Comparative study of aArtificial nNeural nNetwork (ANN) and aAdaptive

Nneuro-fFuzzy iInference sSystem (ANFIS) and mMultiple lLinear

rRegression (MLR) for modelling of Cu (II) adsorption from aqueous
solution using biochar derived from rambutan (Nephelium lappaceum) peel

Yong Jie Wonga,

Senthil Kumar Arumugasamyb*,
[email protected]
Chang Han Chungb,
Anurita Selvarajoob, and
Vasanthi Sethub
a
Research Center for Environmental Quality Management, Graduate School of
Engineering,,
Graduate School of EngineeringKyoto University,,
Kyoto University, 1-2 Yumihama, Otsu, Otsu, Shiga 520-0811, Japan
b
Department of Chemical and Environmental Engineering, Faculty of Engineering,
University of Nottingham Malaysia,
Jalan Broga, 43500 Semenyih, Selangor, Malaysia.

Abstract
Presence of copper within water bodies deteriorates human health and degrades natural
environment. This heavy metal in water is treated using a promising biochar derived
from rambutan (Nephelium lappaceum) peel through slow pyrolysis. This research
compares the efficacies of artificial neural network (ANN), adaptive neuro-fuzzy
inference system (ANFIS), and multiple linear regression (MLR) models and evaluates
their capability in estimating the adsorption efficiency of biochar for the removal of Cu
(II) ions based on 480 experimental sets obtained in a laboratory batch study. The
effects of operational parameters such as contact time, operating temperature, biochar
dosage, and initial Cu (II) ion concentration on removing Cu (II) ions were investigated.
11Eleven different training algorithms in ANN and 8 different membership functions in
ANFIS were compared statistically and evaluated in terms of estimation errors, which
are root mean squared error (RMSE), mean absolute error (MAE), and accuracy. The
effects of number of hidden neuron in ANN model and fuzzy set combination in ANFIS
were studied. In this study, ANFIS model with Gaussian membership function and
fuzzy set combination of [4 5 2 3] was found to be the best method, with accuracy of
90.24% and 87.05% for training and testing dataset, respectively. Contribution of this
study is that ANN, ANFIS, and MLR modeling techniques were used for the first time
to study the adsorption of Cu (II) ions from aqueous solutions using rambutan peel
biochar.

Keywords
Adsorption
Artificial nNeural nNetwork (ANN)
Adaptive neuro-fFuzzy iInference sSystem (ANFIS)
Multiple lLinear rRegression (MLR)
Biochar
Cu (II) ions
Rambutan peel

Introduction
The presence of heavy metals within water bodies around the world attributes
towards the deterioration of human health (Minh-Viet Nguyen 2014). These heavy
metals originate from the major production sectors, such as paper production, chemical
production, mining, electronic circuit board production, metal extrusion, and battery
production industries (Mandu Inyang et al. 2012). Several potent contaminants have
been identified to be regulated in drinking water which include copper, arsenic, lead,
barium, and nickel (USEPA, U. S. E. P. A 2016). Copper, or Cu (II) ions, are of upmost
importance due to the rise in electronic industries throughout the world which sees the
increase in the use of smartphones and electronic devices. Copper is needed in relatively
small amounts in the human body for muscle growth, synthesis of enzymes, and
development of bone structure (Muhammad Bilal et al. 2013).
 On the contrary, high Cu (II) ion intake will pose severe health risks and
damages to human health. Some of the immediate effects of high Cu (II) ion intake
include diarrhoea, nausea, and dizziness (USEPA 2016). USEPA has also found that
long-term exposure to high concentrations of Cu (II) ions leads to liver and kidney
damage due to the accumulation of Cu (II) ion in the liver and failure of the kidney in
regulating concentration of Cu (II) ions in the bloodstream. There are multiple methods
in removing Cu (II) ions that isare present in the water bodies, and every method has its
own advantages and disadvantages as shown in Table 1. Recent studies show that
biosorption could be an effective and low- cost technology in removing heavy metals
from water bodies (Muhammad Bilal et al. 2013). A Pprevious studiesy by Amit
Bhatnagar and Witek-Krowiak (2015) has demonstrated that agricultural waste peels of
orange, pomelo, grapefruit, and banana prove to be highly effective in removing heavy
metals and dyes from water bodies. When these fruit peels are turned into biochar, their
metal ion removal efficiency is known to improve, due to the increased surface area and
exposed active groups (Selvarajoo and Arumugasamy, 2019). Many researchers have
demonstrated that biochar is a very effective low- cost adsorbent in removing heavy
metals from water bodies (Foereid 2015; Krishna et al. 2014; Mandu Inyang et al. 2012;
Minh-Viet Nguyen 2014; Morrow 2013; Pusker Regmi et al. 2011).
Biochar is the carbon-rich product obtained through thermal decomposition of
biomass at high temperature in a closed system with limited or no oxygen (Berger 2012;
Chinmai et al. 2014; Feng Li et al. 2016; Joesph et al. 2014). In this work, biochar is
produced from the slow pyrolysis of rambutan peel (Nephelium lappaceum), which is
used due to its low cost and abundant availability. The rambutan biochar was then tested
for the adsorptive removal of Cu (II) ions from aqueous solutions. The effects of various
operational parameters such as contact time, operating temperature, biochar dosage, and
initial Cu (II) concentration on removing Cu (II) ions were investigated.

Artificial intelligence (AI) techniques, for example artificial neural network (ANN) and
adaptive neuro-fuzzy inference system (ANFIS), are the emerging technologies in the
field of environmental engineering in the past decades. AI models are capable of
analyszing the complex and highly non-linear relationship between variables. Hence, AI
models which do not require abundant knowledge about the process, and rather only
need the input-output data to be inserted into the model. Because of its robustness and
reliability, AI techniques have been implemented on the removal of Cu (II) ions (Bingöl
et al. 2013; Fan et al. 2017; Ghosh et al. 2015; Rezaei et al. 2017; Turan and Ozgonenel
2013; Turan et al. 2011; Shojaeimehr et al. 2014). A multilayer feed-forward neural
network (FANN) model was developed for the pyrolysis of durian rinds (Arumugasamy
et al. 2015). To the best of the authors’ knowledge, there is no published research work
on the effect of different types of training algorithms and the types of membership
functions (MF) and fuzzy set combination on the performance of the ANN and ANFIS
model, respectively, for Cu (II) ion adsorption by biochar derived from rambutan peel.
There has been some work on the application of the feedforward neural network (FANN)
modelling for the biosorption of Cu (II) ions with rambutan peel biochar (Selvanathan,
et al. 2017), but only the Levenberg Marquardt training algorithm was used. Therefore,
the aim of this study was to develop the model of the adsorption behaviour of biochar
derived from rambutan peel by using MLR, ANN, and ANFIS models. A total of 11
different training algorithms were used in the ANN model, and 8 different MFs were
compared and statistically evaluated in terms of estimation errors, which are root mean
squared error (RMSE), mean absolute error (MAE), and symmetrical mean absolute
percentage error (SMAPE) with the experimental results.

