Predictive modeling of methylene blue dye removal by using artificial neural network

Predictive Modeling of Aqueous Dye Removal

by application of Artificial Neural Networks

Ashutosh Tamrakar

ChE 391: Independent Study
Department of Chemical and Biomolecular Engineering
Lafayette College,
Fall 2011-Spring 2012

Ashu Tamrakar
CHE 391: Independent Study

Advisors:

Dr. Polly Piergiovanni, Chemical and Biomolecular Department

Dr. Michael Senra, Chemical and Biomolecular Department

Dr. Chun Wai Liew, Computer Science Department

Lafayette College,

Easton, PA

2

Ashu Tamrakar

Abstract

This paper introduces dye removal processes in general and presents the fundamental and practical
aspects of neural networks techniques including an overview of their structures, strengths, and
limitations. The main objective of this study is to apply neural network modeling to predict the
Methylene Blue dye adsorption onto different commercial and non-conventional adsorbents as well
as to subsequently use the network model in ranking and identifying the best choice of adsorbents
for different wastewater conditions.

This work also illustrates the development of a comprehensive adsorbent database utilizing literature
experimental data available online and methods to prepare the data for use in neural network
modeling. In this project, Matlab Neural Network toolbox developed by the The MathWorks, Inc.
was used to develop the predictive model discussed. This research is unique in that it reports the
comprehensive modeling of dye removal by 22 different adsorbents and provides insight into a
ranking system which will be helpful in establishing a dye effluent treatment plant.

3

Ashu Tamrakar

Contents
Chapter 1. Introduction .................................................................................................................................... 7
1.1 Introduction to Dye Pollution......................................................................................................... 7
1.2 Research Motivation ......................................................................................................................... 8
1.3 Neural Network in adsorption kinetic study ................................................................................. 9
1.4 Research Aim ...................................................................................................................................11
1.5 Paper Organization .........................................................................................................................11
Chapter 2. Dyes and Dye Removal ...............................................................................................................13
2.1 Colorants and Dyes.........................................................................................................................13
2.2 Classification systems for dyes ......................................................................................................13
2.3 Methylene Blue (MB) ......................................................................................................................15
2.4 Methods of removing dye from wastewater................................................................................16
2.5 Adsorption Process .........................................................................................................................18
2.7 Common adsorption isotherms ....................................................................................................19
Chapter 3. Artificial neural networks ............................................................................................................21
3.1 Introduction to ANN .....................................................................................................................21
3.2 Neural Network architecture .........................................................................................................21
3.3 Network Learning ...........................................................................................................................23
3.4 Methods for Validation of Neural Networks..............................................................................24
3.5 Advantages and disadvantages of ANN (4): ...............................................................................25
Chapter 4. Database Collection .....................................................................................................................27
4.1 Variables ...........................................................................................................................................27
4.2 Adsorbents studied .........................................................................................................................27
4.3 Data Source ......................................................................................................................................28
4.4 Data Extraction Process.................................................................................................................30
4.5 Ranges of input data collected.......................................................................................................31
4.6 Data Preparation .............................................................................................................................34
Chapter 5. NN modeling of adsorption .......................................................................................................35
5.1 Model Development .......................................................................................................................35
5.2 Optimization of the NN architecture ..........................................................................................36
5.3 Present Study Results......................................................................................................................39
Chapter 6. NN Applications ..........................................................................................................................41

4

Ashu Tamrakar

6.1 Comparative predictions of variable effect on removal efficiency ..........................................41
6.2 Results ...............................................................................................................................................47
Works Cited .....................................................................................................................................................48

List of Appendices

Appendix A – Raw Dataset for Dye Removal Efficiency
Appendix B – Matlab Fuction file for Neural Network simulation
Appendix C – Detailed Adsorbent ranking

5

Ashu Tamrakar

List of Figures
Figure 1. Molecular Structure of Methylene Blue dye (20) ........................................................................15
Figure 2 Optimal ANN structure, together with a BP algorithm for the prediction of the Pollutant
Removal Efficiency (PRE) (11) .....................................................................................................................22
Figure 3.Training, validation and test mean squared error for the Levenberg-Marquardt algorithm .25
Figure 4. Extraction process for getting data points from a graph ..........................................................30
Figure 5. Range of pH in the dataset collected ..........................................................................................32
Figure 6. Range of initial dye concentration in the dataset collected .......................................................32
Figure 7. Range of contact time in the dataset collected............................................................................33
Figure 8. Range of temperatures in the dataset collected ..........................................................................33
Figure 9. Range of adsorbent dose in the dataset collected.......................................................................33
Figure 10. Neural network schematic for dye adsorption model .............................................................36
Figure 11. Dependence of Neural network performance on the number of neuron at hidden layer
and the distribution of dataset separated for training the neural network. .............................................37
Figure 12. Reproducibility check for the best NN architecture combinations......................................38
Figure 13. Quality of NN predictions for the training, validation and testing dataset. The overall
performance of the neural network shows 89% accuracy between predicted values and literature
data. ....................................................................................................................................................................39
Figure 14. Prediction for the performance of adsorbents at different pH levels.. The circled
adsorbents are commercial carbons while the rest are non-conventional adsorbents...........................42
Figure 15. Prediction for the performance of adsorbents at different initial dye concentration levels.
............................................................................................................................................................................43
Figure 16. Prediction for the performance of adsorbents at different contact times. ...........................44
Figure 17. Prediction for the performance of adsorbents at different temperature conditions. ..........45
Figure 18. Prediction for the performance of adsorbents at different adsorbate dosage conditions. .46

List of Tables
Table 1. Classes of Dyes and Their Chemical Types (13) ..........................................................................14
Table 2. Existing and emerging processes for dye removal (2).................................................................17
Table 3. Examples of adsorbents used in wastewater treatment (2) ........................................................19
Table 4. List of adsorbents and source article for the development of ANN model ............................28
Table 5. Statistical Index of Input and Output Data ..................................................................................31
Table 6. Ranking spectrum of each adsorbent at various variable conditions ........................................47

6

Ashu Tamrakar

Chapter 1. Introduction

1.1 Introduction to Dye Pollution
Synthetic dyes are widely used in textile, rubber, paper, plastic, and cosmetic industries for coloring
purposes due to their color variety, fastness, and ease of production as compared to natural dyes. (1)
Since most of these industries also consume substantial volumes of water, they generate a
considerable amount of colored aqueous waste. More than 100,000 commercially available dyes exist
in the market today with an annual production rate of more than 70,000 tons. (2) With such a large
production rate and good water solubility, these dyes are found frequently in industrial wastewater.
An indication of the scale of the problem is given by the fact that approximately 2% per cent of dyes
that are produced are discharged directly in aqueous effluent (2). Furthermore, even during its
application, about 10 to 15% of the dye is lost from the dyeing process to the effluent (3).

The effects of water contamination with these effluent dyes can be seen on two levels: effects on the
environment and effects on humans. Even if a small amount of dye is present in water (for example,
even less than 1 ppm for some dyes), it is highly visible (3). On an environmental level, the presence
of coloring material in water system reduces the penetration of light, thereby affecting
photosynthesis in aquatic planktons (4). Some of these dyes are also toxic as well as carcinogenic and
this poses a serious hazard to aquatic living organisms. Sulfur dyes, for example, can rapidly reduce
oxygen content of the water and are catastrophic for aqueous organisms. Most dyes increase the
acidity of the water system and as a result not only kill fish and other aquatic life but also damage
agricultural land and crops. In addition, the toxic compounds from dye effluent have been shown to
bioaccumulate through aquatic food chain as well as cause several physiological and biochemical
changes in fish (5). For humans, dyes have been linked to increased heart rate, nausea, vomiting,
shock, cyanosis (blue baby syndrome), jaundice and quadriplegia (6).This problem is exacerbated by
the fact wastewater containing dyes are very difficult to treat, since the dyes are recalcitrant organic
molecules, resistant to aerobic digestion, and are stable to light, heat and oxidizing agents (3).

7

Ashu Tamrakar

In order to combat the increasing dye pollution, several environmental preservation efforts and
federal regulations have been put in place to restrict the industrial effluent. In 1972, title III of the
Federal Water Pollution Control Act established the Effluent Guidelines Program (EGP) which was
amended first in 1977 by the Clean Water Act Amendments and subsequently in 1987 by the Water
Quality Act (7). The EGP, implemented by the Environmental Protection Agency (EPA) sets the
limits for the amount of discharge for toxic compounds including organic and inorganic dye
molecules, legally allowed to be returned to the environment. Another program called the National
Pollutant Discharge Elimination System (NPDES) strives to control water pollution by issuing
permits for industries that discharge pollutant into the waters of United States (7).

