Case Studies in Chemical and Environmental Engineering 11 (2025) 101180

Contents lists available at ScienceDirect

Case Studies in Chemical and Environmental Engineering

journal homepage: www.sciencedirect.com/journal/case-studies-in-chemical-
and-environmental-engineering

Case Report

Prediction of the viscosity of iron-CuO/water-ethylene glycol

non-Newtonian hybrid nanofluids using different machine
learning algorithms
Mohammed Shorbaz Graish a , Ali B.M. Ali b, Murtadha M. Al-Zahiwat c , Saja Mohsen Alardhi d,
Mohammadreza Baghoolizadeh e,** , Soheil Salahshour f,g,h , Mostafa Pirmoradian i,*
a
Department of Chemical Engineering, University of Technology- Iraq, Baghdad, 10066, Iraq
b
Air Conditioning Engineering Department, College of Engineering, University of Warith Al-Anbiyaa, Karbala, Iraq
c
Department of Chemical Engineering, College of Engineering, University of Misan, Amarah, Iraq
d
Nanotechnology and Advanced Materials Research Center, University of Technology – Iraq, Baghdad, Iraq
e
Department of Mechanical Engineering, Shahrekord University, Shahrekord, 88186-34141, Iran
f
Faculty of Engineering and Natural Sciences, Istanbul Okan University, Istanbul, Turkey
g
Faculty of Engineering and Natural Sciences, Bahcesehir University, Istanbul, Turkey
h
Research Center of Applied Mathematics, Khazar University, Baku, Azerbaijan
i
Department of Mechanical Engineering, Khomeinishahr branch, Islamic Azad University, Khomeinishahr, Iran

A R T I C L E I N F O A B S T R A C T

Keywords: Viscosity is a crucial parameter for heat transfer systems, governing pumping power, Rayleigh number, and
Machine learning algorithms Reynolds number; thus, viscosity prediction for hybrid nanofluids is important. Although some studies have
Non-Newtonian hybrid nano- antifreeze employed ML algorithms for predicting viscosity, limited ML algorithms or specific nanofluid types were
Viscosity
examined in previous studies, disregarding the complexities involved in the rheological behavior of a complex
nanofluid system such as non-Newtonian hybrid nanofluids. To overcome this limitation, this study offers a
practical contribution by utilizing 20 different machine-learning models to predict the viscosity of iron-CuO/
water-ethylene glycol non-Newtonian hybrid nanofluids. The influences of the input variables: solid volume
fraction (SVF), temperature, and shear rate on viscosity prediction are systematically assessed. We evaluate the
prediction accuracy and reliability of algorithms using ten performance metrics including RMSE, MAE, R2 and
NSE. Multivariate Polynomial Regression (MPR) outperforms the other algorithms, which is evident in the
highest correlation coefficient (R2 = 0.992) and lowest error metrics. At the other end, is the Extreme Learning
Machine (ELM), which turns out to be the worst performer. A unique contribution of this paper is that we extract
a mathematical equation from the MPR model that allows for straightforward calculation of viscosity, avoiding
non-trivial ML computations. This simplicity aids in practical applications and increases usefulness for engineers
and researchers alike. Using advanced data visualization techniques (heatmaps, box plots, KDE plots and Taylor
diagrams), the relationships between input variables and viscosity as well as the model performance are
explored. These results give a better understanding of the non-Newtonian hybrid nanofluid behavior and a solid
predictor of design-efficient heat transfer systems.

* Corresponding author.
** Corresponding author.
E-mail addresses: mohamadbaghoolizadeh@gmail.com (M. Baghoolizadeh), m.pirmoradian@iaukhsh.ac.ir, mostafaa.pirmoradian@gmail.com (M. Pirmoradian).

https://doi.org/10.1016/j.cscee.2025.101180
Received 12 February 2025; Received in revised form 2 March 2025; Accepted 4 March 2025
Available online 7 March 2025
2666-0164/© 2025 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
M.S. Graish et al. Case Studies in Chemical and Environmental Engineering 11 (2025) 101180

Nomenclature
30–50 ◦ C. An analytical regression model was used to predict the ther-
mal conductivity and dynamic viscosity of an Al2O3-GO HNF. The TC
Solid volume fraction (SVF) φ (%) Viscosity μnf (cP)
improvement of the HNF was around 4.30 % higher than that of pure
Temperature T (◦ C) Shear rate γ
(RPM) Al2O3 and 4.34 % higher than GO mono NF. Khouri et al. [9] investi-
Machine Learning Algorithm MLA Decision Tree DT gated the heat transfer properties of water-based graphene oxide (GO)
Multi-Layer Perceptron MLP Adaptive Neuro-Fuzzy ANFIS nanofluids in a heat exchanger. They analyzed the nanofluid’s thermal
Inference System conductivity, specific heat capacity, and viscosity at different tempera-
Gradient Descent GD Group Method of Data GMDH
Handling
tures and GO concentrations. The TC increased with both temperature
Radial Basis Function RBF Back Propagation Neural BPNN and NP concentration, enhancing heat transfer through conduction and
Network convection. However, the specific heat capacity increased with tem-
BFGS Quasi-Newton BFGS Levenberg-Marquardt LM perature but decreased with higher GO concentrations. Additionally, the
Support Vector Machine SVM Bayesian network BN
viscosity of the nanofluid rose with higher NP concentrations. Ajeena
Extreme Learning Machine ELM Elastic Component ECR
Regression et al. [10] investigated the dynamic viscosity of an HNF with ZrO2–SiC
Multiple Linear Regression MLR Gaussian Process GPR (50 %–50 %) and distilled water. To do so, a two-step process was used
Regression to disperse NPs in the BF. The viscosity of SVFs in the range of 0.025 %–
XGBoost XGB Multivariate Polynomial MPR 0.1 % was measured at temperatures from 20 to 60 ◦ C. The study
Regression
Partial Least Squares PLR Least Absolute Shrinkage Lasso
exhibited a relationship between NF viscosity and temperature, as well
Regression and Selection Operator as SVF parameters. The results showed that there was an increase in the
Evaluation Criteria EC Analysis Plots AP dynamic viscosity value at greater solid concentrations and lower tem-
Mean Absolute Relative Error MARE Root Mean Squared Error RMSE peratures. The dynamic viscosity of NF at 20 and 60 ◦ C was measured
Nash–Sutcliffe Model NSE Correlation Coefficient R
and showed an increase in viscosity by 29.6 % and 64.2 %, respectively,
Efficiency Coefficient
Mean Squared Error MSE Mean Absolute MAPE with 0.025 % NPs. This indicated that NP viscosity was more sensitive at
Percentage Error higher temperatures. The experiments also indicated that the ZrO2–SiC
R squared R2 Kernel density estimate Kde HNF was Newtonian across a range of temperatures. The research also
Standard Deviation STD Root Mean Squared RMSD suggested a new correlation through which the dynamic viscosity of
Deviation
Mean Absolute Error MAE Mean Bias Error MBE
HNF could be calculated based on the experimental parameters (tem-
Coefficient of the Variation of CvRMSE ​ ​ perature and SVF) with a 98.92 % error. Sepehrnia et al. [11] investi-
the Root Mean Square Error gated the rheological efficiency and dynamic viscosity of an HNF made
of SiO2 and MWCNTs NPs (90:10) with 5W30 engine oil as BF at various
SRs (50–1000 rpm) experimentally. The SVFs and temperatures were
assessed within the ranges of 0.05–1.00 vol% and 5–65 ◦ C, respectively.
1. Introduction The HNF exhibited characteristics of a non-Newtonian fluid. Further-
more, the HNF exhibited pseudoplastic characteristics in all SVFs and
Heat exchanger performance is a function of a number of variables temperatures, as the calculated power law index was less than unity. It
including heat transfer coefficient, cross-sectional area, and temperature was noted that the dynamic viscosity decreased as the NF’s temperature
difference. Water, ethylene glycol, and oils are only some of the working increased, whereas the dynamic viscosity increased as the NF’s SVF
fluids being utilized in industry and engineering. The main problem with increased. A three-variable association was found. Dynamic viscosity
these fluids is that they have poor heat transfer and low thermal con- sensitivity to temperature, SVF, and SR was also assessed. With a con-
ductivity. One of the new methods that have been utilized in order to stant SR of 800 rpm, dynamic viscosity sensitivity increased with NF’s
increase the TC of such fluids is suspending nanoparticles (NPs) with temperature and concentration. Wanatasanappan et al. [12] conducted
enhanced thermal properties in the working fluid to create a material experimental investigations on the rheological and viscosity character-
called nanofluid (NF). Choi [1] coined the term “nanofluid” to refer to a istics of an Al2O3–Fe2O3 HNF. After determining a correlation for vis-
fluid that suspends extremely small particles (NPs with a diameter cosity prediction, they investigated how the Al2O3–Fe2O3 mixture ratio
smaller than 100 nm). NFs involve the addition of one or more solid affected the viscosity property. The BF was a 60/40 mixture of water and
phases to a fluid, resulting in an enhanced heat transfer rate and altered ethylene glycol. Five distinct Al2O3–Fe2O3 NP compositions were
viscosity. Certain NPs, like aluminum oxide and magnesium oxide, exist examined at temperatures ranging from 0 to 100 ◦ C. The 40/60
as metal oxides and can be readily distributed and suspended in liquids. Al2O3–Fe2O3 composition exhibited the highest viscosity value at all
Diamond, carbon nanotubes, and other materials will greatly enhance temperatures examined, according to the experimental data.
heat transfer in comparison to oxide NPs. Hence, the simultaneous use of Conversely, the 60/40 composition exhibited the lowest viscosity value.
these nanomaterials can result in a steadfast amalgamation with distinct Additionally, the viscosity decreased by 87.2 % as the temperature was
and coveted thermal characteristics. The heat transfer coefficient is raised from 0 to 100 ◦ C. Furthermore, the Newtonian characteristic was
significantly influenced by viscosity, which is a key factor in the re- observed in all Al2O3–Fe2O3 compositions. Sepehrnia et al. [13] exam-
lationships that govern heat transfer. The incorporation of nanoparticles ined the thermophysical and rheological properties of hydraulic oil HLP
into base fluids significantly alters the thermophysical properties of 68 as the BF with a ternary combination of Fe3O4, TiO2, and GO
nanofluids, particularly their viscosity [2–4]. It has been shown that nano-additives in a variety of nanomaterial mixing ratios (MRs) (1:1:1,
viscosity plays a crucial role in determining pumping power, Rayleigh 2:1:1, 1:2:1, and 1:1:2), SVFs (0–1 %), and temperatures (15–65◦ ). Their
number, Reynolds number, and consequently, heat transfer. Numerous analysis of all MRs showed Newtonian THNF behavior. At 15 ◦ C, the BF
experiments and studies have been conducted to investigate the vis- viscosity increased by 345 %, 1821 %, 1763 %, and 1990 % for MRs of
cosity and heat transfer in NFs. Bashirnezhad et al. [5], Koca et al. [6], 1:1:1, 1:2:1, and 2:1:1, respectively, when a 1 % SVF of GO: Fe3O4: TiO2
and Murshed and Estellé [7] highlighted that the viscosity of NFs is was present. In addition, at the highest SVF, the MRs of 1:1:1, 1:2:2, and
influenced by factors such as temperature, NP type, solid volume frac- 3:1:1 showed a 66 %, 75 %, 60 %, and 70 % rise in viscosity with a
tion (SVF), particle size, and base fluid (BF) type. While investigating temperature decrease from 65 ◦ C to 15 ◦ C, respectively. The experi-
heat transfer, Selvarajoo et al. [8] experimented with different volume mental, statistical, and numerical viscosity of a MWCNT (50 %)-MgO
concentrations of mono NFs Al2O3/water and GO/water and hybrid (50 %)/SAE40 HNF was measured by Esfe et al. [14]. Temperatures,
nanofluid (HNF) Al2O3-GO/water to evaluate their thermophysical SVFs, and shear rates were assessed within the ranges of 25–50 ◦ C,
properties. Viscosity and TC were examined in the temperature range of 0.0625–1 %, and 666.5–9331 s− 1, respectively. According to their