Materials and mMethods

Materials
Rambutan peels were chosen as the raw materials for pyrolysis. Though this
fruit is native to Malaysia, it can be easily found throughout the Southeast Asia region.
Rambutan trees are capable of reaching a height of 6 metersm which are considered
really gigantic for fruit trees (Bhayana 2014). Due to the various benefits of rambutan, it
is widely consumed throughout the world. However, since the rambutan peels are
inedible, they are discarded. These rambutan peels are not of any use to the society so
far, which is why they were discarded,; however, the good news is these rambutan peels
can be converted to a more useful form, which is biochar. This can be seen as a long-
term solution for the increasing rambutan waste peels that goes to the landfill.
A farm in Selangor, Malaysia, provided fresh rambutan peels for this study.
These were cleaned to eliminate any soil present within the peels and then sun-dried for
2 days to remove any excess moisture. A hot air oven (Memmert, Germany) was used
for another day on the rambutan peels to achieve uniform and complete drying. The
moisture content of the dried rambutan peels should be less than 10% on a dry basis to
prevent the proliferation of fungus and organisms (Gary Yeager and R. L. W. 1981).
The dried samples were ground using a rotor beater mill (Retsch SR 300) to fine
particles with size of ≤ 5 mm. The samples were then sent for thermogravimetric
analysis (TGA) to identify the fixed carbon, volatile matter content, moisture content,
and ash content.

Slow pyrolysis
The pyrolysis of the samples was conducted using an electric tubular furnace.
The main component of the furnace is the stainless steel cylindrical reactor of length
1200 mm, inner diameter of 50 mm, and outer diameter of 60 mm. It was placed in the
electric tubular furnace. A thermocouple was placed in the middle of the cylindrical
reactor to determine the sample temperature at any given time. A nitrogen cylinder
which supplies nitrogen gas to the cylindrical reactor was connected to the inlet of the
reactor. The nitrogen gas which serves as purging gas was passed through the reactor at
a flowrate of 2 L/min. The two main reasons for using nitrogen gas in this experiment
waesre to purge out any oxygen that is present in the reactor before the start of the
pyrolysis experiment and to avoid any undesired reactions. The outlet of the cylindrical
reactor was connected to a volumetric flask which contains distilled water, which in turn
was connected to another volumetric flask that contains silica gel. The experiment was
set up in this manner to cool and scrub the gaseous effluents before being released into
the atmosphere.
Each pyrolysis experiment was conducted using 30 g of samples. These samples
were loaded onto the metal support and placed at the center of the cylindrical reactor to
achieve uniform heating on the samples. A layer of vacuum grease was applied onto the
connecting flanges before they are tightened to prevent any leakages that might happen.
The pyrolysis experiment was conducted by subjecting the samples to a temperature of
600 °°C. The furnace was heated at a rate of 5 °°C/min and held for 3 hoursh to allow
complete pyrolysis. The furnace shuts down automatically after the 3 hoursh of holding
time. The resulting biochar from the pyrolysis experiment was left to cool down to room
temperature before being weighing again. The biochar yield was calculated using
Eq.uation (1). The biochar was further grounded using an ultra-centrifugal mill to the
size of 45 μm for efficient adsorption of Cu (II) ions.
𝑚
𝑌𝑖𝑒𝑙𝑑 (%) = 𝑚1 × 100 (1)
2

where,

m1 the mass in grams of the biochar;

m2 the mass in grams of the sample before pyrolysis

Cu (II) ion adsorption

The metal adsorption experiments were conducted in batches. Stock solution of
1000 ppm of copper (II) sulphfate (CuSO4), was prepared by dissolving anhydrous
CuSO4 in deionized water. The CuSO4 solution was further diluted by 10-fold with
deionized water. The experiments were conducted using 500 mL conical flasks. 0.1 g
of biochar was dosed into each conical flask containing 100 mL of 100 ppm CuSO4
solution. The mixture was then agitated at 120 rpm using a water bath shaker at room
temperature. The experiments were repeated using different concentrations of biochar
(0.04 g, 0.08 g, 0.2 g, 0.4 g) to investigate the effect of biochar concentration on Cu
(II) ion adsorption at a given Cu (II) ion concentration.
The optimum biochar concentration was determined based on the tabulated
results. Another set of experiments were conducted at a different concentrations of Cu
(II) ions. CuSO4 solutions of 50 ppm are used in lieu of 100 ppm. The 50 ppm CuSO4
solution was prepared by diluting the stock solution of 1000 ppm by 20-fold with
deionized water. The adsorption experiments were conducted for a period of 1 hourh.
1One milliliter of aliquot is collected from each conical flask at every 10- minutesmin
interval for the first hour and at every 30- minutesmin interval for the second hour. The
1 mL aliquot was filtered using Whitman no. 1 filter paper to remove any residual
biochar and further diluted with additional deionized water to make up aliquot of
10 mL. The aliquots are tested for Cu (II) ion content using atomic absorption
spectrometry (AAS).

Atomic absorption spectroscopy (AAS)

AAS is used for quantitative determination of elements that are present in an aqueous
solution. This instrument relies on flame for atomic absorption determination. In a
flame atomic absorption spectrophotometer, the UV-Vis light from the source lamp is
passed through the flame that contains free atoms which have been generated by heat
dissociation of various species of elements that are present within the aqueous solution.
Any atom of the specific element that is excited at the wavelength of the source of light
will selectively absorb the radiation. The light emitted from the AAS lamp is made up
of a number of discrete emission lines. One of these emission lines is the light which
has been chosen to pass through the flame, and this causes absorption to occur. The
hollow cathode lamp contains traces of a particular element to be analyzed. The heat
from the hollow cathode lamp causes the element to emit light energy. This energy will
be strongly absorbed if the sample contains this element due to similar orbital energy.
This shows that the instrument is highly sensitive and highly selective due to the fact
that elements at different energy levels will not absorb the light emitted.

During the experiment, AAS is used to determine the final concentration of Cu

(II) ion in the treated water sample after the biosorption process using the biochar. The
final concentration of Cu (II) ion in the treated water signifies the efficiency of the
biochar in removing Cu (II) ions from aqueous solution.

2.5 Artificial nNeural nNetwork (ANN)

Artificial neural network (ANN) is a biologically inspired model that consists of
interconnected group of artificial neurons for solving complex problems, including but
not limited to pattern classification, clustering, function approximation, forecasting, and
image completion in many disciplines (Basheer and Hajmeer 2000; Dastorani et al.
2010). ANN will be able to learn from the data, capturing the essential linear and non-
linear relationship between data and providing a reliable prediction for new situations,
even with noise or insufficient information (Iphar 2012).
Among a vast number of ANNs, i.e., radial basis function neural network, Kohonen
self-organizing neural network, recurrent neural network, convolutional neural network,
modular neural network, and feedforward backpropagation network (FFBN), FFBN was
chosen in this study, as this architecture is reported to be most suitable for problems
based on pattern matching and pattern prediction (Esmaeili et al. 2012). Similar to
biological neural network of the human brain, Fig.ure 1 presents an FFBN that is made
up of (i) input layer with nodes that representing input variables (time, temperature,
initial Cu (II) concentration, and biochar dosage); (ii) hidden layer with nodes to
identify the nonlinearity in the data and; (iii) output layer with nodes representing
dependent variable (final Cu (II) concentration) in this study. In FFBN, each neuron in
each layer are is connected unidirectionally, such that the weighted sum of inputs is
combined with a bias, then passing through a transfer function to produce an output
without feedback to the input neurons.