With the need to comply with such stringent discharge standards and to achieve an optimum control
and management of dye effluent, new concepts involving effective dye removal designs have been
developed around the world. During the past three decades, several physical, chemical and biological
decolorization methods have been reported in literature that attempt to showcase the best
techniques for dye removal. Amongst the numerous techniques, adsorption of dye onto different
compounds has been shown to give the best results as it can be used to remove different types of
coloring materials (2). This paper provides an overview of the dye adsorption process and studies
the possibility of developing an efficient predictive model of dye removal that will help in optimizing
the adsorbent selection decisions.

1.2 Research Motivation
As previously mentioned, adsorption is one of the most established unit operations used for the
treatment of contaminated wastewater and as such adsorption boasts of innumerable studies that
target different kinds of dye removal. In general, these studies are usually conducted either in
batches or in adsorbent column experimentations. The batch studies are aimed at determining the
kinetics and isotherm constants for the adsorption process while column studies are performed for
determining the breakthrough curve that represent the dye concentration as the wastewater leaves
the adsorbent bed. In both types of studies, the dye removal efficiency is the most significant output
and the variation of this efficiency depends on several factors such as adsorbent characteristics,

8

Ashu Tamrakar

contact time, initial adsorbate (dye) concentration etc. However, even with the large availability of
experimental studies that provide information on the dye removal efficiency of adsorbents, or
perhaps precisely due to its extensive nature, it is extremely difficult to rank the adsorbents and
hence decide the best one to use with a particular effluent stream. This dilemma is compounded by
the fact that there are no specific guidelines available to assess the suitability of the adsorbent for the
contaminated water treatment. Most of the adsorption studies are not carried out within a standard
range of parameters and when investigators do develop new indigenous adsorbents, more often than
not the physio-chemical characteristics of the absorbents are not reported. Basu et. al. highlights
these underlying issues with the observations and characteristics that are being reported in these
literature studies in terms of (8):

• Inconsistency in the characteristics of several adsorbents that are being reported
• Insufficiency in the information to completely understand the adsorption mechanism when a
database is generated for similar characteristics/ trends in adsorption

To put it simply, the results that are reported by an investigator may be accurate and sufficient for
their study (i.e., on a micro level) but when they are compared on a common basis (i.e., macro level)
to other adsorption studies, with respect to the adsorbate /adsorbent/specific characteristic property
etc., inconsistent trends in the results are observed (8). In fact, majority of the authors report that it
is extremely difficult to compare any indigenous data with that already published in literature. This
difficulty along with the dire need to optimize the dye removal system led to this development of a
model to predict the dye removal efficiency from batch studies database of adsorption from
literature. The core assumption for this investigation is that the database reported in the literature is
accurate at the micro level.

1.3 Neural Network in adsorption kinetic study
Developing a predictive model that will compare the relative dye removal performance of several
adsorbents over a wide range of parameters is not a straight forward task. Although the efficiency of
each adsorbent can be modeled using isotherms, it is difficult to understand the relationship
between the efficiencies of several adsorbents and the interconnectedness of variables that

9

Ashu Tamrakar

determine the removal performance. It becomes, thus, imperative to tackle such problems using
techniques outside the conventional chemical engineering curriculum.

Different techniques are being used in the computational community for prediction tasks such as
the dye removal model. In recent years the concept of artificial neural networks (ANN) which is a
type of artificial intelligence technique has emerged as one of them. A neural network, in general, is
able to work with just input dataset to find patterns and irregularities as well as to detect multi-
dimensional non-linear connections in data. The latter quality is extremely useful for modeling
dynamical systems, such as the dye adsorption process. ANNs have the ability to relate the input and
output variables without having any knowledge on physics of the system provided an accurate and
large amount of data on the system variables to train the networks is available. The neural networks
can thus yield solutions to complex phenomena where the relationships and rules are not known.

The use of neural networks in adsorption studies is not a new concept. Over the last four years
alone, several researches such as Improving the Efficiency of Wastewater Treatment Process by Soft
Computational Methods (1), Modeling of Nitrate compounds on granular activated carbon (CAG) using artificial
neural network (ANN) (9), Neural Network Modeling and Simulation of the Solid/Liquid Activated Carbon
Adsorption Process (10) and Artificial Neural Network (ANN) approach for modeling of Pb(II) adsorption from
aqueous solution by Antep pistachio (Pistacia Veral L.) shells (11) have made use of neural networks to
model the adsorption behavior. The application of neural network in these studies, however, have
been more for extrapolation of experimental results of a single adsorbent-pollutant combination
rather than to develop a comprehensive predictive model of several adsorbents. Trained on the data
obtained from the limited number of physical experiments, the neural networks in these researches
allow for prediction of adsorption efficiencies under novel conditions.

Fortunately, there are a handful of researchers who have attempted to build an adsorption model
with more than one set of adsorbent-pollutant combinations. Basu, Ramakrishna and Chakravarthy
from the Birla Institute of Technology and Science, India are at the forefront of this endeavor and
have applied ANN to the literature data pertaining to adsorption batch studies of 25 adsorbents

10

Ashu Tamrakar

tested over 8 different pollutants namely Chromium, Zinc, Copper, Mercury, Cadmium, Lead, COD
and color (12). The group trained their ANN using 440 literature data points with 7 independent
variables (adsorbent material, pollutant type, equilibrium time, pH, contact time, adsorbate
concentration, and adsorbent dose), and tested the network with 73 data points. However, the focus
of their study is more towards the development of a strong neural network architecture with
optimization of the ANN parameters such as number of hidden layers, learning rate, number of
epoch, etc. rather than finding the optimal adsorbent to use with a given pollutant problem. This
project, on the other hand, will attempt to not only train ANN with a database of wide range of
adsorbents as reported in literature (as many as 22 different adsorbents, only one dye) but also will
apply the establish neural network to rank the adsorbents tested . The ranking developed will help
streamline the decision making process for the selection of best adsorbent to use at a given dye
contamination condition.

1.4 Research Aim
The main objective of this project, therefore, is to formulate a functional relationship between the
efficiency (output) and the adsorption factors (inputs) for the adsorption of Methylene blue dye
from aqueous solutions. The factors considered in this study are material of adsorbent, pH, contact
time, initial concentration of dye in solution, temperature of solution and the adsorbent dosage.
There are both dimensional variables (contact times, temperature, initial concentration of adsorbate,
& adsorbent dosage) and dimensionless groups [material of adsorbent and pH] in these factors.
Because of the complexity of non-linear relationship and incomplete understanding between the
dimensional/ dimensionless variables and the efficiency as well relativity between the adsorbents,
predictive modeling through ANNs is proposed. In addition, the neural network thus developed will
be used to rank the adsorbents in various conditions such as high/ low pH wastewater, high/low
temperature conditions, etc and the best ones to be used in each condition will be identified.

1.5 Paper Organization
In order to accomplish the above mentioned objectives, the project was carried out in two main
phases: Phase I – NN development (carried out in Fall 2011 semester) and Phase II- NN application

11

Ashu Tamrakar

(Spring 2012). The descriptions of these phases as well as the necessary theory related to the project
are detailed in this paper over 6 chapters:

• Chapter 2 presents an overview of dye removal process using adsorbents and the principles
related to the adsorption mechanisms.

• Chapter 3 discusses the architecture and parameters associated with artificial neural network
technique. This chapter also provides a literature review on application of neural networks in
adsorption processes.

• Chapter 4 discusses the first half of Phase I of the project which is database collection for the
neural network development. The methodology of experimental data extraction from literature, the
test for quality of data and standardization is further discussed.

• Chapter 5 presents the steps in development of the artificial neural network. Simulation results are
presented assessing the performance of neural network based on parameters such as number of
hidden layers and amount of training data set. Validation methods using both a validation sample set
as well as visualization methods are also presented and discussed.

• Chapter 6 gives an overview of the application of the neural network model developed, including
effects of variables such as pH, dye concentration, temperature, etc on the dye removal efficiency.

12

Ashu Tamrakar

Chapter 2. Dyes and Dye Removal

2.1 Colorants and Dyes
Colorants are compounds that are characterized by their ability to absorb and emit visible light from
400 to 700 nm (13). Broadly, colorants are classified into either dyes or pigments based on the
manner in which they give color to materials. Pigments are generally insoluble compounds and
hence are used to color substrates like textile, paper, plastic, etc by attaching to the substrates
through use of additional compounds such as polymers. Dyes, on the other hand, are applied to the
substrate via a liquid medium in which they are either completely or partially soluble and attach by
binding with the material. Chemically, dyes are ionic, aromatic organic compounds with structures
including aryl rings with delocalized electron systems (14). The characteristic color of a dye molecule
is given by the chromophore group present in the structure where the energy difference between the
two different molecular orbitals falls within the range of visible spectrum (15).

