applied

sciences
Systematic Review
Response Surface Methodology Using Observational Data: A
Systematic Literature Review
Mochammad Arbi Hadiyat 1,2, * , Bertha Maya Sopha 1 and Budhi Sholeh Wibowo 1

1 Industrial Engineering Program, Department of Mechanical and Industrial Engineering, Universitas Gadjah
Mada, Yogyakarta 55281, Indonesia
2 Industrial Engineering Program, Faculty of Engineering, Universitas Surabaya (Ubaya),
Surabaya 60293, Indonesia
* Correspondence: [email protected] or [email protected]

Abstract: In the response surface methodology (RSM), the designed experiment helps create in-
terfactor orthogonality and interpretable response models for the purpose of process and design
optimization. However, along with the development of data-recording technology, observational
data have emerged as an alternative to experimental data, and they contain potential information on
design/process parameters (as factors) and product characteristics that are useful for RSM analysis.
Recent studies in various fields have proposed modifications to the standard RSM procedures to
adopt observational data and attain considerable results despite some limitations. This paper aims to
explore various methods to incorporate observational data in the RSM through a systematic literature
review. More than 400 papers were retrieved from the Scopus database, and 83 were selected and
carefully reviewed. To adopt observational data, modifications to the procedures of RSM analysis
include the design of the experiment (DoE), response modeling, and design/process optimization.
The proposed approaches were then mapped to capture the sequence of the modified RSM analysis.
The findings highlight the novelty of observational-data-based RSM (RSM-OD) for generating re-
producible results involving the discussion of the treatments for observational data as an alternative
Citation: Hadiyat, M.A.; Sopha, B.M.; to the DoE, the refinement of the RSM model to fit the data, and the adaptation of the optimization
Wibowo, B.S. Response Surface technique. Future potential research, such as the improvement of factor orthogonality and RSM
Methodology Using Observational model modifications, is also discussed.
Data: A Systematic Literature Review.
Appl. Sci. 2022, 12, 10663. https:// Keywords: classic RSM; observational data; RSM-OD; RSM stages; systematic literature review
doi.org/10.3390/app122010663

Academic Editors: Ricardo J. Alves

de Sousa and Mehdi Safari
1. Introduction
Received: 31 August 2022
Accepted: 18 October 2022
Since first introduced by Box and Wilson in 1951 [1], response surface methodology
Published: 21 October 2022
(RSM) has been widely used by scientists and engineers to find optimal parameter settings
to improve a process and equipment designs. The RSM adopts the design of experiment
Publisher’s Note: MDPI stays neutral
(DoE) concept to collect data and identify significant factors and interactions that influence
with regard to jurisdictional claims in
the process response. Next, RSM is used to develop a mathematical model to capture
published maps and institutional affil-
the causal relationships between factors and responses. Thus, the final result of RSM
iations.
is obtaining optimal factor settings by optimizing the causality model as the objective
function. As one of the common techniques for process optimization, this method works
in situations where engineers have complete control over the factor levels and treatments,
Copyright: © 2022 by the authors.
such as in laboratory experiments, scientific method applications, computer experiments,
Licensee MDPI, Basel, Switzerland. and any other research environments that involve controllable factors. For certain industrial
This article is an open access article processes or design optimization, RSM provides a means for engineers to find the best
distributed under the terms and parameter settings to optimize process/product characteristics. As long as engineers have
conditions of the Creative Commons the chance to set the process/equipment parameter, then the RSM can ideally work based
Attribution (CC BY) license (https:// on experimentation activities.
creativecommons.org/licenses/by/ Nevertheless, conducting designed experiments for continuous process/production is
4.0/). challenging. Changing the parameters during the running process can disturb production,

Appl. Sci. 2022, 12, 10663. https://doi.org/10.3390/app122010663 https://www.mdpi.com/journal/applsci

Appl. Sci. 2022, 12, 10663 2 of 23

and increase the number of nonconforming items, hence, resulting in high costs [2]. When
direct experimentation is not feasible, one of the alternative solutions is to use observational
data as the input to the RSM [3,4]. Some high-tech industries are often complemented with
intelligent data-acquiring systems that allow them to record real-time process/equipment
parameter changes. For prediction purposes, these real-time recorded data become the
input for a mathematical model to generate outputs, such as a forecasting system for
maintenance schedules or product quality [5]. Several pieces of research on chemical
engineering and food production [6,7] have demonstrated that observational-data-based
RSM (RSM-OD) provided a fitted mathematical model to find an optimal factor setting.
Other research used the observed data from a running process or equipment as the input
for RSM-OD, as shown in the work of [8] for steel production and [9] with pollutant
removal processes.
However, observational data and their similarities, including real-time recording data
and already conducted experiment data, limits a researcher’s control over their factor levels,
as the DoE ideally affords. There are presumptions that observational data contain a high
volume, high variability, unstructured, and serial-correlated situations [10]. Therefore,
some modifications to selecting the observations are required prior to the use of the data
in RSM analysis, including the adaptive RSM model and optimization techniques, while
still considering the ideal concept of the RSM. The authors of [3] have successfully adopted
observational data for the DoE by selecting a subset of observations and identifying stages
within the data, similarly, Refs. [2,11,12] also giving alternatives by matching the data with
certain DoE to ensure orthogonality. Moreover, the authors of [4,13] apply the RSM to
real-time data acquisition for optimization during continuous processes. It is also worth
noting that the recent development of big data has accelerated the use of observational
data. For instance, Ref. [14] demonstrated real-manufacturing-oriented big data, in which
recorded datasets provide information for process improvement and optimization. The
data-recording technology provides massive datasets in which huge datasets are recorded
along with operations [15–17]. Once the acquired dataset contains the process parameters
and product characteristics, then the RSM-OD should be considered as an optimization
methodology. However, existing pieces of literature on RSM-OD have a unique approach
to treating the observational data and modifying the RSM model or procedures; thus, the
opportunities to develop an established RSM-OD are still open.
Therefore, the paper aims to explore various approaches to develop RSM-OD through
a systematic literature review. The review was based on 82 pieces of literature which were
selected and analyzed using the PRISMA framework [18]. The paper focuses on how obser-
vational data can be considered as input for RSM for process/design optimization purposes.
According to the authors’ best knowledge, the present paper is the first comprehensive
review of the successful implementation of RSM-OD in various research fields. Other
review studies on RSM systematic literature review papers have discussed classic RSM and
DoE in advanced manufacturing optimization [19] and neural network, replacing the DoE
model [20]. Hence, the paper contributes by providing insights into the development of
new procedures in RSM-OD following three stages of analysis in standard classic RSM, i.e.,
the treatment of nondesigned experimental data, the modeling of the relationship between
factors and response, and optimization.
The rest of the paper is structured as follows. Section 2 briefly explains how the classic
and ideal RSM model works based on experimental data and the opportunity to adopt
observational data. Section 3 describes the systematic literature review (SLR) methodology.
Section 4 presents the results of descriptive and bibliometric analysis, which is followed by
synthesis and discussion in Section 5. Lastly, Section 6 concludes by highlighting the main
findings, limitations, and future research.

2. RSM Overview for Response Optimization

This section contains a theoretical perspective of the classic RSM and its applications
in various research fields. Considerable research on the classic RSM showed that this
Appl. Sci. 2022, 12, x FOR PEER REVIEW 3 of 24

Appl. Sci. 2022, 12, 10663

2. RSM Overview for Response Optimization 3 of 23
This section contains a theoretical perspective of the classic RSM and its applications
in various research fields. Considerable research on the classic RSM showed that this
method
method has
has recently
recently provided
provided significant contributions. A
significant contributions. A designed
designed experiment-based
experiment-based
RSM with fulfilled statistical assumptions will give strong theoretically-based
RSM with fulfilled statistical assumptions will give strong theoretically-based analysis analysis
and
and
interpretation. Nevertheless, the consideration of using observational data in RSMinshould
interpretation. Nevertheless, the consideration of using observational data RSM
should not be ignored
not be ignored because because some pieces
some pieces of literature
of literature [4,13]demonstrated
[4,13] have have demonstrated the suc-
the successful
cessful
implementation of observational data in the RSM. The section also presents some ofsome
implementation of observational data in the RSM. The section also presents these
of theseas
papers papers as motivating
motivating examples examples of the rationale
of the rationale for this
for writing writing this paper.
paper.

2.1.
2.1. Classic
ClassicRSM
RSM
As
As mentioned
mentioned above,above,classical
classicalRSMRSM works
worksby by integrating
integratingthree
threetools
tools in
in aa sequential
sequential
analysis
analysis (Figure
(Figure 1).1). In
Inthe
the first
first stage,
stage, classic
classic RSM
RSM implements
implementsthe the DoE.
DoE. In In this
this step,
step, the
the
DoE
DoEplays
plays aa role
role in
in experiment
experiment planning,
planning,datadatacollection,
collection, analysis,
analysis, and
and interpretation
interpretation and and
ensures that the experiment
ensures experiment fulfills
fulfillsits
itspurpose.
purpose.Orthogonality
Orthogonalityfulfillment
fulfillmentininthe DoE
the DoEmatrix
ma-
ensures that the predetermined process parameters can be estimated
trix ensures that the predetermined process parameters can be estimated independently independently among
others. others.
among Second,Second,
classic RSM
classicapplies a specific
RSM applies mathematical
a specific model tomodel
mathematical fit thetodata
fit obtained
the data
by the DoE.
obtained ThisDoE.
by the model captures
This the relationship
model captures between factors
the relationship between orfactors
parameters as inputs
or parameters
and
as responses
inputs as outputs.
and responses Classic RSM
as outputs. usually
Classic RSM prefers
usuallyto adoptto
prefers a adopt
linear model
a linearbecause
model
of its simple
because of itsinterpretation and formal
simple interpretation andstatistical inferenceinference
formal statistical of all its required assumptions
of all its required as-
during the modeling stage. Third, the optimization stage works
sumptions during the modeling stage. Third, the optimization stage works by finding the by finding the factor
(or parameter)
factor setting
(or parameter) to optimize
setting the response.
to optimize the response. Standard
Standardoptimization
optimization tools, such
tools, as
such
mathematical optimization and desirability functions [21], are preferred
as mathematical optimization and desirability functions [21], are preferred in classic RSM, in classic RSM,
along with
along withsome
sometheoretical
theoretical approaches.
approaches. As Asthe
therequired
requiredassumptions
assumptionsin inRSM
RSMare arefulfilled
fulfilled
foreach
for eachstage,
stage,this
thismethodology
methodologyhas hasbecome
becomethe thebest
bestchoice
choicerather
ratherthan
thananyanymodification.
modification.

RSM sequential analysis

1 2
Guide data collection and Express the relationships
the identification of between dependent and
factors and interactions independent variables
influencing the process mathematically

Tools: DoE Tools: Modeling

RSM as integration
of tools

Design of Experiment Modeling

(DoE)

RSM

3
Optimization method
Perform the maximization or
minimization of functions for
improving the process

Tools: optimization method

Figure
Figure 1.
1.Overview
Overviewof
ofRSM
RSM(adopted
(adoptedfrom
from [19]).
[19]).

In addition, an essential prestage in RSM involves determining the factors involved in

the analysis. As the DoE is applied, researchers should subjectively select the factors in RSM.
They need to find the factors with more minor or significant effects on the response based
on previous research, considering the scope and knowledge domain of the researchers.
Appl. Sci. 2022, 12, x FOR PEER REVIEW 4 of 24

In addition, an essential prestage in RSM involves determining the factors involved

Appl. Sci. 2022, 12, 10663 in the analysis. As the DoE is applied, researchers should subjectively select the factors 4 of 23
in
RSM. They need to find the factors with more minor or significant effects on the response
based on previous research, considering the scope and knowledge domain of the research-
ers.an
As Asestablished
an established methodology
methodology forfor designed,
designed, experimentally-basedoptimization,
experimentally-based optimization,thethe
classic RSM has
classic has been
beensuccessfully
successfullyapplied
appliedforfor
years
yearsin many
in many research fields.
research Starting
fields. with
Starting
the concept
with proposed
the concept by [22],
proposed by more than 48,000
[22], more Scopus-indexed
than 48,000 paperspapers
Scopus-indexed appliedapplied
classic
RSM. Figure
classic 2 shows
RSM. Figure that the
2 shows general
that engineering
the general fields fields
engineering dominate the percentage
dominate of RSM
the percentage of
applications,
RSM followed
applications, by chemical
followed engineering,
by chemical chemistry,
engineering, biological
chemistry, sciences,
biological and other
sciences, and
other
appliedapplied sciences.
sciences. It meansIt means
that RSM thatplays
RSManplays an important
important role
role as an as an optimization
optimization method-
methodology
ology in manyinresearch
many research fields,
fields, and andare
there there
alsoare also considerable
considerable developments
developments in RSM in to
RSM to accommodate
accommodate recent research
recent research issues. issues.