2.5.1 Hidden lLayer and hHidden nNeuron sSelection

Network topography in ANN modelling, for instance transfer function, epochs, learning
rate, and especially number of hidden layer and hidden neurons are the most important
parameters that affects the performance of the model (Tokatli et al. 2009; Uzuner and
Cekmecelioglu 2016). Hidden layers are the core in ANN, as the learning capability of
the network comes from here. Thus, selection of a suitable hidden layer is important. In
general, one hidden layer is selected for most of the process as two hidden layers are
generally used for modelling data with discontinuities and it may introduce a great risk
of convergence in a local minima (Heaton 2008). Therefore, in this work, an FFBN with
one hidden layer is used. Besides that, number of neurons will be greatly affecting the
performance of the network too. A high number of hidden neurons will ensure the
training data to be well performed,; however, the training time required will be
excessively long and, worse, the overfitting may occur. On the other hand, if having too
few hidden neurons, underfitting or unsatisfactory convergence may occur. As reported
in literature on the number of neurons used for adsorption modelling, optimum number
of neurons normally lays between 1 and –-20 (Podstawczyk et al. 2015; Bagheri et al.
2016; Chowdhury and Saha 2013; Khan et al. 2017). A few studies have reported the
Cu (II) ions absorption from leachate by pumice,; the optimum number for hidden
neurons was 8 with Levenberg-Marquardt backpropagation training algorithm (N.
Gamze Turan et al. 2011) whereas adsorption of Cu (II) ions using light expended clay
aggregate was 4 (Shojaeimehr et al. 2014). Therefore, in this study, the number of
hidden neurons was chosen based on trial and error method, by varying the neurons
from 1 to 20.

Training aAlgorithms
Training algorithm is a method used to process data to extract patterns and adjust the
network parameters until optimum in order to adapt a system to a specific input-output
transformation task (Salim 2016a, b; Ghaffari et al. 2006). Depending on the data type,
complexity, and number of data available, there is no one algorithm that works best for
every problem. Thus, 11 training algorithms belonging to six classes, namely, (1)
Levenberg-Marquardt; (2) gradient descent with variable learning rate; (3) resilient
backpropagation; (4) conjugate gradient descent; (5) Quasi-Newton algorithm; and (6)
Bayesian regulation, were compared and evaluated. The brief description of the 11
training algorithms is summarisedsummarized in Table 2.

Adaptive neuro-fFuzzy iInference sSystem (ANFIS)

Adaptive neuro-fuzzy inference system (ANFIS) is a hybrid intelligent system that
integrates the learning capability of ANN from given training data, and solution is
mapped out onto a fuzzy inference systems (FIS). FIS are classified into three types:
Mamdani’s system, Tsukamoto’s system, and Sugeno’s system (Amirkhani et al. 2015;
Prasojo et al. 2017; Uzuner and Cekmecelioglu 2016). Among these FIS, the Sugeno’s
system the most popular as it is more compact and efficient. Thus, the Sugeno-type FIS
was used in this study.

FIS is the central part of ANFIS, which is expressed in terms of “If-Then” principle and
can be used to predict the behaviour and nature of the system. ANFIS used in this study
has four inputs and single output and the rule set is defined as follows:

Rule 1: IF w is A1 AND x is B1 AND y is C1 and z is D1

THEN f1 = P1w + Q1x + R1y + S1z + T1.
Rule 2: IF w is A2 AND x is B2 AND y is C2 and z is D2
THEN f2 = P2w + Q2x + R2y + S2z + T2

...and so on.
where w, x, y and z are the inputs; Ai, Bi, Ci, and Di are the fuzzy sets; fi is the output;
and Pi, Qi, Ri, Si and Ti are the design parameters that are determined during training.

ANFIS model has five layers and the schematic diagram of the proposed ANFIS is
indicated in the Fig.ure 2. Each layer is descried as follows (Uzuner and Cekmecelioglu
2016; Amid and Mesri Gundoshmian 2017):

Layer 1: consists input variable MFs and to provide inputs to next layer
Layer 2: computes the weight for each MF
Layer 3: calculates outputs in normalized firing strength
Layer 4: computes parameters function of the outputs of layer 3
Layer 5: computes simple summation of the outputs of layer 4

For selection of best fit ANFIS model, 8 different types of MFs were tested and the
fuzzy set combinations of MFs were varied systematically from [2 2 2 2] to [5 5 5 5] for
input variables, then were compared and evaluated. The input-output mapping
relationship in this study was done through hybrid learning which is the combination of
backpropagation and the least square solution in order to determine the optimal
distribution of the MFs (Amid and Mesri Gundoshmian 2017; Iphar 2012; Walia et al.
2015).

Membership fFunction
Membership function (MF) is the most essential concept of the fuzzy sets theory,; i.e.,
fuzziness in a fuzzy set is determined by its MF. It specifies the relationship of an
element to a set graphically and defines how the feature of fuzzy set is mapped on a
membership value or degree of membership on a continuous scale from 1 (full
membership) to 0 (no membership) (Noor and Hamid 2014; Kurtener et al. 2008; Ali et
al. 2015). To generate maximum accuracy, researchers have developed various shapes
of membership functions for instance triangular, trapezoidal, and π-shaped etc., suitable
for particular problem since they effect on a fuzzy inference system (Talpur et al. 2017).
Five different categories of MFs, namely, (a) piecewise linear functions, (b) Gaussian
distribution functions, (c) bell-shaped function, (d) sigmoidal function, and (e)
polynomial- based function, were used for modelling of Cu (II) adsorption from
aqueous solution using biochar derived from rambutan (Nephelium lappaceum). The
brief description of the 11 training algorithms is summarisedsummarized in Table 3.
 Multiple lLinear rRegression (MLR)
Multiple linear regression (MLR) is one of the earliest statistical techniques for
modelling the linear relationship between two or more independent variables and a
single dependent variable by using least square method. The following represents a
generalisedgeneralized MLR equation.

𝑦 = 𝛽0 + 𝛽1 𝑥1 + 𝛽2 𝑥2 + ⋯ 𝛽𝑖 𝑥𝑖 (12)

where y is the dependent variable, 𝛽0 is the regression constant, 𝛽𝑖 is the

coefficient of the independent variable, and 𝑥𝑖 is the independent variable.

Model dDevelopment for Cu (II) adsorption

The proposed algorithm for constructing the ANN, ANFIS, and MLR models for Cu (II)
adsorption in this study can be divided into three sections,: (i) data prepressing, (ii)
modelling; and (iii) accuracy evaluation as shown in Fig.ure 3.

Data pPre-processing
Data normalization is performed before data insertion into the model as the scaling of
input and output variables are is in different ranges. This action prevents the dominance
of larger value data overriding the smaller ones (Basheer and Hajmeer 2000; Mirrashid
2014). In this study, min-max normalization is adopted, and data were normalized in the
range of [0.1, 0.9], as shown in Eq. (22):
𝑥𝑖 −𝑥min
𝑦𝑖 = 0.8 × + 0.1 (23)
𝑥max −𝑥min

where yi is the = normalized value of certain variable, xi is the original data of

certain variable, and xmin and xmax are minimum and maximum values of xi respectively.