2
M.S. Graish et al. Case Studies in Chemical and Environmental Engineering 11 (2025) 101180

findings, the HNF exhibited properties of a non-Newtonian fluid that and the lowest dynamic viscosity. Four elite algorithms—strong Pareto
was pseudo-plastic. At a specific amount of SVF (0.0625 %), tempera- evolutionary algorithm II, Pareto envelope-based selection algorithm II,
ture (25 ◦ C), and specific rate (3999 s− 1), the maximum viscosity non-dominated sorting GA II, and multi-objective particle swarm opti-
decrease was 5.91 % and the maximum viscosity rise was 28.70 %. In mization—were applied in the selection of the most significant set of
another study by Ajeena et al. [15], the viscosity of ZrO2/DW and input parameters. Zhang et al. [25] proposed a hybrid method
SiC/DW NFs was measured over a temperature range of 20–60 ◦ C with combining ML, MOO, and MCDM to select the optimum parameters of
various SVFs (0.025, 0.05, 0.075, and 0.1 %). The dynamic viscosity of MWCNTs-oxide HNFs based on water. They used two effective ML
the sample was enhanced by increasing the SVF of the NPs while methods, the group method of data handling neural network
lowering the temperature. Consequently, NPs had a more noticeable (GMDH-NN) and the combinatorial (COMBI) algorithm, to simulate four
effect on the viscosity as the temperature increased. The results important TPPs: density ratio, viscosity ratio, specific heat capacity
demonstrated that the viscosity increase for the ZrO2/DW NF topped at ratio, and TC ratio. The optimization variables that were considered
226.3 %, while for the SiC/DW NF, it was 110.5 %. Thermophysical were the type of oxide NP, SVF, and system temperature. Different types
properties of MWCNT-ZnO (30–70 %)/SAE40 HNF were studied by Esfe of oxide-NPs showed substantially different distributions of optimum
et al. [16] over a variety of SVFs, temperatures, and SRs. At T = points over various temperature ranges. To model the relationship be-
31.156 ◦ C, SVF = 0.063 %, and SR = 933.923 s− 1, their results tween the dynamic viscosity of the MgO-SAE 5W30 Oil HNF and three
demonstrated that the ideal viscosity value was 209.53 mPa.sec. At critical parameters, namely the SVF, temperature, and SR, Gao et al.
temperatures ranging from 10 to 50 ◦ C, with five SVFs of magnetocaloric [26] employed an RBF-ANN. Their findings demonstrated that the dy-
NF and varying SRs, Abbasian et al. [17] investigated the rheological namic viscosity of this NF was reduced as the temperature and SR were
properties of CoFe2O4 superparamagnetic NPs dispersed in increased. Contrarily, although this result can be ignored, the produc-
water-ethylene glycol (EG) coolant. Dispersed in a 50:50 mixture of EG tion was directly affected by the volume proportion of NPs. A drop in
and water, the cobalt ferrite metallic complexes were prepared by sol- dynamic viscosity would result from a temperature increase of 5–55 ◦ C.
vothermal methods. The experimental results of the magnetocaloric NF The dynamic viscosity dropped from 400 cP to 25 cP when the SR was
demonstrated non-Newtonian dynamics. An increase in NP mass con- increased from 50 rpm to 1000 rpm. In their study, Fadhl et al. [27]
centration from 0.05 % to 0.8 % increased viscosity by almost 80 % at modeled the dynamic viscosity of NFs containing MgO NPs using two
10 ◦ C. Furthermore, TC was enhanced with increasing temperature, intelligence approaches: ANFIS and GMDH. Their findings indicated
reaching a high of 9.4 % at 50 ◦ C. In their study, Hafeez et al. [18] that GMDH had better precision. Utilizing various machine learning
investigated the effects of Cu–Al2O3 NPs on the rheological properties, models including SVR, ANN, and ANFIS, AbuShanab et al. [28] aimed to
dynamic viscosity, and TC of HNFs based on kerosene oil. For numerical forecast the dynamic viscosity of NFs, namely Polyalpha-
analysis, a mathematical model of an HNF was derived from PDEs, Olefin-hexagonal boron nitride. There were 540 points of experi-
which were then transformed into ODEs using similarity conversion. mental data used to train and evaluate the models. While all three
Both the friction coefficient and the heat transfer rate were shown to be models were able to correctly estimate the viscosity of NFs, the ANFIS
enhanced by the introduction of NPs. Esfe et al. [19] used different RSM and ANN models performed better than the SVR model. Although both
models to figure out the dynamic viscosity of a SiO2-MWCNT HNF in the ANFIS and ANN models performed similarly, the ANN model was
SAE40 oil, where 60 % SiO2 and 40 % MWCNT made up the HNF. The chosen because it trained and computed faster. As the SR parameter was
study was performed under the following conditions: temperature = removed from the input layer, the ANN model became more accurate.
25–50 ◦ C, SVF = 0–1% and SR = 666.5–9331 s− 1. The viscosity of HNFs Their research proved that ANN and other machine learning models
was determined using the correlation functions of several models. The could accurately predict the dynamic viscosity of the NF. Hua et al. [29]
impact of various parameters on HNF viscosity was examined. SR and evaluated the dynamic viscosity of a hybrid antifreeze with MWCNTs,
temperature had the least and most significant effects, respectively. aluminum oxide, and water and ethylene-glycol in response to SVF and
While molecular dynamics simulations and laboratory investigations temperature. The NF’s anticipated viscosity was determined using an
have yielded efficient and practical insights into NF characteristics, the ANN trained on samples at temperatures of 25–50 ◦ C, and SVFs of
time and money required to run these models have prompted academics 0.25–1 %. The NF’s viscosity was impacted in various ways by SR and
to seek other modeling approaches. Fuzzy logic, adaptive neuro-fuzzy SVF. While rising SR greatly lowered viscosity mean and variance,
inference system (ANFIS), genetic algorithm (GA), and artificial neural increasing SVF increased deviation and mean values. The change in
network (ANN) are some of the methods classified under this category. viscosity due to temperature was smaller compared to the variations
Their adaptability, superior effectiveness, and accuracy have been caused by SR and SVF. The two-layer network and thirteen neurons with
demonstrated in several scientific domains [20–23]. The described nonlinear activation functions in the hidden layer of the suggested ANN
predictive models are self-improving and data-driven; thus, they can model predicted viscosity versus inputs accurately. To improve the dy-
anticipate process outcomes with high accuracy. They also have great namic viscosity of an HNF composed of MWCNT-Al2O3 (40:60)/Oil
correlation results and can model nonlinear relationships with a variety 5W50, Esfe et al. [30] employed ANN. To determine which parameters
of input variables. Concurrently, it saves time and money by not of MWCNT-Al2O3 (40:60)/Oil 5W50 HNF were most important for dy-
requiring any more experiments. In this regard, Esfe et al. [23] carried namic viscosity, a sensitivity analysis was performed. The results
out an investigation in which they employed an ANN to predict the showed that dynamic viscosity values were highest at temperatures
viscosity (μnf) of an NF composed of MWCNT-MgO/SAE40 engine oil at lower than 5 ◦ C. Changes in SR between zero and eight hundred revo-
different temperatures, SVF, and SR. The Levenberg-Marquardt (ML) lutions per minute also decreased the dynamic viscosity. Dai et al. [31]
learning algorithm was employed in an MLP ANN consisting of two used ANN to evaluate the effects of SVF, SR, and temperature on the
hidden layers. The first layer contained the optimum structure of 10 dynamic viscosity and torque of SiO2/EG NF. Rheological properties
neurons, while the second layer contained 4 neurons. Concentration, SR, were predicted using many machine learning (ML) models, and the best
and temperature were used as input parameters for ANN modeling, model was picked. All samples showed that torque increased linearly
while the predicted μnf was observed as the output parameter. The op- with SR, steeper at lower temperatures. Gaussian Process Regression
timum ANN, as per the data, possessed the least mean square error and (GPR) models with the Matérn covariance function predicted dynamic
consisted of eight neurons in both of the layers. As per their observation, viscosity well on both datasets. Singh and Ghosh [32] suggested a single
ANN prediction of data was much better compared to correlation. By feed-forward MLPNN to predict the density and dynamic viscosity of
applying a mix of heuristic techniques and an ANFIS, Wang et al. [24] graphene NP/DW, aluminum oxide/DW, and MWCNT NFs over
identified the optimal temperature and SVF pair for an HNF made of 30–80 ◦ C. Dynamic viscosity and density were output from temperature,
silicon oxide, MWCNT, aluminum, and water with the highest TC (knf) volume concentration, and NFs that were input. As NF concentration

3
M.S. Graish et al. Case Studies in Chemical and Environmental Engineering 11 (2025) 101180

grew, density and viscosity decreased, and vice versa as temperature (viscosity). The comprehension of input variable behavior through
climbed. Experimental findings were compared to MLPNN and mathe- advanced data analysis techniques such as heat maps and box plots is a
matical model results. The greatest viscosity and density errors were less requirement, and this study contributes greatly to such an important
than 1 % and 0.2 %, respectively. In another study, Said et al. [33] relevance. These techniques ease feature selection and ensure that useful
generated ethylene glycol (EG) NFs utilizing rGO-Fe3O4–TiO2 ternary data are used to train ML models that could enhance the latter’s pre-
hybrid nanocomposites. To measure density and viscosity, they varied dictive ability. The critical result obtained from this study is the math-
temperatures from 25 to 50 ◦ C and weight fraction from 0.01 to 0.25 %. ematical formula describing the relationship between the dynamic
Modern machine learning techniques including ANN, BRT, and SVM viscosity of hybrid nanofluids and the following parameters; tempera-
were used to build the forecasting model. It was found that SVM, ANN, ture, shear rate, and SVF. All this was made possible by thorough
and BRT could accurately duplicate lab density and viscosity data. Kanti analysis of the employed experimental data along with graphical tools
et al. [34] research examined the effects of concentration and temper- like error histograms, KDE plots, and Taylor diagrams, and sophisticated
ature on the viscosity and thermal conductivity of graphene oxide, sil- error assessment methods like RMSE, MAE, and R. With these in-
icon dioxide, and titanium dioxide water-based nanofluids. Thermal struments, the precision, reliability, and error distribution from opti-
conductivity and viscosity were found to be higher in the nanofluids mization algorithms to machine learning models can be assessed.
compared to water, with graphene oxide recording the best result. Deep
learning, deep neural networks, and a high gradient amplification • In this instance, work scrutinizes, in-depth, twenty different machine
approach were utilized in the study to handle complex data properly. learning algorithms to perhaps more accurately forecast the viscosity
The extreme slope model of the viscosity model had a lower R2 value of a non-Newtonian hybrid nanofluid. Few works manage such
(0.9122) and mean squared error (0.010) than the deep neural painstaking comparisons, since most discussions focus only on a
network-based model (0.3329). In another study, Kanti et al. [35] limited set of algorithms. The research recommends valuable insights
investigated silicon dioxide, graphene oxide, titanium dioxide, and into the best algorithms for viscosity prediction of the nanofluids;
hybrid water-based nanofluids and observed elevated viscosity and this is achieved by investigation across a broad range of alternatives.
thermal conductivity, in addition to improved performance. Optimum • The study also employs heatmaps, box plots, error histograms, KDE
thermal conductivity and viscosity ratio for graphene oxide nanofluids plots, and Taylor diagrams for enhanced data analysis and inter-
were reported at 60 ◦ C and 30 ◦ C, respectively, with an enhancement of pretation. Understanding how the input factors interact also allows
52 and 177 % relative to the base fluid. Graphene oxide-TiO2 hybrid for optimizing nanofluids’ performance in real applications.
nanofluids exhibited thermal conductivity and viscosity ratios of 43 % • This study presents a new approach to formulate a relationship be-
and 144 % higher than the base fluid under the same conditions. tween viscosity and its input parameters (SVF, temperature, and
Thermophysical properties of hybrid nanofluids were predicted by shear rate). It also acts as an ancillary benchmark, in that it can now
state-of-the-art machine learning models, with the random forest model provide viscosity values without the burden of applying complex
being the most versatile. Nanofluid concentration played a significant machine learning models.
role in the prediction of the TC ratio but not the viscosity ratio. Akilu • This research goes beyond basic accuracy measurements to provide a
et al. [36] investigated the heat transfer behavior and friction factor of more detailed picture of how well each machine learning algorithm
silicon dioxide and glycerol-based ethylene glycol nanofluids, which performed by introducing ten assessment variables.
were higher than the base fluid. It was observed from the study that
ethylene glycol-based glycerol and silicon dioxide nanofluid exhibited The majority of previous studies have trained a few machine
maximum heat transfer improvement by 5.4 % and 8.3 %, respectively. learning algorithms (ANN, SVM, or ANFIS) to predict viscosity. On the
The research employed five machine learning methods for the predic- other side, this paper assesses 20 unique machine learning models for
tion of intricate experimental data, which were linear regression, the first time and offers a complete benchmark for approximate vis-
random forest, steep gradient boosting, adaptive boosting, and decision cosity estimation. In previous research, this kind of comprehensive
trees. The models successfully predicted nanofluid thermal perfor- comparison has never before been conducted, leading to more insights
mance, which played a part in the development of efficient cooling into how algorithms perform on different metrics. In previous works,
systems and the stability of energy systems. Sharma et al. [37] explored models were typically evaluated using a small number of performance
the thermophysical characteristics of polydisperse SiO2 nanoparticles in metrics, such as RMSE or R2. Singh and Ghosh, for example, focused
a glycerol-water mixture. A two-step method was used to prepare on using RMSE and MAPE. However, 10 evaluation criteria include not
aqueous glycerol nanofluid containing 30 % glycerol (30 GW) and SiO2 only RMSE, MAE, NSE, MBE, and CvRMSE to comprehensively analyze
with particle sizes of 15, 50, and 100 nm. The study measured the the accuracy and reliability of the models ensuring a deep-
properties of nanofluids at 30–100 ◦ C, revealing a stable single-phase er examination of the strengths and weaknesses of each algorithm.
liquid, decreased viscosity and density, increased thermal conductiv- While some research studies have suggested correlations or models for
ity, and slightly increased specific heat. At 60 ◦ C, the thermal conduc- viscosity prediction, most do not provide an explicit mathematical
tivity and viscosity of a 0.5 % SiO2 solution changed by 11.1 % and 32 %, equation derived from the machine learning outputs. The mathematical
respectively, compared to the base liquid. The Gaussian process equation obtained in this study by employing the MPR algorithm allows
regression optimization Bayesian approach, an explainable artificial us to easily compute the viscosity without the need for endless complex
intelligence technique, was used to develop a predictive model for computation which the ML approaches require; thus greatly improving
nanofluid properties, enhancing predictability and explainability the practical implementation status of this algorithm, and making it very
through kernel functions and historical data representation, resulting in useable for engineers and researchers. Most of the existing studies are
high correlation values and minimal modeling errors. well-trained in predicting the viscosity without examining the link be-
The present study introduces a novel and comprehensive approach to tween the input and output variables like SVF, temperature, and shear
predict the viscosity of iron-CuO/water-Ethylene glycol non-Newtonian rate and viscosity. In this regard, the current study utilizes advanced
hybrid nanofluids, employing a diverse array of machine learning al- data visualization techniques such as heatmaps, box plots, error histo-
gorithms. The innovation resides in the systematic and meticulous grams, KDE plots and Taylor diagrams to illustrate the influence of input
evaluation of twenty distinct machine learning algorithms, each with variables on viscosity, and ultimately offer a more insightful perspecti-
unique strengths and applications, to identify the most optimal model ve regarding the underlying physical phenomena governing nanofluid
for viscosity prediction. This approach not only enhances prediction rheology. Even though many studies have been conducted for predicting
accuracy but also elucidates the complex interactions between input the viscosity of mono-nanofluids or Newtonian fluid, the number of
variables (SVF, temperature, and shear rate) and the output variable researches is limited for predicting viscosity in non-Newtonian hybrid

4
M.S. Graish et al. Case Studies in Chemical and Environmental Engineering 11 (2025) 101180

Fig. 1. The schematic of the stages of this study.

nanofluids. Despite a significant amount of research on hybrid nano- polynomials with one or more variables. However, given their inherent
fluids, this particular study focuses on a less-explored area in the liter- complexity, it’s crucial to recognize that these techniques could not be
ature, iron-CuO/water-ethylene glycol non-Newtonian hybrid as straightforward as supervised algorithms and other neural networks.
nanofluids, which will provide further insight into their complex rheo- In this study, in the most basic section, 20 machine learning algorithms
logical behaviors as evidenced by the results obtained. Some research is (MLA) are introduced. Then 10 indicators of machine learning algorithm
limited to theoretical modeling and not used in real-world applications. evaluation are introduced. The data analysis is then handled initially to
Practical applications of the machine learning regression (MPR) model ascertain how the input data behaves on the result. The data is then sent
in designing heat exchangers, cooling systems, etc. are reported in this to machine learning algorithms and predictive operations are per-
study. Industrial relevance is further augmented by the provision of an formed. Finally, using evaluation indicators and data analysis charts, the
easy prediction methodology that could be implemented in real-time best machine learning algorithm and its mathematical equations are
monitoring systems. ML approaches are also not transparent, and so extracted. The schematic of the stages of this study is according to Fig. 1.
their results can be difficult to interpret. Furthermore, this study em- This research utilizes machine learning methods for several key
ploys explainable AI methods like Taylor diagrams and mathematical reasons. Various parameters, including the SVF (φ), temperature (T),
models to ensure clarity and interpretability, in line with contemporary and shear rate (γ), influence the rheological behavior and viscosity of
trends in machine learning research, where explainability is becoming a non-Newtonian hybrid nanofluids, such as nanofluids containing iron,
vital concern. Our study thus makes several contributions that set it copper oxide, water, and ethylene glycol. The viscosity is influenced in a
apart from previous work: extensive analysis of 20 machine learning complex, nonlinear manner by these parameters, making it challenging
algorithms, thorough model comparison based on ten performance for traditional modeling methods to accurately represent this behavior.
metrics, derivation of a mathematical equation for predicting viscosity, The capacity of machine learning algorithms to describe complicated
utilization of advanced data analysis techniques to investigate variable and nonlinear connections between variables makes them ideal in-
interactions and the consideration of practical applications and real- struments for precise viscosity prediction. It is time-consuming and
world usability of prediction results through a focus on nanofluids expensive to conduct experimental procedures to assess the viscosity of
with strong non-Newtonian properties and hybrid additives as well as nanofluids under various circumstances. To save time and money, ma-
explainable AI methods for greater transparency in the trained models. chine learning algorithms can precisely estimate viscosity, which means
Overall, this enhances the novelty of the study, as well as emphasizes fewer tests are needed. Machine learning algorithms, especially more
the importance of this work to the existing literature on nanofluid vis- complex ones like artificial neural networks (ANN) and multivariate
cosity prediction. polynomial regression (MPR), may learn from experimental data and
provide incredibly precise predictions. This degree of accuracy is
2. Methodology required for engineering equipment design and heat transfer system
optimization. Several dimensions and parameters, such as temperature,
Nanofluids have been used extensively in many contexts and set- shear rate, and SVF, are frequently included in nanofluid data. Machine
tings, greatly advancing both commercial operations and scientific learning algorithms may discover relationships that were previously
research. However, conducting research in a controlled laboratory undiscovered by examining this enormous volume of data. Because
setting requires a large time and money commitment because many machine learning algorithms are flexible, they may be adjusted to
studies are experimental. However, using strong AI systems could alle- accommodate new data or shifting test conditions. Examples of actual
viate these difficulties [38–43]. In data analysis, a wide range of tech- industrial applications that considerably benefit from these techniques
niques and strategies are used to make forecasts, with ongoing are the design of heat exchangers and cooling systems. With accurate
improvements and additions to these tools being common. Within the estimates of nanofluid viscosity, engineers may more effectively design
topic of machine learning, numerous subsets may be made from the heat transfer systems. As a result, engineering systems operate better
techniques provided here. Many different mathematical techniques are overall and use less energy and fuel. By using different evaluation
covered by these subgroups, including algorithms, linear equations, and criteria, such as mean square error (MSE), coefficient of determination