Figure2.2. Research
Figure Researchfields
fieldsapplying
applyingRSM.
RSM.

Many improvements
Many improvements to to classic
classic RSM
RSM have
have been
been performed,
performed, mainly
mainly when
when optimiza-
optimiza-
tion of the target
tion target by bystandard
standardRSM RSMprocedures
proceduresprovides
provides dissatisfactory
dissatisfactory results. Some
results. pa-
Some
pers on RSM improved the linear model to increase its performance in capturing the cau-
papers on RSM improved the linear model to increase its performance in capturing the
causality between
sality between factors
factors andand responses
responses by replacing
by replacing it with
it with nonlinear
nonlinear versions.
versions. For ex-
For example,
ample,
[23,24] [23,24]
appliedapplied neural networks
neural networks and support
and support vector regression
vector regression for RSMfor RSM modeling
modeling to opti-
to optimize
mize surfacesurface
roughness,roughness, respectively,
respectively, in the milling
in the milling and turningand process.
turning process. Other
Other research-
researchers [25] provided
ers [25] provided a similaraapproach
similar approach
that usesthat
the uses
RSMthe RSM
neural neural network
network model tomodel
optimize to
optimize iron extraction from food. The complexity of these modified
iron extraction from food. The complexity of these modified RSM models requires ad- RSM models requires
advanced optimization
vanced optimization techniques
techniques andand adopt
adopt a meta-heuristics
a meta-heuristics method;
method; for for example,
example, the the
au-
authors
thors of [26,27] successfully adopted a genetic algorithm for injection-molding and CNC
of [26,27] successfully adopted a genetic algorithm for injection-molding and CNC
process
processoptimization.
optimization.
Classic
Classic RSM
RSM can can be
be improved
improved by by some
some modifications
modifications in in order
order to
to enhance
enhance thethe perfor-
perfor-
mance of the process being investigated. However, all the methodological
mance of the process being investigated. However, all the methodological improvements improvements
of
of classic
classic RSM
RSM should
should consider
consider the thebasic
basic concept
concept of
of RSM,
RSM, itsits stages,
stages, and
and the
the final
finalpurpose
purpose
of the RSM, i.e., process optimization.
of the RSM, i.e., process optimization.
2.2. RSM-OD
The data-driven concept as a part of smart manufacturing has grown and has become
a recent issue in some research, as proposed in other literature reviews [28,29]. Moreover,
the rapid development of data acquisition systems supports the application of data-driven
Appl. Sci. 2022, 12, 10663 5 of 23

analysis. In the manufacturing process, a data acquisition system, especially those with au-
tomatic sensor-based data recording, will produce massive mounts of data that potentially
contain information about the characteristics of the process/equipment.
This system records data on the equipment parameters and product characteristics; as
examples of the data-acquiring process, as explained by [5,15], some smart sensor devices
can collect data from various types of equipment as a part of the data-driven technology.
Therefore, several researchers argued that data analysis should be applied to obtain useful
information. Other research successfully performed analyses based on these collected data
for industrial application purposes, such as product quality prediction [30], preventive
equipment maintenance [16], the process optimization purposes [4], similar to our topic.
For practical purposes within manufacturing or laboratory scale, with the provided dataset
or data acquiring system, the RSM-OD analysis is preferred because it does not need to
interrupt the ongoing production, nor does it require exceptional equipment parameter
adjustments for experimenting. Other papers argued that it reduces experimental costs [2].
Both sets of authors from [3,4] considered observational data as alternatives to de-
signing experiments and applied them for continuous semiconductor and tire production,
respectively. A large number of recorded data opened up opportunities to use them as a
part of the process optimization system based on the data-driven concept. Both research
papers showed how the RSM concept incorporates observational or historical data as
the basis for process optimization. Specific iterative procedures, such as the selection of
potential factors, the identification of stages in the dataset, and the search for a subset of
observations with similar characteristics to the designed experiment, were proposed to
treat the dataset to become suitable to adopt RSM.
In addition, some papers with laboratory-scope experimentation implemented RSM
based on observational or historical data with a specific approach called historical data
design (HDD), which is provided by Design Expert® v.11 software from Stat-Ease, Inc.,
1300 Godward St NE, Suite 6400, Minneapolis, MN, USA. Although it is more similar
to ordinary multiple regression analysis fitted to observational data, HDD is a type of
observational data-based DoE within the RSM analysis. For details on performing HDD,
refer to the software manual guide from [31] based on a case study by [32]. Both [6,8]
explicitly mentioned and applied HDD, where previous, un-designed, and experimental or
observational data were used as inputs for RSM analysis to optimize energy consumption
and plastic strength.
Another similar paper gave a different perspective; the authors of [33] worked on an
additive manufacturing process to predict surface roughness, and real-time data-driven
modeling techniques were applied to minimize the prediction error. A real-time approach
requires no assumptions for the data and does not need to evaluate the significance of the
factors; its main target is to obtain the minimum error in the predicted response with no
model interpretation required [34]. Meanwhile, the standard RSM proposed by [1] applies
the philosophy of three stages in its analysis (Figure 1), with several required assumptions
in the data, such as factor independence, treatment randomization, and factor significance,
to give a strong interpretation; the final target of this RSM is to obtain the optimum response
by finding the optimal factor setting/level.
The next section explains that this approach treats the dataset’s variables, features,
and responses as input and output. Some papers provided additional filtering procedures
for selecting available observations to be fitted in the RSM model by treating the dataset
so as to become similar to the designed experimental data [3,35]. Moreover, they applied
machine learning models, such as neural networks and support vector machines (SVM),
to replace ordinary linear models. However, this action will increase the risk of black-box
modeling rather than keep the concept of model interpretability.
A number of RSM modifications to accommodate observational data have been con-
ducted. Some modifications focused on data treatment before being used as input for the
RSM. Other modifications develop adaptive mathematical modeling to any data condition,
 to replace ordinary linear models. However, this action will increase the risk of black-box
modeling rather than keep the concept of model interpretability.
Appl. Sci. 2022, 12, 10663 A number of RSM modifications to accommodate observational data have been6con- of 23
ducted. Some modifications focused on data treatment before being used as input for the
RSM. Other modifications develop adaptive mathematical modeling to any data condi-
tion, including
including the
the use of use of machine
machine learninglearning approaches.
approaches. The mostThe most
recent recent modifications
modifications deal with
deal with the
the ability ability of optimization
of optimization techniquestechniques to solveRSM
to solve complex complex RSM models.
models.

3. Methodology
3. Methodology
The systematic
The literature review
systematic literature review conducted
conducted in in the
the present
present study
study follows
follows systematic
systematic
literature review guidance from [36] and conforms to the PRISMA statement
literature review guidance from [36] and conforms to the PRISMA statement in in [18].
[18]. We
We
started by
started byidentifying
identifyingstudies
studiesand followed
and thisthis
followed with database
with searches,
database filtering
searches, processes,
filtering pro-
and content analyses, as shown in Figure 3.
cesses, and content analyses, as shown in Figure 3.

Figure 3. Systematic
Figure 3. Systematic literature
literature review framework based
review framework based on
on PRISMA
PRISMA [18].
[18].

A systematic literature
A systematic literature review
review gives
gives an objective synthesis
an objective synthesis as as it involves a
it involves a decent
decent
number of references based on selected keywords. It follows the identification of studies,
It follows the identification of studies,
and the
thestages
stagesinvolve
involvepaper
papersearches, filtering,
searches, andand
filtering, synthesis.
synthesis.As shown
As shownin Figure 3, iden-
in Figure 3,
tification stepssteps
identification define the problems
define in RSM,
the problems whichwhich
in RSM, are thenare formulated as research
then formulated ques-
as research
tions. By applying
questions. certain
By applying criteria
certain based
criteria on the
based research
on the questions,
research questions,thethe
collected pieces
collected piecesof
literature were
of literature screened
were screenedandand
analyzed withwith
analyzed respect to descriptive,
respect bibliometric,
to descriptive, and com-
bibliometric, and
parative
comparativeanalysis. TheThe
analysis. Scopus
Scopusdatabase
databasewas wasdeployed
deployed because
becauseititprovided
providedbetter
better article
article
searching in terms of source titles, journal impact metrics, and the number of publishers
when compared to others, as shown by by [37].
[37].
The systematic literature review methodology
methodology was was used to achieve a reproducible
result in the development and application of RSM-OD. The analysis and discussion in this
paper focused
focusedon onthose
thoseapproaches
approachesaccommodating
accommodating nondesigned
nondesigned experimental
experimental datadata
in the
in
classic RSM. Moreover, as the context of this paper discusses the development
the classic RSM. Moreover, as the context of this paper discusses the development of RSM- of RSM-OD,
the literature
OD, research
the literature questions
research (LRQs) emphasize
questions how the standard
(LRQs) emphasize how theRSM is modified
standard RSM to is
accept data.
modified to accept data.
• What are
LRQ1: What arethe
therationales
rationalesforfor using
using observational
observational data
data as alternative
as an an alternative to
to con-
ducting a real
conducting RSM
a real experiment?
RSM experiment?
• LRQ2: What condition within observational data can be adapted to RSM?
• LRQ3: How are observational data adopted to RSM?
The descriptive analysis and synthesis stages in this paper attempted to answer those
LRQs associated with the need for well-designed experimentally-based optimization in
Appl. Sci. 2022, 12, x FOR PEER REVIEW 7 of 24

 LRQ2: What condition within observational data can be adapted to RSM?

Appl. Sci. 2022, 12, 10663  LRQ3: How are observational data adopted to RSM? 7 of 23
The descriptive analysis and synthesis stages in this paper attempted to answer those
LRQs associated with the need for well-designed experimentally-based optimization in
various
various fields
fields of of studies.
studies. TheThe practical
practical limitations
limitations of of conducting
conducting the experiments were
the experiments were
raised and prompted the consideration of adopting
raised and prompted the consideration of adopting observational data as an observational data as an alternative.
alternative. As
As
shownshown in Figure
in Figure 3, the3,stages
the stages
start start
with awith a bibliometric
bibliometric analysis
analysis to maptothe map the interrela-
interrelationship
tionship among keywords
among research research keywords
as a referenceas a for
reference for methodological
methodological mapping and mapping and an-
answering the
swering
LRQs. LRQ1 the LRQs. LRQ1 was
was answered byanswered
identifying bytheidentifying
rationalesthe forrationales for using observa-
using observational data for
tional data for RSM
RSM analysis, analysis,
considering theconsidering
limitations the limitations
of classic RSM of in classic RSM
practice but in practice
still referringbuttostill
its
referring to its standard procedures (Figure 1). As LRQ1 was
standard procedures (Figure 1). As LRQ1 was answered, LRQ2 and LRQ3 can parallelly be answered, LRQ2 and LRQ3
can parallelly
processed. Thebeanswers
processed. The answers
to LRQ2 review strictto LRQ2 review strict
assumptions assumptions
of the RSM and how of the RSM
observa-
and
tional data can still be adopted by RSM. As a result, observational data preprocessingdata
how observational data can still be adopted by RSM. As a result, observational and
preprocessing
evaluations to conformand evaluations
to the RSM to conform
analysis were to theidentified.
RSM analysis were identified.
Meanwhile, LRQ3 dealt Mean-
with
while, LRQ3 dealt
the procedures to with
adoptthe procedures to
observational adopt
data intoobservational data into
the RSM analysis, the RSM
subject analysis,
to the classic
subject
statisticalto assumptions
the classic statistical assumptions
within, including DoEwithin, including
properties, DoEmathematical
modified properties, modified
models,
mathematical
and the optimization models,method.
and the Finally,
optimization method. Finally,
the discussion the discussion
was started based on the was started
results of
based on the results of LRQ1 to LRQ3, which focus on
LRQ1 to LRQ3, which focus on the opportunities and gaps for adopting observational the opportunities and gaps for
adopting
data as an observational
alternative to data
DoEas an alternative
in RSM analysis to DoE
and in RSMthe
open analysis and development
potential open the poten- of
further
tial research. of further research.
development
Search strings
strings forforpaper
paperabstracts
abstractsand andtitles
titlesbyby
restricting
restrictingthethe
search
searchreferences
referencesfrom the
from
Scopus
the Scopusdatabase
databasewerewere predetermined
predetermined to ensure that that
to ensure the papers still covered
the papers the proposed
still covered the pro-
topics topics
posed (Figure(Figure4). Similar termsterms
4). Similar related to nonexperimental
related to nonexperimental datadata
usedusedin the reference
in the refer-
papers, such as observational, historical, or retrospective data, were
ence papers, such as observational, historical, or retrospective data, were found. Further, found. Further, these
terms were combined with common keywords in the RSM
these terms were combined with common keywords in the RSM analysis, such as “opti- analysis, such as “optimization”
and “DoE”.
mization” and Some
“DoE”.Boolean
Someoperators were applied,
Boolean operators were considering that RSM-OD
applied, considering analysis
that RSM-OD
should refer
analysis should to the classic
refer to theRSM stages.
classic RSM stages.