After normalization process, the data were divided randomly into two parts, training and
testing dataset which is used to learning and fitting the variable of classifier and to
access the performance of a fully specified classifier respectively. Out of the 480 data
samples, 70% (360) were used as training datasets and the remaining as testing datasets.

Evaluation of the model

In order to evaluate the performance of developed model statistically and into different
error projections (Chai and Draxler 2014), a combination of criteria, namely (1) root
mean squared error (RMSE), (2) mean absolute error (MAE), and (3) symmetrical mean
absolute percentage error (SMAPE), were calculated. (Y. J. Wong et al. 2018; Prasojo et
al. 2017).

1
RMSE = √ ∑𝑁
𝑖=1(𝐴𝑖 − 𝐹𝑖 )
2 (34)
𝑁

RMSE is a measure of the variation between the actual and predicted data and as an
indicator of absolute fit of the model. Lower values of RMSE indicate better fitting. It is
one of the most important evaluation criteriona if the main purpose of the model is
prediction.

1
MAE = 𝑁 ∑𝑁
𝑖=1|𝐴𝑖 − 𝐹𝑖 | (45)

MAE is taking the absolute value of the difference between the actual and predicted
data, making it less sensitive to the outliers.

100% |𝐴𝑖 −𝐹𝑖 |

SMAPE = ∑𝑁
𝑖=1 (6)
𝑁 (|𝐴𝑖 |+|𝐹𝑖 |)/2

Accuracy = 100% − SMAPE (76)

SMAPE is a measure of accuracy based on percentage or relative error. SMAPE is

developed based on mean absolute percentage error (MAPE); to overcome
shortcomings of MAPE, (i) measurement is not defined when there are zero or near-zero
data and (ii) heavier penalty on negative error when Ai < Fi (Hyndman and Koehler
2006).

where Ai = actual final copper (II) concentration, Fi = predicted final copper (II)
concentration, and N = number of samples.

Results and dDiscussion

ANN Mmodelling
As discussed in the earlier section, the number of hidden neurons as well as the type of
training algorithm, haves a great impact in affecting the accuracy of the network.
Numerous studies have been carried out to determine the optimum number of hidden
neuron in the network (Yong Jie Wong et al. 2019; Xinxing et al. 2013; Shibata and
Yusuke 2009),; however, the selection is case dependent in order to prevent over- or
under fitting. Therefore, during data processing, the number of neurons in the hidden
layer wereas adjusted in numerous iterations by trial and error method, ranging from 1
to 20, and different weightages for training and testing datasets were performed.

A box plot is constructed to graphically present the distribution and variation of the
prediction accuracy of ANN models under 11 different training algorithms with
different hidden neurons in this study, as illustrated in Fig.ure 4. The overall
performance in ANN models is satisfactory, and not much variation is observed in most
training algorithms except the traingdm and trainbr algorithms. A relatively large
variation and lower median prediction accuracy in both training and testing dataset is
observed in both training algorithms, which indicating that the training algorithms are
not suitable to extract the relationship among variables in the study. Similar results were
reported by Zhang et al. (2019) and Yang et al. (2019) where the traingdm algorithm is
the least performing algorithm which may be due to that gradient difference update
method in traingdm is insufficient to capture the characteristics of Cu(II) adsorption in
this study. Ttrainlm, traincgb, trainscg, and trainbfg algorithms outperforms other
algorithms in the training dataset in which trainscg algorithm showed the maximum
prediction accuracy.

To identify the optimum ANN structure in this study, the performance of ANN
modelling for 11 different training algorithms with its optimum number of neurons for
both training and testing datasets is presented in Table 4. SCG algorithm with 12
hidden nodes is selected as the best method for ANN in predicting the Cu (II)
adsorption by biochar derived from rambutan peel, with maximum R2 and accuracy and
minimum RMSE and MAE values. For training dataset, the R2 and accuracy values
obtained were 0.93 and 88.3%, whereas for testing dataset, the R2 and accuracy values
obtained were 0.90 and 86.6%. As shown in Fig.ure 5, even though the training
algorithm is suitable for the nature/behaviour of dataset, the effect of hidden neuron
number is significant.
 The graphical comparison between the simulation results and experimental data for
Cu (II) adsorption using the SCG method by FFBN is illustrated in Fig. 6. It can be
observed that there is a good agreement between the simulated results and experimental
data, and the differences between them are represented by Fig.ure 7. For the remaining
training algorithms, the comparison plot and residual plot between simulated results and
experimental data are illustrated in Figure S1-S10 in the supplementary material.

ANFIS Mmodelling
As the number of fuzzy set and type of MFs are important in the construction of the
ANFIS architecture, therefore, a systematic method was developed to investigate these
factors in this study. The number of MFs for each input variable was varied from [2 2 2
2] to [5 5 5 5] with a total of 256 combinations for one MF. Figure .8 presents a box
plot of ANFIS models under 8 different training algorithms with different fuzzy set
combinations. Compared towith ANN models, ANFIS models show a large variation of
prediction accuracy especially in the testing dataset. This suggests that the fuzzy set
combination is one of the key factors influencing the prediction accuracy.

To identify the optimum ANFIS structure in this study, the performance for 8 different
types of MF with its optimum fuzzy set combination for input variables is presented in
Table 5. It can be observed that GaussMF with fuzzy set combination of [4 5 2 3] was
the best method for ANFIS in predicting the final Cu (II) concentration, with maximum
R2 and accuracy and minimum RMSE and MAE values. For training dataset, the R2 and
accuracy values obtained were 0.95 and 90.24%, whereas for testing dataset, the R2 and
accuracy values obtained were 0.89 and 87.06%. As compared towith ANN modelling,
the fuzzy set combination in ANFIS has higher effect in prediction accuracy as shown
in Fig.ure 9.

The graphical comparison between the simulation results and experimental data for Cu
(II) adsorption using the Gaussian MF by ANFIS is illustrated in Fig.ure 10. Both
results are almost identical to each other, and the errors between simulated results and
experimental data are illustrated in Fig.ure 11. For remaining MFs, the comparison plot
and residual plot between simulated results and experimental data are illustrated in
Figure S11-S17.
 Comparison between ANN, ANFIS, and MLR
In this section, a prediction performance comparison is made between MLR model with
the best ANN and ANFIS models selected in the previous section (Fig. 12). The MLR
model constructed includes the independent variables similar to the ANN and ANFIS
models, and the best linear equation is presented by Eq. (6). Table 6 presents the MLR
performance in predicting the final Cu (II) concentration.

𝑌Final Cu(II) Conc = 0.427𝑋1 + 20.106𝑋2 − 0.004𝑋3 − 0.02𝑋4 − 8.157 (68)

where 𝑌Final Cu(II) Conc is the final Cu(II) concentration and 𝑋1 , 𝑋2 , 𝑋3 , and 𝑋4
represents the time, temperature, biochar discharge, and initial Cu(II) concentration,
respectively.