5
M.S. Graish et al. Case Studies in Chemical and Environmental Engineering 11 (2025) 101180

(R2), and mean absolute error (MAE), the performance of multiple al- defined. The phenomenon of both underfitting and overfitting is well-
gorithms can be compared, and the best approach for viscosity predic- known to occur in the multiple regression area. Making regulations is
tion can be selected. This comparison can be used by researchers to a practical way to handle and lessen these problems. This entails
choose the most accurate and error-free model for practical uses. The measuring the regression model’s parameters and penalizing them to get
results of this work could be expanded upon by researchers who are an optimal value. The result is a reduction in complexity while main-
interested in heat transfer systems and nanofluids. Additionally, the taining the model’s efficacy. Under these conditions, the decision is
provided models may be advantageous for both technical equipment fairly favorable.
and cutting-edge cooling systems in the workplace. The intricacy of
nanofluid behavior, the need to reduce the time and cost of experimental • Remarkably high number of descriptive variables
tests, the aim to increase prediction accuracy, the ability to handle data • Because of their enormous quantity, there are more variables than
with multiple dimensions, the study’s adaptability, and the study’s real- observations.
world industrial applications all make the use of machine learning al- • The descriptive variables exhibit one or more collinearities.
gorithms in this study completely justified. These techniques not only
improve viscosity prediction but also offer a workable solution for en- A “ridge regression” is a sort of linear regression model whereby the
gineering equipment design and heat transfer system optimization. coefficients are not estimated using ordinary least squares (OLS) but
The present work adopts a novel approach by leveraging 20 different rather a biassed estimator known as the “ridge estimator,” with a
machine learning algorithms for the prediction of non-Newtonian reduced variance than the OLS estimator. Sometimes the mean squared
hybrid nanofluids viscosity. These algorithms are selected based on error of the ridge estimator—the result of its variance times square of
their strengths and applications, as they improve prediction accuracy bias—is less than the OLS estimator’s. Ridge Regression and Lasso
and offer insights into the complex relationships between the input Regression punish the absolute value of the regression coefficients in
variables (solid volume fraction (SVF), temperature and shear rate) and somewhat similar ways. Another name for Lasso regression is L1 regu-
the target variable (viscosity). The selected algorithms include a variety larization, a frequently used technique in statistical modeling and ma-
of machine learning techniques, such as regression models, e.g. Ridge chine learning that forecasts and assesses the correlations between
Regression, Lasso Regression, and Elastic Component Regression (ECR), variables. The basic goal of LASSO regression is to compromise model
tree-based models, e.g. Decision Tree (DT) and XGBoost (XGB), neural accuracy with simplicity. This is achieved by adding a penalty compo-
networks (e.g. Multi-Layer Perceptron (MLP) and Back Propagation nent to the traditional linear regression model, therefore encouraging
Neural Network (BPNN)) and ensemble methods (e.g. Gaussian Process sparse solutions whereby some coefficients must be exactly zero. Lasso is
Regression (GPR) and Support Vector Machine (SVM)). This makes both quite useful for feature selection because of this quality since it can
linear and nonlinear relationships in the data explored. The rheological automatically identify and remove extraneous or redundant variables.
behavior of non-Newtonian hybrid nanofluids is complex and nonlinear Lasso and Ridge Regression are used in the linear regression method
as it depends on various parameters. The intricate interactions observed known as ECR. ECR incorporates a regularization factor and, like Lasso
in these datasets can be identified by algorithms like Multivariate and Ridge Regression, seeks to minimize the sum of squared errors be-
Polynomial Regression (MPR), which outperformed others in our test as tween observed and predicted values. But what distinguishes ECR is its
they can model polynomial behavior between inputs and outputs. Al- combination of L1 and L2 regularization methods. PLR is a covariance-
gorithms like MPR, ANFIS (Adaptive Neuro-Fuzzy Inference System), based regression technique that is quick, effective, and ideal. It is
and GMDH (Group Method of Data Handling) had prevailed in previous advised in regression scenarios with a high number of explanatory fac-
studies with similar datasets, providing accurate and robust predictions tors and a high probability of multicollinearity, or the correlation be-
with computational efficiency. The study compares algorithms by tween the explanatory variables. The PLR is a technique that narrows
testing 20 algorithms, providing researchers with a comprehensive down the collection of predictors from the original set of variables. A
benchmark for relatively simple and advanced algorithms and high- regression is then carried out using these predictors. Similar to k-Nearest
lighting the most suitable algorithm(s) for any application. We centered Neighbors, GPR makes predictions using training data. It delivers a
the evaluation of the performance of each algorithm on ten metrics prediction together with a quantification of uncertainty and performs
including, but not limited to, Mean Absolute Error (MAE), Root Mean well with tiny data sets. It is necessary to specify the previous mean and
Squared Error (RMSE), Coefficient of Determination (R2), and metrics prior covariance. It specifies the covariance as a kernel object. One kind
related to recall, precision, f-measure. This thorough evaluation not of probabilistic model that can be applied to regression tasks is the GPR
only highlights the best algorithm but also its strengths, weaknesses, model. This non-parametric approach is predicated on the idea that the
and where improvements can be made. The algorithms are commonly function that needs to be learned comes from a Gaussian process. The
adopted for industrial cases such as heat transfer systems and cooling model can produce predictions with a well-defined uncertainty thanks to
technologies. SVM and ANN models’ expressiveness has made them this assumption, which is helpful for tasks like active learning and
popular choices for predicting nanofluids’ thermal properties, making uncertainty-aware decision-making. Multiple regression, or MLR for
them relevant for this study. Moreover, MPR and DT provide inter- short, is a statistical method that predicts a response variable’s value
pretable results via mathematical equations and decision rules, using several explanatory variables. Modeling the linear relationship
respectively, which is key for engineers and scientists who want to between the explanatory (independent) factors and response (depen-
know which mechanisms affect viscosity. Thus, the choice of these 20 dent) variables is the aim of multiple linear regression. Since multiple
algorithms was motivated by their potential to accurately and consis- regression uses more than one explanatory variable, it is essentially an
tently predict viscosity for hybrid nanofluids while also providing a extension of ordinary least-squares (OLS) regression. MPR, sometimes
benchmark for further research in the area. Overall, the ranges do a referred to as simple polynomial regression, can be used to examine one
thorough job of exploring the published requirements, and emphasizing variable in a univariate scenario. Nevertheless, multiple polynomial
the balance of the MPR as the best algorithm – this says a lot about why regression allows it to be applied to analyses involving many variables.
it is vital to choose the correct tool for the job. Making selections will- The structure of real trees, which have a root that spreads out until it
become even better in future work as more influences like computa- reaches a leaf, is the model for DT algorithms. The above-described
tional cost and scalability are added. method is widely used for both the classification of discrete and
continuous data, as well as for regression methods that use real or
2.1. MLA continuous input data [44,45]. The reduction of the subgroups is the
first requirement for data segmentation. This technique performs well
Ridge Regression cannot be discussed unless regularization is when dealing with complex data; but, when the data grows in size, more

6
M.S. Graish et al. Case Studies in Chemical and Environmental Engineering 11 (2025) 101180

branches and leaves develop in the data, causing the data processing to The minimum value of an objective function is found using a
slow down. A method for reducing the leaf count was added to the al- gradient descent optimizing method. The gradient reduction or inverse
gorithm to circumvent this issue. Other methods could be used, like gradient reduction principle—which holds that the algorithm moves in
branch trimming and segmenting the total number of branches. In ma- the direction of the inverse gradient of the objective function at each
chine learning, support vector machines (SVM) [46,47] are frequently step—defines the basis of this method’s operation. The salient features
utilized for classification tasks. However, SVM may also be utilized for of the gradient descent method are as follows.
regression tasks using Support Vector Regression (SVR). SVR and SVM
share similar principles, but instead of classifying data points, SVR • Ability and simplicity implementation: Relatively simple on this al-
concentrates on forecasting continuous outputs. In this lesson, the gorithm to be able to understand and easy to implement.
principles of SVR will be examined, with a focus on sigmoid kernels, and • Optimization efficiency: Gradient descent is a powerful optimization
quadratic, and radial basis functions. SVR is capable of handling com- technique for finding the minimum of highly complex functions.
plex, non-linear relationships in data by utilizing these kernels. One of • Broad Range of Services: This technique is used in the fields of
the key varieties of deep neural networks is multi-layer neural networks, machine learning, neural networks, parameter optimization, and
or MLPs (Multi-Layer Perceptrons), which include three levels mini- optimization problems.
mum: an input layer, a hidden layer (or hidden layers), and an output • Local minimum: One of the problems with this algorithm is that it
layer. Each neuron, or unit, in these networks, is made up of several often gets stuck at local minimums. This is corrected by
different weights and is in charge of translating input into output. MLP applying better heuristics like stochastic gradient descent.
consists mostly of the following elements.
The gradient Descent method is a powerful optimization technique
• Despite its multi-layered architecture and great number of neurons, and is used extensively for parameter optimization and in machine
MLP can learn intricate and non-linear patterns [48]. learning. The least value of an objective function is what we want to
• MLP can fit fresh data and environmental changes [49]. find. To solve the parameter estimation optimization problem presented
• MLP moves data across layers using activation functions including in the model, a nonlinear optimization algorithm, the Levenberg-Mar-
sigmoid or ReLU [50]. quardt method, is employed. This method builds upon the gradient
• The MLP training method uses [51] the error backpropagation Lloyd method to overcome issues of fitting parameters to data. The
method to update network weights. Ultimately, MLP is a useful features of the Levenberg-Marquardt algorithm are as follows.
neural network model capable of spotting complex patterns and
applied for a variety of problems including picture recognition and • Great efficiency: This algorithm is highly accurate and fast in getting the
price prediction. solutions to parameter optimization problems.
• Levenberg-Marquardt is a solution for estimating parameters in
A machine learning method functioning as a nonlinear model is the nonlinear models and determining nonlinear problems.
Radial Basis Function (RBF) algorithm [52,53]. This method rests on • Stability: As a result of integrating gradient reduction and other tech-
radial basis functions, used as activation functions in the buried layers of niques, this algorithm solves hard problems with a high degree
the neural network. Important features of the RBF algorithm are as of stability.
follows. • Flexibility: Levenberg-Marquardt models can be improved with new data
over time.
• Radial Basis Functions: RBF uses the Gaussian function among others
to translate input to output. BFGS (for Broyden–Fletcher–Goldfarb–Shanno) is an optimization
• Capacity to learn intricate patterns: radial basis functions let RBF method for solving unconstrained optimization problems. This strategy
learn non-linear and complex patterns. requires Quasi-Newton methods which is an upgrade of Newton’s
• Dimension reduction technique: RBF usually reduces input di- method which solves non-linear optimization problems. The essen-
mensions and increases algorithm efficiency by using PCA or another tial components of the BFGS algorithm are as follows.
dimension-reducing technique.
• Usually utilized in RBF training are adaptive techniques including • High efficiency: BFGS is among the high-performance algorithms to solve
the adaptive nearest neighbor (KNN) algorithm. parameter optimization problems with high speed and accuracy.
• Application in non-linear problems: This is the second area in which,
Simply said, the RBF approach is a non-linear machine learning estimating parameters in non-linear models and solving non-linear opti-
method based on radial basis functions to address challenging pattern mization problems is performed efficiently.
issues. Rapid and efficient training of neural networks is accomplished • Stability: BFGS, being a quasi-Newton method, provides a higher level of
using an Extreme Learning Machine, or ELM, machine learning method. stability when working with hard-to-solve problems.
This method consists of a single hidden layer whose weights are selected • Customization: As time passes, the algorithm can optimize its models and
at random then support vector amplification (SVR) or linear regression make needed modifications to the data sets.
to maximize the weights of the output layer. The main elements of the
ELM algorithm are. Thus, the BFGS algorithm is an Unconstrained optimization algo-
rithm that is configured based on the Quasi-Newton method to optimize
• Fast training: ELM can train models quickly due to its use of random parameters. Hence, due to its excellent efficiency, stability, and adapt-
weights and a short training procedure. ability, it can be employed to address complex problems. A back-
• High efficiency: By employing the support vector amplification propagation perceptron neural network (BPNN) is an artificial
technique or linear regression, ELM can forecast and adapt to new neural network that takes in data to solve classification and prediction
data with high efficiency. problems. This method utilizes a multilayer neural network (MLP) ar-
• ELM runs automatically and does not need the configuration of chitecture which contains at least three layers: an input layer, a hidden
challenging factors like the number of layers or neurons. layer, and an output layer. The key ingredients of BPNN algorithm are
• Reiterability in several issues: ELM is simple and efficient so it may as follows.
be used in many different fields, from prediction to pattern
recognition. • Capacity for learning: BPNN can learn complex patterns and relation-
ships in data to improve prediction.

7
M.S. Graish et al. Case Studies in Chemical and Environmental Engineering 11 (2025) 101180

• Feature Selection: A Large Amount of Flexibility in the tuning of the al- • It makes accurate predictions: The power of neural networks and
gorithm can be applied to a given problem (or challenge) & alteration in fuzzy systems, gives ANFIS the ability to make accurate predictions,
the data. regarding data.
• Parallelism: The ability of BPNN to process information in parallel allows • Versatility: The algorithm can be applied in various fields, including
for rapid training and prediction. control, decision-making, prediction, and optimization.
• Wide application: This method is used in multiple domains like pattern
recognition, image processing, text processing, and data analysis. Simply put, ANFIS is a predictive model that predicts an interme-
diary through a sum-out of the output fuzzy set from the input neural
The ability to carry out nonlinear mapping through multi- networks by combining fuzzy systems and Artificial Neural Networks
layer networks with learning, adaptability, and parallel processing (ANNs). This algorithm finds applications in several areas like as opti-
gives the backpropagation algorithm an advantage in classification and mization, control and prediction, etc. It is highly adaptive and precise
prediction problems. XGBoost, a tree-based machine learning method, is in predicting data. The Group Method of Data Handling, or GMDH al-
used on prediction and classification problems [54,55]. This method gorithm is a machine learning algorithm for data modeling and pre-
uses gradient boosting architecture and offers an improvement over diction. This program is used to create complex and lethal prediction
more traditional boosting techniques such as AdaBoost. The main fea- models using hybrid techniques. The main components of the
tures of the XGBoost algorithm are. GMDH algorithm are.

