TYPE Original Research

PUBLISHED 10 January 2023

DOI 10.3389/fceng.2022.1055744

Artificial intelligence models for

OPEN ACCESS yield efficiency optimization,
EDITED BY
Fausto Gallucci,
Eindhoven University of Technology,
prediction, and production
Netherlands

REVIEWED BY
scalability of essential oil extraction
Anping Wang,
Guizhou Normal University, China
Juan Gabriel Segovia Hernandez,
processes from citrus fruit
University of Guanajuato, Mexico

*CORRESPONDENCE
exocarps
Sandra E. Fajardo Muñoz ,
[email protected]
Sandra E. Fajardo Muñoz 1*, Anthony J. Freire Castro 1,
SPECIALTY SECTION
This article was submitted to
Michael I. Mejía Garzón 1, Galo J. Páez Fajardo 2 and
Computational Methods Galo J. Páez Gracia 1
in Chemical Engineering,
a section of the journal 1
Facultad de Ingeniería Química, Universidad de Guayaquil, Guayaquil, Ecuador, 2WMG, University of
Frontiers in Chemical Engineering Warwick, Coventry, United Kingdom
RECEIVED 28 September 2022
ACCEPTED 14 December 2022
PUBLISHED 10 January 2023

CITATION Introduction: Excessive demand, environmental problems, and shortages in market-

Fajardo Muñoz SE, Freire Castro AJ, leader countries have led the citrus (essential) oil market price to drift to
Mejía Garzón MI, Páez Fajardo GJ and
Páez Gracia GJ (2023), Artificial unprecedented high levels with negative implications for citrus oil-dependent
intelligence models for yield efficiency secondary industries. However, the high price conditions have promoted market
optimization, prediction, and production incentives for the incorporation of new small-scale suppliers as a short-term supply
scalability of essential oil extraction
processes from citrus fruit exocarps. solution for the market. Essential oil chemical extraction via steam distillation is a
Front. Chem. Eng. 4:1055744. valuable option for these new suppliers at a lab and small-scale production level.
doi: 10.3389/fceng.2022.1055744 Nevertheless, mass-scaling production requires prediction tools for better large-
COPYRIGHT scale control of outputs.
© 2023 Fajardo Muñoz, Freire Castro,
Mejía Garzón, Páez Fajardo and Páez Methods: This study provides an intelligent model based on a multi-layer perceptron
Gracia. This is an open-access article
(MLP) artificial neural network (ANN) for developing a highly reliable numerical
distributed under the terms of the Creative
Commons Attribution License (CC BY). dependency between the chemical extraction output from essential oil steam
The use, distribution or reproduction in distillation processes (output vector) and orange peel mass loading (input vector).
other forums is permitted, provided the
In a data pool of 25 extraction experiments, 14 output–input pairs were the in training
original author(s) and the copyright
owner(s) are credited and that the original set, 6 in the testing set, and 5 cross-compared the model’s accuracy with traditional
publication in this journal is cited, in numerical approaches.
accordance with accepted academic
practice. No use, distribution or Results and Discussion: After varying the number of nodes in the hidden layer, a
reproduction is permitted which does not
1–9–1 MLP topology best optimizes the statistical parameters (coefficient of
comply with these terms.
determination (R2) and mean square error) of the testing set, achieving a
precision of nearly 97.6%. Our model can capture non-linearity behavior when
scaling-up production output for mass production processes, thus providing a
viable answer for the scalability issue with a state-of-the-art computational tool
for planning, management, and mass production of citrus essential oils.

KEYWORDS

multi-layer perceptron neural network, orange peels, citrus oil, steam distillation,
optimization, extraction yield