The performance of these models was compared referring to Tables 4, 5, and to 6. As

shown in Fig.ure 9, both ANN and ANFIS models were capable in predicting the final
Cu (II) concentration,; however, the prediction capacity of MLR model is relatively low.
MLR model is one of the most commonly used regression analysis due to its simplicity
and ease of interpretation as compared towith the ANN and ANFIS models,; however, it
has a significant limit, which is poorly suited to modelling the complex nonlinear
relationships in the adsorption process. Table 7 presents the Pearson’ correlation
coefficient between variables of the study, which ranges in [− -1, 1] where 0 indicates
minimum linear correlation and − -1 and 1 indicate the maximum negative and positive
linear correlation, respectively. It can be observed that most input variables
(temperature, biochar discharge, and initial Cu(II) concentration) hasve a very weak
correlation (< ± 0.3) with the output variable except the variable “time,”, which has a
moderate positive correlation (0.65) with the output variable (Ratner 2009). This is
implying that the data exhibits both linear and non-linearity, leading to the poor
prediction of MLR model.

In this study, it is found that the ANFIS model performs better than ANN model in both
training and testing stages. The main reason is that ANFIS model integrates both
advantages of neural network learning capabilities and fuzzy logic principles into one
framework, making it capable to establish a rating relationship between input variables
with the final Cu (II) concentration in the adsorption process. Besides that, one of the
major limitations of ANN is the lack of explanatory power, also be known as “black
box problem,” and some of these shortcomings are removed by the ANFIS model
(Dastorani et al. 2010). Numerous researches conducted in the different fields of science
also reported that the performance of ANFIS model is better and more capable in
dealing with complex problem (Uzuner and Cekmecelioglu 2016; Bilgehan 2011;
Esmaeili et al. 2012; Iphar 2012; Emamgholizadeh et al. 2014).

Besides that, referring to Figs.ure 5b and 8b, a large difference can be observed at
certain points. In both ANN and ANFIS models, the largest difference in the training
dataset is observed at the 239th sample which is 20.59 and 16.45 respectively. The large
difference can be caused by the human or technical error, which could be due to
presence of contaminant during experiment. Thus, to further improvinge the accuracy of
the models, more experiments should be conducted, and crosschecking should be
carried out for the inaccurate results.

Conclusion
This research was designed to evaluate the capability of ANN, ANFIS, and MLR in Cu
(II) adsorption modelling from aqueous solution using biochar derived from rambutan
peel. These techniques are rather new for heavy metal adsorption process using biochar
produced from agricultural wastes. The input variables used in both models were
contact time, operating temperature, biochar dosage, and initial Cu (II) concentration,
while the output was the final Cu (II) concentration from 480 datasets collected in a
laboratory batch study. In this study, it can be observed that systematic trial and error
approach for varying number of hidden neuron and fuzzy set combination is necessary
for both ANN and ANFIS models, respectively, in order to develop the best network
structure. SCG backpropagation training algorithm with 12 hidden neurons is selected
the best ANN model whereas Gaussian MF with fuzzy set combination of [4 5 2 3] is
selected as the best ANFIS model in this study. Both the optimum ANN and ANFIS
models are reliable and can be considered as a prediction tools for Cu (II) adsorption
process, however not for MLR model due to the complex relationship among variables
in this study. ANFIS model was found to be the best model, followed by ANN model
and MLR model, with values of RMSE, MAE, and accuracy of 3.29, 2.16, and 90.24%,
respectively, for training dataset and 4.50, 2.84, and 87.05%, respectively, for testing
dataset. The novelty of this study is that ANN, ANFIS, and MLR modeling techniques
were used for the first time in the adsorption behaviour of Cu (II) ions using biochar
derived from rambutan peel.

Compliance with ethical standards

Conflict of interest
There authors declare that they have is no conflict of interest between any of the authors
of this manuscript.

References

Ahmad, I., Ansari, M. A., & Mohsin, S. (2008). Performance comparison between
backpropagation algorithms applied to intrusion detection in computer
network systems. Paper presented at the pProceedings of the 9th WSEAS
iInternational cConference on nNeural nNetworks, Sofia, Bulgaria.

Ali, O. A. M., Ali, A. Y., & Sumait, B. S. (2015). Comparison between the eEffects of
dDifferent tTypes of mMembership fFunctions on fFuzzy lLogic cController
pPerformance. Int J Emerg Eng Res Technol, 3(3), 8.

Amid, S., & Mesri Gundoshmian, T. (2017). Prediction of output energies for broiler
production using linear regression, ANN (MLP, RBF), and ANFIS models.
Environ Prog Sustain Energy, 36(2), 577–-585, doi:10.1002/ep.12448.

Amirkhani, S., Nasirivatan, S., Kasaeian, A. B., & Hajinezhad, A. (2015). ANN
and ANFIS models to predict the performance of solar chimney power plants.
Renew Energy, 83, 597–-607, doi:10.1016/j.renene.2015.04.072.

Amit Bhatnagar, M. S., Anna Witek-Krowiak (2015). Agricultural waste peels as

versatile biomass for water purification - aA review. Chem Eng J, 244–-271.
Bagheri, A. R., Ghaedi, M., Asfaram, A., Hajati, S., Ghaedi, A. M.,
Bazrafshan, A., et al. (2016). Modeling and optimization of simultaneous
removal of ternary dyes onto copper sulfide nanoparticles loaded on activated
carbon using second-derivative spectrophotometry. J Taiwan Inst Chem Eng,
65, 212–-224, doi:10.1016/j.jtice.2016.05.004.

Basheer, I. A., & Hajmeer, M. (2000). Artificial neural networks: fundamentals,

computing, design, and application. J Microbiol Methods, 43(1), 3–-31,
doi:10.1016/s0167-7012(00)00201-3.

Berger, C. (2012). Biochar and activated carbon filters for greywater treatment - comparison
of organic matter and nutrients removal. Sweden: Swedish University of
Agricultural Sciences.

Bhayana, S. (2014). Top 15 hHealth bBenefits and uUses of Rambutan.

Bilgehan, M. (2011). Comparison of ANFIS and NN models-wWith a study in critical buckling
load estimation. Appl Soft Comput, 11(4), 3779–-3791,
doi:10.1016/j.asoc.2011.02.011.

Bingöl, D., Inal, M., & Çetintaş, S. (2013). Evaluation of cCopper bBiosorption onto
dDate pPalm (Phoenix dactylifera L.) sSeeds with MLR and ANFIS mModels. Ind
Eng Chem Res, 52(12), 4429–-4435, doi:10.1021/ie400484c.

Chai, T., & Draxler, R. R. (2014). Root mean square error (RMSE) or mean absolute error
(MAE)? – arguments against avoiding RMSE in the literature. Geosci Model
Dev, 7(3), 1247–-1250, doi:10.5194/gmd-7-1247-2014.

Chinmai K., H. B. C, Kerwyn Donald D'souza, Mahesh Chandra B. R. and

Shilpa B. S. (2014). Feasibility sStudies on sSpent cCoffee gGrounds
bBiochar as an aAdsorbent for cColor rRemoval. Int J Appl Innov Eng Manag,
3(10), 9–-13.

Chowdhury, S., & Saha, P. D. (2013). Artificial neural network (ANN) modeling of
adsorption of methylene blue by NaOH-modified rice husk in a fixed-bed
column system. [journal article]. Environ Sci Pollut Res, 20(2), 1050–-1058,
doi:10.1007/s11356-012-0912-2.
Dastorani, M. T., Moghadamnia, A., Piri, J., & Rico-Ramirez, M. (2010).
Application of ANN and ANFIS models for reconstructing missing flow data.
Environ Monit Assess, 166, 14, doi:10.1007/s10661-009-1012-8.