Figure 4.
Figure Search string
4. Search string and
and Boolean
Boolean operators.
operators.

Figure 44 shows
Figure showsthe
thesearch
searchprocess. Based
process. onon
Based thethe
research questions,
research the key
questions, the terms were
key terms
“RSM”, “non-experimental data”, and “optimization”. The search queries involved
were “RSM”, “non-experimental data”, and “optimization”. The search queries involved all of
these, along with the use of the Boolean operator “AND”. To enrich the search process,
all of these, along with the use of the Boolean operator “AND”. To enrich the search pro-
we identified some synonyms within each of the key terms based on the mentioned terms
cess, we identified some synonyms within each of the key terms based on the mentioned
in reference papers. For example, several papers used different terms when mentioning
terms in reference papers. For example, several papers used different terms when men-
the nonexperimental data but gave similar meanings, i.e., observational, historical, or
tioning the nonexperimental data but gave similar meanings, i.e., observational, historical,
retrospective data. One of these similar terms was then selected with the Boolean operator
“OR” to complete the search string.
The search result with the determined search strings and Boolean operators yielded
more than 400 papers from the SCOPUS database. However, not all of these papers
discussed RSM with regard to observational data. Some mentioned similar keywords, but
Appl. Sci. 2022, 12, x FOR PEER REVIEW 8 of 24

or retrospective data. One of these similar terms was then selected with the Boolean oper-
Appl. Sci. 2022, 12, 10663
ator “OR” to complete the search string. 8 of 23
The search result with the determined search strings and Boolean operators yielded
more than 400 papers from the SCOPUS database. However, not all of these papers dis-
cussed RSM with regard to observational data. Some mentioned similar keywords, but
thethe topics were outside of this paper’s scope. The filtering process was then applied with
topics were outside of this paper’s scope. The filtering process was then applied with
the inclusion criteria in Table 1. The selected research in this paper was considered to follow
the inclusion criteria in Table 1. The selected research in this paper was considered to
the RSM concepts, consisting of different stages (Figure 1). The final 82 selected papers
follow the RSM concepts, consisting of different stages (Figure 1). The final 82 selected
led to the synthesis stages, with additional references to the standard RSM, such as those
papers led to the synthesis stages, with additional references to the standard RSM, such
within [1,19].
as those within [1,19].
Table 1. Inclusion criteria for filtering papers.
Table 1. Inclusion criteria for filtering papers.
Paper
Paper Inclusion
Inclusion Criteria
Criteria Paper Paper Exclusion
Exclusion Criteria
Criteria
Theshould
The RSM RSM should not conduct
not conduct a designed
a designed experiment
Application of observational or historical to experiment
obtain data to
(however, obtain data
some (however,
papers still some papers
referred to
Application of observational or historical
datadata
as anasalternative to the still referred to nondesigned
an alternative to DoE in RSM
the DoE in RSMnondesigned experiments/non-DoE with a rationale of
experiments/non-DoE
hard-to-control with aare
factors; the details rationale
in Figure of 8)
hard-to-control factors; the details are in Figure 8)
Involving previous experimental data for
Involving previous experimental The RSM entirely refers to the dataset without com-
data for
RSM, some papers referred to combined The RSM entirely refers to the dataset without
RSM,from
datasets some papers referred
previous to combinedpleting it, with new additional experiments.
experiments completing it, with new additional experiments.
datasets from previous experiments
Involvement of the three stages of standard
RSMInvolvement
analysis (DoE, of the three stages
modeling, of standard
and optimi- One ofOne
the of
stages of standard
the stages RSM analysis
of standard RSM analysisis missing
RSM analysis (DoE, modeling, and
zation) is missing
RSMoptimization)
analysis involves searching for influ- A direct prediction system with real-time data record-
encing factors, similar to the original RSM ing and A modeling
direct prediction
is not asystem
part of with real-time
this SLR data
because no
RSM analysis involves searching for
concept such recording
analysis of and modeling
significant is not a part
influencing of this
factors SLR
exists.
influencing factors, similar to the original
because no such analysis of significant influencing
RSM concept
factors exists.
4. Descriptive and Bibliometric Analysis
The descriptive analysis in this section explained the research trends associated with
4. Descriptive and Bibliometric Analysis
the topics in this paper, and the bibliometric analysis focused on the methods involved in
The descriptive
the RSM-OD and researchanalysis
fieldsin
to this section
which it hasexplained the research
been applied. Since thetrends associated
early 2000s with
(Figure
5), the increased number of indexed publications in the SCOPUS database with searchthe
the topics in this paper, and the bibliometric analysis focused on the methods involved in
RSM-OD
strings and 4)
(Figure research
showsfields to which
that the it has been
application applied.
of the RSM-ODSince theoccurred
has early 2000sin (Figure
various 5),
the increased number of indexed publications in the SCOPUS database
research fields. Table 2 shows some of the research fields where the method has been with search strings
(Figure
applied. 4) shows that the application of the RSM-OD has occurred in various research fields.
Table 2 shows some of the research fields where the method has been applied.

12

10
(SCOPUS indexed)

8
Number of papers

0
1970

2010

2018
2000
2001
2002
2003
2005
2007
2009

2011
2012
2013
2014
2015
2016
2017

2019
2020
2021

Year of publication

Figure 5. Paper trends for RSM-OD.

Figure 5. Paper trends for RSM-OD.
The pharmacy/chemistry/chemical engineering fields commonly deal with laboratory-
scope experiments. They can be improved with the use of standard RSM, but they use
already provided data as the input for RSM. Meanwhile, manufacturing, petroleum, and
similar engineering fields with modern equipment mostly have a data-acquiring system.
 Table 2. Distribution of papers based on research fields.

Field of Application of RSM-OD Percentage

Appl. Sci. 2022, 12, 10663 pharmacy/chemistry/chemical engineering 22.50% 9 of 23
manufacturing process 18.75%
petroleum/coal/mining 11.25%
cleaner production/waste 10.00%
Thus, using the provided dataset rather than experimental data is more reasonable. Fur-
material & mechanical engineering 7.50%
thermore, using a treemap (Figure 6) to categorize the journal quartile shows that the
energy 6.25%
highly impacted journal (Q1/Q2) gives the highest percentage among other quartiles,
food 5.00%
which means that the RSM-OD has supported high-quality research.3.75%
civil engineering For Q1 journals, the
research field
medical science of (pharmacy/chemistry/chemical) engineering (10.00%) and manufacturing
3.75%
processes
aerospace (11.25%) still dominated regarding the application of RSM-OD, followed by other
2.50%
fields.
biologyA similar interpretation is also drawn for Q2 and the others.2.50% Thus, scholars have
opportunities
methodologicaltodevelopment
develop RSM-OD procedures required by various research 2.50% fields involving
designed, experiment-based optimization processes at any level of the
waste processing impacted journals.
2.50%
social science 1.25%
Table 2. Distribution of papers based on research fields.

FieldThe pharmacy/chemistry/chemical
of Application of RSM-OD engineering fields commonly deal with labora-
Percentage
tory-scope experiments. They can be improved with the use of standard RSM, but they
pharmacy/chemistry/chemical engineering 22.50%
use already provided data as the input for RSM.
manufacturing process
Meanwhile, manufacturing, petroleum,
18.75%
and similar engineering fields with modern equipment
petroleum/coal/mining 11.25% mostly have a data-acquiring sys-
tem. Thus,
cleaner using the provided dataset rather 10.00%
production/waste than experimental data is more reasonable.
Furthermore, using a engineering
material & mechanical treemap (Figure 6) to categorize
7.50% the journal quartile shows that the
energy impacted journal (Q1/Q2) gives the highest
highly 6.25% percentage among other quartiles,
food
which means that the RSM-OD has supported5.00% high-quality research. For Q1 journals, the
civil engineering 3.75%
research field of (pharmacy/chemistry/chemical)
medical science
engineering (10.00%) and manufacturing
3.75%
processes
aerospace (11.25%) still dominated regarding2.50% the application of RSM-OD, followed by
other
biologyfields. A similar interpretation is also drawn
2.50% for Q2 and the others. Thus, scholars
methodological
have development
opportunities 2.50% required by various research fields
to develop RSM-OD procedures
waste processing
involving 2.50% processes at any level of the impacted
designed, experiment-based optimization
social science 1.25%
journals.

Figure6.6. Journal
Figure Journal quartile
quartile by
by research
researchfield.
field.

VOSviewer ®
VOSviewer® v.1.6.15
v.1.6.15software provided
software providedby Centre for Science
by Centre and Technology
for Science Studies
and Technology
of Leiden
Studies ofUniversity was used
Leiden University to obtain
was used tothe graphical
obtain networknetwork
the graphical in Figurein 7. In the7.figure,
Figure In the
the author’s
figure, keywords
the author’s represent
keywords variousvarious
represent terms incorporated in the RSM-OD.
terms incorporated The figure
in the RSM-OD. The
also gives
figure alsoinsights into theinto
gives insights development of the integrated
the development RSM tools/methods
of the integrated to handle
RSM tools/methods to
nonexperimental data, particularly for certain RSM-related methodological terms, although
specific research field-related keywords were still included. The bibliometric analysis
consisted of nodes and the links connecting them. Large nodes represent high keyword
occurrences, and the links indicate co-occurrences in the same papers. Table 3 shows
the results of the complete bibliometric analysis, including the total link strength, which
exhibited a high number of co-occurrences between the keywords.
 handle nonexperimental data, particularly for certain RSM-related methodological terms,
although specific research field-related keywords were still included. The bibliometric
analysis consisted of nodes and the links connecting them. Large nodes represent high
keyword occurrences, and the links indicate co-occurrences in the same papers. Table 3
Appl. Sci. 2022, 12, 10663 shows the results of the complete bibliometric analysis, including the total link strength,
10 of 23
which exhibited a high number of co-occurrences between the keywords.

Figure 7.
Figure Graphical network
7. Graphical network of
of bibliometric
bibliometric analysis.
analysis.