Frontiers in Chemical Engineering 01 frontiersin.org

Fajardo Muñoz et al. 10.3389/fceng.2022.1055744

1 Introduction expense of prediction capability away from the local domain—that is,
extrapolation, a prediction feature required to enable the scalability
The essential oil market is a well-established industry with of the extraction processes.
interconnections with other industries, such as processed food, Artificial intelligence is a numerical approach that recognizes
perfume fragrance, and strategic sectors (Maurya et al., 2021; internal patterns in data to identify and classify data into larger
Sharmeen et al., 2021; Israfi et al., 2022; Navarra et al., 2015; Bora sets. Artificial intelligent algorithms can then handle non-linearity
et al., 2020). The world market has not yet recovered from the massive and complexity in data structure and predict outputs of numerical
collapse of commerce during the COVID-19 pandemic and still shows processes (Meuwly, 2021). Given the complex nature of the variables
signs of deceleration, with orange production dropping from the in chemical processes, artificial intelligence is a suitable tool for
principal producers worldwide (Brazil, Florida, and Spain) formulating prediction models in this field. Steam distillation
(United States Department of Agriculture, 2022). Consequently, extraction models with artificial neural network algorithms have
prices of citrus essential oils have reached unprecedented high achieved remarkable accuracy in yield prediction of oil extraction
levels (Technavio, 2022), making it attractive for newcomer from the soil (Daryasafar et al., 2014). Optimization of operating
producers to enter the market at least to cover local demand. The conditions in energy-intensive distillation processes can be explored
city of Las Naves in the Ecuadorian province of Bolivar is an attractive via a machine learning-based predictive model concluding with
hub for citrus fruit production, with ambitious signs of expansion. recommendations for optimal steam flow that minimize energy
Ecuador’s national institution for statistics and census (Instituto consumption and maximize production yield (Park et al., 2022).
Nacional de Estadística y Censos—INEC) and the Ecuadorian We focus on developing an artificial intelligence model with a tool
minister’s office for agriculture and livestock farming (Ministerio for steam distillation yield prediction of essential oil extraction from
de Agricultura y Ganadería—MAG) estimate a rise in orange orange peel.
production to 110.472 tons in upcoming years. At a local level (in Using state-of-the-art artificial intelligence algorithms, this
other provinces in Ecuador), citrus fruit from Las Naves attracts high work introduces a technique-independent predictive neural
demand and consumption, guaranteeing a vast source of citrus waste network model for essential oil extraction from orange peel. We
that could ultimately serve in secondary product production. present a multi-layer perceptron (MLP) artificial neural network
Essential oils are organic compounds that provide the (ANN) with supervised learning. The model topological
characteristic odor to citrus fruits and are a principal architecture achieves its best prediction via internal node
component of orange rind. The main component of citrus adjusting of the model structure, while statistical error
essential oil is limonene; its empirical formula (C 10H16) is a descriptors of the testing set are optimized. The MLP ANN
monoterpene that exhibits two D- and L-limonene optical handles input (orange peel mass) and output (essential oil mass)
isomers and the racemic combination known as dipentene data from steam distillation experiments. Steam distillation is the
(González-Mas et al., 2019). Limonene is present in more than most common technique at the laboratory level for essential oil
49 volatile organic compounds (90% classified as terpenoid esters), extraction (Ferhat et al., 2006) and works on the principle that oil
with concentrations depending on the fruit variety: 30%–40% in molecules diffuse when attached to water molecules in the vapor
bergamot, 40%–75% in lemon, and 68%–98% in sweet orange state (Chandler, 2002). The mean square error (MSE) and the
(Moufida and Marzouk, 2003). Most of its chemical and coefficient of determination (R2) are the error descriptors of choice
physical extraction processes rely on essential oil molecular to guide network optimization, as used elsewhere (Park et al.,
volatility (Balboa Laura, 2011), although the quality and 2022). A single hidden layer with nine neurons, back-
quantity of the essential oils in the fruit peel can impact the propagation, and neural weights adjusted using the
extraction efficiency levels independently of the extraction Levenberg–Marquardt algorithms optimizes the MSE and R2 of
method (García, 2014). Microwave-assisted hydrodiffusion the oil extraction model. This model approach paves the way for
stands out because it is solvent-free and highly efficient with a planning and designing oil extraction processes at larger scales
short extraction time (Bustamante et al., 2016; Bora et al., 2020). than in the laboratory.
However, for any extraction method, a highly accurate predictive
model capable of determining extraction efficiency is very
important for the optimization and design of scalable extraction 2 Methodology
pathways at various levels, such as in the laboratory or industry.
The wide set of variables and their non-linear and complex inter- 2.1 Sample collection
relations make predicting the extraction yield of essential oil a
challenging task. A lack of generality for mathematical models Local farmers from the Ecuadorian province of Bolívar in Las
also contributes to this difficulty, with models being extraction Naves city, 88 km northwest of Guaranda city, provided the Citrus
technique-dependent (Berna et al., 2000; Sovová, 2005; Lainez- sinensis L. raw peel material. This raw peel is waste from the
Cerón et al., 2022; El Ouaddari et al., 2022). A prediction model production processes in the local citrus farms. The oranges used in
for steam distillation can even become obsolete with a variation in production processes at local farms in Las Naves have a °Brix/
the essential oil source (Gawde et al., 2014). The inadequacy of a Acidity index of 7. “°Brix” is a measure of the dissolved sugar in
clear quality standard for a mathematical tool for predictions has led aqueous substances. One °Brix is traditionally defined as 1 g of
researchers to use polynomial, exponential, and logarithmic fitting sucrose in 100 g of water. “Acidity” is defined as titratable acidity,
tools to help at least with predictions in the locality (local space) of which measures the total acid concentration of food (also called
the data points (Zlatev and Shivacheva, 2018; Fakayode and Abobi, total acidity). The reader can refer to Jayasena and Cameron
2018). The local space of these basic fitting types comes at the (2008) for further reading. °Brix/Acidity is a classification step