Emamgholizadeh, S., Moslemi, K., & Karami, G. (2014). Prediction the

gGroundwater lLevel of Bastam Plain (Iran) by aArtificial nNeural nNetwork
(ANN) and aAdaptive neuro-fFuzzy iInference sSystem (ANFIS). [journal article].
Water Resour Manag, 28(15), 5433–-5446, doi:10.1007/s11269-014-0810-0.

Esmaeili, M., Osanloo, M., Rashidinejad, F., Aghajani Bazzazi, A., & Taji, M.
(2012). Multiple regression, ANN and ANFIS models for prediction of
backbreak in the open pit blasting. Eng Comput, 30(4), 549–-558,
doi:10.1007/s00366-012-0298-2.

Fan, M., Hu, J., Cao, R., Xiong, K., & Wei, X. (2017). Modeling and prediction of
copper removal from aqueous solutions by nZVI/rGO magnetic
nanocomposites using ANN-GA and ANN-PSO. Sci Rep, 7(1), 18040,
doi:10.1038/s41598-017-18223-y.

Feng Li, K. S., Xiaolin Long, Jiasheng Wen, Xiaojie Xie, Xiangyun Zeng,
Yanyan Liang, Yansha Wei, Zefeng Lin, Wenrou Huang, Ruida
Zhong (2016). Preparation and cCharacterization of bBiochars from Eichornia
crassipes for cCadmium rRemoval in aAqueous sSolutions. PLoS OnePLOS ONE,
1–-13.

Foereid, B. (2015). Biochar in nNutrient rRecycling-tThe eEffect and iIts uUse in

wWastewater tTreatment. Open J Soil Sci, 39–-44.

J. Gary Yeager, R. L. W. (1981). Effects of mMoisture cContent on long-tTerm sSurvival

and rRegrowth of bacteria in wWastewater sSludge. Appl Environ Microbiol,
41(5), 1117–-1122.

Ghaffari, A., Abdollahi, H., Khoshayand, M. R., Bozchalooi, I. S., Dadgar, A.,
& Rafiee-Tehrani, M. (2006). Performance comparison of neural network
training algorithms in modeling of bimodal drug delivery. Int J Pharm, 327(1–-
2), 126–-138, doi:10.1016/j.ijpharm.2006.07.056.
Ghosh, A., Das, P., & Sinha, K. (2015). Modeling of biosorption of Cu(II) by alkali-
modified spent tea leaves using response surface methodology (RSM) and
artificial neural network (ANN). [journal article]. Appl Water Sci, 5(2), 191–-199,
doi:10.1007/s13201-014-0180-z.

Heaton, J. (2008). Introduction to neural networks for Java, 2nd Edition: Heaton Research,
Inc.

Hyndman, R. J., & Koehler, A. B. (2006). Another look at measures of forecast

accuracy. Int J Forecast, 22(4), 679–-688, doi:10.1016/j.ijforecast.2006.03.001.

Iphar, M. (2012). ANN and ANFIS performance prediction models for hydraulic impact
hammers. Tunn Undergr Space Technol, 27(1), 23–-29,
doi:10.1016/j.tust.2011.06.004.

Joesph M. Kimetu, J. M. H., Maen Husein, Joule Bergerson, David B.

Layzell (2014). Using activated biochar for greenhouse gas mitigation and
industrial water treatment. Mitig Adapt Strateg Glob Change, 1–-16.

Khan, T., Mustafa, M. R. U., Isa, M. H., Manan, T. S. B. A., Ho, Y.-C., Lim,
J.-W., et al. (2017). Artificial nNeural nNetwork (ANN) for modelling
aAdsorption of lead (Pb (II)) from aAqueous sSolution. Water Air Soil Pollut,
228(11), 426, doi:10.1007/s11270-017-3613-0.

Krishna R. Reddy, F. A., Tao Xie, Sara Dastgheibi (2014). Evaluation of biochar as
a potential filter media for the removal of mixed contaminants from urban
storm water runoff. Journal of Envionmental Engineering, 1–-10.

Kurtener, D., Torbert, H. A., & Krueger, E. (2008). Evaluation of agricultural land
suitability: application of fuzzy indicators. In Berlin, Heidelberg, (pp. 475–-490,
Computational Science and Its Applications – ICCSA 2008): Springer Berlin
Heidelberg.

Mandu Inyang, B. G., Ying Yao, Yingwen Xue, Andrew R. Zimmerman,

Pratap Pullammanappallil, Xinde Gao (2012). Removal of heavy
metals from aqueous solution by biochars derived from anaerobically digested
biomass. Bioresource Technology, 1–-7.
Minh-Viet Nguyen, B.-K. L. (2014). Cu2+ ion adsorption from aqueous solutions by
amine activated poultry manure biochar. Journal of Selcuk University Natural
and Applied Science, 877-884.

Mirrashid, M. (2014). Earthquake magnitude prediction by adaptive neuro-fuzzy inference

system (ANFIS) based on fuzzy C-means algorithm. Nat Hazards, 74(3), 1577–-
1593, doi:10.1007/s11069-014-1264-7.

Morrow, P. R. (2013). Biochar: saturated hHydraulic cConductivity and mMethylene bBlue

sSorption cCharacteristics as aApplied to sStorm wWater tTreatment. Oregon:
Oregon State University.

Muhammad Bilal, J. A. S., Tayyab Ashfaq, Syed Mubashar Hussain Gardazi,

Adnan Ahmad Tahir, Arshid Pervez, Hajira Haroon, Qaisar
Mahmood (2013). Waste biomass adsorbents for copper removal from
industrial wastewater-aA review. J Hazard Mater, 322–-333.

Noor, N. M. M., & Hamid, S. H. A. (2014). Comparison of triangular and trapezoidal

membership functions for improving decision making in crime prevention. In G.
Phillips-Wren;, S. Carlsson;, A. Respício;, & P. Brézillon (Eds.), DSS 2.0 –
Supporting Decision Making with New Technologies (Vol. 261, pp. 11): IOS
Press.

Podstawczyk, D., Witek-Krowiak, A., Dawiec, A., & Bhatnagar, A. (2015).

Biosorption of copper(II) ions by flax meal: Empirical modeling and process
optimization by response surface methodology (RSM) and artificial neural
network (ANN) simulation. Ecol Eng, 83, 364–-379,
doi:10.1016/j.ecoleng.2015.07.004.

Prasojo, R. A., Diwyacitta, K., Suwarno, & Gumilang, H. (2017). Transformer

pPaper eExpected lLife eEstimation uUsing ANFIS bBased on oOil
cCharacteristics and dDissolved gGases (cCase sStudy: Indonesian
tTransformers). Energies, 10(8), 1135, doi:10.3390/en10081135.

Pusker Regmi, J. L. G. M., Doris Hamill, Sandeep Kumar, Gary Schafran

(2011). Biochar: a rRenewable mMaterial for rRemoving cContaminants from
 wWater. Hampton: Department of Civil and Environmental Engineering, Old
Dominion University.

Ratner, B. (2009). The correlation coefficient: Its values range between +1/−1, or do they?
Journal of Targeting, Measure Anal Market, 17(2), 139–-142,
doi:10.1057/jt.2009.5.

