were centrifuged and the concentrations were analyzed in a UV-Vis spectrophotometer.
Equilibrium and kinetic study of DB-86: The monolayer adsorption capacity of the prepared activated carbon was determined against direct blue-86 (DB-86) by the equilibrium study.
The equilibrium study was carried out by adding 0.1 g of adsorbent into a series of 250 mL conical flask containing 100 mL solution of direct blue 86 and was shaken for 36 h at room temperature. After this the samples were centrifuged and the concentrations were analyzed in a UV-Vis spectrophotometer. The equilibrium adsorption capacity was calculated from the relationship:
where, qe (mg g-1) is the equilibrium adsorption capacity,
Ce is the dye concentration at equilibrium (mg L-1), V
is the volume of solution (l) and w is the weight of adsorbent (g). In order
to establish an appropriate relationship between the adsorption capacity (qe)
and dye concentration (Ce) at equilibrium by using AC750NMW5, the
data were fitted to Freundlich, Langmuir and Temkin isotherms. The Langmuir
isotherm represents the uni-molecular adsorption of the adsorbate molecule on
the adsorbent surface (Song et al., 2010). The
model can be expressed as:
where, Kl is the Langmuir constant related to the energy of adsorption (L mg-1) and Q0 is the maximum amount of adsorption corresponding to complete the monolayer coverage on surface (mg g-1).
Similarly, the Freundlich isotherm can be used for non-ideal sorption that
involves heterogeneous surface energy systems (Ozturk and
Kavak, 2005) and is expressed by the following equation:
where, KF is a rough indicator of the adsorption capacity and 1/n
is the adsorption intensity. In general, with the increase of the adsorption
capacity KF value increases. The Temkin isotherm describes the heat
of adsorption and interaction between adsorbent-adsorbate molecules (Ergene
et al., 2009). The Temkin isotherm can be expressed as:
where, kT and B1 are Temkin isotherm constant. The equilibrium
plot of direct blue 86 is shown in Fig. 1 and values of different
adsorption isotherm constants are also shown in Table 1. It
was seen from Fig. 1 that the equilibrium behavior of Db 86
was successfully predicted by Langmuir adsorption isotherm. Besides, it was
observed from Table 1 that the Langmuir adsorption isotherm
gave the most satisfactory result for DB 86 adsorption with a higher correlation
coefficient value (R2) of 0.98. The monolayer adsorption capacity
was found to be 500.16 mg g-1 for an initial dye concentration of
500 mg L-1. The chi-square value (χ2) for different
isotherms are also shown in Table 1. The chi-square value
is a test statistic that is calculated as the sum of the squares of observed
values minus expected values divided by the expected values. Lower value of
chi-square signifies more accurate system. It was observed from Table
1 that Langmuir isotherm had the lowest value of chi-square which signified
a better fit.
The artificial neural networks: ANN is an advanced mathematical modeling
procedure inspired by biological neuron system. In the present investigation,
Neural Network Toolbox V5.0 of MATLAB R2006a mathematical software was used.
A three-layer (one input layer, one hidden layer and one output layer) feed
forward ANN network with tangent sigmoid transfer in the hidden layer linear
function in the output layer was constructed for predicting the adsorption capacity
of the prepared adsorbent. A typical ANN is shown in Fig. 2
where various input variables namely initial solution pH, initial adsorbate
concentration, adsorbent concentration, reaction time and reaction temperature
were fed to the input layer whereas, the dye adsorption capacity qt
(mg g-1) was considered as the output of the respective network.
The ranges of different input variables are given in Table 2.
Altogether 166 experimental data points were fed to the ANN network. The data
points were divided for training, test and validation purpose in a ratio of
|| The equilibrium curve for DB 86
|| Equilibrium model parameters
|| The ranges of different input variables used for experimental
|| Schematic diagram of the three layers ANN network with three
input and one output layer
The training of the ANN network depends on the number of neurons in the hidden
layer. The performance of the network can be developed by increasing the number
of neurons in the hidden layer but too many neurons may affect the performance
conversely in the over fitting (Kumar and Porkodi, 2009).
RESULTS AND DISCUSSION
Comparison of surface properties of activated carbons: The BET surface area (SBET), total pore volume (Vtot) and average pore size of C750N and AC750NMW5 were determined from the physical adsorption data of N2 at 77 K and the values are shown in Table 3.
It can be seen from Table 3 that AC750NMW5 has the highest
surface area and total pore volume. Inside the microwave oven a high temperature
could be reached in comparatively shorter period of time resulting dissipation
of huge amount of energy at a molecular level. Consequently, the roughness of
the pore walls may also be increased due to rapid heating with the formation
of additional active sites (Huang et al., 2011).
Besides, rapid heating could accelerate the release of tar or volatile matter
from the pore interior which results into higher pore volume (Liu
et al., 2010).
Surface morphology of activated carbons: The surface morphology of activated
carbons was investigated through Scanning Electron Microscope (SEM) analysis.
The SEM images of various activated carbons are shown in Fig.
3a and b. It was observed from Fig. 3a
that a typical honeycomb structure with pores of different size was formed on
the surface of char (C750N) when it was treated at optimum condition. The similar
surface morphology could also be observed when activated carbon was prepared
from corncob by chemical activation (Tseng, 2006). The
roughness of the interior pores of activated carbon was increased when it was
further treated in a domestic microwave oven (Fig. 3b).
Optimization by using artificial neural network (ANN): In the present
work Levenberg-Marquardt back-propagation was implied to predict the rate of
DB 86 adsorption.
|| The SEM image of (a) C750N and (b) AC750NMW5
|| Comparison of surface properties of different adsorbents
|| The Levenberg-Marquardt backpropagation of various algorithm
|NHL: Neurons in the hidden layer, LR: Learning rate
Trial and error method was followed to find out the optimum number of neurons
in the hidden layer. It had been found that 8-12 neurons produce minimum Mean
Square Error (MSE).
Therefore, 8-12 neurons were considered for this study. The Levenberg-Marquardt
back-propagation algorithm along with its corresponding correlation coefficient
values (R2) is shown in Table 4. The number of
iterations required to find out the optimum number of neurons was also mentioned.
The simulation result of Levenberg-Marquardt back-propagation algorithm
is shown Fig. 4. It was seen from Fig. 4
that the network was successfully trained by using Levenberg-Marquardt
back-propagation algorithm. The similar transfer function was also used
by Oguz and Ersoy (2010) where removal of Cu2+
was carried out in a fixed bed column.
Prediction of the adsorption capacity through ANN: The actual and ANN
model predicted values for adsorption capacity were compared through Levenberg-Marquardt
back-propagation algorithm (Fig. 5). It was found that
both of these values were in reasonable agreement with each other which is in
commensurate with the result obtained by Aber et al.
(2007). The values were compared for training, test and validation purpose.
The values of corresponding correlation coefficient were also provided in the
||Simulation curve for Levenberg-Marquardt backpropagation algorithm
(best validation performance is 0.0085922 at epoch 24)
||Comparison of experimental and model predicted values of adsorption
capacity by Levenberg-Marquardt back-propagation algorithm
In the present scenario ANN was successfully implemented for predicting the
adsorption capacity in a solid-liquid adsorption system. The value of MSE was
also calculated for Levenberg-Marquardt algorithm. It was observed that 8 numbers
of neurons in the hidden layer was required for the above algorithm. A reasonable
agreement between the simulated results and experimental results was found with
a correlation coefficient value of 0.982. Therefore, it can be said that ANN
model could be used efficiently to run an automated water treatment process.
Pinaki Ghosh and Monal Dutta has conducted the experiments and Jayanta K Basu has developed the idea.