Frontiers in Chemical Engineering 02 frontiersin.org

Fajardo Muñoz et al. 10.3389/fceng.2022.1055744

TABLE 1 Classification scheme to consistently spot similar peel features for oil
extraction experiments.

Parameters (units) Standard

pH 4.84

Total solid (%) 22.92

Humidity (%) 77.08

Reducing sugar (%) 4.42

taken by local farmers before they use the oranges for food
processing.
After we carefully transported the raw material to the city of
Guayaquil, a visual inspection selected peel material with undamaged FIGURE 1
surfaces for the extraction experiments. We used cold water and a Extracted oil in terms of mass loading showing an extraction rate
piece of soft tissue paper to wash off strange stains on the orange rind. faster than expected. The extraction rate scales at a non-linear rate
demonstrated by the first principal component in a PCA analysis.
The washed peel chunks went through a cut-down process to obtain
smaller regular pieces of approximately 1.5 × 1.1 cm2 to facilitate
essential oil extraction.

after the process is the total solid parameter (manual N° 925, 10,
2.2 Steam distillation AOAC, 1997), and the mass difference with respect to the initial mass
is the humidity measure.
The steam distillation equipment consisted of three main Potentiometric pH measurements were used to probe the acidity
parts: a Clevenger-type apparatus, a heat source, and a glass of the peel samples (pH). The tried-and-tested 780 pH meter by
spiral condenser. The glass tower consisted of two chambers: a Metrohm was employed. By comparing a known voltage with an
lower chamber containing 1250 L of distilled water and a unknown voltage according to the norm N° 981.12E, G, AOAC, 1997,
subsequent upper chamber containing the peel chunks. The the pH is calculated for the sample.
bottom chamber served as a steam source under heat. We used Spectrophotometry (PHARMACIA model Ultrospec 3000)
five loading mass points (sample weight) of orange peel at 200, measured the reducing sugar components of the orange peel in
300, 350, 400, and 500 g. Five steam distillation experimental collaboration with UBA-laboratories (Uba Lab, 2022). The
repetitions per mass point were the statistical baseline of the fundamentals of the principle can be found in Haldar et al.
artificial intelligence approach. This work performed a total of (2017).
25 distillation processes.
In each distillation experiment, the water vapor came from
heating the 1250 L of distilled water in the lower chamber at 120°C. 2.4 Computational methods
Vapor entered the upper chamber, separating the essential oil from
the chunks, given the molecular volatility of the desired A multi-layer perceptron (MLP) neural network algorithm is the
compounds. We later let it condense using the glass spiral artificial intelligence model of choice for our experiments due to its
condenser at 21°C in a mixture of water and essential oil compatibility with our data structure. MLP is a supervised learning
molecules. An elapse time of 30 min for distillation proved algorithm implemented and readily accessible in Matlab. A traditional
optimal, based on previous hydro-distillation studies model architecture consists of numerous hidden layers between the input
(Golmohammadi et al., 2018), to guarantee enough material and output layers, with some coded neurons per layer. We fixed the
after the first condensed drop. The Clevenger-type apparatus number of hidden layers to one with a sigmoid activation function
separated the resulting solution after condensation with the between layers. The optimization process of the single hidden layer’s
extracted essential oil at the top of a marked two-phase liquid architecture consisted of varying the internal number of neurons to
given by the oil’s lower relative density. optimize a group of error descriptors: the MSE (minimization) and R2
(maximization) (Park et al., 2022)
1 n exp pred 2
2.3 Orange peel characterization MSE