Dyes are widely used in a variety of industries including textile, paper, plastic foodstuff, cosmetics,
pharmaceutical and mineral processing industries to color their products. In fact, the scale and
growth of the dyes industry has been completely linked to that of the industries using the product.
For instance, the annual production of textiles around the world is about 30 million tons and the
amount of dye needed for this production rate is about 700,000 tons (16). Other estimations carried
out in 1994 predict the world dye production be closer to 1 million tons per annum (17).

2.2 Classification systems for dyes
There are several established ways of classifying dyes including differentiating dyes based on their
chemical structure (chemical classification) or according to the method of application (dyeing
method) or by nature of the electronic excitation of the dye molecule. The most popular
classification however is the one advocated in the US International Trade Commission (USITC)
which divides dyes into the following types (13):

13

Ashu Tamrakar

Table 1. Classes of Dyes and Their Chemical Types (13)

Class Substrate Method of Application Chemical types

Azo (including
Nylon, wool, silk, premetallized),
Usually from neutral to acidic
Acid paper, inks and anthraquinone,
bath.
leather. tryphenylmethane, azine,
xanthene, nitro and nitroso.
cyanine, hemicyanine,
Paper, diazahemicyanine,
polyacrylonitrile, diphenylmethane,
Basic Applied from acidic dye baths.
modified nylon, triarylmethane, azo, azine,
polyester and inks. xanthene, acridine, oxazine
and anthraquinone.
Reactive site on dye reacts
with functional group on fiber Azo, anthraquinone,
Cotton, wool, silk
Reactive to bind dye covalently under phthalocyanine, formazan,
and nylon.
influence of heat and pH oxazine and basic.
(alkaline).
Applied from neutral or
Cotton, rayon,
slightly alkaline baths Azo, phthalocyanine,
Direct paper, leather and
containing additional stilbene, and oxazine.
nylon.
electrolyte.
Fine aqueous dispersions
often applied by high
Polyester, temperature/pressure or
Azo, anthraquinone, styryl,
Disperse polyamide, acetate, lower temperature carrier
nitro and benzodifuranone.
acrylic and plastics. methods; dye maybe padded
on cloth and baked on or
thermo fixed.
Plastics, gasoline,
Azo, triphenylmethane,
varnishes, lacquers,
Solvent Dissolution in the substrate anthraquinone, and
stains, inks, fats, oils,
phthalocyanine
and waxes.
Aromatic substrate vatted
with sodium sulfide and
Sulfur Cotton and rayon Indeterminate structures
reoxidized to insoluble sulfur-
containing products on fiber
Water-insoluble dyes
solubilized by reducing with Anthraquinone (including
Cotton, rayon and
Vat sodium hydrogensulfite, then polycyclic quinines) and
wool
exhausted on fiber and indigoids
reoxidized

14

Ashu Tamrakar

2.3 Methylene Blue (MB)
The dye being examined in this project is Methylene Blue (also known as Basic Blue 9) which falls
under the Basic dye category and is a dark green crystalline solid (18). The molecular structure of
MB is illustrated in Figure 1. Solutions of MB in water or alcohol have a characteristic deep blue
color from which the compound derives its name. The annual production of MB was estimated by
the National Institute of Health (NIH) to be around 12-120 thousand pounds in 1977 (19).

Figure 1. Molecular Structure of Methylene Blue dye (20)

The common applications of MB are listed below (19):
A. For Therapeutic Uses:
i. Treatment of methemoglobinemia
ii. Antidote for cyanide poisoning
iii. Treatment of manic-depressve psychosis
iv. Formerly used as urinary antiseptic (currently more effective agents are available)
v. Formerly used as an analgesic, antipyretic and antiparasitic
B. Use as a dye/stain:
i. Bacteriologic stain
ii. Indicator dye
iii. Surgical and medical marking
iv. Coloring paper, cotton, wool and leather
v. Temporary hair colorant

The National Fire Protection Agency (NFPA) rates the Methylene Blue (MB) dye at a health risk of
2 and it was officially nominated for carcinogenicity by the National Cancer Institute (NCI) in 1989
(6). Data from the National Occupational Exposure Survey (NOES) indicate that 69,563 workers,

15

Ashu Tamrakar

including 42,026 female employees were potentially exposed to Methylene blue between 1981and
1983. The toxicological effects of MB on humans are summarized below (6):

• Acute exposure to MB has been found to cause increased heart rate, cyanosis, vomiting,
shock, Heinzbody formation, jaundice, quadriplegia and tissue necrosis in humans. In
addition, corneal and conjunctival injury has been reported following acute exposure to this
compound. Intravenous administration of methylene blue has been found to cause bluish
discoloration of the urine and stool especially for the newborn.

• Chronic application of Methylene blue-containing eyedrops has been found to result in
staining of the bulbar and palpebralconjunctiva, the lid margins and slight staining of the
cornealepithilium

2.4 Methods of removing dye from wastewater
More than 100,000 different types of commercial dyes are manufactured every year and it is
estimated that about 2% of total dye produced annually is released in the effluent stream (14). The
situation is even worse with textile industries where about 10% of dyes used are discharged into
waste stream (14). In order to safely treat the sewage to comply with government regulations several
biological, chemical and physical methods of dye removal have been developed. These include
physical-chemical flocculation, electroflotation, membrane filtration, electrokinetic coagulation,
precipitation, ozonation , adsorption, etc. Table 2 shows the advantages and disadvantages of each
of these processes.

The biological treatment method is the most economic option compared to physical and chemical
processes and involves decolorization through use of fungal/ microbial degradation and adsorption
by living or dead biomass (13). The application of biological methods, however, is limited because of
the sensitivity of the microorganisms. Chemical methods of dye removal, on the other hand, consists
of coagulation, electroflotation and oxidation processes which are generally very expensive and also
lead to accumulation of concentrated sludge (13). The physical treatment of dye contaminated
wastewater use membrane filtration and adsorption techniques for decolorization. However, the
physical methods often suffer from limited membrane/adsorbent lifetime issues which make the
processes expensive since periodic replacements are necessary (13).

16

Ashu Tamrakar

Table 2. Existing and emerging processes for dye removal (2)

Treatment Advantages Disadvantages

Coagulant Simple, economically High sludge production, handling and
Floccutant feasible disposal problems
Slow process, necessary to create an
Economically attractive,
optimal favorable environment,
Conventional Biodegradation publicly acceptable
maintenance and nutrition
treatment treatment
requirement
process
The most effective
Ineffective against disperse and vat
Adsorption on adsorbent, great,
dye, the regeneration is expensive and
activated capacity, produce a
result in loss of the adsorbent, non-
carbon high-quality treated
destructive process
effluent
Remove all dye types,
Membrane High pressure, expensive, incapable of
produce a high-quality
separation treating large volumes
treated effluent
Established
recovery No loss of sorbent on Economic constraints, not effective
Ion-exchange
Process regeneration, effective for disperse dye

Rapid and efficient
Oxidation High energy cost, chemical required
process

No sludge production,
Advanced
little or no consumption Economically unfeasible, formation of
oxidation
of chemicals, efficiency by-products, technical constraints
process
for recalcitrant dyes

Economically attractive,
Emerging Selective regeneration is not Requires chemical modification,
removal Bioadsorbents necessary, high nondestructive process
processes selectivity

Low operating cost,
good efficiency and
Slow process, performance depends
Biomass selectivity, no toxic
on some external factors (pH, salts)
effect on
microorganisms

17

Ashu Tamrakar

2.5 Adsorption Process

As seen from Table 2, adsorption techniques are one of the most established physical methods of
dye removal. Adsorption of dye onto various types of adsorbents is gaining this favor due to their
efficiency in the removal pollutants too stable for conventional methods such as biological treatment
(14). In addition, this removal technique produces a high quality product (complete removal of dye
is possible) with sludge free clean operation which has further increased its popularity (13). In
general, adsorption processes work through either physical adsorption or ion-exchange. Physical
adsorption occurs when weak intermolecular bonds such as Van der Waals, hydrogen and dipole-
dipole develop between the dye and adsorbent while ion-exchange which is a chemical adsorption
occurs when stronger bonds such as covalent and ionic bonds are established between the dye
molecules and adsorbents (14). The decolorization processes in both mechanisms depends on many
physio-chemical factors such as dye/adsorbent interaction, adsorbent surface area, particle size,
temperature, pH and contact time.