Table 3.
Table Occurrences and
3. Occurrences and link strength of
link strength of graphical
graphical keyword
keyword networks
networks in
in Figure
Figure 8.
8.
Author’s
Author’s Selected
Selected Methodological
Methodological Keywords
Keywords Total
Total
Occurrences
Occurrences Links Links Link Strength
(Excluding Specific Research Field
(Excluding Specific Research Field Keywords)Keywords) Link Strength
RSMRSM 33 33 130130 144
144
optimization 11 42 51
optimization 11 42 51
HDD only 7 27 29
HDD only
historical data 6 7 26 27 3229
historical data
neural networks 6 6 23 26 2432
neural
DoE networks 5 6 23 23 2724
DoEgenetic algorithm 3 5 15 23 1527
observational
genetic algorithmdata 3 3 13 15 1315
Analysis of variance
observational data (ANOVA) 2 3 15 13 1613
quality of
Analysis byvariance
design (ANOVA) 2 2 14 15 1416
modeling
quality by design 2 2 9 14 1014
statistical analysis 2 9 10
modeling 2 9 10
Taguchi method 2 9 9
statistical analysis
process optimization 2
2 8
9 9
10
Taguchi method
experimental design 2 2 8 9 89
process optimization
retrospective data 2 2 6 8 109
experimental design
intelligent systems 1 2 7 8 78
retrospective data
machine learning 1 2 7 6 710
response-surface
intelligent systems designs 1 1 7 7 77
six sigma
machine learning 1 1 7 7 77
support vector machine
response-surface designs 1 1 7 7 77
industrial-scale optimization 1 6 6
six sigma 1 7 7
RSM historical data modeling 1 5 5
support vector machine 1 7 7
causality 1 5 5
industrial-scale optimization
data-driven modeling 1 1 5 6 56
RSM historical data
meta-heuristic modeling
optimization 1 1 5 5 55
causality
Note: The red highlighted portion represents common RSM terms, the 1 yellow highlighted
5 5
part denotes high
occurrences, and
data-driven the blue highlighted section denotes low occurrences 1in Figure 7. 5
modeling 5
meta-heuristic optimization 1 5 5
Note: Keywords from research
The red highlighted portionthat appliedcommon
represents the standard RSM the
RSM terms, mainly
yellowconsisted of common
highlighted part de-
termshigh
notes in the analysis stages,
occurrences, and the such as DoE andsection
blue highlighted optimization
denotes low(Table 3, with red,
occurrences highlighted
in Figure 7.
rows). By ignoring specific terms related to research fields, only methodological terms
are shown in Table 3, including those with high (yellow highlighted) and low occurrences
(blue highlighted). The largest cluster with the highest occurrences had “RSM” as the
main keyword, followed by “optimization” and “DoE”; these three keywords represent the
common terms in classic RSM analysis. Therefore, their high occurrences were expected.
The analysis also focused on other clusters supporting them and denotes the development
of RSM-OD (the yellow highlighted portion in Table 3).
Appl.
Appl.Sci. 2022,12,
Sci. 2022, 12,10663
x FOR PEER REVIEW 11
11 of
of 23
24

Figure 8.
Figure 8. Distribution
Distribution of
of the
the rationales
rationales based
based on
on data
datatypes.
types.

The keyword
Keywords from“HDD” gives
research a applied
that high linkthestrength
standard with
RSMRSM because
mainly it is a term
consisted taken
of common
from ® v.11 software by Stat-Ease Inc., and the word design is related to
termsthein Design-Expert
the analysis stages, such as DoE and optimization (Table 3, with red, highlighted
arows).
designed experiment
By ignoring based
specific on historical
terms related toorresearch
observational data.methodological
fields, only The historical word termsdata
are
with similar link strengths were also located near RSM and were
shown in Table 3, including those with high (yellow highlighted) and low occurrencesstrengthened by word
observational data, although
(blue highlighted). The largestthey showed
cluster a low
with the occurrence. Thus, thehad
highest occurrences RSM“RSM”analysisasper-
the
formed in the papers
main keyword, applied
followed observational/historical
by “optimization” and “DoE”; datathese
as thethree
input. Subsequently,
keywords the
represent
keyword
the common “neural networks”
terms in classicformed a cluster near
RSM analysis. the “genetic
Therefore, algorithm”,
their high and these
occurrences werekey-ex-
words were located alongside the RSM. These keywords corresponded
pected. The analysis also focused on other clusters supporting them and denotes the de-with the RSM model
that was replaced
velopment by neural
of RSM-OD (the networks, and the optimization
yellow highlighted techniques
portion in Table 3). adopted a genetic
algorithm.
The keyword “HDD” gives a high link strength with RSM because it is a termanaly-
Furthermore, the blue-highlighted keywords completed the bibliometric taken
sis,
from thespecific
with methodological
Design-Expert keywords
® v.11 software from various
by Stat-Ease research
Inc., and fields.
the word Theseiskeywords
design related to
still showedexperiment
a designed a relationship withonthe
based RSM stages,
historical i.e., DoE, modeling,
or observational data. The and optimization,
historical and
word data
offered insights
with similar into
link the development
strengths were also of RSM-OD.
located near RSM and were strengthened by word
observational data, although they showed a low occurrence. Thus, the RSM analysis per-
5. Synthesis and Discussion
formed in the papers applied observational/historical data as the input. Subsequently, the
The use
keyword of observational
“neural data in RSM
networks” formed is not
a cluster without
near its critics.
the “genetic This practice
algorithm”, andcontradicts
these key-
the golden standard in RSM and runs considerable risks of being used as an alternative
words were located alongside the RSM. These keywords corresponded with the RSM
to experimental data. The authors of [38] wrote that using observational data as a replace-
model that was replaced by neural networks, and the optimization techniques adopted a
ment for experimental data is risky because of the absence of controllable factors, spurious
genetic algorithm. Furthermore, the blue-highlighted keywords completed the biblio-
correlation, and the rise of potential multicollinearity or nonorthogonality. This finding
metric analysis, with specific methodological keywords from various research fields.
was similar to the problem of semiconductor production in the work of [3] and the slurry
These keywords still showed a relationship with the RSM stages, i.e., DoE, modeling, and
thickening process in [39], where observational data contained undetected and uninter-
optimization, and offered insights into the development of RSM-OD.
pretable multicollinearity, given the application of typical observational-based regression
analysis and the need for careful handling ([40]). This opinion was also strengthened by
5. Synthesis and Discussion
the authors of [19], who wrote a systematic literature review of classic RSM development
The usethat
and showed of observational
orthogonalitydata in RSM
between is not
factors without
should its critics.toThis
be reached practice
perform contra-
individual
dicts the golden standard in RSM and runs considerable risks of being
analyses of each factor. Moreover, the ideal experimentally-based RSM accommodates used as an alterna-
tiveprocedure
the to experimental data. The
of the steepest authors
ascent of [38]
for the wrote
shifting of that
factorusing
levelsobservational data as a
in a specific direction
replacement for experimental data is risky because of the absence of controllable
toward a stationary optimum response point [1,19]. The use of observational data presents factors,
a challenge in conducting this procedure, given the limited range of factor levels. This
spurious correlation, and the rise of potential multicollinearity or nonorthogonality. The
finding was similar
optimization region istoalso
the limited
problem toof semiconductor
these production
available level ranges, asinobserved
the workinofall[3]
of and
the
the slurryreferences.
RSM-OD thickening process in [39], where observational data contained undetected and
uninterpretable
The literaturemulticollinearity,
review questions given
in the
the application of typical
previous section observational-based
served as a guide for the re-
gression analysis and the need for careful handling ([40]). This opinion was also
writing order, starting with descriptive and bibliometric analyses. Later, the synthesis stage strength-
was performed in line with answering LRQ1 to LRQ3 and continued with the discussion
Appl. Sci. 2022, 12, 10663 12 of 23

section. Some treemaps used in this paper simplified the interpretation of the descriptive
and research questions answered. Treemaps are used to hierarchically graph structured
information, which uses 100% of the available graph space [41], and acted as an excellent
application for the supporting systematic literature review in [20].
LRQ1: What are the rationales for using observational data as an alternative to conducting
a real RSM experiment?
Approximately 70.51% of the papers employed observational data as the input for
RSM, 23.08% were based on previous experimental data, and the remaining 6.41% referred
to real experiment data (Figure 8). Observation-based data were obtained depending on the
kind of data-acquiring system in the process being studied, and previous-experiment-based
data were collected from associated research. The data contained the factor (or X variables)
and response (Y variables) with continuous scales, as required by the RSM analysis. The ra-
tionale with the highest percentage in Table 4 is potential information that may exist within
the observational data. The next highest percentage is the flexible factor level (or design
space), where an RSM analysis should be flexible enough to accommodate uncontrolled
factor levels within observational data. Moreover, the difficulties in fully controlling the
factor levels during a continuous process showed the limitations in conducting designed
experiments and provided data that were a better option. This rationale also revealed a
high percentage.

Table 4. Rationales for selecting RSM-OD.

Rationales from Papers Percentage

potential information from observational data 33.33%
flexible factor level or design space (using the data as provided) 30.77%
difficult to control process parameters 21.79%
historical data contain DoE 5.13%
conducting experiments can be highly expensive 3.85%
additional experiment points to standard DoE experiments 2.56%
avoid disruption to the production process 2.56%

Several papers acquired real experiment data but used RSM-OD because of difficulties
in controlling the factor levels. They assumed the real experimental data as being observa-
tional and argued that modifying the RSM approach based on a nondesigned experiment
was easier than conducting a formal standard RSM.
LRQ2: What condition within historical data can be adapted to RSM?
Conducting the DoE experiment ensures the orthogonality between the factors, and
the ANOVA can separate each variance for the independent interpretation of their effects
([42]). On the other hand, observational data violate common assumptions in designed
experiment data, such as treatment randomization and interfactor orthogonality, as the
researcher cannot fully control each factor’s level (see an editorial by [38]). Therefore,
this section evaluated each reference paper to capture how they explain the condition of
data before treating them as the input for RSM modeling based on different approaches in
adopting data, i.e., using all observations or obtaining their subsets (Figure 9).
Table 5 shows that more than 70% of papers did not mention specific raw-data condi-
tions. Therefore, they adopted observational data directly as the input for this RSM-OD.
The mathematical model and optimization technique were previously determined without
evaluating data conditions because they forced the data to fit the model, whether linear or
nonlinear, even ignoring the absence of randomization within data. Meanwhile, 5.19% of
the papers followed the data condition as it was, which means that the RSM-OD model
and optimization techniques were adjusted to adapt to the data condition, and a linear
or nonlinear model was selected to give the best fit to the data. A total of 23.38% of the
papers explicitly mentioned other conditions, such as factor independencies, ensuring
orthogonality, and outlier removal, and considered assumptions as if they were a DoE
 Appl. Sci. 2022, 12, 10663 13 of 23
022, 12, x FOR PEER REVIEW 13 of 24

experiment. Some papers used orthogonality criteria for evaluating data conditions, such
before treatingasthem as inflation
variance the input for (for
factors RSMexample,
modeling based
a paper on and
by [43] different approaches
a data matrix in used
subsetting
adopting data,toi.e., usingorthogonality
achieve all observations
in [3]).or obtaining their subsets (Figure 9).

Figure 9. Data conditions based

Figure 9. Data on the number
conditions based on of
thedata involved
number of dataininvolved
RSM-OD.in RSM-OD.

Table 5. Required data condition for RSM-OD.