m − mi  , (1)
n i
1 i
exp pred 2
We followed the 930.15 method described by the Asociación de ni
1 mi − mi 
Químicos Analíticos Oficiales (AQAO, 2019) for humidity R2
1 − exp exp 2
, (2)
ni
1 mi
i 
−m
measurement. The underlying measurement principle was to detect
the weight lost due to evaporation when exposing the sample to 70°C respectively. At a given mass loading level, n is the number of data
exp
constant heating after a fixed amount of time. We used heating points in the mass loading set, mi the observed yield values, mpred the
intervals of 2 h at a fixed 70°C and measured the mass of the neuron network predicted value for the mass loading point, and m
exp
washed peel pieces before and after the process. The mass value is the mass loading set mean. Our model used back-propagation to

Frontiers in Chemical Engineering 03 frontiersin.org

Fajardo Muñoz et al. 10.3389/fceng.2022.1055744

TABLE 2 Extraction yield of each mass loading with five repetitions per set. Values in parenthesis are part of the t-distribution confidence interval at the last significant
figure.

Sample weight (g) Extracted oil (g) Yield (%) Mean yield (%)

A B 100*A/B
200 0.1240 0.0620 0.0623(8)

0.1245 0.0623

0.1234 0.0617

0.1265 0.0633

0.1255 0.0628

300 0.2935 0.0978 0.097(1)

0.2925 0.0975

0.2945 0.0982

0.2885 0.0962

0.2902 0.0967

350 0.4058 0.1159 0.116(6)

0.4214 0.1204

0.3981 0.1137

0.4296 0.1227

0.3721 0.1063

400 0.5834 0.1459 0.147(2)

0.5850 0.1463

0.5819 0.1455

0.5960 0.1490

0.5932 0.1483

500 0.8413 0.1683 0.168(1)

0.8315 0.1663

0.8417 0.1684

0.8412 0.1682

0.8342 0.1668

reinforce the model learning process, and we weighed the neuron 2.5 Principal component analysis
interconnection using the Levenberg–Marquardt algorithm
(Levenberg, 1944; Marquardt, 1963). Say we have a data matrix X, of n×p dimension, where the
The model required a training and testing data set. The MLP columns are the data features of interest, and the rows are the
algorithm processes the mass loading values as input and the essential number of measurements for each feature. A principal component
oil extraction yield values as output. Of the 20 extraction experiments analysis focuses on determining a reduced number of features that
(5 per each 200, 300, 400, and 500 g mass loading point), data (mass encapsulate most of the information (variance) of the data set. We
loading and extraction yield) from 70% (14) of the experiments comprised achieve this feature-dimensional-reducing property by computing
the training data set. The remaining 30% (from the other six experiments) the equation
corresponded to the neural network testing set. We also used data from
M
XV, (3)
the testing set to compute the MSE and R2 descriptor in every iteration of
the architectural restructuring. The 350 g set (data from five experiments where M is the reduced components representing most of the original
are 350 g of mass loading) benchmarked the model’s predictive information and V is the normalized eigenvector matrix of the
capabilities and compared it against traditional linear, logarithmic, and ^TX
variance matrix σ, σ
X ^ and T means the real matrix transpose
polynomial fitting curves. operation.

Frontiers in Chemical Engineering 04 frontiersin.org

Fajardo Muñoz et al. 10.3389/fceng.2022.1055744

TABLE 3 Neuron network prediction capabilities against other common models.