Most commercial systems currently use activated carbon as sorbent to remove dyes in wastewater
because of its excellent adsorption ability. Activated carbon adsorption has been reported by the US
Environmental Protection Agency as one of the best available control technologies for dye effluent
treatment (2). The efficiency of activated carbon in adsorption is due to its structural characteristics,
porous texture and chemical nature (3). The dye removal performance for all activated carbons is
dependent on the type of activated carbon used and the characteristics of the wastewater. Thus, like
many other dye-removal treatments, it is well suited for one particular waste system and ineffective
in another.

2.6 Commercial Adsorbents versus Nonconventional Adsorbents

Although activated carbon is a preferred sorbent, its widespread use is restricted due to high cost
which gets steeper yet with increases with its quality. In order to decrease the cost of treatment,
attempts have been made around the world to find inexpensive alternative adsorbents. Recently,
numerous approaches have been studied for the development of cheaper and effective adsorbents
using non-conventional, low-cost materials including natural supplies, biosorbents, and waste

18

Ashu Tamrakar

materials from industry and agriculture. Table 3 below shows example of commercial adsorbents as
well as the non-conventional adsorbents reported in literature. These materials could be used as
adsorbents for the removal of dyes from solution. Some of the reported sorbents include clay
materials (bentonite, kaolinite), zeolites, siliceous material (silica beads, alunite, perlite), agricultural
wastes (bagasse pith, maize cob, rice husk, coconut shell), industrial waste products (waste carbon
slurries, metal hydroxide sludge), biosorbents (chitosan, peat, biomass) and others (starch,
cyclodextrin, cotton).

Table 3. Examples of adsorbents used in wastewater treatment (2)

Adsorption
Type Adsorbent supplier or material Dye
capacity (mg/g)
Filtasorb Corporation (USA) Reactive orange 107 714
Commerical Merck Co. (Taiwan) Reactive red 2 712.3
activated carbons Chemviron carbon (UK) Acid blue 40 133.3
Calgon Corporation (USA) Direct red 28 7.69
Pine wood Acid blue 264 1176
Bagasse Basic red 22 942
Non Rice husk Basic Green 4 511
conventional Treated sawdust Basic green 4 26.9
carbon materials Fly ash Alizarin sulfonic 11.21
Neem sawdust Basic violet 3 3.78
Activated bentonite Acid blue 193 740.5

2.7 Common adsorption isotherms

The adsorption of dye is conventionally modeled in literature using various adsorption isotherms:
the two common types of isotherms encountered are Langmuir and Freundlich isotherms.
Regardless of the type, both of these isotherms describe the capacity of the adsorbent as a function
of equilibrium concentration of dye in solution at a constant temperature. Brief description of each
model is described below (21):

19

Ashu Tamrakar

A. Langmuir Isotherm: This isotherm is the most widely used model to obtain the maximum
adsorption capacity produced from complete monolayer coverage of adsorbent surface. The
isotherm equation gives the fractional coverage (Θ) in the form:

𝜃= = 1+𝑏𝐶
𝑞𝑒 𝑒 𝑏𝐶
𝑄𝑚 𝑒
[1]

where, Θ = fractional coverage
b = adsorption equilibrium constant, l/mg
Qm = quantity of adsorbate required to form a single monolayer of adsorbent, mg/g
Qe = amount adsorbed on unit mass of adsorbent, mg/g
and Ce = equilibrium concentration, mg/ l

Equation 1 can be rearranged to get linear form as shown in Equation 2. If the adsorption behavior
follows a straight line for a plot of (Ce/qe) vs. Ce, the adsorption obeys Langmiur isotherm.

= + � 𝑄 � 𝐶𝑒
𝐶𝑒 1 1
𝑞𝑒 𝑏𝑄 𝑚 𝑚
[2]

B. Freundlich Isotherm: The Freundlich isotherm is a semi-empirical equation which is
widely used to represent adsorption equilibrium data for low to intermediate range of

𝑞 𝑒 = 𝐾 𝑓 𝐶 𝑒𝑛
concentration. The Freundlich equation is characterized below:
[3]

where, n = Freundlich coefficient
and, Kf = Freundlich constant, mg1-1/n l 1/n g-1

20

Ashu Tamrakar

Chapter 3. Artificial neural networks

3.1 Introduction to ANN
The neural network technique is an artificial intelligence technique that attempts to mimic the
human brain’s problem solving capabilities. ANNs are analogous to the biology of a human brain,
where billions of neurons are interconnected to process a variety of complex information (9). When
presented with data patterns, sets of historical input and output data that describe the problem to be
modeled, ANNs map the cause-and-effect relationships between the model input data and output
data. This mapping of input and output relationships in the ANN model architecture allows
developed models to be used to predict the value of the model output parameter, given any
reasonable combination of model input data, with satisfactory accuracy (4). With the advances in
computing power, ANNs have become extremely fast and flexible.

Presently, artificial neural networks have been successfully applied in many fields, which include
character recognition, speech recognition, image processing, and stock performance prediction. In
chemical engineering, ANNs were found to be successfully applied to predict adsorption equilibrium
of solid/liquid systems, activity coefficients of aromatic organic compounds, and solubility of
proteins. (10)

3.2 Neural Network architecture
In general, a neural net, as shown in Figure 2, is parallel interconnected structure consisting of input
layer of neuron (independent variables), a number of hidden layers, and an output layer (dependent
variables). The number of input and output neurons is fixed by the nature of the problem. However,
the hidden layers, which act like feature detectors, can be adjusted as needed.

21

Ashu Tamrakar

Figure 2 Optimal ANN structure, together with a BP algorithm for the prediction of the Pollutant
Removal Efficiency (PRE) (11)

Besides the layers, an ANN model also consists of a pattern of connectivity or weights between
units (illustrated by the red lines in Figure 1) , a propagation rule for propagating patterns of
activities through the weights, and a learning rule whereby weights are modified by experience (4).
Depending on the ANN software being used, some or all of these components may be adjusted.
Commonly employed neural network are generally feed-forward networks where the model input
data is processed forward through the network in sequential fashion. The network prediction error
information may, however, be propagated in a backward direction through the network.

22

Ashu Tamrakar

3.3 Network Learning
Artificial neural networks learn by reorganizing their internal structure according to a learning rule or
algorithm to minimize the error between the actual output value and the model-predicted output
value for the entire set of data patterns (4). Because the error is propagated backwards to adjust the
weights on the input, the learning is commonly referred to as Back Propagation (BP) algorithm.
Considering the model illustrated in Figure 2, the learning process for the particular network follows
the following sequence:

a. Initially, the input parameter information (from the 5 input variables) is scaled in the input
layer according to a scaling function. Typical scaling functions are linear and scale the values
of all the input parameters to a common range, generally 0 to 1.
b. Each input layer neuron is then connected to each of the 11 hidden layer neurons by a
connection weight. Weights here are mathematical constructs that assign a numerical value
to the importance of the connection between neurons. The output from each input neuron
is thus multiplied by the appropriate connection weight and the resulting products are
transferred to the hidden layer neurons.
c. In the hidden layer, each neuron sums the value of the incoming products and processes the
sum through a predefined activation function, which defines the neuron’s state of activity.
d. Output values from each of the hidden layer neurons are then multiplied by the appropriate
weights, as before, and the resulting products are transferred to the output layer. In the
output layer, which has one neuron for each output parameter, each neuron sums the value
of incoming products and maps the sum into an output value according to a predefined
scaling function.
e. The resulting model predicted value of the output is then compared with the actual value of
the model output parameter from the data pattern. The output units then backpropagate the
prediction error to the hidden layer according to a learning algorithm.
f. Finally, the hidden layer units modify their incoming connection weights according to the
learning algorithm to reduce the prediction error.

The entire process is repeated over several iterations each of which are termed epochs until the
ANN produces a sufficiently small error, or other conditions as determined by the user are met.

23

Ashu Tamrakar

There are two main termination steps that can be controlled to stop the iterations: late and early
stopping. Late stopping refers to the termination condition where the network is trained until a
minimum pre-specified error on the training set is reached. The minimum error on the training set
does not always indicate the best results as more often than not the network is clearly over fitted
which gives it a poor generalization ability. The concept of early stopping, on the other hand, refers
to the condition where the learning progression is monitored over each epoch and the training is
terminated as soon as signs of over fitting appear. The presence of over fitting is generally tested by
using a novel set of data points on the network after each epoch and measuring the error for its
prediction. As soon as the error stops decreasing for this new set of test data, the training is
terminated. A clear advantage with early stopping is that the time of training is relatively short.