• XGBoost is one of the machine learning algorithms with excellent per- • Creating groups helps GMDH to create complex and all-
formance that might be applied to handle challenging problems due to its encompassing models. This approach creates prediction models by
remarkable speed and accuracy. aggregating many feature groups gradually.
• Resistance to overfitting: This method uses reasonable parameter selection GMDH can automatically update its settings and adjust to changes in
and expectations (regularization) as control systems against overfitting. data and other situations.
• XGBoost can adapt to many facts and difficulties and generate better • Accurate prediction: This algorithm can forecast data with accuracy
forecasts with a great degree of accuracy. utilizing group building and algorithms combined.
• Two traits this method supports are reproducibility and parallel distri- • Wide application for GMDH spans control, optimization, prediction,
bution, which define features. and decision-making among other fields.

In conclusion, the XGBoost algorithm is a well-liked machine The GMDH algorithm is, all things considered, a machine-learning
learning algorithm due to its high performance, resistance to overfitting, method that groups data to create sophisticated and exact prediction
adaptability, and support for repeatability features. It is an incremental models. Wide-ranging applications for this method abound in many
gradient architecture-based high-performance algorithm for prediction different industries; it also has the ability to learn dynamically and yield
and classification problems. A probabilistic model for modeling and precise forecasts.
displaying probabilistic interactions between variables is the Bayesian
network. Conditional probabilities between variables and prior proba- 2.2. Tr and HP
bilities are represented in this model by the use of Bayesian probability
theory [56–58]. The following are the Bayesian network’s key Setting some parameters (like the activation function) to designated
characteristics. values helps a machine learning model to generalize to different data
usage patterns [48–50,63]. The goal is to optimize the model’s solution
• Visual representation: A directed graph with nodes for variables and by adjusting these parameters, also known as hyperparameters. The
edges for probabilistic correlations between variables is used to selection pressure coefficient, neuron count, and number of layers in a
illustrate a Bayesian network. GMDH neural network are referred to as hyper-parameters. The optimal
• Complex probabilistic relationships between variables can be prop- number of neurons and layers, as well as the activation function of the
erly displayed and modeled by this approach, which can also fully hidden layers for the LM, BFGS, BPNN, MLP, BN, and GD algorithms, are
represent conditional probabilities. tansig and purelin, respectively. The kind and quantity of membership
• Application of Bayesian rules: To update posterior probabilities in functions affect the ANFIS algorithm’s performance. The SVM algorithm
response to fresh data, a Bayesian network employs Bayesian rules. is a polynomial kernel function type, and its HP support vectors are
• Use in decision-making: This model works effectively when making determined by the epsilon value. HPs for other algorithms are set in the
decisions in complex and unpredictable situations. same manner. Seventy percent of the data are utilized as Tr data to begin
the regression, with the remaining portion being used as test and vali-
To put it briefly, a Bayesian network is a probabilistic model that dation data. Aiming to estimate the objective functions, different re-
represents and models probabilistic relationships between variables gressions use different paths and establish various mathematical
using Bayesian probability theory. This model is appropriate for relationships between the input variables to do so.
modeling and analyzing data in complicated and uncertain situations
because it makes use of visual representation, probabilistic connection 2.3. EC
modeling, Bayesian rule application, and application in decision-
making. The Adaptive Neuro-Fuzzy Inference System, or ANFIS [59, Regression algorithms’ performance needs to be assessed using a
60], is a hybrid model that models and predicts data by combining fuzzy range of indicators. The NSE, MARE, R, MSE, R2, MAPE, MAE, RMSE,
systems and neural networks [61]. This algorithm can learn from data MBE, and CvRMSE were the ten criteria that were used in this study. The
and is automatically adaptive [62]. The following are ANFIS’s key MSE quantifies how far the model deviates from actual values. This
characteristics. criterion aids in selecting the ideal model. A lower mean square error
(MSE) indicates that the model is doing better when there is less of a
• ANFIS combines fuzzy systems and neural networks: Fuzzy systems difference between the predicted and actual outputs. The RMSE criteria
allow for the accurate representation of fuzzy concepts, while neural is the square root of the error that is obtained by taking the square root
networks can learn complex patterns. of the MSE. Although MAE and MSE are similar, MAE calculates the
• Flexibility: This approach can self-adjust its parameters in reaction to error’s absolute value as opposed to its mean squared error. MAE offers a
both data changes and environmental changes. more accurate picture of the overall error since it rates mistakes

8
M.S. Graish et al. Case Studies in Chemical and Environmental Engineering 11 (2025) 101180

differently than MSE. Dividing the mistake by the real amount and then n ⃒
⃒ ⃒
1 ∑ ⃒yi,pred − yi,Experiment ⃒
⃒
multiplying the outcome by 100 yields the MAPE—a percentage vari- MARE = * ⃒ ⃒ (36)
ance. This percentage technique offers a consistent image of the ex- n i=1 ⃒ yi,Experiment ⃒
pected mistakes and helps one understand more easily. MAPE gives a
more exact estimate of the error rate than MSE. The linear correlation RMSE
CvRMSE = *100 (37)
between two variables is measured by R2. It computes the fraction of the yi,Experiment
dependent variable’s fluctuations that might be ascribed to the inde-
∑n ( )
pendent variable. In currently in-use definitions, R2 is referred to as the i=1 yi,Experiment − yi,predicte
coefficient of determination or the coefficient of detection. Thus, a high MBE = (38)
n
or low correlation coefficient indicates a strong or weak relationship
between the input variables and the goal functions. The lack of a link is
indicated by a number close to 0. The Nash-Sutcliffe efficiency is 2.4. AP
calculated as the ratio of the observed time-series variance divided by
the variance of the modeled time-series error variance. The Nash- Taylor diagram is one of the most significant and generally imple-
Sutcliffe Efficiency that arises when a model is perfect and has no esti- mented tools in scientific and engineering data analysis, thus making an
mated error variance is 1 (NSE = 1). An indicator of the average bias in a important contribution in forecasting the behavior of nanofluids. Such a
model’s predictions is the MBE. Even while MBE isn’t usually used as the diagram is often utilized by researchers when investigating properties of
only way to assess model error since it might not be able to catch very nanofluids that have gained special interest on account of their distinct
high individual prediction errors, it’s important for determining and thermal and physical properties in a wide array of applications ranging
measuring the average bias in the model’s outputs. A variable with from electronics to medicine and the energy sectors. There are multiple
positive bias represents an overestimation of data from datasets, reasons why the Taylor diagram is highly significant for the prediction
whereas a variable with negative bias represents an underestimation. of nanofluids.
Combining the evaluation of MBE with other metrics, such as correlation
coefficients, contributes to a more thorough comprehension of model • Very accurate and efficient: Taylor’s diagram provides a com-
performance. Better model accuracy is shown by lower error values and pound representation of multiple statistics such as the correlation
greater correlation coefficients, especially for directional variables. The coefficient, standard deviation, and RMSD (root mean square error)
coefficient of the Variation of the Root Mean Square Error is referred to with high accuracy in the evaluation of forecasting models. This
as CVRMSE. When calibrating models for measured nanofluid perfor- feature allows researchers to further assess their model performance
mance, the CVRMSE is employed. This statistic reveals the unstable more thoroughly and accurately.
observed correlation between the baseline period’s variables. It is the • Integrated visualization: A key characteristic of Taylor’s diagram
coefficient of variance of the expected input series concerning the is the ability of integrated visualization. This graph allows you not
observed one. If the CV(RMSE) is 10 %, the average distance between a only to compare both models visually but also to observe differences
point and the fit line is 10 % of the fit line. The Efficiency Valuation and similarities. This ability is particularly advantageous in studies
Organization (EVO) advises that for linear regressions the coefficient of of nanofluids, which contain many complexities.
determination (RMSE) should be less than half of the expected savings • Ability to compare models: Taylor’s chart offers a comparison of
fraction. Equations 29–38 [64,65] form the mathematical formulations different forecasting models. By using this tool, researchers can
for these evaluation indices. pinpoint and fine-tune the models that yield the most accurate pre-
dictions. This is a great success in improving the accuracy of pre-
∑n ( )( )
dictions for the type of nanofluid behavior.
i=1 yi,Experiment − yi,Experiment yi,predicte − yi,predicte
R = √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
)2̅ (29) • Other applications: Besides using the Taylor diagram for pre-
∑n ( )2 ∑ (
i=1 yi,Experiment − yi,Experiment
n
i=1 yi,predicte − yi,predicte
dictions about various nanofluid thermal and physical behavior,
applications in the simulation of fluid flows, heat transfer analysis
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
√∑ and other complex systems also utilize this method. The spectrum of
√n ( )2
these applications demonstrates the significance and high potential
√ y − yi,Experiment
√i=1 i,predicte of this tool in the research for nanofluids.
RMSE = (30)
n
The stability of the Taylor diagram [66] requires the geometric
n ⃒
∑ ⃒
⃒ ⃒
⃒yi,predicte − yi,Experiment ⃒ relationship of RMSD, STD, and R. The RMSD indices and standard
MAE = i=1 (31) deviation are retrieved using Equations (39) and (40) [66].
n
N (
1 ∑ )2
n (
∑ )2 σ ANN,Exp 2 = ypred,Exp (i) − ypred,Exp (i) (39)
yi,predicte − yi,Experiment N n=1
MSE = i=1 (32)
n
RMSD2 = σ Exp 2 + σpred 2 − 2 σ Exp σpred R (40)
( )2
n
∑ yi,predicte − yi,Experiment The Taylor diagram comes in two varieties [66]: a semicircle that
R2 = 1 − (33) displays both positive and negative correlation, and a quarter circle that
yi,Experiment 2
i=1 only shows positive correlation. In both cases, the R values are repre-
⃒ ⃒ sented by the circle’s radius on its arc, and the STD values are repre-
n ⃒ ⃒
∑ ⃒yi,predicte − yi,Experiment ⃒
⃒ yi,Experiment ⃒ sented by concentric circles in the circle’s center. The RMSD values are
i=1 ⃒ ⃒ displayed on the faces of concentric circles drawn concerning the hori-
MAPE = *100 (34)
n zontal reference point on the axis of the hollow circle. The reference
point is the value of the experimental data as shown by the STD. The
∑n ( )2
image indicates that the position of the data is determined by the RMSD,
i=1 yi,Experiment − yi,predicte
NSE = 1 − ∑ ( )2 (35) the STD, and the R between the analyzed and experimental data. Any
n
i=1 yi,Experiment − yi,Experiment experimental data that is found nearer the diagram’s reference point is
considered to be more precise when determining its value.

9
M.S. Graish et al. Case Studies in Chemical and Environmental Engineering 11 (2025) 101180

data to forecast the nanofluid set-back. The dataset for this study com-
prises of the output group itself as well as the input elements influencing
behavior of the output. The study outputs μnf and has φ, T, and γ as its
inputs specifically. First of all, one must look at the link between the
inputs and outputs. One good way to show the scope of these in-
teractions is with a heat map graph. The output is least affected if a
desired input has little to no link with the output. On the other hand, a
negative correlation value—that which is near to one—indicates that
the input clearly affects the output. Fig. 2 displays the heatmap that il-
lustrates how inputs influence outputs. The correlation matrix between
the input variables— SVF (φ), temperature (T), and shear rate (γ)—and
the output variable, viscosity (μnf), is graphically shown in Fig. 2’s
heatmap. Each pair of variables in the heatmap is given a numerical
value and a color gradient, with − 1 representing a strong negative
connection and +1 representing a strong positive correlation. If the
value is positive, then there is a positive correlation between the two
variables, and if the value is negative, then there is an inverse correla-
tion, where a reduction in one variable is correlated with a rise in the
other. The correlation value of 0.79 indicates that φ has the greatest
positive relationship with μnf, as seen by the heatmap. This proves that
Fig. 2. Diagramming data input and output behavior with a heatmap. the viscosity of the fluid grows exponentially with the SVF. We antici-
pate this result because the introduction of nanoparticles increases the
The following formula can be used to compute the μnf error analysis barrier to flow by strengthening intermolecular forces and obstructing
to assess the precision of values predicted by MLA [67]: the velocity of the fluid. On the other hand, there is a modest negative
association shown by the correlation coefficient of − 0.13 between
Error = μnf Exp − μnf pred (41) temperature (T) and μnf. Consistent with the idea that greater temper-
atures lessen intermolecular tensions and enhance fluidity, the negative
Another graph is the error histogram. The higher the frequency of the
sign indicates that viscosity reduces somewhat with increasing tem-
error column is closer to zero, which means that the machine learning
perature. Likewise, γ is negatively correlated with μnf, with a value of
algorithm has been able to perform the prediction well.
− 0.47, indicating that a considerable decrease in viscosity occurs when
the shear rate is increased. Because the nanofluid’s particles align under
3. Results and discussion
shear, the flow resistance is reduced, and this behavior is a reflection of
that characteristic. Since temperature and shear rate are inversely
3.1. The study of data behavior
related to viscosity, the negative sign in their correlations is a result of
this inverse connection. A decrease in viscosity and a weakening of
This study incorporates the method and findings of Bahrami et al.
intermolecular cohesion result from an increase in the thermal energy of
[68]. Fe–CuO/water-ethylene glycol hybrid NFs flow was all thoroughly
fluid molecules as the temperature rises. The negative association be-
investigated in their thermophysical property studies. Important metrics
tween shear rate and viscosity is a result of nanoparticle structural
such as DV is tested within the 25–50 ◦ C temperature range, shear rate
rearrangement caused by strong shear, which reduces internal friction.
(γ) range of 3–100 Rpm and VF range of 0–1.5 %. To be more specific,
If the correlation coefficient (R) is between − 1 and +1, then the
204 points were taken from Bahrami et al.’s experimental dataset [68].
value represents both the strength and direction of the degree of the
This section of the work explores the behavior of the dataset used to
linear relationship between two random variables. Positive value (R >
predict the viscosity of a hybrid nanofluid devoid of Newtonian char-
0) means that as the first variable rises, the second variable rises too,
acter. The output variable in the dataset is viscosity (μnf), which addi-
while negative value (R < 0) means that as the first variable rises the
tionally incorporates shear rate (γ), temperature (T), and SVF(φ).
second variable falls. These characteristics are seen in hybrid-
Analysis of this data is necessary for numerous reasons.
nanofluids. The negative correlation between temperature (T) and
viscosity (R = − 0.13) represents the established inverse relationship,
• The study particularly shows how the input variables are linked with
wherein the viscosity is known to decline as temperature increases.
the output variable (viscosity) via a heatmap. This helps one to
However, as the temperature increases, due to the increase in thermal
identify the most influencing input parameters on viscosity.
energy of the fluid molecules, the attractive forces between the mole-
• Box charts are a helpful tool for a better knowledge of how the input
cules are overcome, thus reducing resistance to flow. The shear rate (γ) is
elements influence viscosity. These graphs show the effects on the
also more strongly inversely correlated with viscosity (R = − 0.47),
viscosity of the nanofluid of temperature, shear rate, and nano-
which corresponds to the shear-thinning behavior of non-Newtonian
particle concentration.
fluids. Nanoparticles orient with velocity field and this orientation re-
• Knowledge of the distributions and correlations of the input vari-
duces the internal friction, thus decreasing the viscosity at high shear
ables helps one identify the optimal features to teach machine
rates. On the other hand, the volume fraction of nanoparticles, ϕ highly
learning models. In this sense, the models will be more accurate
influences the viscosity (R = 0.79); as the number of nanoparticles in-
predictors since they might be trained on data relevant for the
creases, intermolecular interaction and hindrance within the flow path
output.
also increase and lead to increasing viscosity.
• Understanding the effect of input parameters on viscosity will help
The heat map of relationships between inputs and outputs is illus-
one to maximize the efficiency of nanofluids in useful environments
trated in Fig. 2, where correlation coefficients between input variables,
like heat exchangers and cooling systems. Engineers must find the
solid volume fraction (ϕ), temperature (T), and shear rate (γ), and
primary factors influencing viscosity if they are to maximize the
output variable, viscosity (μnf), are shown. The main outcomes indicate
enhanced thermal properties of nanofluids.
that ϕ correlates positively (R = 0.79) with viscosity due to the nano-
particle concentration, which if increased, viscosity substantially.
Machine learning techniques must first examine the behavior of the
Temperature (T) has a weak inverse correlation (R = − 0.13), meaning