Model R2 Predicted yield for 350 g Error (%)

Lineal 0.9792 0.130 12.608

Logarithm 0.9795 0.125 8.031

Polynomial 0.9862 0.126 8.377

Neural network 0.9929 0.118 2.443

• (5.99646, −11.061, −0.27701, −8.0537, 1)

• (50.9817, 30.863, −0.515781, −4.28649, 1)

We chose V to be a column vector with the normalized eigenvector

for the eigenvalue 163.681, which is the first principal component.
With V defined, the matrix XTV reduced the number of
measurements to the one representing the most data so that we
could obtain a reliable representation of the data structure as a
function of mass loading. We referred to the reduced measurement
as the first component measurement.

3 Results and discussion

A homogeneous sample quality guaranteed enhanced statistical
FIGURE 2
confidence in the data and thus improved the model’s prediction
Results of the MLP neural network learning process with (A) the
optimization process of the fitting parameters MSE and R2 for the testing accuracy. We implemented a quality control scheme where, after the
set and (B) yield prediction for the verification set - 350 g. peel washing step described in Section 2, the regular peel chunks that
fell within a 5% confidence interval relative to the standards in Table 1
were selected for the extraction experiments. High humidity was
The matrix X ^ is obtained after subtracting the average value of a desirable so that an even molecular mixture (water + limonene oil
feature from its corresponding feature column. molecules in the orange peel) would boil at lower boiling points than
For our data set, the entries in matrix X are the extracted essential the separate liquids while steam flowed through the sample in the
oil percentages. The matrix is steam distillation experiments.
Figure 1 and Table 2 summarize the experimental results at 200 g,
0.062 0.0978 0.1159 0.1459 0.1683 300 g, 350 g, 400 g, and 500 g mass loading levels with the washed
⎜
⎛
⎜
⎜ 0.0623 0.0975 0.1204 0.1463 ⎟
0.1663 ⎞
⎟
⎟ regular chunks of orange rind after the extraction treatment. Yield
⎜
⎜ ⎟
⎟
X
⎜
⎜ 0.0617
⎜
⎜ 0.0982 0.1137 0.1455 ⎟
⎟,
0.1684 ⎟
⎟
⎟ (4) figures are comparable with literature reports for steam distillation of
⎜
⎜
⎝ 0.0633 ⎟
⎟
⎠
0.0962 0.1227 0.149 0.1682 essential oil from C. sinensis peel (Blanco Tirado et al., 1995; de Moraes
0.0628 0.0967 0.1063 0.1483 0.1668 Pultrini et al., 2006).
where the columns are, from left to right, the orange peel mass As the mass loading increases, the extracted essential oil mass should
loadings 200 g, 300 g, 350 g, 400 g, and 500 g. The rows are the five linearly augment in the distillation process as depicted in Figure 1.
experiments we performed per mass loading. However, the extraction yield scales up with the mass loading at a
After subtracting the average-extracted essential oil percentage for a faster rate than the linear trend, rendering a positive over-extraction.
given mass loading value from its corresponding column, we find that σ is This analysis is further supported by a principal analysis component
developed in Section 2.5, with data increasing non-linearly with respect to
1.628 −2.078 3.38 3.88 −0.49 the mass loading. This behavior makes the data structure ideal for non-
⎜
⎛
⎜
⎜ −2.078 2.668 −2.81 −5.02 0.63 ⎞⎟
⎟
⎟
⎜
⎜ ⎟
⎟ linear numerical processes such as in the MLP algorithm. A t-distribution
106 σ
⎜
⎜
⎜
⎜ 3.38 −2.81 163.44 1.27 4.15 ⎟⎟
⎟
⎟
⎟. (5)
⎜
⎜ ⎟
−0.9 ⎟
⎝ 3.88 ⎠ confidence interval analysis with 95% confidence and four degrees of
−5.02 1.27 9.64
−0.49 0.63 4.15 −0.9 3.82 freedom further supports the over-linear increase of the extraction yield.
The t-distribution estimations (Column 4 in Table 2) suggest a more
We introduce factor 106 to ease the matrix visualization for the reader. efficient extraction with sample weight.
The eigenvalues of σ are 163.681, 13.934, 3.569, 0.012, and 0 with their Sun et al. (2017) also reported an over-linear dependency between
respective eigenvectors. the extraction yield and sample weight. Although the reason behind
this dependency is beyond our current scope, we theorize that larger
• (0.8237, −0.695018, 38.801, 0.3574, 1) sample weights render an enhanced molecular diffusion across the
• (−2.6767, 3.46971, 0.15437, −6.63946, 1) orange peel–water interface. A larger loading mass occupies a larger
• (0.0000611, −0.013081, −0.027326, 0.143118, 1) space in the mass loading chamber, reducing the mixture’s boiling