3.4 Methods for Validation of Neural Networks
Since the internal computation of model function is not reported by a neural network, it becomes
imperative to validate the network once it is trained. One of the simplest methods of validation that
are used by most developers is to separate a set of available data into three sets: training, validation,
and testing sets (22). The training data set is then used as the primary set of data that is fed to the
neural network during the training phase for learning and adaptation. The second set of data,
confusingly called the validation dataset, is used to further refine the neural network learning by
determining the termination step for the iterations (as described in the section above) rather than
actually validating the network. The main validation of the neural network is, thus, carried out by
application of the developed neural network onto the testing dataset. Since the testing data points
are novel to the network they can be used to determine the performance of the neural network by
observing the computation of an error between the networks’s predicted output and experimental
result. . Figure 3 shows a representative performance plot of the neural network developed.

In Figure 3, the blue line illustrates the decreasing error on training data while the green line shows
the error in the validation set. The training for the network stops when validation error stops
decreasing or when it reaches the desired error tolerance. The red line indicates the error on the test
data which gives the information on how well the network will generalize to new data.

24

Ashu Tamrakar

Figure 3.Training, validation and test mean squared error for the Levenberg-Marquardt algorithm

3.5 Advantages and disadvantages of ANN (4):

Advantages:

a. Since the data processing occurs purely from the data inputs to the system, no preexisting
mathematical/ mechanistic models or assumptions are needed.
b. The ANN technique is fault-tolerant in model development and thus accommodates
discontinuities in the data, different levels of data precision, noise, and data scatter are easily
accommodated.

25

Ashu Tamrakar

c. The technique is also extremely fast and flexible; advances in computing power have
minimized the time required to develop models, as well as the time required to re-train
models to incorporate new data and to reflect process modifications.

Disadvantages:

a. Since ANNs do not yield explicit mathematical formulae, many researchers consider the
developed models to be “black-box” models.
b. Little is known about the applicability of the models to data that lie outside the domain on
which the models were trained.
c. No set protocol for developing ANN models exists; each modeler may incorporate different
modeling techniques.
d. The ANN technique is data intensive and is best suited to problems where large data sets exist.

26

Ashu Tamrakar

Chapter 4. Database Collection
4.1 Variables
In order to model dye adsorption kinetics the following variables were examined:

pH of the solution

Initial concentration of dye (mg/L)

Independent Variable (Input) Contact time (min)

Adsorbent dose (g/L)

Temperature of solution (oC )

Dependent Variable (Output) Dye Removal Efficiency (%)

4.2 Adsorbents studied
In the present study, thirteen different low cost adsorbents such as activated wheat bran, activated
clay minerals, sawdust, bio-adsorbents, etc used to remove MB from wastewater are investigated in
this report (#1-13, see Table 1). In addition, their adsorption capacity is compared with the
commercially available activated carbon. Additionally, a brief cost analysis is carried out for the
developed and commercial adsorbent which shows an economic feasibility of developed adsorbents
for the removal of MB.

Since the quality of the neural network is dependent on the quantity and quality of data it is trained
on, sufficient data mining for adsorption kinetics is a large part of the project. The main criteria used
to specify the boundaries of a source dataset is that it must be fully representative of the full
spectrum of possible conditions to which the model will be applied. In modeling dye removal
processes to treat different volumes of effluent industrial water for example, the source data set
should be selected to encompass the levels of initial dye concentration encountered. This section
discusses the methods of data extraction used and the subsequent quality of data.

27

Ashu Tamrakar

4.3 Data Source
Artificial neural network are initially developed and trained using historical data. The datasets used
for training chemical process models, such as adsorption kinetics, are usually collected by physical
experimentation by the researcher. Fortunately, there are abundant quantities of research carried out
around the world to study MB dye kinetics that is completely feasible to develop an ANN by using
only information published in scientific journals. A complete list of articles used as source data for
this project is provided in Table 4.

Table 4. List of adsorbents and source article for the development of ANN model
ID # Adsorbent used Journal Article Title Authors

Kinetic and Equilibrium Studies of Methylene O.S. Bello; O. M. Adelaide;
Treated
1 Blue Removal from aqueous Solutions by M.A. Hammed; O. A.M.
Sawdust
adsorption on treated sawdust Popoola
Removal of Methylene Blue, a basic dye from
S. Patil; S. Renukdas; N.
2 Teak Tree bark aqueous solutions by adsorption using teak
Tatel
tree bark powder
Fast Removal of Methylene blue from aqueous
Chitosan CTS- L. Wang; J. Zhang; A.
3 solution by adsorption onto chitosan-g-poly
g-PAA Wang
(acrylic acid)/ attapulgite composite

Fast Removal of Methylene blue from aqueous
Chitosan CTS- L. Wang; J. Zhang; A.
4 solution by adsorption onto chitosan-g-poly
g-PAA/APT Wang
(acrylic acid)/ attapulgite composite

Experimental study of Methylene blue
Carbon Z. Shahryani; A. S.
5 adsorption from aqueous solutions onto
Nanotubes Goharrizi; M. Azadi
carbon nanotubes

Acid-activated Removal of Methylene Blue from aqueous
B. Karima; B. L. Mossab;
6 Algerian solutions using an acid activated Algerian
M. A-Hassen
Bentonite Bentonite: Equilibrium and kinetic Studies

Removal of Methylene Blue from aqueous Oualid Hamdaoui, Mahdi
7 Wheat Bran
solutions by Wheat Bran Chiha

8 Biosolids Removal of Methylene Blue by using biosolids M. Sarioglu, U.A. Atay

The removal of cationic dyes using coconut
9 Coconut husk K.S. Low and C.K. Lee
husk as an adsorbent

28

Ashu Tamrakar

Removal of basic Dye Methylene Blue fro
Walnut R. Ansari and Z.
10 aqueous solutions using sawdust and sawdust
Sawdust Mosayebzadeh
coated polypyrrole
Walnut
Removal of basic Dye Methylene Blue fro
Sawdust coated R. Ansari and Z.
11 aqueous solutions using sawdust and sawdust
with Mosayebzadeh
coated polypyrrole
Polypyrrole
Water-washed A kinetic, thermodynamic and mechanistic
K.M. Parida, Swagatika
manganese approach toward adsorption of Methylene
12 Sahu, K.H. Reddy and P.C.
module leached blue over water-washed manganese nodule
Sahoo
residue leached residues
Sunflower stalks as adsorbents for color Gang Sun and Xiangjing
13 Sunflower stalk
removal from textile wastewater Xu
Titanium Sportive removal of dyes using titanium Kalpana C. Maheria and
14
Phosphate phosphate Uma V. Chudasama
Araceli Rodriguez, Gabriel
Removal of Dyes from wastewaters by Ovejero, Maria Mestanza
15 Sepoilote
adsorption on Sepiolote and Pansil and Juan Garcia

Araceli Rodriguez, Gabriel
Removal of Dyes from wastewaters by
16 Pansil Ovejero, Maria Mestanza
adsorption on Sepiolote and Pansil
and Juan Garcia
Evaluation of Loofa as a sorbent in the N.A. Oladoja, C.O.
17 Loofa (fruit) decolorization of basic dye contaminated Aboluwoye and A.O.
aqueous system Akinkugbe
Raw Sugarcane Kinetics study of methylene blue dye S. P. Raghuvanshi1, r.
18
baggase bioadsorption on baggase Singh, c. P. Kaushik
Chemically
Treated Kinetics study of methylene blue dye S. P. Raghuvanshi1, r.
19
Sugarcane bioadsorption on baggase Singh, c. P. Kaushik
baggase
P. Waranusantigul; P.
Kinetics of basic dye (Methylene blue)
Giant Duck Pokethitiyook; M.
20 biosorption by giant duckweed (Spirodela
weed Kruatrachue; E.S.
polyrrhiza)
Upatham
Sulfuric acid Basic dye (Methylene blue) removal from
treated Indian simulated wastewater by adsorption using V.K. Garg; M. Amita, R.
21
rosewood Indian Rosewood sawdust: a timber industry Kumar; R. Gupta
sawdust waste
Formaldehyde Basic dye (Methylene blue) removal from
treated Indian simulated wastewater by adsorption using V.K. Garg; M. Amita, R.
22
Rosewood Indian Rosewood sawdust: a timber industry Kumar; R. Gupta
Sawdust waste
Note: Adsorbents #3,4,5 and 14 are commercial type of adsorbents while the rest of the adsorbents
were developed indigenously by the researcher.

29

Ashu Tamrakar

4.4 Data Extraction Process
The difficult part about collecting data from published articles is that the information on the
adsorbent performances is usually presented graphically. In order to the extract the experimental
data points from the graphs presented in the articles, GetData Digitizer evaluation version 2.24 was
used. This software essentially maps out the graph into a grid and collects the coordinates for the
points/ line selected. The steps for extracting data points from a sample graph of temperature
effects on the kinetics of dye removal by wheat bran, for example are explained below:

Figure 4. Extraction process for getting data points from a graph

Step 1: Set the scale for the grid by setting four points Xmin, Xmax, Ymin and Ymax, and by assigning
logical coordinates to these points .In Figure 4, Xmin and Xmax are assigned values 0 and 200
respectively while Ymin and Ymax are assigned values 1.5 and 2.3.