10
M.S. Graish et al. Case Studies in Chemical and Environmental Engineering 11 (2025) 101180

behavior of composite nanofluids is closely related to particle aggre-

gation, Brownian motion, and van der Waals forces, which is an inherent
characteristic of composite nanofluids. The MPR is a good model for
such data because it allows us to model the complex interactions be-
tween these variables. First, the model is grounded in mathematical
equations and can be well generalized to the dataset. Simples’ linear
models would not be able to capture many complicated dependencies
between the variables.
According to the sensitivity analysis carried out in this work, shear
volume fraction (SVF) has the most impact on viscosity, followed
closely by shear rate and then temperature. The heat map (Fig. 2) con-
firms this result with correlation coefficients. The correlation coefficient
between viscosity (μnf) and nanoparticle volume fraction (ϕ) is 0.79 and
with increasing nanoparticle volume fraction, intermolecular forces are
strengthened, resulting in increasing viscosity. The shear rate (γ) shows
a moderate negative coefficient of correlation of − 0.47, in which at
higher shear rates the nanoparticles were arranged with the flow and the
value of internal friction was less, where viscosity is decreased. Tem-
perature (T) has a correlation coefficient of − 0.13; that is, it has a weak
positive correlation; Viscosity decreases as temperature increases; As
intermolecular cohesion weakens, its effect is small compared with the
volume fraction of nanoparticles and shear rate. Thus, the volume
fraction of nanoparticles is the dominant factor influencing viscosity,
followed by shear rate and temperature.
Fig. 3 shows box plots that show how the input variables, heat (T),
shear rate (γ), and SVF, behave and how they affect the output variable,
viscosity (μnf). You may learn a lot about the inputs’ effects on the vis-
cosity from box plots, which show the distribution, variability, and re-
lationships of the data. Fig. 3 shows that when the temperature
increases, μnf drops. The reason for this tendency is that when the
temperature of the fluid molecules rises, their thermal energy also in-
creases, which weakens the intermolecular interactions and reduces the
flow resistance. As a result, the viscosity of the fluid decreases. On the
contrary, there is a significant increase in nanofluid viscosity as φ be-
comes larger. The presence of nanoparticles intensifies the interactions
between molecules, leading to the formation of bigger clusters as a result
of van der Waals forces. These clusters, in turn, increase the viscosity of
the fluid and hinder its movement. It has been noted that the impact of γ
on μnf is not linear. According to the box plot, viscosity initially rises as γ
increases, but after reaching an ideal value, it steadily decreases. The
shear-thinning behavior of non-Newtonian fluids—in which the align-
ment of nanoparticles at higher shear rates decreases internal resistance
to flow—can be used to explain these phenomena. You can see the data’s
spread and variability clearly in the box plots as well. The higher vari-
ation of viscosity values for φ indicates the high effect of particle con-
centration on the behavior of the fluids. T has a smaller but significant
effect, as indicated by its smaller range and constant decrease
in viscosity. Thus, the relationship between γ and viscosity is condi-
tional on the shear rate and other mixing factors, which results in a
large data dispersion. These trends illustrate relationships between
factors-influencing input and viscosity where φ is the most sensitive
input when compared to γ and T, which are less correlated with vis-
Fig. 3. Box plot to see the behavior of a) φ b) Temperature c) γ cosity; in some cases, the box plots also reflect outliers representing
abnormal behavior of viscosity caused by high φ, T, or γ values. This
that an increased temperature reduces viscosity by loosening intermo- proves the fact that the rheological properties of the nanofluid are
lecular forces. Also, shear rate (γ) was negatively correlated (R = − 0.47) nonlinear and complex. To optimize the performance of non-Newtonian
which is more negative than the second and third ones, but for non- hybrid nanofluids in real-world applications, it is vital to consider the
Newton fluids such as paint till now we handle them as Newton fluids viscosity of these fluids, and Fig. 3 shows how these elements interact
to do so when viscosity becomes smaller with the increase of γ. Some ML critically.
models work better because they are based on the fundamental prop- These figures demonstrate that the lowest μnf happens at the highest
erties of the data. For instance, a mad scientist’s right algorithm like temperature together with the lowest γ and φ. This phenomenon is
multivariate polynomial regression (MPR) works well due to its ability explained by the following theories.
to model polynomial relationships between input variables to output
variables. In comparison, many non-linear models like elastic compo- • The random movement of nanoparticles inside the base fluid, or
nent regression (ECR) fail when the interactions are higher-degree, the Brownian motion, is one of the elements influencing viscosity. The
relationships being considered. To some extent, the non-Newtonian

11
M.S. Graish et al. Case Studies in Chemical and Environmental Engineering 11 (2025) 101180

Table 1
Values of 10 EC for 20 MLA algorithms.
Algorithm R R2 RMSE NSE MAE MBE MSE MARE MAPE CvRMSE

ANFIS 0.9827 0.9657 3.498 0.975 1.338 − 2.132 12.233 0.075 11.5 23.14
MPR 0.9961 0.992 1.66 0.994 1.146 − 5*10-12 2.753 0.129 7.61 9.98
BN 0.9888 0.9777 2.84 0.983 2.25 5.2 8.056 0.228 25.32 18.775
BFGS 0.9927 0.985 2.28 0.989 1.777 − 9.53 5.22 0.198 22.582 15.11
GD 0.9888 0.9776 2.822 0.984 2.15 2.766 7.965 0.223 24.93 18.67
LM 0.9888 0.9778 2.81 0.984 2.194 5.015 7.89 0.228 25.13 18.581
BPNN 0.9087 0.826 7.97 0.87 3.97 − 49.89 63.58 0.184 21.143 52.744
DT 0.9808 0.962 3.676 0.972 1.707 − 6*10− 13 13.52 0.083 8.4 24.32
ECR 0.229 0.227 16.6 0.435 10.874 − 2*10− 12 275.567 0.443 112.11 109.81
PLR 0.8069 0.651 11.15 0.75 7.072 9*10− 11 124.4 0.03 63.23 73.777
ELM 0.9641 0.929 17.06 0.402 16.365 − 1631 291.28 0.619 221.76 112.9
GMDH 0.9903 0.981 2.65 0.986 1.941 30.22 7.026 0.19 20.55 17.53
GPR 0.984 0.968 3.4 0.976 1.877 0.393 11.575 0.144 12.49 22.5
Lasso 0.7996 0.64 12.43 0.683 6.714 − 3*10− 12 154.5 0.534 51.17 82.22
MLP 0.9877 0.976 3.01 0.981 2.34 − 18.9 9.056 0.23 24.31 19.91
MLR 0.8068 0.651 11.153 0.745 7.071 0.0184 124.395 0.031 63.232 73.777
RBF 0.9913 0.983 2.505 0.987 1.691 19.4 6.27 0.021 18.61 16.57
Ridge 0.9708 0.942 10.283 0.783 4.6 444 105.73 0.291 19.55 68
SVM 0.931 0.867 8.93 0.836 2.703 193.43 79.73 0.099 8.18 59.06
XGB 0.9946 0.942 4.55 0.958 2.48 224.4 20.695 0.166 14.25 30.09

constant collisions between the molecules of the base fluid and the clusters that give rise to an increase in viscosity, whereas at high shear
nanoparticles are what create this chaotic motion. rates, alignment of the particles leads to a reduction of drag. Such dy-
• The addition of the nanoparticles causes the nanomaterials to scatter namics require flexible predictors for accurate prediction. Data-wise,
and aid in the formation of bigger, symmetrical nanoclusters through ELM (Extreme Learning Machine) and ECR (Elastic Component
the base fluid’s van der Waals contact with the nanoparticles. Vis- Regression) are over-simplified (parameter-inefficient), resulting in
cosity rises because ethylene-glycol molecules are prevented from poor generalizability, while MPR falls somewhere in the middle, able to
rubbing against one another by these nanoclusters. reach high accuracy whilst also remaining computationally efficient by
• Nanostructures lose some of their key properties, such as density, due employing a polynomial model that directly incorporates the governing
to their small size and relatively low mass, making them primarily physical relationships.
dependent on surface and intermolecular interactions. This results in
an extraordinarily high surface-to-volume ratio.
3.2. Results for EC
• Viscosity rises in a base fluid containing nanomaterials as intermo-
lecular forces rise.
As mentioned, this study aims to predict laboratory data by machine
• Viscosity, a feature of liquids that depends on temperature, is caused
learning algorithms. For this purpose, 20 machine learning algorithms
by cohesive forces between molecules. A rise in temperature causes
ANFIS, BFGS, BN, GD, LM, BPNN, DT, ECR, PLR, ELM, GMDH, GPR,
liquids to become less viscous. At higher temperatures, liquid mol-
Lasso, MLP, MLR, MPR, RBF, Ridge, SVM, and XGB were used to achieve
ecules have more energy and can overcome the forces of molecular
this goal. A 204*4 dataset is introduced to these algorithms, the first 3
cohesiveness. Consequently, molecules that are energetic move more
columns of which are inputs (T, γ, and φ) and the final column is the
quickly. The intermolecular connections become weaker and the
target or output function (μnf). Then the data was divided and 70 % of it
flow resistance reduces as temperature rises. Newtonian nanofluid
was assigned to the training data, 15 % to the validation data and the
viscosity consequently falls as temperature rises.
rest to the test data. In the following, hyperparameters were set and
• It makes sense to consider how the viscosity of the nanofluid is
changed for each algorithm to determine the best state of each algo-
impacted by the temperature-dependent Brownian motion of the
rithm. Finally, a machine learning algorithm predicts a value for each
nanoparticles.
data. To evaluate the performance of each algorithm, evaluation indices
• Vessel viscosity and flow resistance decrease with increasing tem-
or EC should be used. In this study, 10 evaluation indicators were used.
perature because of an increase in the intermolecular distance be-
Table 1 shows the value of these indicators.
tween the nanoparticles and the base fluid.
Twenty different machine learning algorithms were tested for their
ability to predict the viscosity of non-Newtonian hybrid nanofluids. The
Boxplots are presented in Fig. 3 to examine how the input variables
results of these tests are shown in Table 1. Several metrics are used to
(ϕ, T, γ) affect viscosity (μnf). The results, revealed at the same time, are
evaluate the models’ accuracy and reliability. These include R-squared,
that the increase of viscosity is qualitative with increasing solid volume
R2, RMSE, NSE, MAE, MSE, MPE, and CvRMSE, among others. With
fraction (ϕ) and that there is little overlap between levels of ϕ. Rising T
higher numbers indicating greater correlations, the R shows the degree
uniformly produces lowered viscosity, but T (compared to ϕ) has a
and direction of the linear link between expected and actual values. ECR
much smaller range of behavior. For the shear rate (γ), a nonlinear trend
has the lowest R-value of 0.229, indicating weak correlation, in contrast
was observed that showed the shear rate first increased and then
to MPR, which obtains the greatest R-value of 0.9961, indicating greater
decreased with viscosity with a specific optimal point, which is in line
prediction accuracy among the methods. R2 is a measure of how much of
with the characteristic of nozzle shear of non-Newtonian fluids. Machine
the observed data variation can be explained by the model. ECR has the
learning models capable of modeling multilinear relationships, like
lowest R2 of 0.227, showing that it cannot adequately predict the vis-
multivariate polynomial regression (MPR), Gaussian process regression
cosity data, while MPR scores best with an R2 of 0.992, showing that it
(GPR), radial basis functions (RBF), for predictions, outperform models
captures virtually all data variability. Measuring the average size of
constrained to linear assumptions – partial least squares regression
mistakes, smaller numbers indicate greater performance according to
(PLR). DTs also perform well because they can split data along thresh-
the root mean squared error (RMSE) and the mean squared error (MSE).
olds ideal for the piecewise trends observed. Non-Newtonians are
There is very little discrepancy between the predicted and actual values
known for their complexity; At low shear rates particles may induce
when using MPR, as its RMSE is 1.66 and its MSE is 2.753. Alternatively,

12
M.S. Graish et al. Case Studies in Chemical and Environmental Engineering 11 (2025) 101180

ECR demonstrates substantial prediction errors with an RMSE of 16.6 providing very reliable and accurate predictions. In comparison to MPR,
and an MSE of 275.567, the highest values. The NSE measures how well other algorithms like BFGS, BN, and GPR do decent work; but they do
the models can predict outcomes in comparison to the average of the not achieve MPR’s level of accuracy. On the other hand, algorithms like
data points. With an NSE of 0.994, MPR is a very effective model, but ECR, PLR, and ELM perform poorly because they have a hard time
ECR’s 0.435 shows that it is not very good at making predictions. As accurately modeling the dataset’s nonlinear connections. Choosing the
with RMSE, a smaller number indicates better accuracy when calcu- right algorithm for predictive modeling is crucial, as shown by MPR’s
lating the MAE, which gives an average of the absolute errors. The MAE ability to capture complicated dynamics and provide very precise vis-
of 1.146 for MPR shows that it makes accurate predictions, but the MAE cosity forecasts.
of 10.874 for ECR shows that it is quite inaccurate. If the MBE is nega- According to Table 1, the value of R is different for each algorithm
tive, then the predictions are underestimated, and if it is positive, then and ranges from 0.229 to 0.9961. If this value is close to 1, the best MPR
the predictions are overestimated. This statistic displays the average bias algorithm and the worst one should be related to the ECR algorithm. The
in predictions. While ECR shows a greater bias, MPR provides unbiased value of R2 is an index like R, whose value varies from 0.227 to 0.992 for
predictions with a minimal MBE of − 5*10− 12. One way to compare the algorithms, the best algorithm is MPR and the worst one is ECR algo-
performance of different models is using the mean absolute percentage rithm. Also, the NSE index is following these two indices. Other in-
error (MAPE), which represents mistakes as a percentage. In comparison dicators are related to the amount of error between the predicted points
to ECR’s 112.11 % MAPE, which indicates exceptionally high prediction and the experimental data, the smaller the value, the better. Only the
errors, MPR’s MAPE is the lowest at 7.61 %. As a last metric, the coef- MPR algorithm has a CvRMSE value below 10 %, which is optimal, and
ficient of variation of RMSE (CvRMSE) indicates the consistency of the rest of the algorithms are more than this value, and this index rea-
performance compared to the error; lower values show more stable ches 112.89 for the ELM algorithm, which is the worst value. MSE and
predictions. With a CvRMSE of 9.98 %, MPR has the greatest prediction RMSE are close indicators and the only difference between them is that
accuracy, while ECR’s CvRMSE of 109.81 % shows that its forecasts are one of them is squared. Obviously, if the value of this index is close to
quite unpredictable. These numbers show that MPR is the top algorithm, zero, the machine learning algorithm has less error, the lowest error is

Fig. 4. Histogram plot error for a) ANFIS b) BFGS c) BN d) GD e) LM f) BPNN g) DT h) ECR i) PLR j) ELM k) GMDH l) GPR m) Lasso n) MLP o) MLR p) MPR q) RBF r)
Ridge s) SVM t) XGB algorithm.