Rezaei, H., Rahmati, M., & Modarress, H. (2017). Application of ANFIS and MLR
models for prediction of methane adsorption on X and Y faujasite zeolites:
effect of cations substitution. Neural Comput Appl, 28(2), 301–-312,
doi:10.1007/s00521-015-2057-y.

Salim, L. (2016a). On sSimulation pPerformance of fFeedforward and NARX nNetworks

uUnder dDifferent nNumerical tTraining aAlgorithms. In M. Francisco, & A.

Carlos (Eds.), Handbook of rResearch on cComputational sSimulation and

mModeling in eEngineering (pp. 171–-183). Hershey: IGI Global.

Salim, L. (2016b). Prediction of iInternational sStock mMarkets bBased on hHybrid

iIntelligent sSystems. In M. Jorge Tiago, & M. Andreea (Eds.), Handbook
of rResearch on iInnovations in iInformation rRetrieval, aAnalysis, and
mManagement (pp. 110–-124). Hershey: IGI Global.

Shibata, K., & Yusuke, I. (2009). Effect of number of hidden neurons on learning in large-
scale layered neural networks. In 2009 ICCAS-SICE, 18–-21 Aug. 2009 (pp.
5008–-5013).

Shojaeimehr, T., Rahimpour, F., Khadivi, M. A., & Sadeghi, M. (2014). A

modeling study by response surface methodology (RSM) and artificial neural
network (ANN) on Cu2+ adsorption optimization using light expended clay
aggregate (LECA). J Ind Eng Chem, 20(3), 870–-880,
doi:10.1016/j.jiec.2013.06.017.

Talpur, N., Salleh, M. N. M., & Hussain, K. (2017). An investigation of membership

functions on performance of ANFIS for solving classification problems. IOP
Conference Series: Mater Sci Eng, 226, 012103, doi:10.1088/1757-
899x/226/1/012103.
Tokatli, F., Tari, C., Unluturk, S. M., & Gogus Baysal, N. (2009). Modeling of
polygalacturonase enzyme activity and biomass production by Aspergillus
sojae ATCC 20235. J Ind Microbiol Biotechnol, 36(9), 1139–-1148,
doi:10.1007/s10295-009-0595-y.

Turan, N. G., & Ozgonenel, O. (2013). The dDesign and iImplementation of aAdsorptive
rRemoval of Cu(II) from lLeachate uUsing ANFIS. Sci World J, 2013, 590267,
doi:10.1155/2013/590267.

Turan, N. G., Mesci, B., & Ozgonenel, O. (2011). The use of artificial neural networks
(ANN) for modeling of adsorption of Cu(II) from industrial leachate by pumice.
Chem Eng J, 171(3), 1091–-1097, doi:10.1016/j.cej.2011.05.005.
USEPA, U. S. E. P. A. (2016). Table of Regulated Drinking Water Contaminants.

Uzuner, S., & Cekmecelioglu, D. (2016). Comparison of aArtificial nNeural nNetworks

( ANN ) and aAdaptive neuro-fFuzzy iInference sSystem ( ANFIS ) mModels in
sSimulating polygalacturonase pProduction. BioResources, 11(4), 10.

Walia, N., Singh, H., & Sharma, A. (2015). ANFIS: adaptive neuro-fFuzzy iInference
sSystem- aA sSurvey. Int J Comput Appl, 123(13), 7,
doi:10.5120/ijca2015905635.

Wong, Y. J., Arumugasamy, S. K., & Jewaratnam, J. (2018). Performance

comparison of feedforward neural network training algorithms in modeling for
synthesis of polycaprolactone via biopolymerization. Clean Techn Environ
Policy, doi:10.1007/s10098-018-1577-4, 20, 1971, 1986.

Wong, Y. J., Arumugasamy, S. K., & Mustapha, K. B. (2019). Development of a

computational predictive model for the nonlinear in-plane compressive
response of sandwich panels with bio-foam. Compos Struct, 212, 423–-433,
doi:10.1016/j.compstruct.2019.01.039.

Xinxing, P., Lee, B., & Chunrong, Z. (2013) A comparison of neural network
backpropagation algorithms for electricity load forecasting. In 2013 IEEE
International Workshop on Inteligent Energy Systems (IWIES), 14–-14 Nov.
2013 (pp. 22–-27). doi:10.1109/IWIES.2013.6698556.

Yang, J., Du, L., Shi, S., Gong, W., Sun, J., & Chen, B. (2019). Potential of
fFluorescence iIndex dDerived from the sSlope cCharacteristics of lLaser-
 iInduced cChlorophyll fFluorescence spectrum for rice lLeaf nNitrogen
cConcentration eEstimation. Appl Sci, 9(5), 916.

Zhang, Y., Dai, M., Liu, K., Peng, C., Du, Y., Chang, Q., et al. (2019). Appraisal of
Cu( II ) adsorption by graphene oxide and its modelling via artificial neural
network. RSC Adv, 9, 30240–-30248, doi:10.1039/C9RA06079K.

Fig.ure 1: Proposed structure of ANN

 Fig.ure 2: Proposed structure of ANFIS

Fig.ure 3: Flowchart for modelling Cu(II) adsorption estimations

Fig.ure 4: Box plot of overall prediction accuracy of ANN models for 11 different
training algorithms in (a) training dataset and (b) testing dataset.

Fig.ure 5: Effect of number of hidden neuron on ANN accuracy for Cu(II) adsorption
prediction using biochar derived from rambutan peel

Fig.ure 6a: Graph of experimental data versus predicted data under SCG
backpropagation training algorithm with 12 hidden nodes for both training and testing
datasets

Fig. 7 Residual plot showing difference between experimental data versus predicted
data under SCG backpropagation training algorithm for both training and testing dataset

Fig.ure 8: Box plot of overall prediction accuracy of ANFIS models for 8 different
membership functions in (a) training dataset and (b) testing dataset.

Fig.ure 9: Effect of fuzzy set combination on ANFIS accuracy for Cu(II) adsorption
prediction using biochar derived from rambutan peel

Fig.ure 10 Graph of experimental data versus predicted data under Gaussian MF with
MF numbers of [4 3 2 3] for both training and testing datasets

Fig. 11 Residual plot showing difference between experimental data versus predicted
data under Gaussian MF for both training and testing dataset

Fig.ure 12: Graph of actual data versus predicted data generated by ANN, ANFIS, and
MLR

Table 1: Various methods for copper metal removal (Muhammad Bilal 2013)

Technology Brief Advantages Disadvantages

description
Chemical precipitation Precipitants Able to achieve Require huge amount of
by addition low copper chemical, sensitive to
of chemicals. concentration in pH and thus optimum
the effluent pH has to be
stream. Suitable maintained. Various
for handling huge factors such as pH,
amount of temperature, and
wastewater. Easy precipitate
to handle. concentration have to
be monitored carefully.
Ion exchange Synthetic ion Higher rate of Synthetic resins are
exchange removal of copper very expensive. Not
resin or ions. High suitable for large scale
natural ion treatment capacity. removal of copper ions.
exchange
resin.
Membrane filtration Membranes, Very effective in Membrane fouling is a
surfactants fulfilling stringent major problem. High
improve and effluent limits. capital and maintenance
support the Reuse of cost. Not efficient at
process. wastewater is treating water bodies
possible. Capable with low concentration
of preventing of copper ions.
severe
environmental
damages.
Coagulation/flocculation Reagents like Suitable for large Chemical reagents are
alum and scale wastewater expensive. Have to deal
ferric salt. treatment. with sludge production
and sludge disposal.
Electrolytic recovery Utilizes Efficient removal High energy
 electrical of copper ions. consumption. Not
energy. Does not consume suitable for high
much chemicals. amount of wastewater.
Adsorption Fluid Extremely Chemical regeneration
handling unit efficient at copper needed. Require
or ions removal. disposal of used
regenerating adsorbents. Loss of
media. adsorption capacity
after certain cycles of
adsorption.
Reverse osmosis Membrane Efficient at copper High cost of membrane.
removal Fouling of membrane
will incur high
maintenance cost.
 28

Table 2: Training algorithms employed in this study.