Table 5 shows that more than 70% of papers did not mention specific raw-data
conditions. Therefore, they adopted
Observational observational data directly as the input forPercentage
Data Condition this RSM-
OD. The mathematical model and optimization technique were previously
No specific data condition requirement (model and optimization stage determined
71.43%
without evaluating
were data conditions
determined withoutbecause
consideringthey forced
data the data to fit the model,
condition) whether
Assuming
linear or nonlinear, evenindependence
ignoring the of absence
factors 12.99%
of randomization within data. Meanwhile,
Ensure orthogonality between factors 9.09%
5.19% of the papers followed the data condition as it was, which means that the RSM-OD
Follow data condition as it is (specify RSM-OD model and
model and optimization techniques were adjusted to adapt to the data condition, 5.19% and a
optimization-based data condition)
linear or nonlinear model was selected to give the best fit to the data. A total of1.30%
No outliers 23.38% of
the papers explicitly mentioned other conditions, such as factor independencies, ensuring
orthogonality,LRQ3:
and outlier removal,
How historical dataandareconsidered assumptions as if they were a DoE
adopted to RSM?
experiment. Some As papers used orthogonality criteria
shown in Figure 1, the three stages forofevaluating
RSM form thedataintegrated
conditions, such for
procedures
as variance inflation
DoE-based factors (for example,
optimization. Ideally,athe
paper
RSM-ODby [43]
withand a data
similar matrix subsetting
optimization purposes should
used to achieve orthogonality in [3]).
also adopt these stages. The answers to LRQ2 explain how standard RSM stages with
modifications adapt to observational data. Especially at the designed-experiment stage,
Table 5. Required data approaches
several condition for RSM-OD.
show how the RSM-OD treats data as the input to the RSM analysis.
At the DoE stage (Figure 10), two approaches were used to adopt observational data:
Observational Data Condition
the first type used all observations (80.52%), and the second type Percentage
selected a subset (19.48%),
No specific datawith
condition
some requirement (model and
required conditions. optimization
Those that usedstage
all provided observations mostly did
71.43%
were determined without considering data condition)
not need to adopt a DoE. All observations were treated as an input for the RSM model,
Assuming independence of factorswas found based on this input. A few 12.99%
and the optimum of these papers filtered data
Ensure orthogonality between
to remove factors
unusual observations before RSM modeling. As 9.09%
for the other types, some
Follow data condition as it is were
observations (specify RSM-OD
selected model
as the and
subset optimization-
data for RSM modeling based on specific criteria.
5.19%
Mainly, the requirement of orthogonality between factors was one of the reasons for
based data condition)
No outliers selecting observations into a subset; these criteria are required in standard RSM analysis
1.30%
and fulfilled by the DoE. Thus, a certain DoE-like adaptation is needed in the RSM-OD
analysis, including
LRQ3: How historical common assumptions,
data are adopted to RSM? such as treatment randomization and interfactor
independence (orthogonality). Figure 10 shows that standard DoEs, such as factorial,
As shownoptimal,
in Figureand 1, the three
Taguchi stages
designs, wereofused
RSM as form the integrated
references in selecting procedures forsubset.
the observation
DoE-based optimization. Ideally, the RSM-OD with similar optimization purposes should
also adopt these stages. The answers to LRQ2 explain how standard RSM stages with
modifications adapt to observational data. Especially at the designed-experiment stage,
several approaches show how the RSM-OD treats data as the input to the RSM analysis.
At the DoE stage (Figure 10), two approaches were used to adopt observational data:
 observations were selected as the subset data for RSM modeling based on specific criteria.
Mainly, the requirement of orthogonality between factors was one of the reasons for se-
lecting observations into a subset; these criteria are required in standard RSM analysis
and fulfilled by the DoE. Thus, a certain DoE-like adaptation is needed in the RSM-OD
analysis, including common assumptions, such as treatment randomization and interfac-
Appl. Sci. 2022, 12, 10663 14 of 23
tor independence (orthogonality). Figure 10 shows that standard DoEs, such as factorial,
optimal, and Taguchi designs, were used as references in selecting the observation subset.

Figure 10.
Figure 10. DoE
DoE stage.
stage.

In
In the
themodeling
modelingstage
stage(Figure
(Figure11),
11),almost
almostallall
papers
papers (90.54%) applied
(90.54%) a linear
applied model;
a linear the
model;
others usedused
the others a neural network
a neural (6.76%),
network and the
(6.76%), andrestthe
usedrestother
usedmodels, such as the
other models, Taguchi
such as the
and support
Taguchi andvector model
support (2.7%).
vector modelAs (2.7%).
a commonAs linear
a common modellinear
in RSM, this in
model approach
RSM, thisworksap-
as the standard
proach works asRSM completed
the standard RSMby completed
typical statistical
by typicalanalyses, such
statistical as factorsuch
analyses, significance
as factor
and R-square.
significance ForR-square.
and the neural Fornetwork
the neuralapproach,
network most of the papers
approach, most ofimplemented it for
the papers imple-
modeling and optimization purposes. As the neural networks are
mented it for modeling and optimization purposes. As the neural networks are close to a close to a black-box
model
black-boxwithout
model any statistical
without any analysis,
statisticalthe authors
analysis, thestill performed
authors ANOVA and
still performed ANOVAR-square
and
analysis to evaluate significant factors and show an interpretable result.
R-square analysis to evaluate significant factors and show an interpretable result. Alter- Alternatively, the
Appl. Sci. 2022, 12, x FOR PEER REVIEW
Taguchi 15 of 24
natively,method approach,
the Taguchi method which was proposed
approach, which was in [2,11],
proposedwas in
also applied
[2,11], wasbased on the
also applied
typical
based on signal-to-noise ratio in its analysis.
the typical signal-to-noise ratio in its analysis.

Figure 11.
Figure 11. RSM
RSM modeling
modeling stage.
stage.

For
Forthe
theoptimization
optimization stage (Figure
stage 12),12),
(Figure as the
as highest percentage
the highest showed
percentage a linearamodel,
showed linear
amodel,
standard local search
a standard local optimization algorithm
search optimization was preferred
algorithm and commonly
was preferred provided
and commonly pro-
vided in some software. Moreover, several papers with linear models adopted metaheu-
ristics algorithms to find an optimum response. Notably, the graph in Figure 13 shows
that some papers excluded the optimization process, and they only considered the first
two stages of RSM-OD for prediction or factor investigation.
 Figure 11. RSM modeling stage.
Appl. Sci. 2022, 12, 10663 15 of 23

For the optimization stage (Figure 12), as the highest percentage showed a linear
model, a standard local search optimization algorithm was preferred and commonly pro-
vided in some in some software.
software. Moreover,Moreover, severalwith
several papers papers with
linear linearadopted
models models metaheu-
adopted metaheuristics
algorithms
ristics algorithms to find
to find an an optimum
optimum response.
response. Notably,
Notably, the graph
the graph in Figure
in Figure 13 shows that some
13 shows
papers excluded the optimization process, and they only considered
that some papers excluded the optimization process, and they only considered the first the first two stages of
RSM-OD for prediction or factor investigation.
two stages of RSM-OD for prediction or factor investigation.

Appl. Sci. 2022, 12, x FOR PEER REVIEW 16 of 24

Figure 12. Optimization stage.

Figure 12. Optimization stage.

Stages in classic RSM

Pre-processing Step 1: Observational data adoption to DoE Step 2: RSM model Step 3: Optimization
Data type Recorded data Adopted dataset DoE adopted
Code: A Code: B Code: C Code: D Code: E Code: F

Observational or Classic DoE

Process equipment Subset of observation
historical data from a (Factorial, Mixture, etc) Linear model (E1) Local search (F1)
parameter setting (B1) (C1)
process (A1) (D1)

Previous experiment Process material

Optimal DoE (D2) Neural networks (E2) Metaheuristics (F2)
data (A2) composition (B2)

Real experiment Mixed of equipment

Other model (SVM, Evo-
without following parameter and meterial Taguchi (D3) Taguchi S/N (F3)
NN, S/N ratio) (E3)
standard DoE (A3) composition (B3)

Other method (NLP,

Other DoE (D4) montecarlo)
(F4)

Prediction purposes
All observations (C2) No DoE adopted (D5)
(F5)

Method/Tools combinations of RSM-OD

Figure 13. Combination of the methods adopted in RSM-OD.

Figure 13. Combination of the methods adopted in RSM-OD.

5.1. Comparative
Comparative Analysis
Several
Several approaches
approachestotohandling
handlingobservational
observationaldatadataforfor
RSM RSMwere proposed,
were proposed,andandra-
tionales
rationaleswere provided
were provided for for
each based
each on specific
based references.
on specific Figure
references. 13 represents
Figure the com-
13 represents the
combination
bination of tools
of tools or methods
or methods applied
applied to RSM-OD,
to RSM-OD, andandthe the
mapsmaps based
based on on
thethe stages
stages in
in classic
classic RSMRSM analysis
analysis are are
shown shown in Figure
in Figure 1. By1.reading
By reading from
from the the
left lefteach
side, side,boxeach
in box
the
in the figure
figure represents
represents the tools theortools or methods
methods used inused in RSM-OD,
RSM-OD, and the and thedenote
lines lines denote the
the other
other tools/methods at each stage of the RSM analysis. Various modifications
tools/methods at each stage of the RSM analysis. Various modifications in RSM-OD in the in RSM-OD
in the reference
reference papers papers still obeyed
still obeyed the basicthe basic principles
principles of RSM,
of classical classical
andRSM,
any RSMand improve-
any RSM
improvement
ment should notshould
be muchnot be much different.
different.
The method combinations started with the identification of nondesigned-experi-
mental data (Code A). Several papers referred to observational data from this type of con-
tinuous process, and others referred to previous experimental data or conducted an actual
nondesigned experiment. Code B categorizes the recorded variables that will be the fac-
Appl. Sci. 2022, 12, 10663 16 of 23

The method combinations started with the identification of nondesigned-experimental

data (Code A). Several papers referred to observational data from this type of continuous
process, and others referred to previous experimental data or conducted an actual nonde-
signed experiment. Code B categorizes the recorded variables that will be the factors in
RSM. Primarily, the studied process records its equipment parameters, the composition
of the materials, or both. Code C represents how the provided data will be treated by
considering all observations or selecting its subset. Code D categorizes the standard DoE
adopted while treating the provided data within the RSM analysis. Code E shows the RSM
model adopted, and Code F represents the optimization technique.
The combinations of Code A–F provided many options. However, the main concept of
the three-stage RSM became the focus of grouping each paper. As is shown in Table 6, based
on Codes C (stage 1: treating data), E (Stage 2: RSM model), and F (Stage 3: optimization),
only seven types of approaches, which were represented by seven clusters, were obtained,
and the references are shown in Table 6.

Table 6. References for Figure 13.

Three Stages of RSM Additional Stage

Clusters Stage 1 Stage 2 Stage 3 References
Code A Code B Code D
(Code C) (Code E) (Code F)
D1 [44]
F1 B1
D3 [2]
D2 [45]
F2
Cluster 1: B2 D2 [35]
Subset—Linear model—local D1 [46]
C1 E1 A1
search
(12.05%) F4 D3 [11]
B1
D3 [47]
D1 [3]
F5
B2 D2 [13,48]
Cluster 2: B2 [49]
Subset—NN F2
model—metaheuristics. C1 E2 A1 D1 [50]
(3.61%) F5 B3 [12]
Cluster 3:
Subset—other models—other
C1 E3 F5 A1 B3 D3 [51]
purposes.
(1.20%)
B1 [8,43,52–60]
A1
B2 [7,61–65]
B3 [66,67]
Cluster 4:
All obs.—linear model—local B1 [6,68–73]
C2 E1 F1 D5
search B2 [74–77]
A2
(55.42%)
B3 [9,78–82]
B1 [83,84]
A3
B2 [85]
D4 [86]
Cluster 5: B1
All obs—linear D5 [4,87–91]
C2 E1 F2 A1
model—metaheuristics B2 D1 [92]
(10.84%)
B3 D4 [93]
Appl. Sci. 2022, 12, 10663 17 of 23

Table 6. Cont.

Three Stages of RSM Additional Stage

Clusters Stage 1 Stage 2 Stage 3 References
Code A Code B Code D
(Code C) (Code E) (Code F)
F4 A3 B2 D4 [2,94]
Cluster 6:
All obs.—linear model—other B1 [95–97]
C2 E1
optimization technique A1 B2 [98,99]
F5 D5
(8.43%)
A3 B3 [100]
Cluster 7: F2 A1 [101]
All obs.- NN
model—metaheuristics C2 E2 B1 D5
(7.23%) F5 A2 [102]

The most preferred was cluster 4, with 55.42% relativity to all the selected reference
papers; it used all observations as the input to the linear RSM model and applied the
ordinary local search optimization method. It is similar to standard RSM, but risks may arise
during the analysis by selecting all observations. Cluster 1, which was similar to Cluster 4,
had the second highest value: 12.05%; the difference is that this cluster selected a subset
of observations that fulfilled a particular DoE and guaranteed interfactor orthogonality.
Next, Cluster 5 (10.84%) was similar to Cluster 4 but replaced the local optimization
method with a metaheuristic technique. A more complex RSM model with all observations
as the input became the rationale for this replacement. Cluster 6 (8,43%) applied other
optimization techniques, such as Taguchi S/N ratio, linear programming, and the Monte
Carlo method [2,44,45]. In Clusters 2 and 7, the linear model was replaced with neural
networks to handle the nonlinearity of the observational data, all observations, or the
subset data. Moreover, such a complicated black-box neural-network model applied the
metaheuristics method to find the optimum. Concerning the three stages of the RSM, a
summary of the method combinations (Figure 13 and Table 6) is rewritten in Table 7.