13
M.S. Graish et al. Case Studies in Chemical and Environmental Engineering 11 (2025) 101180

Fig. 4. (continued).

14
M.S. Graish et al. Case Studies in Chemical and Environmental Engineering 11 (2025) 101180

Fig. 4. (continued).

15
M.S. Graish et al. Case Studies in Chemical and Environmental Engineering 11 (2025) 101180

Fig. 4. (continued).

16
M.S. Graish et al. Case Studies in Chemical and Environmental Engineering 11 (2025) 101180

Fig. 4. (continued).

17
M.S. Graish et al. Case Studies in Chemical and Environmental Engineering 11 (2025) 101180

Fig. 4. (continued).

18
M.S. Graish et al. Case Studies in Chemical and Environmental Engineering 11 (2025) 101180

Fig. 4. (continued).

19
M.S. Graish et al. Case Studies in Chemical and Environmental Engineering 11 (2025) 101180

Fig. 4. (continued).

20
M.S. Graish et al. Case Studies in Chemical and Environmental Engineering 11 (2025) 101180

Fig. 4. (continued).

21
M.S. Graish et al. Case Studies in Chemical and Environmental Engineering 11 (2025) 101180

Fig. 4. (continued).

related to the MPR algorithm, and the highest error is related to the ELM Fig. 4, which sheds light on how well they performed in terms of pre-
algorithm. MAE and MARE indices are other indices that are close to diction. With a concentration of error values near zero being the ideal
each other, which according to two indices, RMSE and MSE, are related condition, the histogram shows the frequency distribution of prediction
to the error of the points, and the smaller the error, the better. The MPR mistakes. To show the performance disparities, we have underlined the
algorithm has the lowest error and the ELM algorithm has the highest greatest error for each method. To illustrate its superior accuracy, the
error. The MAPE index is related to the error percentage of the points, Multivariate Polynomial Regression (MPR) method displays the lowest
and the lower the error percentage, the better the MPR algorithm with maximum error among competing algorithms. The little dispersion of
7.5 % error, and the ELM algorithm with 221.76 % error, are the best the error bars centering on zero in the MPR histogram shows that the
and worst algorithms. According to other indicators, MBE is related to experimental and projected viscosity values are quite close to one other.
the error and according to the previous, the best value is related to the The Extreme Learning Machine (ELM) method, on the other hand, has
MPR algorithm, and the worst value is related to the ELM algorithm. the most error values, which are shown by a larger and more distributed
Therefore, according to EC analysis, the best algorithm is MPR and the histogram. This indicates that its predictions are very inaccurate. The
worst algorithm is ELM. histograms of the Adaptive Neuro-Fuzzy Inference System (ANFIS) and
Gradient Descent (GD) algorithms reveal error distributions that are
3.3. Results for AP somewhat wider than MPRs but smaller than ELMs, indicating inter-
mediate performance. Although it isn’t quite as good as MPR, the De-
In this part, three parts of graphical analysis are discussed. The first cision Tree (DT) method still shows a low error distribution, which
part will deal with the graphical analysis of the error histogram. Ac- means it’s successful. Bayesian Network (BN), Broyden-Fletcher-
cording to the previous explanations, the histogram chart is more Goldfarb-Shanno (BFGS), and Gaussian Process Regression (GPR) are
optimal as long as its error column is close to zero. The error histogram three algorithms with histograms that have modest error distributions.
diagram for all algorithms can be seen in Fig. 4. The error histogram for This indicates that these algorithms strike a good compromise between
each of the twenty ML algorithms included in this research is shown in computing efficiency and accuracy. Algorithms having a broader error

22
M.S. Graish et al. Case Studies in Chemical and Environmental Engineering 11 (2025) 101180

distribution, such as Partial Least Squares Regression (PLR) and Elastic quantitative measures given in Table 1, our study confirms that MPR is a
Component Regression (ECR), are less effective at accurately predicting good fit for forecasting the viscosity of hybrid nanofluids. According to
viscosity. With MPR showing the best results for viscosity prediction, Fig. 4, the lower the maximum error, the more optimal the algorithm.
and ELM and ECR the worst, the general trend in the histograms high- For ANFIS, BFGS, BN, GD, LM, BPNN, DT, ECR, PLR, ELM, GMDH, GPR,
lights the algorithmic performance variability. The dependability of the Lasso, MLP, MLR, MPR, RBF, Ridge, SVM and XGB algorithms, the
findings is directly affected by the distribution and quantity of mistakes; maximum error value is 23, 10, 23, 10, 9, 37, 9, 72, 50, 73, 13, 17, 63, 9,
hence, the error histograms graphically show how important it is to pick 50, 6.7, 11, 46, 53 and 25. According to Fig. 4, the lower the maximum
the correct technique for predictive modeling. Consistent with the error, the more optimal the algorithm. Therefore, the MPR algorithm has

Fig. 5. Kde plot for a) ANFIS b) BFGS c) BN d) GD e) LM f) BPNN g) DT h) ECR i) PLR j) ELM k) GMDH l) GPR m) Lasso n) MLP o) MLR p) MPR q) RBF r) Ridge s)
SVM t) XGB algorithm.

23
M.S. Graish et al. Case Studies in Chemical and Environmental Engineering 11 (2025) 101180

Fig. 5. (continued).

24
M.S. Graish et al. Case Studies in Chemical and Environmental Engineering 11 (2025) 101180

Fig. 5. (continued).

25
M.S. Graish et al. Case Studies in Chemical and Environmental Engineering 11 (2025) 101180

Fig. 5. (continued).

26
M.S. Graish et al. Case Studies in Chemical and Environmental Engineering 11 (2025) 101180

Fig. 5. (continued).

27
M.S. Graish et al. Case Studies in Chemical and Environmental Engineering 11 (2025) 101180

Fig. 5. (continued).

28
M.S. Graish et al. Case Studies in Chemical and Environmental Engineering 11 (2025) 101180

Fig. 5. (continued).

29
M.S. Graish et al. Case Studies in Chemical and Environmental Engineering 11 (2025) 101180

Fig. 5. (continued).

30
M.S. Graish et al. Case Studies in Chemical and Environmental Engineering 11 (2025) 101180

Fig. 5. (continued).

31
M.S. Graish et al. Case Studies in Chemical and Environmental Engineering 11 (2025) 101180

Fig. 5. (continued).

the lowest error and the ELM algorithm has the highest error. So, the between input and output. MPR is highly accurate, because the algo-
best algorithm in terms of histogram is the MPR algorithm and the worst rithm itself directly models the interaction between variables in poly-
is the ELM algorithm. nomial equation form that are derived from the data. Methods like
The error histograms in Fig. 4 provide insights into the distribution of ANFIS (Adaptive Neuro-Fuzzy Inference System) and GPR (Gaussian
predicted errors across the SVM, RF, and MLP algorithms. The Process Regression) also have good accuracy but with comparably lower
results indicate a close distribution of error by means of multivariate than that of MPR because they can handle high-cost complicated
polynomial regression (MPR) centered at zero (i.e. when predicted and nonlinear data owing to their versatility and ability to learn higher-
actual values deviate the least). While, the error distribution of ELM dimension functions with the help of kernel. Physically, this means
(Extreme Learning Machine) and ECR (Elastic Component Regression) that we need to have a good understanding of the interaction between
is wide and scattered, which shows that predictions have more differ- ϕ, T, and γ. MPR does a great job delivering this need with mathematical
ences and the uncertainties of predictions. The better performing ma- equations that reflect these interactions explicitly. On the computa-
chine learning models actually capture more complex relationships tional side, random initialization of the weights leads to a sub-

32
M.S. Graish et al. Case Studies in Chemical and Environmental Engineering 11 (2025) 101180

Table 2
The performance of 20 machine learning algorithms in this study.
Algorithm Peak Density Width of Error Max Error Overall
(Near Zero) Distribution Magnitude Accuracy

MPR Highest Narrowest 6.7 Excellent

ANFIS High Moderately 22 Good
Narrow
GPR High Moderate 17.65 Good
DT Moderate Moderate 9 Good
BN Moderate Moderate 23 Average
BFGS High Moderately 10.15 Good
Narrow
GD Moderate Moderate 10.15 Average
MLP Moderate Moderate 9.51 Average
BPNN Moderate Wide 37 Below
Average
RBF High Moderately 11.55 Good
Narrow
ECR Lowest Widest 96 Poor
PLR Low Wide 51.84 Poor
ELM Lowest Widest − 11.7 Poor
Lasso Low Wide − 64.74 Poor Fig. 6. Taylor diagram for 20 MLA
XGB Moderate Moderately 25 Average
Narrow
to zero but with wider dispersion. These trends indicate that these al-
Ridge Moderate Wide 46.78 Below
Average gorithms can make educated guesses, but they can’t compare to MPR in
SVM Moderate Wide 54.2 Below terms of accuracy and reliability. The wide variation in the error dis-
Average tributions across the algorithms is shown by the overall analysis in
GMDH High Moderate − 13.36 Good
Fig. 5. SVF, temperature, and shear rate are the input factors, and MPR’s
LM Moderate Moderate 9.47 Good
ECR Lowest Widest 96 Poor
narrow and clearly defined KDE plot demonstrates its ability to accu-
rately represent the nonlinear interactions between these variables and
the output variable, viscosity. On the other side, ELM and ECR’s scat-
optimal solution, for example, in algorithms like the extreme learning tered KDE plots show that they do a poor job of capturing the data’s
machine (ELM). In contrast, MPR solved this issue and achieved much complicated dynamics. Since MPR consistently underperformed in this
more precise prediction due to its derivation of constant equations that investigation, it is clear that using the right algorithm is crucial for
were tuned directly to the data set. making accurate and dependable viscosity predictions. According to
Among other analysis charts, there is a Kde chart with error values Fig. 5, the lower the error rate, the more optimal the algorithm. For
indicated. According to this chart, the number of points close to zero ANFIS, BFGS, BN, GD, LM, BPNN, DT, ECR, PLR, ELM, GMDH, GPR,
indicates that the chart has less error and is a better algorithm. The error Lasso, MLP, MLR, MPR, RBF, Ridge, SVM, and XGB algorithms, the
value is minimal for an algorithm with proper correlation. These values maximum error value is − 22, 10.15, 9.53, 10.15, 9.47, − 80.48, 9.84, 96,
can be seen in Fig. 5. Fig. 5 shows the error value kernel density esti- 51.84, − 11.7, − 13.36, − 17.65, − 64.74, 9.51, 51.84, − 6.59, − 11.55,
mation (KDE) plots for twenty machine learning algorithms. The 46.78, 54.2 and 26.27, respectively. Therefore, the MPR algorithm has
explanation for this figure emphasizes the prediction accuracy and the least error and the ECR algorithm has the most error. Therefore, the
dependability of the techniques. For each method, these KDE graphs best algorithm in terms of the correlation diagram is the MPR algorithm
show the probability density of mistakes; a peak close to zero indicates and the worst is the ECR algorithm. Table 2 shows the performance of
little prediction error. The KDE plot of the Multivariate Polynomial each algorithm.
Regression (MPR) method shows a tiny peak near zero, further Another analysis diagram is the Taylor diagram. In this diagram,
demonstrating its excellent performance. This shows that it is quite ac-
curate and reliable, as most of its predictions match the observed vis-
cosity values very well. However, the Extreme Learning Machine (ELM)
algorithm’s prediction performance and considerable departures from
the experimental data are shown by its wide and scattered KDE plot,
which has a peak that is far from zero and a significant error density at
larger magnitudes. As their KDE plots are modestly concentrated around
zero but with larger bases compared to MPR, algorithms like Adaptive
Neuro-Fuzzy Inference System (ANFIS), Gradient Descent (GD), and
Gaussian Process Regression (GPR) show moderate performance. Pre-
dictions made by these algorithms are not as accurate as MPRs, but they
do a decent job of capturing patterns in viscosity data. Bayesian Network
(BN) and Decision Tree (DT) algorithms also do a respectable job; KDE
plots reveal acceptable peaks around zero, albeit error variability is
somewhat larger. Alternatively, KDE graphs generated by Elastic
Component Regression (ECR) and Partial Least Squares Regression
(PLR) are distorted and uneven, indicating poor accuracy and large
predicted errors. Particularly inaccurate is the ECR KDE figure; the error
density there is very non-zero and spans a wide range. In comparison to
the best algorithms, the Radial Basis Function (RBF), Multi-Layer Per-
ceptron (MLP), and Backpropagation Neural Network (BPNN) algo-
rithms display intermediate behavior; their KDE plots show peaks close Fig. 7. Plot of experimental data with the points predicted by the
MPR algorithm.