Frontiers in Chemical Engineering 05 frontiersin.org

Fajardo Muñoz et al. 10.3389/fceng.2022.1055744

point (water vapor + essential oil molecules), thus promoting more a supervised learning approach to achieve an accuracy beyond the
material extraction. We will investigate this in more detail in standard predicting models in the literature. By simultaneously
subsequent research. optimizing two fitting parameters (R2 and MSE), we demonstrate
The MLP neural network shows promising results in Figure 2, that the MLP neural network achieves its optimal architecture with
predicting the extraction oil yield as a function of the neuron numbers one hidden layer and nine neurons, obtaining accuracy and predicting
for the 350-g sample set. Figure 2A shows the evolution of the error capabilities at R2 = 0.9929 and EMC% = 0.0040 for the validation
descriptors (MSE and R2) while the number of neurons increases in the sample set. The model developed in this work has the potential to
network’s hidden layer. R2 stabilizes after six neurons, reaching a unlock prediction capabilities in line with scalable design to bring lab-
plateau afterward, while MSE bounces during the neuron number based production prototypes to an industrial level, thus being a viable
increment. Figure 2B describes the evolution of the predicted values path for a temporary solution for the citrus oil market.
from the network given a denser hidden layer.
We observe a clear correlation between the MLP’s predictions in
Figure 2B and MSE in 2A. Numerical analysis demonstrates the MSE’s Data availability statement
correlation with the model predictions stems from the MSE’s sensitivity to
spiking values in the MLP model. We define a spiking value as mpred falling The original contributions presented in the study are included in
into the range of more than three sigma errors about m
exp . Conversely, the article/Supplementary Material; further inquiries can be directed
2
although R exhibits some spiking behavior at low neuron numbers to the corresponding author.
(Figure 2A), this error descriptor stabilizes as the neuron number
increases, given the weighing denominator in Equation 2. While the
neuron number increases, R2 compounds the effects of both the predicted Author contributions
and mean values of the sample weight sets, diluting any spiking behavior
from the predicting algorithm. The different sensitivity levels to the SF: supervision, writing—review and editing, project
statistical data of both R2 and MSE show the importance of utilizing administration, and conceptualization. AF: writing—original draft,
more than one error descriptor during the network outcome investigation, and formal analysis. MM: methodology. GPF: writing.
optimization. The optimization process demonstrates that the optimal GPG: validation, supervision, and writing—review and editing.
neuron number to use in the hidden layer is 9, based on two findings: a
stable R2 and MSE and predictive values within the mean yield confidence
interval for the 350 g set - (0.110, 0.122). Acknowledgments
The predictive capacity of the network outperforms other more
common mathematical models widely used in the literature (Zlatev We acknowledge the support of the chemistry department at the
and Shivacheva, 2018; Fakayode and Abobi, 2018), as observed in Universidad de Guayaquil in allowing us to publish this work.
Table 3, with the highest R2 for the MLP neuron network. Our MLP
model shows clear advantages over other models in accurately
predicting the extraction yield for the 350-g mass loading set Conflict of interest
located in the vicinity of the sample weight range (from 200 to
500 g). With larger training sets, MLP should accurately predict the The authors declare that the research was conducted in the
extraction values far from the training sample weight range. This absence of any commercial or financial relationships that could be
feature is relevant for planning and designing scalable industrial construed as a potential conflict of interest.
extraction processes and is a valuable approach for initiating the
leap from the laboratory to industrial scale-up.
Publisher’s note
4 Conclusion All claims expressed in this article are solely those of the
authors and do not necessarily represent those of their affiliated
We developed an innovative and alternative approach for an oil organizations, or those of the publisher, the editors, and the
extraction yield predictor using state-of-the-art artificial intelligence reviewers. Any product that may be evaluated in this article, or
algorithms. We use steam-distilled essential oil extraction data from claim that may be made by its manufacturer, is not guaranteed or
orange peel to train a multi-layer perceptron (MLP) neural network in endorsed by the publisher.