Table 7. Summary of method combinations in consideration of the three stages of RSM.

Advantages Disadvantage
Selecting a subset based on specific criteria A number of of observations will be
subset
increases inter-factor orthogonality excluded from the RSM analysis
Stage 1
RSM As a potential source of information, all
potential multicollinearity between factors
all observation observations will be included in the RSM
and the possibility of outlier observations
analysis
strong foundation with clear inference and
linear model strictly statistical assumptions
interpretation
Stage 2 no model interpretation and potential
RSM Neural-net model black-box model free of assumptions
garbage-in-garbage-out
Similar to neural networks, the SVM model has no required assumptions, and the Taguchi
other models
method works without a pre-specified mathematical model.
local search fast iterative algorithm potential local optimum
Stage 3 metaheuristics accommodate global optimum highly depends on initial conditions
RSM Some papers with prediction purposes exclude optimization techniques; the others involve
other technique
linear programming and Monte-Carlo.

5.2. Advantages and Disadvantages of RSM-OD

Based on the synthesis above, the discussion emphasized how the classic RSM concept
methodologically adopts nonexperimental data as an alternative to the DoE experiment.
The classic RSM has strong scientific references when integrating the three stages of its
Appl. Sci. 2022, 12, 10663 18 of 23

analysis (Figure 1), and each stage also gives a clear theoretical basis. Therefore, any
development in RSM, including the fulfillment of assumptions during the analysis, should
remain within these stages. Thus, the discussion will explore the methods and combinations
used in the reference paper (Figure 13).
According to Table 7, those options that combined the methods within the three stages
of RSM raised some advantages and limitations. In stage 1 of the RSM, contradictions
existed between the selection of all observations or their subsets. One problem relates
to interfactor orthogonality and the other deals with the justification of selecting only a
subset from several potentially informative observations. In stage 2, different types of RSM
models, i.e., linear (or polynomial) or machine-learning type models, provided different
modeling approaches with each of their consequences. The powerful and interpretable
linear model works with several strict assumptions, whereas the free-assumption machine-
learning-based model contains potential over-fitting and is noninterpretable. In stage 3,
the ordinary local search algorithm works best for a single-optimum point linear model,
whereas the metaheuristics algorithm provides a larger search area with local and global
optima.
By referring to the papers needing observational data, RSM can be developed with
alternatives to conduct a real experiment. Notably, observational data will not give pieces
of information that are as perfect as within the designed experiments because of the
assumptions of violations within. However, numerous references in this paper have
shown the success of RSM-OD, although some ignored the concept behind the classic RSM.
Therefore, a new procedure must be developed for this type of RSM to fulfill all the required
assumptions of the standard classic RSM.

5.3. Potential Gaps and Future Research

With all the explained descriptive and synthesis analyses, we identified opportunities
and gaps in the development of new RSMs in consideration of adopting observational
data (Table 8). Stage 1 deals with how the developed procedures work, according to the
concept of classic DoE, including the concept of orthogonality and randomization. Stage 2
developments can be improved when considering model interpretation, including factor
significance and goodness-of-fit. Stage 3 deals with the capability of finding the global
optimum based on the fitted model in Stage 2. All these opportunities are expected to give a
stronger theoretical basis for implementing RSM-OD to complete its practical applications,
assuring the users regarding its use.

Table 8. Opportunities and gaps for further development.

Development Opportunities
RSM Stages Potential Gaps in References
for Future Research
Procedure development to:
1. fulfill factor orthogonality and its
evaluation/measurement
Develop procedures to adopt observational 2. Improve orthogonality of observational data
Stage 1 3. handle non-randomized treatment within observational
data considering the concept of classic DoE
data
4. pre-process observational data
(cleaning/filtering/subsetting)
5. Dividing variation for each factor, similar to ANOVA
Model development to:
Develop an adaptive RSM mathematical 1. accommodate un-designed/unpatterned observational
Stage 2 model to adapt observational data data
concerning required assumptions 2. fulfill model-fitting assumptions, or ignore them
3. enhance of model interpretability
Appl. Sci. 2022, 12, 10663 19 of 23

Table 8. Cont.

Development Opportunities
RSM Stages Potential Gaps in References
for Future Research
Optimization technique to:
Develop an optimization algorithm referring
Stage 3 1. provide a comprehensive optimum search area
to a pre-defined RSM model
2. avoid local optimum

6. Conclusions
Using observational data within RSM is promising, particularly when data-recording
technology (big data) exists. It was found that the main rationales for adopting observa-
tional data within RSM are the existence of historical data and avoiding interruptions in
continuous production. However, due to the unstructured, highly variable, and serial-
correlated nature of the observational data, data modifications prior to use in the RSM
is necessary. Therefore, the paper aims to explore the various methods/approaches for
incorporating observational data in RSM through a systematic literature review using the
PRISMA framework, from which 83 studies were analyzed. Based on the three stages of
classic RSM, modifications can be conducted at each stage, i.e., data treatment, modeling,
and optimization. With respect to the first stage (data treatment), the modification involves
selecting an observation subset or pretreating the data to increase acceptance in the RSM
based on specific criteria, such as orthogonality and treatment randomization. In the second
stage, adaptive RSM mathematical models are selected to handle nonideal observational
data. Complex nonlinear machine learning models are common approaches for adapting
RSM models, for example, the neural network and SVM models. In the last stage, an
alternative optimization method suitable for such a complex RSM model is also highlighted.
Metaheuristic optimization techniques perform well when finding the optimal factor levels
modeled using a nonlinear RSM model. The combinations of the proposed methods for the
RSM stages reveal insights into the fact that there is an open potential for developments in
RSM-OD as an alternative to classic RSM.
Despite the deviation from standard RSM techniques, the proposed RSM-OD methods
in the literature can still achieve their design/process optimization purpose with reasonable
results. However, the methods also raised some limitations, such as data orthogonality
issues, statistical assumptions, model specifications, model interpretability, and the need
for advanced optimization methods.
This paper contributes to the RSM literature by providing the advantages or dis-
advantages of using observational data for process/design optimization, demonstrating
opportunities to further improve the proposed methods in RSM-OD, and coping with their
theoretical limitations and unexpressed assumptions. Once those issues are well addressed,
RSM-OD may be a promising alternative to classic RSM.

Author Contributions: Conceptualization, M.A.H.; methodology, M.A.H., B.M.S. and B.S.W.; formal
analysis, M.A.H., B.M.S. and B.S.W.; writing—original draft preparation, M.A.H.; writing—review
and editing, B.M.S. and B.S.W.; supervision, B.M.S. and B.S.W.; funding acquisition, M.A.H. All
authors have read and agreed to the published version of the manuscript.
Funding: This research was funded by Universitas Surabaya (Contract No. 1186/PKD-SL/SDM/IX/2020).
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: Not applicable.
Conflicts of Interest: The authors declare no conflict of interest.
Appl. Sci. 2022, 12, 10663 20 of 23

Abbreviations

Abbreviations Full Form

(Alphabetical Order)
DoE Design of experiment
HDD Historical data design
LRQ Literature review questions
NN Neural network model
PRISMA Preferred reporting items for systematic reviews and meta analyses
RSM Response surface methodology
RSM-OD Observational data-based RSM
SLR Systematic literature review
SVM Support vector machine model

References
1. Myers, R.H.; Montgomery, D.C.; Anderson-Cook, C.M. Response Surface Methodology: Process and Product Optimization Using
Designed Experiments; Wiley Series in Probability and Statistics; Wiley: Hoboken, NJ, USA, 2016.
2. Sukthomya, W.; Tannock, J.D.T. Taguchi Experimental Design for Manufacturing Process Optimisation Using Historical Data and
a Neural Network Process Model. Int. J. Qual. Reliab. Manag. 2005, 22, 485–502. [CrossRef]
3. Chien, C.F.; Chang, K.H.; Wang, W.C. An Empirical Study of Design-of-Experiment Data Mining for Yield-Loss Diagnosis for
Semiconductor Manufacturing. J. Intell. Manuf. 2014, 25, 961–972. [CrossRef]
4. Sadati, N.; Chinnam, R.B.; Nezhad, M.Z. Observational Data-Driven Modeling and Optimization of Manufacturing Processes.
Expert Syst. Appl. 2018, 93, 456–464. [CrossRef]
5. Cerquitelli, T.; Pagliari, D.J.; Calimera, A.; Bottaccioli, L.; Patti, E.; Acquaviva, A.; Poncino, M. Manufacturing as a Data-Driven
Practice: Methodologies, Technologies, and Tools. Proc. IEEE 2021, 109, 399–422. [CrossRef]
6. Hussain, S.; Khan, H.; Khan, N.; Gul, S.; Wahab, F.; Khan, K.I.; Zeb, S.; Khan, S.; Baddouh, A.; Mehdi, S.; et al. Process Modeling
toward Higher Degradation and Minimum Energy Consumption of an Electrochemical Decontamination of Food Dye Wastewater.
Environ. Technol. Innov. 2021, 22, 101509. [CrossRef]
7. Fazeli Burestan, N.; Afkari Sayyah, A.H.; Taghinezhad, E. Mathematical Modeling for the Prediction of Some Quality Parameters
of White Rice Based on the Strength Properties of Samples Using Response Surface Methodology (RSM). Food Sci. Nutr. 2020, 8,
4134–4144. [CrossRef]
8. Garg, H.K.; Singh, R. Investigations for Obtaining Desired Strength of Nylon6 and Fe Powder-Based Composite Wire for FDM
Feedstock Filament. Prog. Addit. Manuf. 2017, 2, 73–83. [CrossRef]
9. Mahmoodi, N.M.; Taghizadeh, M.; Taghizadeh, A. Activated Carbon/Metal-Organic Framework Composite as a Bio-Based Novel
Green Adsorbent: Preparation and Mathematical Pollutant Removal Modeling. J. Mol. Liq. 2019, 277, 310–322. [CrossRef]
10. Demchenko, Y.; De Laat, C.; Membrey, P. Defining Architecture Components of the Big Data Ecosystem. In Proceedings of the
2014 International Conference on Collaboration Technologies and Systems (CTS), Minneapolis, MN, USA, 19–23 May 2014; pp.
104–112.
11. Khoei, A.R.; Masters, I.; Gethin, D.T. Design Optimisation of Aluminium Recycling Processes Using Taguchi Technique. J. Mater.
Process. Technol. 2002, 127, 96–106. [CrossRef]
12. Loy, C.; Goh, T.N.; Xie, M. Retrospective Factorial Fitting and Reverse Design of Experiments. Total Qual. Manag. 2002, 13,
589–602. [CrossRef]
13. Berni, R. The Use of Observational Data to Implement an Optimal Experimental Design. Qual. Reliab. Eng. Int. 2003, 19, 307–315.
[CrossRef]
14. Kong, W.; Qiao, F.; Wu, Q. Real-Manufacturing-Oriented Big Data Analysis and Data Value Evaluation with Domain Knowledge.
Comput. Stat. 2020, 35, 515–538. [CrossRef]
15. Tao, F.; Qi, Q.; Liu, A.; Kusiak, A. Data-Driven Smart Manufacturing. J. Manuf. Syst. 2018, 48, 157–169. [CrossRef]
16. Harding, J.A.; Shahbaz, M.; Srinivas; Kusiak, A. Data Mining in Manufacturing: A Review. J. Manuf. Sci. Eng. Trans. ASME 2006,
128, 969–976. [CrossRef]
17. Kuo, Y.H.; Kusiak, A. From Data to Big Data in Production Research: The Past and Future Trends. Int. J. Prod. Res. 2019, 57,
4828–4853. [CrossRef]
18. Page, M.J.; McKenzie, J.E.; Bossuyt, P.M.; Boutron, I.; Hoffmann, T.C.; Mulrow, C.D.; Shamseer, L.; Tetzlaff, J.M.; Akl, E.A.;
Brennan, S.E.; et al. The PRISMA 2020 Statement: An Updated Guideline for Reporting Systematic Reviews. Int. J. Surg. 2021,
88, 105906. [CrossRef]
19. de Oliveira, L.G.; de Paiva, A.P.; Balestrassi, P.P.; Ferreira, J.R.; da Costa, S.C.; da Silva Campos, P.H. Response Surface Methodology
for Advanced Manufacturing Technology Optimization: Theoretical Fundamentals, Practical Guidelines, and Survey Literature
Review. Int. J. Adv. Manuf. Technol. 2019, 104, 1785–1837. [CrossRef]
20. Arboretti, R.; Ceccato, R.; Pegoraro, L.; Salmaso, L. Design of Experiments and Machine Learning for Product Innovation: A
Systematic Literature Review. Qual. Reliab. Eng. Int. 2021, 38, 1131–1156. [CrossRef]
Appl. Sci. 2022, 12, 10663 21 of 23