30

Ashu Tamrakar

Step 2: Since multiple experiments are presented in the graphs, in order to maintain the quality of
data, the each trial should be manually digitized by clicking on the respective point. Figure 4
shows all the data points for the trail run at 50 C highlighted.

Step 3: All the respective coordinates are displayed on the right column and can be exported to an
Excel document.

The data extraction process was carried out for all of the experiments, the range and quality of data
extracted is detailed in the next section.

4.5 Ranges of input data collected
In total 1363 data points from various combinations of inputs and their respective dye removal %
were collected for the 16 adsorbents. Table 5 shows the statistical indexes of input and output data
collected. The original data collected for each adsorbent is available in Appendix A.

Table 5. Statistical Index of Input and Output Data

Inputs Outputs
Initial
Contact time Temperature Adsorbent % Dye
Concentration pH
(min) (C ) Dose (g/L ) Removal
(mg/L)
Min 5.0 1.1 0 20.0 0.0 0
Max 2201.5 12.3 2889.8 65 15.0 100
Mean 195.9 6.7 167.4 25.9 3.3 73
Median 100.0 7.0 60.0 25 4.0 81.3
S.D 364.3 1.6 371.1 6.4 2.6 25.3
95% CI (215.3,176.6) (6.8,6.6) (187.1,147.7) (26.3, 25.6) (3.4,3.1) (74.4,71.1)

From Table 5, it is clear that most of the data collected from the adsorbents lie in a very small range
of pH, initial concentration, contact time, temperature as well as adsorbent dose values. However,

31

Ashu Tamrakar

the spread of the data, which is more important as it establishes the connection between each
variable, is not visible in the statistical indices. Figures 5-9 below shows the distribution of each
variable for each adsorbent.

11.0
10.0
9.0
8.0
Figure 5. 7.0
pH range

Range of pH in 6.0
the dataset 5.0
collected 4.0
3.0
2.0
1.0
0.0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
Adsorbent Type

500.0
Initial Dye concentration (mg/L)

450.0
Figure 6. 400.0
Range of initial 350.0
dye 300.0
concentration 250.0
in the dataset 200.0
collected 150.0
100.0
50.0
0.0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
Adsorbent Type

32

Ashu Tamrakar

500.00
450.00
400.00
Contact Time (min)

350.00
300.00 Figure 7.
250.00 Range of
200.00
contact time in
the dataset
150.00
collected
100.00
50.00
0.00
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
Adsorbent Type

50.0

45.0
Temperature Range (C)

Figure 8.
40.0 Range of
temperatures in
35.0
the dataset
collected
30.0

25.0

20.0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
Adsorbent Type

16.00
14.00
Adsorbent Dosage (g/L)

12.00
Figure 9.
10.00 Range of
8.00 adsorbent dose
in the dataset
6.00
collected
4.00
2.00
0.00
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
Adsorbent Type

33

Ashu Tamrakar

As can be seen from the Figures above, despite the poor statistical indices, the distribution of data
across pH, dye concentration, contact time and adsorbent dosage ranges is actually reasonable.
However, it must be noted that the adsorption experiments across temperature ranges were very
difficult to find and hence the results of the adsorbent performance across temperature conditions
might not be as accurate as other variables.

4.6 Data Preparation
A database (see Appendix A) of 1363 data points comprising of the above variables is collected and
compiled. The final step of the database collection phase is data preparation for neural network
through normalization. The database is normalized to suit the input requirements of ANN using the
formula:

𝑋 =
(𝑋 𝑖 −𝑋 𝑖,𝑚𝑖𝑛 )
𝑖,𝑛𝑜𝑟𝑚 (𝑋 𝑖,𝑚𝑎𝑥 −𝑋 𝑖,𝑚𝑖𝑛 )
[4]

where, Xi= original value of the variable i

Xnorm= normalized value of the variable

Xmax = maximum value of the variable

and, Xmin = minimum value of the variable

This normalized data is then used for training the network such that, the data will lie in the range of
0 to 1.0. Arbitrary numbers are assigned to all 22 adsorbents to facilitate in normalizing the data
available for the type of adsorbent.

34

Ashu Tamrakar

Chapter 5. NN modeling of adsorption

5.1 Model Development
The first step in the development of neural network is the formulation of the adsorption model to
use with the dye removal behavior. As previously discussed, many adsorption isotherm systems such
as the Langmuir and Freundlich isotherms have been developed over the years to describe the
adsorption capacity of sorbates. Although these classic adsorption isotherm models are capable of
representing equilibrium adsorption data sets the variable nature of adsorption behavior across the
type of adsorbents being used in this project presents a challenge to the development of an equation
that can be used to model the behavior of all adsorption processes. Thus, a general adsorption
model must be considered to successfully predict the trends between the dye removal efficiency of
different adsorbents.

For batch studies of adsorption data extracted from literature, the MB dye removal efficiency, DRE
(%) is described by the following function for this study:

DRE % = f (adsorbent material, pH, T, tc, Cd, Co) [5]

where, T = Temperature of the solution, C

tc = contact time of adsorption, minutes

Cd = adsorbent dosage, g/mL

and, Co = initial concentration of the dye, mg/mL

In this study, Neural Network Toolbox v.4.0 of MATLAB® mathematical software was used to
predict the adsorption efficiency using the normalized dataset prepared from the database collection
phase.

35

Ashu Tamrakar

5.2 Optimization of the NN architecture

Figure 10. Neural network schematic for dye adsorption model

For the present problem, there are six inputs and one output (DRE) variable. The back propagation
network is systematically trained and tested using various combinations of dataset distribution (see
section 3.4) and number of hidden layer to identify the optimum architecture for the network. The
optimal architecture of the ANN model and its parameter variation were determined based on the
minimum value of the R2 of the testing data set prediction set varying following parameters:

• Number of neurons in the Hidden layer:
• Distribution of data set into training set, validation set and test set

Figure 11 shows the contour plot of the neural network performance in terms of R2 correlation of
network predicted and actual experimental results of the testing data set. The larger the value of the
R2 in the figure the better the neural network is adept at predicting the value of DRE. It is evident
that if the number of neurons in the hidden layer is increased, the performance of the network
increases. However, after a limit the network becomes complicated and the performance plateaus
out illustrated by the green contours after the hidden layer is increased beyond 10 layers. In terms of
the dataset distribution, the 90% separation of dataset into training set hands down gave the best
results.

The best NN performances were observed at 8, 200 and 400 number of hidden layers and 90%
training dataset distribution.

36

37
Figure 11. Dependence of Neural network performance on the number of neuron at hidden
layer and the distribution of dataset separated for training the neural network.
Ashu Tamrakar

Ashu Tamrakar

Since the dataset distribution is being controlled randomly in the Neural network toolbox, it
becomes essential to check the reproducibility of the results obtained for the best NN architecture
determined. Consequently, 12 trials of neural networks were developed for each of the three hidden
layer- dataset distribution combination. Figure 12 show the performance of each combination in
terms of the R2 value of the testing data correlation.

Figure 12. Reproducibility check for the best NN architecture combinations

From figure, the best NN performance are obtained for 200 and 400 hidden layer terms, however,
since the amount of computation time increases exponentially with the increase of hidden layers,
200 was chosen to be the optimal number of terms in the hidden layer for dye adsorption model.

38

Ashu Tamrakar

5.3 Present Study Results
Using 200 hidden layers and by using 90% of the normalized dataset collected as training data, a
neural network model for the adsorption of MB dye by the 22 adsorbents was developed. Figure 13
shows the performance of the network thus developed.

Figure 13. Quality of NN predictions for the training, validation and testing dataset. The overall
performance of the neural network shows 89% accuracy between predicted values and literature
data.

39

Ashu Tamrakar

There are some important trends visible in the performance graphs that must be noted when
moving forward with the neural network model

i. The most important regression value to observe in the performance graphs is the
correlation coefficient (R2) value for the test data. Since these data points were not
trained, they represent the best accuracy of the model to predict new conditions. The
network currently shows 86.4% accuracy in predicting the dye removal efficiency.

ii. As can be seen in the training dataset graph (top left graph), the network seems to have
the hardest time fitting values close to the low and high point for the variables. This is
evident by the large discrepancy in the predicted output at target values close to 0 and 1.
Since the exact data points that have caused the error are unknown, it is important to
hence understand that the network predictions at upper and lower bound for any
variable is not as accurate as the predictions for the range in between.

This trend is also visible in the testing dataset where the prediction for the central region
is significantly better than at the bounds. This trend might have caused the lower R2
values for the training and the test data.