References
AQAO (2019). The AOAC 21st edition of the official methods of analysis available at Blanco Tirado, C., Stashenko, E., Combariza, M., and Martinez, J. (1995). Comparative study
official methods of analysis. Rockville, Maryland: AOAC International. of colombian citrus oils by high-resolution gas chromatography and gas chromatography-
mass spectrometry. J. Chromatogr. A 697, 501–513. doi:10.1016/0021-9673(94)00955-9
Balboa Laura, M. H. (2011). Obtención experimental de aceite esencial y subproducto a
partir de la cascara de naranja (citrus sinesis). La Paz, Bolivia: Bachelor’s thesis, Bora, H., Kamle, M., Mahato, D. K., Tiwari, P., and Kumar, P. (2020). Citrus essential
Universidad Mayor de San Andrés. oils (CEOs) and their applications in food: An overview. Plants (Basel, Switz. 9, 357. doi:10.
3390/plants9030357
Berna, A., Tárrega, A., Blasco, M., and Subirats, S. (2000). Supercritical CO2 extraction of
essential oil from orange peel; effect of the height of the bed. J. Supercrit. Fluids 18, Bustamante, J., van Stempvoort, S., García-Gallarreta, M., Houghton, J. A., Briers, H.
227–237. doi:10.1016/S0896-8446(00)00082-6 K., Budarin, V. L., et al. (2016). Microwave assisted hydro-distillation of essential oils