Training algorithm Function Brief description

(i) Levenberg-Marquardt (Lourakis 2005; Cömert and Kocamaz 2017)
Levenberg-Marquardt trainlm Most widely applied method for solving nonlinear least square problems in a broad
backpropagation spectrum of disciplines due to its adaptive behaviour according to the nature of the
problem
(ii) Gradient descent with variable learning rate (Cömert and Kocamaz 2017)
Gradient descent with traingdm Momentum coefficient is introduced that will utilize the past weight changes as
momentum backpropagation well as the current error in the relevant direction
Gradient descent with traingda Learning rate of the backpropagation is adjusted according to the error calculated to
adaptive learning rate keep the learning size as large as possible to ensure learning is stable
backpropagation
Gradient descent with traingdx Both the momentum coefficient and adaptive learning rate are utilized to derive
momentum and adaptive more accurate results.
learning rate backpropagation
(iii) Resilient backpropagation (Ahmad et al. 2008; Mastorocostas 2004)
Resilient backpropagation trainrp Direct adaptation of the weights and biases based on the backpropagated gradient
of the network, instead of the magnitude of the gradient.
(iv) Conjugated gradient descent (Sharma and Venugopalan 2014)
Conjugate gradient traincgb Search direction is periodically reset to the negative of the gradient when there is
 29

backpropagation very little orthogonality left between the current gradient and the previous gradient.
with Powell-Beale restarts
Conjugate gradient traincgp Constant in conjugate gradient algorithm, 𝛽𝑡 is updated as the dot product between
backpropagation the previous change in the gradient vector with the current gradient vector and
with Polak-Ribiére updates divided by the norm squared of the previous gradient vector.
Scaled conjugate gradient trainscg Step size scaling mechanism is used instead of line-search at each iteration
backpropagation
(v) Quasi-Newton algorithm (Sharma and Venugopalan 2014; Lewis and Overton 2012)
One-step secant trainoss Minimum computational memory is required as the hessian matrix is not stored and
backpropagation the new search direction can be calculated without calculating a matrix inverse.

BFGS quasi-Newton trainbfg Hill-climbing optimization technique is used to seek for a stationary point of a
backpropagation function.
(vi) Bayesian regulation (Burden and Winkler 2009)
Bayesian regularization trainbr Objective Bayesian criterion is provided for stopping training and effectively
backpropagation updating the network parameter
 30

Table 3: Membership functions employed in this study. (Wong et al. 2019)

Membership function Function Description

(i) Piecewise linear functions
Trapezoid-shaped MF trapmf Function of vector, x is mapped onto a trapezoidal curve.
Triangular-shaped MF trimf Function of vector, x is mapped onto a triangular curve.
(ii) Gaussian distribution functions
Gaussian MF gaussmf Function of vector, x is mapped onto Gaussian distribution curve.
Gaussian Combination MF gauss2mf Function of vector, x is mapped onto the superimposition of two Gaussian
distribution curves.
(iii) Bell-shaped function
Bell-shaped MF gbellmf Function of vector, x is mapped onto a bell-shaped curve.
(iv) Sigmoidal function
Difference sigmoidal MF dsigmf Function of vector, x is mapped onto a deviation of two sigmoidal curves.
Product sigmoidal MF psigmf Function of vector, x is mapped onto a multiplication of two sigmoidal curves.
(v) Polynomial- Bbased function
Polynomial-Pi MF pimf Function of vector, x is mapped onto a π-shaped curve.

Table 4: Summary of performance criteria for different training algorithms with its optimisedoptimized number of neuron.

Training OptimisedOptimized R2 RMSE MAE Accuracy (%)

 31

algorithm no. of neuron Training Testing Training Testing Training Testing Training Testing
trainlm 5 0.90 0.89 4.34 4.37 3.10 3.11 85.70 85.38
traingdm 11 0.91 0.90 4.04 4.29 2.90 3.01 85.45 84.73
traingda 18 0.90 0.87 4.28 4.82 3.22 3.64 83.03 81.61
traingdx 18 0.92 0.87 3.98 4.86 2.82 3.56 85.91 83.40
trainrp 18 0.92 0.85 3.86 5.02 2.77 3.65 86.33 83.13
traincgb 11 0.94 0.88 3.52 4.81 2.53 3.05 87.65 85.89
traincgp 10 0.92 0.90 3.85 4.38 2.66 3.03 86.68 86.09
trainscg 12 0.93 0.90 3.75 4.44 2.49 3.00 88.27 86.60
trainoss 10 0.93 0.90 3.82 4.18 2.69 2.99 86.07 85.52
trainbfg 12 0.91 0.89 4.15 4.41 2.83 3.09 86.61 85.87
trainbr 8 0.92 0.90 3.95 4.18 2.71 3.00 86.69 85.46

Table 5: Summary of performance criteria for different types of membership function with its optimisedoptimized number of
membership function for each parameter.

R2 RMSE MAE Accuracy (%)

MF type V1 V2 V3 V4
Training Testing Training Testing Training Testing Training Testing
TrapMF 2 4 2 3 0.95 0.87 3.09 5.22 2.01 3.26 90.92 84.33
TriMF 3 4 2 3 0.95 0.87 3.09 5.22 2.01 3.26 90.92 84.33
 32

GaussMF 4 3 2 3 0.95 0.89 3.29 4.50 2.16 2.85 90.24 87.06

Gauss2MF 2 5 2 3 0.95 0.87 3.19 5.09 2.04 3.19 90.85 84.37
GbellMF 5 5 2 2 0.94 0.87 3.52 5.10 2.27 3.10 89.80 85.53
DsigMF 4 5 2 3 0.95 0.87 3.11 5.23 2.01 3.24 90.92 84.45
PsigMF 4 5 2 3 0.95 0.87 3.11 5.23 2.01 3.24 90.92 84.45
PiMF 2 4 2 3 0.95 0.86 3.09 5.37 2.00 3.29 90.95 84.35

Table 6: Summary of MLR model performance

R2 RMSE MAE Accuracy (%)

Training Testing Training Testing Training Testing Training Testing
0.60 0.62 8.42 8.50 10.35 10.78 59.24 58.73

Table 7 Pearson’s correlation matrix

Biochar Initial Cu(II) Final Cu(II)

Variables Time Temperature
discharge concentration concentration
Time 1.00
Temperature 0.00 1.00
Biochar discharge 0.00 0.00 1.00
Initial Cu(II) concentration 0.00 0.00 0.00 1.00
Final Cu(II) concentration 0.57 0.18 − 0.04 − 0.06 1.00