In general, this neural network provides a reasonable estimate of the individual adsorbent
performance of Methylene blue dye removal.

40

Ashu Tamrakar

Chapter 6. NN Applications

6.1 Comparative predictions of variable effect on removal efficiency
One of the simplest ways in which the neural network developed can be utilized is by predicting the
performance of adsorbents at different variable ranges. With the ability to now forecast the relative
efficiencies of each adsorbent at the same parameter range, a new possibility of ranking the
adsorbents is available. A ranking thus developed would be very helpful in determining the best
sorbent to use in a particular waste water condition hence optimizing the decision process for
effluent treatment systems.

For this project, a simple ranking model was developed by analyzing the dye removal efficiency (%)
of each adsorbent at the given conditions. The ranks are as follows:

1st preference adsorbents: 80% - 100% dye removal

2nd preference adsorbents: 60% - 80% dye removal

3rd preference adsorbents: 40% - 60% dye removal

4th preference adsorbents: 20% -40% dye removal

5th preference adsorbents: 0% - 10% dye removal

This ranking was used to categorize the 22 adsorbents tested on the basis of their performance
under each variable condition. The performances were first predicted by applying the neural
network model to each adsorbent type by changing the variable of interest over a desired range
while keeping the other conditions constant. For instance, to look at how all the adsorbents would
fare in a waste water of varying acidity, the pH input conditions from 1-11 were tested while keeping
the initial dye concentration at 195.9 mg/l, contact time at 167 mins, temperature at 25.9 and
adsorbent dose at 3.3 mg/l. The contour plots of the dye removal efficiency for changes in the 5
input conditions are illustrated in Figures 14-19 along with the best choice of adsorbent to use at
high and low conditions of the particular variable.

41

42
Figure 14. Prediction for the performance of adsorbents at different pH levels.. The circled
adsorbents are commercial carbons while the rest are non-conventional adsorbents.
Ashu Tamrakar

43
Figure 15. Prediction for the performance of adsorbents at different initial dye concentration levels.
Ashu Tamrakar

44
Ashu Tamrakar

Figure 16. Prediction for the performance of adsorbents at different contact times.

45
Ashu Tamrakar

Figure 17. Prediction for the performance of adsorbents at different temperature conditions.

46
Figure 18. Prediction for the performance of adsorbents at different adsorbate dosage conditions.
Ashu Tamrakar

Ashu Tamrakar

6.2 Results
The spectrum below visulizes the capacity of each adsorbent at different variable conditions
presented in the above figures. The color spectrum is dependant on the ranking of the adsorbent at
the condition specified. The detailed ranking is provided in Appendix C.

Table 6. Ranking spectrum of each adsorbent at various variable conditions
ID # pH Dye conc. (mg/l) Contact time (min) Temperature(C) Dose (g/l)
1-3 4-7 8-11 0 -20 30 - 2000 - 0 -20 30 - 2000 - 20-26 29-38 41-50 0-2 3-6 7-10
150 500 150 500
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
Legend:
1st preference 4th preference
2nd preference 5th preference
3rd preference

47

Ashu Tamrakar

Works Cited
1. Improving the Efficinecy of Wastewater Treatment Process by Soft Computational Methods. Attarzadeh, S.
and Jalalinia, F. 3, 2011, International Journal of Modeling and Optimization, Vol. 1, pp. 180-184.
2. Non-Conventional low-cost adsorbents for dye removal: A review. Crini, Gregorio. 2005, Elsevier, pp.
1061-1085.
3. Evaluation of the use of loofa activated carbons as potential adsorbent for aquesous solutions containing dye.
Abdelwahab, Ola. 2008, Desalination 222, pp. 357-367.
4. Database Mining UsingSoft Computing Techniques. An Integrated Neural Network- Fuzzy Logic-Genetic
Algorithm Aproach. Cundari, T.R. and Russo, M. 2001, J. Chem. Inf. Comput. Sci, Vol. 41, pp. 281-
287.
5. Toxicity Assessment and microbial degradation of azo dyes. Puvaneswari, N., Muthukrishnan, J and
Gunasekaran. Madurai, India : Indian Journal of Experimental Biology, 2006, Vol. 44, pp. 618-626.
6. Little, A.D. Executive Summary of Methylene Blue . National Toxicology Program: Deparment of
Health and human Services. [Online] [Cited: 4 25, 2012.]
http://ntp.niehs.nih.gov/?objectid=03DB4384-0364-AB0B-5C71EF4A37D6888A.
7. Goetz, Charity. Textile Dyes: Techniques and their effects on the Environment with a Recommendation for
Dyers Concerning the Green Effect. Lynchburg, VA : Liberty University, 2008.
8. Gupta, S. and Babu, B.V. Economic Feasilibilty analysis of low cost adsorbents for the removal of Cr(VI)
from wastewater. Birla Institute of Technology and Science (BITS). Rajasthan, India : s.n. p. 7,
Research article.
9. Modeling of Nitrate Adsorption on Granular Activated Carbon (GAC) using Artificial Neural Network
(ANN). Khataee, A. and Khani, A. 2009, International Journal of Chemical Reactor Engineering,
Vol. 7, p. Article A5.
10. Neural Network Modeling and Simulation of the Solid/Liquid Activated Carbon Adsorption Process .
Kumar, K.V., porkodi, K. Rondon, R.L.A. and Rocha, F. 2008, Ind. Eng. Chem. Res., Vol. 47,
pp. 486-490.
11. Artificial Neural Network (ANN) approach for modeling of Pb(II) adsorption from aqueous solution by Antep
Pistachio (Pistacia Vera L.) shells. Yetilmezsoy, K. and Demirel, S. 2008, Journal of Hazardous
Materials, Vol. 153, pp. 1288-1300.
12. Babu, B.V., Ramakrishna, V. and Chakravarthy, K.K. Artificial Neural Networks for Modeling of
Adsorption. Pilani-Rajasthan : Birla Institute of Technology & Science.
13. Hasan, M.B. Adsorption of reactive azo dyes on Chitosan/oil-palm ash composite adsorbent: batch and
continuous studies. Malaysia : Universiti sains Malayisa, 2008.
14. Decolorization of water/wastewater using adsorption. Allen, S.J. and Koumanova, B. 3, Belfast, UK :
Journal of the University of Chemical Technology and Metallurgy, 2005, Vol. 40, pp. 175-192.
15. IUPAC. Compendium of Chemical Terminology. Chromophore. [Online] IUPAC, March 23,
2012. [Cited: April 28, 2012.] http://goldbook.iupac.org/C01076.html.

48

Ashu Tamrakar

16. Biological Remediation of dyes in Textile effluent: a review on current treatment technologies. Miao, Yongjie.
s.l. : Water, 1992.
17. Saruya, K. and Sandhya, S. Potential of facultative microorganisms for biotreatment of textile
wastewater. Environmental Information System Centre. [Online] Department of Zoology, University of
Madras, India, April 2009. [Cited: April 28, 2012.]
http://www.envismadrasuniv.org/nl20092%20articles%20Potential%20of%20facultative.htm.
18. National Center for Biotechnology Information. Methylene Blue- Compound Summary.
PubChem Compound Summary. [Online] U.S. National Library of Medicine. [Cited: April 28, 2012.]
http://pubchem.ncbi.nlm.nih.gov/summary/summary.cgi?cid=6099.
19. National Toxicology Program. Production and Use of Methylene Blue. Executive Summary:
Methylene Blue. [Online] Department of Health and Human Services. [Cited: April 28, 2012.]
http://ntp.niehs.nih.gov/ntp/htdocs/Chem_Background/ExecSumm/MethyleneBlue/MethBlue_p
rod.html.
20. ProvenPharma. ProveBlue: The first pure pharma grade Methylene Blue produced under
cGMP in the World. Drug Substance. [Online] Proven Technologies Group, 2007. [Cited: April 28,
2012.] http://www.provepharm.com/substance.php.
21. Gupta, S. and Basu, B.V. Removal of Toxic Metal Cr(VI) from industrial wastewater using sawdust as
adsorbent: Equilibrium, kinetics and regeneration studies. Rajasthan, India : Birla Institute of Technology
and Science.
22. Verification and validation of neural networks: a sampling of research in progress. Taylor, B., Darrah, M.
and Moats, C. Fairmont, WV : s.n., 2003. Intelligent Computing: Theory and Application.
23. Industrial sludge based adsorbents/ industrial by-products in the removal of reactive dyes- A review.
Geethakarthi, A. and Phanikumar, B.R. 1, 2011, International Journal of Water Resources and
Environmental Engineering , Vol. 3, pp. 1-9.
24. Low Cost Biodegradable Adsorbent Material for the removal of dissolved Dyes from Aqueous solutions: An
Economical Process. Bozlur, R.M., et al. 5, 2010, International Journal of Engineering and
Technology, Vol. 2, pp. 468-473.
25. Removal of Methylene Blue Dye from simulated wastewater by alginic acid fiber as adsorbent: Equilibrium,
Kinetic and Thermodynamic Studies. C., Guo., et al. 2011, The Canadian Journal of Chemical
Engineering, Vol. 9999, pp. 1-9.
26. Rabunal, J.R. and J., Dorado. Artificial Neural Networks in Real-Life Applications. [ed.] M. Potter,
et al. Hershy,PA : Idea Group Publishing, 2006.
27. Drinking water quality and treatment: the use of artificial neural networks. Baxter, C.W., et al. Suppl. 1,
2001, Can. J. Civ. Eng, Vol. 28 , pp. 26-35.
28. Genetic Algorithms and Neural Networks based Optimization applied to the wastewater Decolorization by
Rhotocatalytic Reaction. Suditu, G.D., et al. 2008, Rev. Chim., Vol. 7, pp. 816-825.
29. Modling of the kinetics of a photochemical water treatment process by means of artificial neural networks. Gob,
S., et al. 1999, Chemical Engineering and Processing, Vol. 38, pp. 373-382.
30. RSM and ANN modeling for electrocoagulation of copper from simulated wastewater: Multi objective
optimization using genetic algoritm approach. Bhatti, M.S., et al. 2011, Desalination, Vol. 274, pp. 74-80.