Frontiers in Chemical Engineering 06 frontiersin.org

Fajardo Muñoz et al. 10.3389/fceng.2022.1055744

from wet citrus peel waste. J. Clean. Prod. 137, 598–605. doi:10.1016/j.jclepro.2016. Lainez-Cerón, E., Ramírez-Corona, N., López-Malo, A., and Franco-Vega, A. (2022). An
07.108 overview of mathematical modeling for conventional and intensified processes for
extracting essential oils. Chem. Eng. Process. - Process Intensif. 178, 109032. doi:10.
Chandler, D. (2002). Hydrophobicity: Two faces of water. Nature 417, 491. doi:10.1038/
1016/j.cep.2022.109032
417491a
Levenberg, K. (1944). A method for the solution of certain non-linear problems in least
Daryasafar, A., Ahadi, A., and Kharrat, R. (2014). Modeling of steam distillation
squares. Q. Appl. Math. 2, 164–168. doi:10.1090/qam/10666
mechanism during steam injection process using artificial intelligence. Sci. World J.
2014, 1–8. doi:10.1155/2014/246589 Marquardt, D. W. (1963). An algorithm for least-squares estimation of nonlinear
parameters. J. Soc. Industrial Appl. Math. 11, 431–441. doi:10.1137/0111030
de Moraes Pultrini, A., Almeida Galindo, L., and Costa, M. (2006). Effects of the essential
oil from citrus aurantium l. in experimental anxiety models in mice. Life Sci. 78, Maurya, A., Prasad, J., Das, S., and Dwivedy, A. K. (2021). Essential oils and their
1720–1725. doi:10.1016/j.lfs.2005.08.004 application in food safety. Front. Sustain. Food Syst. 5. doi:10.3389/fsufs.2021.653420
El Ouaddari, A., El Amrani, A., Jamal Eddine, J., and Antonio Cayuela-Sánchez, J. Meuwly, M. (2021). Machine learning for chemical reactions. Chem. Rev. 121,
(2022). Rapid prediction of essential oils major components by vis/nirs models using 10218–10239. doi:10.1021/acs.chemrev.1c00033
compositional methods. Results Chem. 4, 100562. doi:10.1016/j.rechem.2022.100562
Moufida, S., and Marzouk, B. (2003). Biochemical characterization of blood orange,
Fakayode, O. A., and Abobi, K. E. (2018). Optimization of oil and pectin extraction from sweet orange, lemon, bergamot and bitter orange. Phytochemistry 62, 1283–1289. Reports
orange (citrus sinensis) peels: A response surface approach. J. Anal. Sci. Technol. 9, 20. on Structure Elucidation. doi:10.1016/S0031-9422(02)00631-3
doi:10.1186/s40543-018-0151-3
Navarra, M., Mannucci, C., Delbò, M., and Calapai, G. (2015). Citrus bergamia essential
Ferhat, M. A., Meklati, B. Y., Smadja, J., and Chemat, F. (2006). An improved microwave oil: From basic research to clinical application. Front. Pharmacol. 6, 36. doi:10.3389/fphar.
clevenger apparatus for distillation of essential oils from orange peel. J. Chromatogr. 2015.00036
APlant Analysis 1112, 121–126. doi:10.1016/j.chroma.2005.12.030
Park, H., Kwon, H., Cho, H., and Kim, J. (2022). A framework for energy optimization of
García, R. T. (2014). Obtención de aceite esencial de citronela (cymbopogon winterianus) distillation process using machine learning-based predictive model. Energy Sci. Eng. 10,
extraído por arrastre con vapor a escala piloto: Estudio de la influencia de variables en el 1913–1924. doi:10.1002/ese3.1134
rendimiento y la calidad del aceite. Master’s thesis, Universidad Tecnológica Nacional -
Sharmeen, J. B., Mahomoodally, F. M., Zengin, G., and Maggi, F. (2021). Essential oils as
Facultad Regional Resistencia.
natural sources of fragrance compounds for cosmetics and cosmeceuticals. Molecules 26,
Gawde, A., Cantrell, C. L., Zheljazkov, V. D., Astatkie, T., and Schlegel, V. (2014). Steam 666. doi:10.3390/molecules26030666
distillation extraction kinetics regression models to predict essential oil yield, composition,
Sovová, H. (2005). Mathematical model for supercritical fluid extraction of natural
and bioactivity of chamomile oil. Industrial Crops Prod. 58, 61–67. doi:10.1016/j.indcrop.
products and extraction curve evaluation. J. Supercrit. Fluids 33, 35–52. doi:10.1016/j.
2014.04.001
supflu.2004.03.005
Golmohammadi, M., Borghei, A., Zenouzi, A., Ashrafi, N., and Taherzadeh, M. J. (2018).
Sun, S.-H., Chai, G.-B., Li, P., Xie, J.-P., and Su, Y. (2017). Steam distillation/drop-by-
Optimization of essential oil extraction from orange peels using steam explosion. Heliyon
drop extraction with gas chromatography–mass spectrometry for fast determination of
4, e00893. doi:10.1016/j.heliyon.2018.e00893
volatile components in jujube (ziziphus jujuba mill.) extract. Chem. Central J. 11, 101.
González-Mas, M. C., Rambla, J. L., López-Gresa, M. P., Blázquez, M. A., and Granell, A. doi:10.1186/s13065-017-0329-6
(2019). Volatile compounds in citrus essential oils: A comprehensive review. Front. Plant
Technavio (2022). Citrus oils market by application and geography - forecast and analysis
Sci. 10, 12. doi:10.3389/fpls.2019.00012
2022-2026. Elmhurst, IL, United States: Technavio.
Haldar, D., Sen, D., and Gayen, K. (2017). Development of spectrophotometric method
Uba Lab (2022). Further information about uba-lab. Guayaquil, Ecuador: Analytical
for the analysis of multi-component carbohydrate mixture of different moieties. Appl.
Laboratories Testing & Consulting. Available at: https://www.uba-lab.com/.
Biochem. Biotechnol. 181, 1416–1434. doi:10.1007/s12010-016-2293-3
Israfi, N. A. M., Ali, M. I. A. M., Manickam, S., Sun, X., Goh, B. H., Tang, S. Y., et al. United States Department of Agriculture (2022). Citrus: World markets and trade.
(2022). Essential oils and plant extracts for tropical fruits protection: From farm to table. Washington, United States: United States Department of Agriculture, Foreign Agricultural
Front. Plant Sci. 13, 999270. doi:10.3389/fpls.2022.999270 Service.
Jayasena, V., and Cameron, I. (2008). °Brix/Acid ratio as A predictor of consumer Zlatev, Z., and Shivacheva, G. (2018). Yield analysis of essential oils extracted by
acceptability of crimson seedless table grapes. J. Food Qual. 31, 736–750. doi:10.1111/j. steam distillation. Appl. Res. Tech. Technol. Educ. 6, 300–305. doi:10.15547/artte.
1745-4557.2008.00231.x 2018.04.003

Frontiers in Chemical Engineering 07 frontiersin.org