33
M.S. Graish et al. Case Studies in Chemical and Environmental Engineering 11 (2025) 101180

three indicators R, STD and RMSD are specified. Taylor’s diagram can be 3.4. Derivation of mathematical equations between inputs and outputs
seen in Fig. 6. For ANFIS, BFGS, BN, GD, LM, BPNN, DT, ECR, PLR, ELM,
GMDH, GPR, Lasso, MLP, MLR, MPR, RBF, Ridge, SVM and XGB algo- Machine learning models can address nonlinear viscosity behavior
rithms, the value of R is equal to 0.9827, 0.9927, 0.9888, 0.9927, by capturing complex relationships between the input (nanoparticle
0.9889, 0.7054, 0.9948, 0.2299, 0.8069, 0.9641, 0.9903, 0.9839, volume fraction, temperature, shear rate) and output variables (viscos-
0.7996, 0.9877, 0.8069, 0.9961, 0.9913, 0.9708, 0.931 and 0.9946. For ity). MPR directly captures nonlinear relationships (such as γ2 and T3)
ANFIS, BFGS, BN, GD, LM, BPNN, DT, ECR, PLR, ELM, GMDH, GPR, and interactions between independent variables (e.g., T⋅γ) and can
Lasso, MLP, MLR, MPR, RBF, Ridge, SVM and XGB algorithms, the value closely define the properties of the hybrid nanofluid’s complex viscosity.
of RMSD is equal to 3.4975, 2.2826, 2.8378, 2.2826, 2.8085, 13.66, Neural Networks like MLP and BPNN approximate non-linear mappings
1.9187, 18.955, 11.1533, 5.02, 2.6335, 3.4022, 12.4295, 3.0034, from input to output via a hidden layer or to avoid overfitting they need
11.1533, 1.6593, 2.4972, 9.2743, 8.7171 and 3.9572. For ANFIS, BFGS, to be tuned carefully. Decision tree-based models (ex, DT, and XGBoost)
BN, GD, LM, BPNN, DT, ECR, PLR, ELM, GMDH, GPR, Lasso, MLP, MLR, construct local trends by recursively partitioning the data into partial
MPR, RBF, Ridge, SVM and XGB algorithms, the value of STD is equal to subsets and then modeling the nonlinear relationship. e.g. Gaussian
18.6925, 18.7, 15.3369, 18.7, 18.71, 16.054, 18.783, 8.995, 15.234, regression (GPR) and support vector machine (SVM): These methods
18.4112, 18.4683, 18.1558, 10.01, 19.189, 15.234, 18.807, 18.997, take advantage of transformations by translating the data into higher-
10.236, 12.241 and 15.337. Therefore, the best algorithm in terms of the dimensional spaces, which lead to nonlinear relationships highlighted
correlation diagram is the MPR algorithm and the worst is the ELM by linear dependencies, thus providing very good approximations of the
algorithm. viscosity as a function of input variables. This also promotes the use of
One of the most versatile implicit functions providing a plot of adaptive systems like ANFIS, which merges fuzzy and neural networks to
multiple values of statistical metrics (correlation coefficient (R), stan- maximize flexibility in modeling the nonlinear nature and uncertainties
dard deviation (STD), root mean square error (RMSD), etc.) is the Taylor in nanofluid behavior. These methods enable nonlinear trends in vis-
plot. Furthermore, Taylor plots are useful to assess and compare the cosity prediction to be modeled with high accuracy. Of these, the
predictive performances of machine learning models in predicting multivariate polynomial regression (MPR) method has been identified as
complex physical properties, e.g. viscosity. Overall, the significance of the most efficient model as regressions express polynomial interactions
this plot is that it gives a condensed general overview of how well the explicitly.
model predicts, how much it varies around that prediction, and what After it is determined that the MPR algorithm is superior to other
systematic bias may exist. In this plot, the correlation coefficient (R) is algorithms, it is possible to extract the mathematical relationships be-
the radial distance to each point from the origin. And a reading of val- tween its inputs and outputs. Equation (42) shows this mathematical
ue in a range of 1 reflects a strong relationship between the predicted relationship.
values and actual values. As shown in Fig. 6, the MPR algorithm (R =
0.9961) is more effective than ECR (R = 0.2299) and other algorithms.
μnf = 542.28 + 2.370462*γ+0.059*γ2-1.9*e− 3*γ3+2.44*e− 5*γ4-
An STD represents how much variability there is between the predicted 72.24*T-0.37*T*γ+2.2*e− 4*T*γ2+5.3*e− 6T*
γ3+7.12*e− 8*T*γ4+3.94*T2+0.02*T2*γ-1.32*e− 5*T2*γ2-
and the actual values. The nearer points to the reference point (empirical
values) are a more significant fit of the model to the data variations. The 2.55*e− 7*T2*γ3-0.11*T3-3.1*e− 4*T3*γ+4.05*e− 7*T3*γ2+
1.47*e− 3*T4+2.02*e− 6*T4*γ-188.01*φ-3.44*φ*γ+0.1*φ*γ2-
value of the standard deviation of MPR (18.807) in the Taylor plot is
nearly identical to the true value, while ELM and ECR are far from the 1.3*e− 3*φ*γ3+6.58*e− 6*φ*γ4+26.37*φ*T-
0.19*φ*T*γ+5.7*e− 4*φ*T*γ2+3.51*e− 6*φ*T*γ3-
reference value, indicating little ability to reproduce the variation of the
data. The root mean square error (RMSD) describes the mean deviation 0.88*φ*T2+5.52*e− 3*φ*T2*γ-1.3*e− 5*φ*T2*γ2+1.11*e− 2*φ*T3-
4.74*e− 5*φ*T3*γ-4.24*e− 5*φ*T4-205.36*φ2+15.98*φ2*γ-0.11*φ2*γ2-
of the predictions, where lower values indicate more accuracy of
the model. The error is lowest in MPR with RMSD = 1.6593, whereas 3.88*e− 4*φ2*γ3-6.86*φ2*T-0.06*φ2*T*γ-
7.04*e− 5*φ2*T*γ2+0.28*φ2*T2+6.45*e− 4*φ2*T2*γ-
ELM (5.02) and ECR (18.955) present a much larger error in this
comparison. From a geometric relationship perspective, in a Taylor plot, 2.82*e− 3*φ2*T3+367.62*φ3-17.57*φ3*γ+0.11*φ3*γ2-
3.73*φ3*T+0.01*φ3*T*γ+0.02*φ3*T2-
the closer the models’ points are to the reference point, the better their
performance, given that proximity signifies high correlation, low root 95.55*φ4+4.58*φ4*γ+0.85*φ4*T-5.43*φ5-8.05*e− 6*T5-
1.23*e− 7*γ5 (42)
mean square deviation (RMSD), and STD around a true value. The
distance of the model point from the reference point shows its better It is also possible to draw a graph of the predicted points next to the
performance, and the position of the MPR close to the reference experimental points, which is according to Fig. 7. According to this di-
point justifies its ability of viscosity prediction. Models with high STD agram, it is clear that the MPR algorithm has a very high performance
or RMSD are inaccurate or over/under-predicting viscosity. For and the predicted points are very close to the experimental points and
instance, a significant distance of the ECR from the reference point have a small error. In Fig. 7, we can see how the experimental data
suggests inherent issues with the accuracy and variability of its pertur- points and the anticipated points from the MPR algorithm are visually
bative predictions. Some models might outperform (2) on one criterion, compared. The figure’s value is in the way it graphically shows how well
but fail on another. Specifically, ANFIS reported R = 0.9827 but with a the MPR algorithm predicts the non-Newtonian hybrid nanofluid’s vis-
larger error (RMSD = 3.4975) compared to MPR, which shows the trade- cosity (μnf) from data collected using iron, copper, water, and ethylene
off between the accuracy of the correlation and the amount of error. glycol. The prediction points are more accurate when they are closer to
This Taylor plot sees which models learn complex and nonlinear re- the experimental points, as shown in Fig. 7, which compares the actual
lationships. The powerful performance of the MPR demonstrates that viscosity values with the predicted viscosity values generated by the
the proposed model is a good way to model the complex interactions of MPR algorithm. The MPR algorithm accurately predicts the nanofluid
input factors, such as volume fraction of nanoparticles, temperature, and viscosity, as seen in Fig. 7, where the predicted points are very
shear rate and viscosity. Models like ELM and ECR that have low R, high congruent with the observed sites. The ability to visually compare ex-
RMSD, and disproportionate STD are bad choices and easily spotted thus pected and experimental results gives a clear picture of how well the
emphasizing a key issue about model choice. model is doing, which is a crucial confirmation of the model’s correct-
According to the analysis of EC and AP, the best algorithm is the MPR ness. Fig. 7’s tightly packed dots demonstrate that the MPR algorithm
algorithm. The next goal is to extract the mathematical equations be- achieves a low prediction error, in line with the quantitative error
tween the inputs and the output and plot the experimental data along metrics (RMSE and MAE) mentioned in Table 1.
with the data predicted by the MPR algorithm.

34
M.S. Graish et al. Case Studies in Chemical and Environmental Engineering 11 (2025) 101180

From Fig. 7, we can observe a good agreement between the experi- that could affect viscosity predictions under specific conditions. Range
mental viscosity (μnf) values and those predicted by the MPR algorithm. of Conditions Although the experimental data cover a relatively narrow
The diagonal distribution of the points implies an excellent agreement range of SVF, temperature, and shear rate values, the model is not able
between the experimental and predicted values. The low scatter of data to make predictions outside the ranges of the training data. If Multi-
indicates little prediction error, confirming the high performance of variate Polynomial Regression (MPR) is modeled very successfully
MPR. MPR models the system behavior with great accuracy since it within fitted limits, the model might tend to lose its ability to capture
exploits the data structure and its physical foundations making non-linear trends at the tails, where complex interactions have domi-
it outperform its competitors. Although other algorithms can perform nance. At very high temperatures for example, the rheological properties
adequately, some are unable to meet the level of accuracy due to slight of the nanofluid may become erratic due to thermal degradation or
biases or poor variance. Physically, the mathematical equation resul- phase changes. At sufficiently high shear rates, shear thinning may
ting from MPR describes the fundamental laws of nanofluids’ viscosity enter more complex regimes that MPR cannot do justice to. While MPR
with high accuracy. From a computational perspective, MPR avoids does not grapple with polynomial overfitting like polynomial regression,
overfitting by avoiding excessive complexity while allowing just about there can be an overfitting risk if the underlying relationships are
the right amount of freedom to model the specifics of the data set. ECR particularly complex and/or noisy outside the training data range.
(Elastic Component Regression) and PLR (Partial Least Squares Since important factors like pumping power, Reynolds number, and
Regression) algorithms suffer more errors relying on a simplification of heat transfer coefficients are directly affected by viscosity, it is crucial to
the problem. accurately anticipate viscosity to optimize heat transfer systems. To help
engineers create coolers and heat exchangers that work better, the MPR
4. Conclusion model provides a trustworthy and effective technique for predicting
viscosity. The simplified prediction method made possible by the
This study systematically developed twenty machine-learning algo- developed mathematical equation also makes it applicable to real-world
rithms to predict the viscosity of non-Newtonian iron-CuO/water- scenarios. There are a few caveats to be aware of, even if this work does
ethylene glycol hybrid nanofluids. The impact of the solid volume provide some great improvements in viscosity prediction. The research
fraction (SVF), temperature, and shear rate on viscosity was studied, and only looked at one kind of hybrid nanofluid—iron-CuO/water-ethylene
it was found that SVF is the most dominant parameter affecting vis- glycol—so the findings may not apply to other types of nanofluids. The
cosity, followed by shear rate and temperature. Of the tested algorithms, SVF, temperature, and shear rate were the only variables that could be
Multivariate Polynomial Regression (MPR) yielded the highest corre- entered. To improve the accuracy of predictions, future research may
lation coefficient (R2 = 0.992) and lowest error metrics (RMSE = 1.66, take into account other characteristics including the size, shape, and
MAE = 1.146). In terms of accuracy, however, although the Extreme surface chemistry of nanoparticles. Improving prediction accuracy
Learning Machine (ELM) algorithm fared poorly, this serves to reinforce might be a future research focus by investigating advanced machine
that appropriate algorithms must be selected for tasks such as these. The learning approaches like ensemble methods and deep learning. To
main contribution of this study is the mathematical equation derived confirm the MPR model’s applicability, the research should be expanded
from MPR, to express the relationship between input variables and to include a broader variety of nanofluid compositions and operating
viscosity in a few words. This is used for engineers and researchers to be circumstances. One way to make the MPR model more useful in indus-
able to find the viscosity of a fluid without computational models. The trial heat transfer applications is to include it in control and monitoring
analysis and interpretation of model performance and results required systems that work in real-time. The paper highlights the possibility of
the use of advanced data visualization techniques, including heatmaps, using machine learning methods, namely Multivariate Polynomial
box plots, error histograms, kernel density estimation (KDE) plots and Regression, to forecast the viscosity of hybrid nanofluids that do not
Taylor diagrams. While it was invaluable to see the strengths and adhere to Newton’s laws. This study helps improve heat transfer tech-
weaknesses of each particular algorithm, it was even more valuable to nology by developing a strong mathematical model and determining the
realize the features of the tools available for the complete evaluation of main components that affect viscosity. The results provide valuable in-
predictive models. Accurate prediction of viscosity affects all parameters sights into the behavior of nanofluids and provide practical methods for
such as pumping power, Reynolds number, and heat transfer coefficients improving engineering systems, which in turn leads to better energy
used when optimizing heat transferring systems. Though this study efficiency and overall performance. By utilizing deep learning models
targeted iron-CuO/water-ethylene glycol hybrid nanofluids, the meth- like Convolutional (RNNs) or CNNs, it can capture those relatively
odologies developed here are applicable to be employed in other complex higher-order interactions and temporal dependencies which
nanofluid systems. In future studies, more features like nanoparticle might enhance the prediction performance. These models work best for
size, shape, and surface chemistry may be included, and more sophis- large data sets containing complex patterns. Challenges: Since deep
ticated approaches, such as ensemble methods and deep learning, may learning requires a lot of data and time to train molds, it does If one does
be used to achieve better prediction accuracy. In summary, the current not apply data correctly, it may be said to be unnecessary. Larger models
paper demonstrates the suitability of MPR (machine learning) as a tool with additional layers might not generalize effectively given the small
for predicting the viscosity of non-Newtonian hybrid nanofluids which amount of available data (204 samples), so MPR might train sufficiently
could have significant theoretical and practical engineering system and will outcompete more complex methods (deep learning approaches)
implications. for this amount of data. Better Predictive Power: A model trained on
another but closely related dataset can achieve better performance on
5. Limitation and future direction small datasets (transfer learning). As an example, for a prediction of the
viscosity of mono-nanofluids, the same pre-trained neural network can
The analysis performed in this study only targets iron-CuO/water- be fine-tuned for hybrid nanofluids. This tactic could enable more pre-
ethylene glycol hybrid nanofluids, so it might not be extended to cise techniques while not asking for vast quantities of additional data.
other incipient nanofluids or base fluids. Nanofluids with different Transfer learning would be especially beneficial in cases where one
nanoparticle materials (e.g., graphene oxide, silver) or even completely would need to extend the model to new combinations of nanofluid
different base fluids (e.g., oil-based systems) may have different compositions or new operating regimes. The mathematical formula
behavior. Generally, y is called input parameters in the study, the study obtained from the experimental results from MPR was verified and
takes only three input parameters, solid volume fraction (SVF), tem- confirmed for use in practical real-time control for heat exchangers,
perature and shear rate. The model does not account for factors like cooling systems, and other industrial applications. This enables engi-
nanoparticle size, shape, surface chemistry, and aggregation effects1 neers to change the parameters (flow rate, temperature, etc.)