21. Akteke-Ozturk, B.; Koksal, G.; Weber, G.W. Nonconvex Optimization of Desirability Functions. Qual. Eng. 2018, 30, 293–310.
[CrossRef]
22. Box, A.G.E.P.; Wilson, K.B. On the Experimental Attainment of Optimum Conditions. J. R. Stat. Soc. Ser. B 1951, 13, 1–45.
[CrossRef]
23. Venkata Rao, K.; Murthy, P.B.G.S.N. Modeling and Optimization of Tool Vibration and Surface Roughness in Boring of Steel
Using RSM, ANN and SVM. J. Intell. Manuf. 2018, 29, 1533–1543. [CrossRef]
24. Mia, M.; Dhar, N.R. Prediction and Optimization by Using SVR, RSM and GA in Hard Turning of Tempered AISI 1060 Steel under
Effective Cooling Condition. Neural Comput. Appl. 2019, 31, 2349–2370. [CrossRef]
25. Alian, E.; Semnani, A.; Firooz, A.; Shirani, M.; Azmoon, B. Application of Response Surface Methodology and Genetic Algorithm
for Optimization and Determination of Iron in Food Samples by Dispersive Liquid–Liquid Microextraction Coupled UV–Visible
Spectrophotometry. Arab. J. Sci. Eng. 2018, 43, 229–240. [CrossRef]
26. Chen, W.C.; Nguyen, M.H.; Chiu, W.H.; Chen, T.N.; Tai, P.H. Optimization of the Plastic Injection Molding Process Using the
Taguchi Method, RSM, and Hybrid GA-PSO. Int. J. Adv. Manuf. Technol. 2016, 83, 1873–1886. [CrossRef]
27. Hazir, E.; Ozcan, T. Response Surface Methodology Integrated with Desirability Function and Genetic Algorithm Approach for
the Optimization of CNC Machining Parameters. Arab. J. Sci. Eng. 2019, 44, 2795–2809. [CrossRef]
28. Yin, S.; Ding, S.X.; Xie, X.; Luo, H. A Review on Basic Data-Driven Approaches for Industrial Process Monitoring. IEEE Trans. Ind.
Electron. 2014, 61, 6418–6428. [CrossRef]
29. Tseng, M.L.; Tran, T.P.T.; Ha, H.M.; Bui, T.D.; Lim, M.K. Sustainable Industrial and Operation Engineering Trends and Challenges
Toward Industry 4.0: A Data Driven Analysis. J. Ind. Prod. Eng. 2021, 38, 581–598. [CrossRef]
30. Tsang, K.F.; Lau, H.C.W.; Kwok, S.K. Development of a Data Mining System for Continual Process Quality Improvement. Proc.
Inst. Mech. Eng. Part B J. Eng. Manuf. 2007, 221, 179–193. [CrossRef]
31. Design-Expert 13; Stat-Ease, Inc.: Minneapolis, MN, USA, 2022; Available online: https://www.statease.com/docs/v11/tutorials/
(accessed on 10 January 2022).
32. Anderson, M.J.; Whitcomb, P.J. RSM Simplified: Optimizing Processes Using Response Surface Methods for Design of Experiments, 2nd
ed.; CRC Press: Boca Raton, FL, USA, 2017.
33. Wu, D.; Wei, Y.; Terpenny, J. Predictive Modelling of Surface Roughness in Fused Deposition Modelling Using Data Fusion. Int. J.
Prod. Res. 2019, 57, 3992–4006. [CrossRef]
34. Maimon, O.; Rokach, L. Data Mining and Knowledge Discovery Handbook; Springer: New York, NY, USA, 2010.
35. Berni, R.; De March, D.; Stefanini, F.M. T-Optimality and Neural Networks: A Comparison of Approaches for Building
Experimental Designs. Appl. Stoch. Model. Bus. Ind. 2013, 29, 454–467. [CrossRef]
36. Xiao, Y.; Watson, M. Guidance on Conducting a Systematic Literature Review. J. Plan. Educ. Res. 2019, 39, 93–112. [CrossRef]
37. Pranckutė, R. Web of Science (Wos) and Scopus: The Titans of Bibliographic Information in Today’s Academic World. Publications
2021, 9, 12. [CrossRef]
38. Montgomery, D. Exploring Observational Data. Qual. Reliab. Eng. Int. 2017, 33, 1639–1640. [CrossRef]
39. Oulhiq, R.; Benjelloun, K.; Kali, Y.; Saad, M.; Kali, Y. A Data Mining Based Approach for Process Identification Using Historical
Data. Int. J. Model. Simul. 2021, 42, 335–349. [CrossRef]
40. Draper, N.R.; Smith, H. Applied Regression Analysis, 3rd ed.; John Wiley & Sons: Hoboken, NJ, USA, 1998.
41. Johnson, B.; Shneiderman, B. Tree-Maps: A Space-Filling Approach to the Visualization of Hierarchical Information Structures. In
Proceedings of the Visualization; IEEE: Manhattan, NY, USA, 1991.
42. Montgomery, D.C. Design and Analysis of Experiments, 9th ed.; Wiley: Hoboken, NJ, USA, 2017.
43. Zakria, M.H.; Nawawi, M.G.M.; Rahman, M.R.A. Ethylene Yield from a Large Scale Naphtha Pyrolysis Cracking Utilizing
Response Surface Methodology. Pertanika J. Sci. Technol. 2021, 29, 791–808. [CrossRef]
44. Cox, J.R. A Design of Experiments Approach to Turbine Engine Aeromechanical Ground Testing. In Proceedings of the 45th
AIAA/ASME/SAE/ASEE Joint Propulsion Conference & Exhibit, Denver, CO, USA, 1–6 August 2009.
45. Adeyinka, A.; Olatunde, F.; Bodunrin, A. Deepwater Infill Drilling Evaluation Using Experimental Design: The Agbami Case
Study. In Proceedings of the SPE Nigeria Annual International Conference and Exhibition, Lagos, Nigeria, 31 July–2 August 2017;
pp. 300–313.
46. Galí, A.; Ascaso, M.; Nardi-Ricart, A.; Suñé-Pou, M.; Pérez-Lozano, P.; Suñé-Negre, J.M.; García-Montoya, E. Robustness
Optimization of an Existing Tablet Coating Process Applying Retrospective Knowledge (Rqbd) and Validation. Pharmaceutics
2020, 12, 743. [CrossRef]
47. Khoei, D.A.R.; Masters, D.I.; Gethin, P.D.T. Historical Data Analysis in Quality Improvement of Aluminum Recycling Process. In
Recycling of Metals and Engineered Materials; The Minerals, Metals and Materials Society: Warrendale, PA, USA, 2000; pp. 1063–1074.
48. Rudisill, T.S.; Hobbs, D.T.; Edwards, T.B. Plutonium Solubility in Simulated Savannah River Site Waste Solutions. Sep. Sci. Technol.
2010, 45, 1782–1792. [CrossRef]
49. Liu, Y.C.; Yeh, I.C. Using Mixture Design and Neural Networks to Build Stock Selection Decision Support Systems. Neural Comput.
Appl. 2017, 28, 521–535. [CrossRef]
50. Vlassides, S.; Ferrier, J.G.; Block, D.E. Using Historical Data for Bioprocess Optimization: Modeling Wine Characteristics Using
Artificial Neural Networks and Archived Process Information. Biotechnol. Bioeng. 2001, 73, 55–68. [CrossRef]
Appl. Sci. 2022, 12, 10663 22 of 23