49

Ashu Tamrakar

31. Simulation of an industrial wastewater treatment plant using artificial neural networks and principle components
analysis. Oliveira-Esquerre, K.P., Mori, M. and Bruns, R.E. 4, 2002, Brazilian Journal of
Chemical Engineering, Vol. 19, pp. 365-370.
32. The use of artificial neural network (ANN) for modeling of COD removal from antibiotic aquoeus solution by
the Fenton process. Elmolla, E.S., Chaudhuri, M. and Eltoukhy, M.M. 2010, Journal of Hazardous
Materials, Vol. 179, pp. 127-134.
33. Rajkumar, T. and Bardina, J. Training data requirement for a neural network to predict aerodynamic
coefficients. SAIC @ NASA Ames Research Center. Moffett Field, CA : s.n. Research Article.
34. Visualizing the function computed by a feedforward neural network. Plate, T., et al. Springer Verlag : s.n.,
1997. The Fourth International Conference on Neural Information Processing (ICONIP'97). pp.
206-309.
35. Jiwalak, Naparat. A Study of the Adsorption of Indigo and Indigo derivatives onto silk. Suranaree
University of Technology. 2010. Phd Thesis.
36. National Facility for Marine Cynobacteria. Role of marine cynobacterial antioxidants in azo
dye decolorization and degradation. [Online] January 30, 2010. [Cited: April 28, 2012.]
http://www.clri.org/LERIG2010/Lectures/Dr.UMA.pdf.

50

Ashu Tamrakar

Appendix A

Raw data collected for the 22 adsorbent

Adsorbent 1: Treated Sawdust

51

Ashu Tamrakar

Adsorbent 2: Teak Tree

52

Ashu Tamrakar

Adsorbent 3 and 4: Chitosan

53

Ashu Tamrakar

Adsorbent 5: Carbon nanotubes

54

Ashu Tamrakar

Adsorbent 6: Acid Activated Algerian Bentonite

55

Ashu Tamrakar

Adsorbent 7: Wheat Bran

56

Ashu Tamrakar

Adsorbent 8: Biosolids

57

Ashu Tamrakar

Adsorbent 9: Coconut Husk

58

Ashu Tamrakar

Adsorbent 10 and 11: Walnut Sawdust (SD) and sawdust coated with Polypyrrole

59

Ashu Tamrakar

Adsorbent 12: Water-washed manganese module leached residue

60

Ashu Tamrakar

Adsorbent 13: Sunflower Stalk

Adsorbent 14: Titanium Phosphate

61

Ashu Tamrakar

Adsorbent 15 and 16: Sepoilite and Pansil

62

Ashu Tamrakar

Adsorbent 17:Loofa

63

Ashu Tamrakar

Adsorbent 18 and 19:Raw and Chemically treated sugarcane bagasse

64

Ashu Tamrakar

Adsorbent 20: Giant duck weed

65

Ashu Tamrakar

Adsorbent 21 and 22: Sulfuric acid treated and Formaldehyde treated Indian Rosewood
sawdust

66

Ashu Tamrakar

Appendix B

Matlab function file for creating the neural network:

To make NN can either use the neural net GUI in Matlab or use the following function file:

function net = create_fit_net(inputs,targets)
%CREATE_FIT_NET Creates and trains a fitting neural network.
% NET = CREATE_FIT_NET(INPUTS,TARGETS) takes these arguments:
% INPUTS - RxQ matrix of Q R-element input samples
% TARGETS - SxQ matrix of Q S-element associated target samples
% arranged as columns, and returns these results:
% NET - The trained neural network
% For example, to fit the latest dataset problem with this function:
% load NewInputs
% load NewOutputs
% net = create_fit_net(NewInputs,NewOutputs);
% NewTargets = sim(net,Newnputs);
% To reproduce the results you obtained in NFTOOL:
%
% net = create_fit_net(New_inputs,New_outputs);

% Create Network
numHiddenNeurons = 100; % 100 seems to be the best right now
net = newfit(inputs,targets,numHiddenNeurons);
net.divideParam.trainRatio = 70/100; % can be adjusted as desired
net.divideParam.valRatio = 15/100; % can be adjusted as desired
net.divideParam.testRatio = 15/100; % can be adjusted as desired

% Train and Apply Network
[net,tr] = train(net,inputs,targets);
outputs = sim(net,inputs);

% Plot
plotperf(tr)
plotfit(net,inputs,targets)
plotregression(targets,outputs)

67

Ashu Tamrakar

To use the NN developed to produce new results:

Simply use:
Estimate=sim(NET1, Condition) %where condition is the matrix to define the condition to be simulated,

% To see the graph of the result
plot(PA(1,:),A), xlabel('Type of adsorbent'), ylabel('% Dye Removal'), legend('same pH, conc, dose and time')

68

Ashu Tamrakar

Appendix C

Detailed ranking of each adsorbent on various parameter ranges

pH Dye conc. (mg/l) Contact time (min) Temperature(C) Dose (g/l)
ID
# 1-3 4-7 8-11 0- 30 - 2000 0- 30 - 2000 20- 29- 41- 0-2 3-6 7-
20 150 -500 20 150 -500 26 38 50 10
1 5 4 1 4 3 2 4 2 1 4 1 1 3 3 5
2 5 4 1 3 3 3 5 3 1 4 1 1 4 3 5
3 5 4 1 1 2 4 5 3 1 4 1 1 5 3 5
4 5 4 1 1 1 4 5 3 1 4 1 1 5 2 3
5 5 4 1 1 1 4 5 3 1 4 2 1 5 1 1
6 5 4 1 1 1 4 4 2 1 4 3 3 5 1 1
7 5 3 1 1 1 4 3 1 1 3 5 4 3 1 1
8 5 3 1 1 1 3 3 1 1 3 5 5 2 1 1
9 5 3 1 1 1 2 3 1 1 3 5 5 1 1 1
10 5 4 1 2 2 2 3 2 1 3 5 5 1 1 2
11 5 4 1 2 2 1 4 2 1 3 5 5 1 2 3
12 5 4 1 2 2 1 4 2 1 3 5 5 1 3 2
13 5 4 1 1 1 1 3 1 1 2 5 5 1 2 2
14 5 3 1 1 1 1 2 1 1 2 5 5 1 2 1
15 5 3 1 1 1 1 2 1 1 2 4 3 1 1 1
16 5 3 1 1 1 2 2 1 1 3 2 1 1 1 1
17 5 3 1 1 1 2 3 2 1 3 1 1 1 1 1
18 3 2 1 1 1 2 3 2 1 4 1 1 2 1 1
19 1 2 1 1 1 2 3 2 1 4 1 1 3 1 1
20 1 2 1 1 1 2 3 2 1 4 1 1 4 1 1
21 1 2 1 1 1 2 2 2 1 4 1 1 5 1 2
22 2 2 1 1 1 3 2 2 1 4 1 1 5 2 2

69

Predictive modeling of methylene blue dye removal by using artificial neural network

More Related Content

What's hot (15)

Viewers also liked (20)

Similar to Predictive modeling of methylene blue dye removal by using artificial neural network (20)

Recently uploaded (20)

Predictive modeling of methylene blue dye removal by using artificial neural network