35
M.S. Graish et al. Case Studies in Chemical and Environmental Engineering 11 (2025) 101180

dynamically depending on predicted viscosity values. It is vital to [12] V.V. Wanatasanappan, et al., Viscosity and rheological behavior of Al2O3-Fe2O3/
water-EG based hybrid nanofluid: a new correlation based on mixture ratio, J. Mol.
validate over a range of operational scenarios to deploy this model in
Liq. 375 (2023) 121365.
wide-scale industrial use. This involves testing at different flow re- [13] M. Sepehrnia, et al., Experimental study on the dynamic viscosity of hydraulic oil
gimes, pressure levels, and contamination effects. Integrating into HLP 68-Fe3O4-TiO2-GO ternary hybrid nanofluid and modeling utilizing machine
Monitoring Systems: You could integrate the model into monitoring learning technique, J. Taiwan Inst. Chem. Eng. 145 (2023) 104841.
[14] M.H. Esfe, D. Toghraie, E.M. Ardeshiri, Experimental study of rheological behavior
systems with feedback loops to continuously optimize system perfor- of MWCNT (50%)-MgO (50%)/SAE40 hybrid nanofluid: dynamic viscosity
mance. For instance, you could set alerts or automatically adjust con- optimization and numerical simulation of turbulent flow, Ann. Nucl. Energy 182
ditions in real time based on viscosity predictions to manage the process (2023) 109575.
[15] A.M. Ajeena, I. Farkas, P. Víg, A comparative experimental investigation of
for optimal operation. Answers could refer to historical data recorded dynamic viscosity of ZrO2/DW and SiC/DW nanofluids: characterization,
during operation, which could retrain that model. rheological behavior, and development of new correlation, Heliyon 9 (10) (2023)
e21113.
[16] M.H. Esfe, et al., Presenting a model based on knowledge management in modeling
CRediT authorship contribution statement and optimizing the dynamic viscosity of MWCNT (30%)-ZnO (70%)/SAE 40 oil
hybrid nano-lubricant with response surface methodology (RSM) for industrial
Mohammed Shorbaz Graish: Funding acquisition, Software, uses, Tribol. Int. 191 (2024) 109135.
[17] A.R. Abbasian, et al., Experimental study of preparing the CoFe2O4 magnetic
Formal analysis, Writing – original draft, Writing – review & editing. Ali nanofluid and measuring thermal-fluid characteristics of the stabilized
B.M. Ali: Funding acquisition, Software, Formal analysis, Writing – magnetocaloric nanofluid, Mater. Sci. Eng., B 306 (2024) 117462.
original draft, Writing – review & editing. Murtadha M. Al-Zahiwat: [18] A. Hafeez, et al., Heat transfer performance in a Hybrid nanofluid (Cu-Al2O3/
kerosene oil) flow over a shrinking cylinder, Case Stud. Therm. Eng. 52 (2023)
Funding acquisition, Software; Formal analysis, Writing – original draft, 103539.
Writing – review & editing. Saja Mohsen Alardhi: Writing – review & [19] M.H. Esfe, et al., Using the RSM to evaluate the rheological behavior of SiO2
editing, Conceptualization, Data curation, Formal analysis, Supervision, (60%)-MWCNT (40%)/SAE40 oil hybrid nanofluid and investigating the effect of
different parameters on the viscosity, Tribol. Int. 184 (2023) 108479.
Investigation, Writing – original draft, Writing – review and editing.
[20] G.A. Bekey, K.Y. Goldberg, Neural Networks in Robotics, vol. 202, Springer Science
Mohammadreza Baghoolizadeh: Writing – review & editing, & Business Media, 2012.
Conceptualization, Data curation, Formal analysis, Supervision, Inves- [21] M. Zeinaddini, M. Pirmoradian, F. Azimifar, Optimizing the torque of knee
movements of a rehabilitation robot, J. Simul. Anal. Novel Technol. Mech. Eng. 11
tigation, Writing – original draft, Writing – review & editing. Soheil
(1) (2018) 5–14.
Salahshour: Funding acquisition, Software, Formal analysis, Writing – [22] A. Alfaleh, et al., Predicting thermal conductivity and dynamic viscosity of
original draft, Writing – review & editing. Mostafa Pirmoradian: nanofluid by employment of Support Vector Machines: a review, Energy Rep. 10
Writing – review & editing, Conceptualization, Data curation, Formal (2023) 1259–1267.
[23] M.H. Esfe, F. Amoozadkhalili, D. Toghraie, Determining the optimal structure for
analysis, Supervision, Investigation, Writing – original draft, Writing – accurate estimation of the dynamic viscosity of oil-based hybrid nanofluid
review & editing. containing MgO and MWCNTs nanoparticles using multilayer perceptron neural
networks with Levenberg-Marquardt Algorithm, Powder Technol. 415 (2023)
118085.
Declaration of Competing Interest [24] Z. Wang, et al., Using different Heuristic strategies and an adaptive Neuro-Fuzzy
inference system for multi-objective optimization of Hybrid Nanofluid to provide
an efficient thermal behavior, Swarm Evol. Comput. 86 (2024) 101536.
The authors declare that they have no known competing financial [25] T. Zhang, et al., Optimization of thermophysical properties of nanofluids using a
interests or personal relationships that could have appeared to influence hybrid procedure based on machine learning, multi-objective optimization, and
the work reported in this paper. multi-criteria decision-making, Chem. Eng. J. 485 (2024) 150059.
[26] J. Gao, et al., An RBF-based artificial neural network for prediction of dynamic
viscosity of MgO/SAE 5W30 oil hybrid nano-lubricant to obtain the best
Data availability performance of energy systems, Mater. Today Commun. 38 (2024) 107836.
[27] B.M. Fadhl, et al., Dynamic viscosity modeling of nanofluids with MgO
nanoparticles by utilizing intelligent methods, Energy Rep. 9 (2023) 5397–5403.
No data was used for the research described in the article. [28] Y. AbuShanab, et al., Accurate prediction of dynamic viscosity of polyalpha-olefin
boron nitride nanofluids using machine learning, Heliyon 9 (6) (2023) e16716.
References [29] S. Hua, et al., Dynamic viscosity prediction using artificial intelligence for an
antifreeze containing MWCNT–alumina hybrid nanopowders, Eng. Appl. Artif.
Intell. 126 (2023) 107046.
[1] S.U. Choi, J.A. Eastman, Enhancing Thermal Conductivity of Fluids with
[30] M.H. Esfe, et al., A novel integrated model to improve the dynamic viscosity of
Nanoparticles, Argonne National Lab.(ANL), Argonne, IL (United States), 1995.
MWCNT-Al2O3 (40: 60)/Oil 5W50 hybrid nano-lubricant using artificial neural
[2] S.Z. Heris, H. Bagheri, S.B. Mousavi, S. Hosseini Nami, Optimizing nanofluid
networks (ANNs), Tribol. Int. 178 (2023) 108086.
additives for enhanced thermophysical properties in anionic crude oil for EOR
[31] X. Dai, et al., Using Gaussian Process Regression (GPR) models with the Matérn
applications, Can. J. Chem. Eng. 102 (7) (2024) 2418–2431.
covariance function to predict the dynamic viscosity and torque of SiO2/Ethylene
[3] Z. Karimi Shoar, H. Pourpasha, S. Zeinali Heris, S.B. Mousavi,
glycol nanofluid: a machine learning approach, Eng. Appl. Artif. Intell. 122 (2023)
M. Mohammadpourfard, The effect of heat transfer characteristics of
106107.
macromolecule fouling on heat exchanger surface: a dynamic simulation study,
[32] S. Singh, S.K. Ghosh, A unique artificial intelligence approach and mathematical
Can. J. Chem. Eng. 101 (10) (2023) 5802–5817.
model to accurately evaluate viscosity and density of several nanofluids from
[4] H. Pourpasha, S.Z. Heris, S.B. Mousavi, Thermal performance of novel ZnFe2O4
experimental data, Colloids Surf. A Physicochem. Eng. Asp. 640 (2022) 128389.
and TiO2-doped MWCNT nanocomposites in transformer oil, J. Mol. Liq. 394
[33] Z. Said, et al., Synthesis, stability, density, viscosity of ethylene glycol-based
(2024) 123727.
ternary hybrid nanofluids: experimental investigations and model-prediction using
[5] K. Bashirnezhad, et al., Viscosity of nanofluids: a review of recent experimental
modern machine learning techniques, Powder Technol. 400 (2022) 117190.
studies, Int. Commun. Heat Mass Tran. 73 (2016) 114–123.
[34] P.K. Kanti, et al., Explainable machine learning techniques for hybrid nanofluids
[6] H.D. Koca, et al., Effect of particle size on the viscosity of nanofluids: a review,
transport characteristics: an evaluation of shapley additive and local interpretable
Renew. Sustain. Energy Rev. 82 (2018) 1664–1674.
model-agnostic explanations, J. Therm. Anal. Calorim. 149 (21) (2024)
[7] S.S. Murshed, P. Estellé, A state of the art review on viscosity of nanofluids, Renew.
11599–11618.
Sustain. Energy Rev. 76 (2017) 1134–1152.
[35] P.K. Kanti, et al., Experimental and explainable machine learning approach on
[8] K. Selvarajoo, V.V. Wanatasanappan, N.Y. Luon, Experimental measurement of
thermal conductivity and viscosity of water based graphene oxide based mono and
thermal conductivity and viscosity of Al2O3-GO (80: 20) hybrid and mono
hybrid nanofluids, Sci. Rep. 14 (1) (2024) 30967.
nanofluids: a new correlation, Diam. Relat. Mater. 144 (2024) 111018.
[36] S. Akilu, et al., Machine learning analysis of thermophysical and thermohydraulic
[9] O. Khouri, H.R. Goshayeshi, S.B. Mousavi, S. Hosseini Nami, S. Zeinali Heris, Heat
properties in ethylene glycol-and glycerol-based SiO2 nanofluids, Sci. Rep. 14 (1)
transfer enhancement in industrial heat exchangers using graphene oxide
(2024) 14829.
nanofluids, ACS Omega 9 (22) (2024) 24025–24038.
[37] K. Sharma, et al., Prognostic modeling of polydisperse SiO2/Aqueous glycerol
[10] A.M. Ajeena, I. Farkas, P. Víg, Characterization, rheological behaviour, and
nanofluids’ thermophysical profile using an explainable artificial intelligence (XAI)
dynamic viscosity of ZrO2-SiC (50–50)/DW hybrid nanofluid under different
approach, Eng. Appl. Artif. Intell. 126 (2023) 106967.
temperatures and solid volume fractions: an experimental study and proposing a
[38] R. Rostamzadeh-Renani, et al., Multi-objective optimization of rheological
new correlation, Powder Technol. 431 (2024) 119069.
behavior of nanofluids containing CuO nanoparticles by NSGA II, MOPSO, and
[11] M. Sepehrnia, et al., Experimental study, prediction modeling, sensitivity analysis,
MOGWO evolutionary algorithms and Group Method of Data Handling Artificial
and optimization of rheological behavior and dynamic viscosity of 5W30 engine oil
neural networks, Mater. Today Commun. 38 (2024) 107709.
based SiO2/MWCNT hybrid nanofluid, Ain Shams Eng. J. 15 (1) (2024) 102257.

36
M.S. Graish et al. Case Studies in Chemical and Environmental Engineering 11 (2025) 101180

[39] R. Rostamzadeh-Renani, et al., Prediction of the thermal behavior of multi-walled [54] M. Gumus, M.S. Kiran, Crude oil price forecasting using XGBoost, in: 2017
carbon nanotubes-CuO-CeO2 (20-40-40)/water hybrid nanofluid using different International Conference on Computer Science and Engineering (UBMK), IEEE,
types of regressors and evolutionary algorithms for designing the best artificial 2017.
neural network modeling, Alex. Eng. J. 84 (2023) 184–203. [55] J. Chen, et al., Improved XGBoost model based on genetic algorithm, Int. J.
[40] M. Baghoolizadeh, et al., Prediction and extensive analysis of MWCNT-MgO/oil Comput. Appl. Technol. 62 (3) (2020) 240–245.
SAE 50 hybrid nano-lubricant rheology utilizing machine learning and genetic [56] A. El-Saleh, et al., Mean opinion score estimation for mobile broadband networks
algorithms to find ideal attributes, Tribol. Int. (2024) 109582. using bayesian networks, CMC-COMPUTERS MATERIALS & CONTINUA 72 (3)
[41] M. Baghoolizadeh, et al., Using different machine learning algorithms to predict (2022).
the rheological behavior of oil SAE40-based nano-lubricant in the presence of [57] B.G. Marcot, T.D. Penman, Advances in Bayesian network modelling: integration of
MWCNT and MgO nanoparticles, Tribol. Int. 187 (2023) 108759. modelling technologies, Environ. Model. Software 111 (2019) 386–393.
[42] M. Baghoolizadeh, et al., Using of artificial neural networks and different [58] B. Cai, et al., Application of Bayesian networks in reliability evaluation, IEEE
evolutionary algorithms to predict the viscosity and thermal conductivity of silica- Trans. Ind. Inf. 15 (4) (2018) 2146–2157.
alumina-MWCN/water nanofluid, Heliyon 10 (4) (2024) e26279. [59] J.-S.R. Jang, Fuzzy modeling using generalized neural networks and kalman filter
[43] W. Jin, et al., Regression modeling and multi-objective optimization of rheological algorithm, in: AAAI, 1991.
behavior of non-Newtonian hybrid antifreeze: using different neural networks and [60] J.-S. Jang, ANFIS: adaptive-network-based fuzzy inference system, IEEE
evolutionary algorithms, Int. Commun. Heat Mass Tran. 155 (2024) 107578. transactions on systems, man, and cybernetics 23 (3) (1993) 665–685.
[44] M. Sabah, et al., Application of decision tree, artificial neural networks, and [61] A. Abraham, Adaptation of fuzzy inference system using neural learning, in: Fuzzy
adaptive neuro-fuzzy inference system on predicting lost circulation: a case study Systems Engineering, Springer, 2005, pp. 53–83.
from Marun oil field, J. Petrol. Sci. Eng. 177 (2019) 236–249. [62] D. Karaboga, E. Kaya, Adaptive network based fuzzy inference system (ANFIS)
[45] Y. Chen, et al., Evaluation efficiency of hybrid deep learning algorithms with training approaches: a comprehensive survey, Artif. Intell. Rev. 52 (4) (2019)
neural network decision tree and boosting methods for predicting groundwater 2263–2293.
potential, Geocarto Int. 37 (19) (2022) 5564–5584. [63] C. Nwankpa, C. Nwankpa, et al., Activation functions: comparison of trends in
[46] J.S. Raj, J.V. Ananthi, Recurrent neural networks and nonlinear prediction in practice and research for deep learning, arXiv preprint (2018) arXiv:1811.03378,
support vector machines, J. Soft Comput. Paradigm (JSCP) 1 (1) (2019) 33–40. https://doi.org/10.48550/arXiv.1811.03378.
[47] M. Sharifzadeh, A. Sikinioti-Lock, N. Shah, Machine-learning methods for [64] H. Zhou, et al., Combination of group method of data handling neural network
integrated renewable power generation: a comparative study of artificial neural with multi-objective gray wolf optimizer to predict the viscosity of MWCNT-TiO2-
networks, support vector regression, and Gaussian Process Regression, Renew. oil SAE50 nanofluid, Case Stud. Therm. Eng. 64 (2024) 105541.
Sustain. Energy Rev. 108 (2019) 513–538. [65] S.A. Hussein, et al., Applying different machine learning algorithms to predict the
[48] I. Goodfellow, Deep Learning, MIT press, 2016. viscosity behavior of MWCNT–alumina/water–ethylene glycol (80: 20) hybrid
[49] C.M. Bishop, N.M. Nasrabadi, Pattern Recognition and Machine Learning, vol. 4, antifreeze, Int. J. Thermofluids 24 (2024) 100966.
Springer, 2006. [66] K.E. Taylor, Summarizing multiple aspects of model performance in a single
[50] S. Haykin, Neural Networks: a Comprehensive Foundation, Prentice Hall PTR, diagram, J. Geophys. Res. Atmos. 106 (D7) (2001) 7183–7192.
1998. [67] M.H. Esfe, S. Saedodin, An experimental investigation and new correlation of
[51] D.E. Rumelhart, G.E. Hinton, R.J. Williams, Learning representations by back- viscosity of ZnO–EG nanofluid at various temperatures and different solid volume
propagating errors, Nature 323 (6088) (1986) 533–536. fractions, Exp. Therm. Fluid Sci. 55 (2014) 1–5.
[52] R.R. Kouser, T. Manikandan, V.V. Kumar, Heart disease prediction system using [68] M. Bahrami, M. Akbari, A. Karimipour, M. Afrand, An experimental study on
artificial neural network, radial basis function and case based reasoning, rheological behavior of hybrid nanofluids made of iron and copper oxide in a
J. Comput. Theor. Nanosci. 15 (9–10) (2018) 2810–2817. binary mixture of water and ethylene glycol: non-Newtonian behavior, Exp.
[53] H. Azimi, H. Bonakdari, I. Ebtehaj, Design of radial basis function-based support Therm. Fluid Sci. 79 (2016) 231–237.
vector regression in predicting the discharge coefficient of a side weir in a
trapezoidal channel, Appl. Water Sci. 9 (4) (2019) 1–12.

37