51. Rahman, M.M.; Imtiaz, S.A.; Hawboldt, K. A Hybrid Input Variable Selection Method for Building Soft Sensor from Correlated
Process Variables. Chemom. Intell. Lab. Syst. 2016, 157, 67–77. [CrossRef]
52. Ighalo, J.O.; Adeniyi, A.G. Thermodynamic Modelling of Dimethyl Ether Steam Reforming. Clean Technol. Environ. Policy 2021,
23, 1353–1363. [CrossRef]
53. Ekpotu, W.F.; Ighalo, J.O.; Nkundu, K.B.; Ogwo, P.; Adeniyi, A.G. Analysis of Factor Effects and Interactions in a Conventional
Drilling Operation by Response Surface Methodology and Historical Data Design. Pet. Coal Artic. 2020, 62, 1356–1368.
54. Raina, A.K. Influence of Joint Conditions and Blast Design on Pre-Split Blasting Using Response Surface Analysis. Rock Mech.
Rock Eng. 2019, 52, 4057–4070. [CrossRef]
55. Komaravolu, Y.; Dama, V.R.; Maringanti, T.C. Novel, Efficient, Facile, and Comprehensive Protocol for Post-Column Amino Acid
Analysis of Icatibant Acetate Containing Natural and Unnatural Amino Acids Using the QbD Approach. Amino Acids 2019, 51,
295–309. [CrossRef]
56. Salam, K.; Agarry, S.; Arinkoola, A.; Shoremekun, I. Optimization of Operating Conditions Affecting Microbiologically Influenced
Corrosion of Mild Steel Exposed to Crude Oil Environments Using Response Surface Methodology. Br. Biotechnol. J. 2015, 7,
68–78. [CrossRef]
57. Pirmohammad, S.; Esmaeili-Marzdashti, S.; Eyvazian, A. Crashworthiness Design of Multi-Cell Tapered Tubes Using Response
Surface Methodology. J. Comput. Appl. Res. Mech. Eng. 2019, 9, 57–72.
58. Faleiro, R.M.R.; Velloso, C.M.; De Castro, L.F.A.; Sampaio, R.S. Statistical Modeling of Charcoal Consumption of Blast Furnaces
Based on Historical Data. J. Mater. Res. Technol. 2013, 2, 303–307. [CrossRef]
59. Kockal, N.U.; Ozturan, T. Optimization of Properties of Fly Ash Aggregates for High-Strength Lightweight Concrete Production.
Mater. Des. 2011, 32, 3586–3593. [CrossRef]
60. Nookaraju, B.C.; Sohail, M. Experimental Investigation and Optimization of Process Parameters of Hybrid Wick Heat Pipe Using
with RSM Historical Data Design. Mater. Today Proc. 2020, 46, 36–43. [CrossRef]
61. Zullaikah, S.; Putra, A.K.; Fachrudin, F.H.; Utomo, A.T.; Naulina, R.Y.; Utami, S.; Herminanto, R.P.; Ju, Y.H. Experimental Investi-
gation and Optimization of Non-Catalytic In-Situ Biodiesel Production from Rice Bran Using Response Surface Methodology
Historical Data Design. Int. J. Renew. Energy Dev. 2021, 10, 804–810. [CrossRef]
62. Luga, E.; Peqini, K. The Influence of Oxide Content on the Properties of Fly Ash/Slag Geopolymer Mortars Activated with NaOH.
Period. Polytech. Civ. Eng. 2019, 63, 1217–1224. [CrossRef]
63. Wulff, R.; Leopold, C.S. Coatings from Blends of Eudragit® RL and L55: A Novel Approach in PH-Controlled Drug Release. Int.
J. Pharm. 2014, 476, 78–87. [CrossRef]
64. Jeirani, Z.; Mohamed Jan, B.; Si Ali, B.; Noor, I.M.; See, C.H.; Saphanuchart, W. Prediction of the Optimum Aqueous Phase
Composition of a Triglyceride Microemulsion Using Response Surface Methodology. J. Ind. Eng. Chem. 2013, 19, 1304–1309.
[CrossRef]
65. Jeirani, Z.; Mohamed Jan, B.; Si Ali, B.; Mohd Noor, I.; See, C.H.; Saphanuchart, W. Prediction of Water and Oil Percolation
Thresholds of a Microemulsion by Modeling of Dynamic Viscosity Using Response Surface Methodology. J. Ind. Eng. Chem. 2013,
19, 554–560. [CrossRef]
66. Shakor, Z.M.; AbdulRazak, A.A.; Shuhaib, A.A. Optimization of Process Variables for Hydrogenation of Cinnamaldehyde
to Cinnamyl Alcohol over a Pt/SiO2 Catalyst Using Response Surface Methodology. Chem. Eng. Commun. 2021, 209, 1–17.
[CrossRef]
67. Widyaningsih, T.D.; Widjanarko, S.B.; Waziiroh, E.; Wijayanti, N.; Maslukhah, Y.L. Pilot Plant Scale Extraction of Black Cincau
(Mesona Palustris BL) Using Historical-Data Response Surface Methodology. Int. Food Res. J. 2018, 25, 712–719.
68. Petrotos, K.; Giavasis, I.; Gerasopoulos, K.; Mitsagga, C.; Papaioannou, C.; Gkoutsidis, P. Optimization of the Vacuum Microwave
Assisted Extraction of the Natural Polyphenols and Flavonoids from the Raw Solid Waste of the Pomegranate Juice Producing
Industry at Industrial Scale. Molecules 2021, 26, 1033. [CrossRef]
69. Kasim, M.S.; Harun, N.H.; Hafiz, M.S.A.; Mohamed, S.B.; Mohamad, W.N.F.W. Multi-Response Optimization of Process Parameter
in Fused Deposition Modelling by Response Surface Methodology. Int. J. Recent Technol. Eng. 2019, 8, 327–338. [CrossRef]
70. Nemati, N.; Eslamlueyan, R. Development of RSM Statistical Model for Methanol Carbonylation Rate for Acetic Acid Synthesis
by Using Cativa TM Technology. Chem. Prod. Process Model. 2019, 14, 1–13. [CrossRef]
71. Salam, K.K.; Arinkoola, A.O.; Aminu, M.D. Application of Response Surface Methodology (RSM) For the Modelling and
Optimization of Sand Minimum Transport Condition (MTC) in Pipeline Multiphase Flow. Pet. Coal 2018, 60, 339–348.
72. Singh, R. Modelling of Micro Hardness in Cold Chamber Pressure Die Casting Process. Adv. Mater. Process. Technol. 2017, 3,
438–448. [CrossRef]
73. Mahmoodi, N.M.; Keshavarzi, S.; Rezaei, P. Synthesis of Copper Oxide Nanoparticle and Photocatalytic Dye Degradation Study
Using Response Surface Methodology (RSM) and Genetic Algorithm (GA). Desalin. Water Treat. 2017, 72, 394–405. [CrossRef]
74. Babu, S.K.; Rao, M.V.; Babu, S.P.; Chakka, M.V.V.S. Chemometric Assisted Development and Validation of a Stability-Indicating
Lc Method for Determination of Related Substances in Haloperidol Decanoate Injection. Indian J. Pharm. Educ. Res. 2021, 55,
904–915. [CrossRef]
75. Ghiasi, E.; Malekzadeh, A. Removal of Various Textile Dyes Using LaMn(Fe)O3 and LaFeMn0.5 O3 Nanoperovskites; RSM
Optimization, Isotherms and Kinetics Studies. J. Inorg. Organomet. Polym. Mater. 2020, 30, 2789–2804. [CrossRef]
Appl. Sci. 2022, 12, 10663 23 of 23

76. Olia, M.S.J.; Azin, M.; Sepahy, A.A.; Moazami, N. Feasibility of Improving Carbohydrate Content of Chlorella S4, a Native Isolate
from the Persian Gulf Using Sequential Statistical Designs. Biofuels 2019, 13, 1–9. [CrossRef]
77. Rao, P.D.; Kiran, C.U.; Prasad, K.E. Mathematical Model and Optimisation for Tensile Strength of Human Hair Reinforced
Polyester Composites. Int. J. Comput. Mater. Sci. Surf. Eng. 2019, 8, 76–88. [CrossRef]
78. Samadi, A.; Sharifi, H.; Ghobadi Nejad, Z.; Hasan-Zadeh, A.; Yaghmaei, S. Biodegradation of 4-Chlorobenzoic Acid by Lysini-
bacillus Macrolides DSM54T and Determination of Optimal Conditions. Int. J. Environ. Res. 2020, 14, 145–154. [CrossRef]
79. Zainal, B.S.; Danaee, M.; Mohd, N.S.; Ibrahim, S. Effects of Temperature and Dark Fermentation Effluent on Biomethane
Production in a Two-Stage up-Flow Anaerobic Sludge Fixed-Film (UASFF) Bioreactor. Fuel 2020, 263, 116729. [CrossRef]
80. Muhamad, M.S.; Hamidon, N.; Salim, M.R.; Yusop, Z.; Lau, W.J.; Hadibarata, T. Response Surface Methodology for Modeling
Bisphenol A Removal Using Ultrafiltration Membrane System. Water. Air. Soil Pollut. 2018, 229, 222. [CrossRef]
81. Wan Azelee, I.; Goh, P.S.; Lau, W.J.; Ismail, A.F. Facile Acid Treatment of Multiwalled Carbon Nanotube-Titania Nanotube Thin
Film Nanocomposite Membrane for Reverse Osmosis Desalination. J. Clean. Prod. 2018, 181, 517–526. [CrossRef]
82. Chen, L.; Liu, Z.; Sun, P.; Huo, W. Formulation of a Fuel Spray SMD Model at Atmospheric Pressure Using Design of Experiments
(DoE). Fuel 2015, 153, 355–360. [CrossRef]
83. Mutalib, N.A.A.; Jaswir, I.; Akmeliawati, R.; Latief, M.; Octavianti, F.; Alkahthani, H. Abstract. Optimization of Lard Compound
Analysis Using Portable Electronic Nose Based upon Response Surface Methodology. Malays. J. Consum. Fam. Econ. 2018, 21,
125–138.
84. Ajav, E.A.; Akogun, O.A. The Performance of a Combined Dewatered Cassava Mash Lump Pulverizer and Sifter under Some
Operational Factors. Agric. Eng. Int. CIGR J. 2015, 17, 82–92.
85. Mohammed, B.S.; Adamu, M.; Liew, M.S. Evaluating the Static and Dynamic Modulus of Elasticity of Roller Compacted
Rubbercrete Using Response Surface Methodology. Int. J. Geomate 2018, 14, 186–192. [CrossRef]
86. Peces, D.P.; García-Montoya, E.; Manich, A.; Suñé-Negre, J.M.; Pérez-Lozano, P.; Miñarro, M.; Ticó, J.R. Approach to Design Space
from Retrospective Quality Data. Pharm. Dev. Technol. 2016, 21, 26–38. [CrossRef]
87. Fellaou, S.; Harnoune, A.; Seghra, M.A.; Bounahmidi, T. Statistical Modeling and Optimization of the Combustion Efficiency in
Cement Kiln Precalciner. Energy 2018, 155, 351–359. [CrossRef]
88. Chan, C.H.; Kasim, M.S.; Izamshah, R.; Bakar, H.A.; Sundi, S.A.; Zakaria, K.A.; Haron, C.H.C.; Ghani, J.A.; Hafiz, M.S.A. Analysis
of Face Milling Performance on Inconel 718 Using FEM and Historical Data of RSM. IOP Conf. Ser. Mater. Sci. Eng. 2017,
270, 012038. [CrossRef]
89. Irudayaraj, S.; Charles, S. RSM Based Prediction of Process Parameters in the Grinding Process of Portland Pozzolana Cement.
Int. J. Appl. Eng. Res. 2015, 10, 15513–15522.
90. Irudayaraj, S.; Charles, S. Optimization of Ball Mill Operating Parameters for Their Effect on Mill Output and Cement Fineness
by Using RSM Method. Int. J. Appl. Eng. Res. 2014, 9, 19959–19967.
91. Šibalija, T.; Majstorovic, V.; Sokovic, M. Taguchi-Based and Intelligent Optimisation of a Multi-Response Process Using Historical
Data. Stroj. Vestnik/J. Mech. Eng. 2011, 57, 357–365. [CrossRef]
92. Kostić, S.; Vasović, N.; Marinković, B. Robust Optimization of Concrete Strength Estimation Using Response Surface Methodology
and Monte Carlo Simulation. Eng. Optim. 2017, 49, 864–877. [CrossRef]
93. Gagliardi, F.; Ambrogio, G.; Ciancio, C.; Filice, L. Metamodeling Technique for Designing Reengineered Processes by Historical
Data. J. Manuf. Syst. 2017, 45, 195–200. [CrossRef]
94. Fatoni, R.; Elkamel, A.; Simon, L.; Almansoori, A. A Computer-Aided Framework for Product Design with Application to Wheat
Straw Polypropylene Composites. Can. J. Chem. Eng. 2015, 93, 2141–2149. [CrossRef]
95. Karami, H.R.; Keyhani, M.; Mowla, D. Experimental Analysis of Drag Reduction in the Pipelines with Response Surface
Methodology. J. Pet. Sci. Eng. 2016, 138, 104–112. [CrossRef]
96. Mohamed, M.S.; Mohamad, R.; Ramanan, R.N.; Manan, M.A.; Ariff, A.B. Modeling of Oxygen Transfer Correlations for Stirred
Tank Bioreactor Agitated with Atypical Helical Ribbon Impeller. Am. J. Appl. Sci. 2009, 6, 848–856. [CrossRef]
97. Galí, A.; García-Montoya, E.; Ascaso, M.; Pérez-Lozano, P.; Ticó, J.R.; Miñarro, M.; Suñé-Negre, J.M. Improving Tablet Coating
Robustness by Selecting Critical Process Parameters from Retrospective Data. Pharm. Dev. Technol. 2016, 21, 688–697. [CrossRef]
98. Liou, J.Y.; Wang, H.Y.; Tsou, M.Y.; Chang, W.K.; Kuo, I.T.; Ting, C.K. Opioid and Propofol Pharmacodynamics Modeling during
Brain Mapping in Awake Craniotomy. J. Chin. Med. Assoc. 2019, 82, 390–395. [CrossRef]
99. Teng, W.N.; Tsou, M.Y.; Chen, P.T.; Liou, J.Y.; Yu, L.; Westenskow, D.R.; Ting, C.K. A Desflurane and Fentanyl Dosing Regimen
for Wake-up Testing during Scoliosis Surgery: Implications for the Time-Course of Emergence from Anesthesia. J. Formos. Med.
Assoc. 2017, 116, 606–612. [CrossRef]
100. Hubadillah, S.K.; Dzarfan Othman, M.H.; Harun, Z.; Ismail, A.F.; Iwamoto, Y.; Honda, S.; Rahman, M.A.; Jaafar, J.; Gani, P.;
Mohd Sokri, M.N. Effect of Fabrication Parameters on Physical Properties of Metakaolin-Based Ceramic Hollow Fibre Membrane
(CHFM). Ceram. Int. 2016, 42, 15547–15558. [CrossRef]
101. Chi, H.M.; Ersoy, O.K.; Moskowitz, H.; Altinkemer, K. Toward Automated Intelligent Manufacturing Systems (AIMS). INFORMS
J. Comput. 2007, 19, 302–312. [CrossRef]
102. Shin, S.J.; Woo, J.; Rachuri, S.; Meilanitasari, P. Standard Data-Based Predictive Modeling for Power Consumption in Turning
Machining. Sustainability 2018, 10, 598. [CrossRef]