ES Energy Environ.

, 2020, 9, 1-14

ES Energy & Environment

DOI: https://dx.doi.org/10.30919/esee8c795

Progress in the Application of Machine Learning in Combustion

Studies
Zhi-Hao Zheng,1, 2, † Xiao-Dong Lin,1, 2, † Ming Yang,1, 2, † Ze-Ming He,1, 2 Ergude Bao,3,* Hang Zhang1, 2,*
and Zhen-Yu Tian1, 2,*

Abstract
Combustion is the main source of energy and environmental pollution. The objective of the combustion study is to improve
combustion efficiency and to reduce pollution emissions. In the past decades, machine learning (ML), as a branch of artificial
intelligence, has attracted increasing interests, especially in the combustion field. In the present work, the definition, current
status and recent progress in the applications of ML on researches related to combustion are briefly reviewed. Combustion
studies combined with ML can be divided into theoretical and industrial aspects. Studies of combustion theory include
computational fluid dynamics (CFD) simulation, combustion phenomenon and fuel. ML is used to reduce the cost of CFD,
including reducing the scale of combustion mechanism, saving the memory storage of the probability density function table
and optimizing Large Eddy Simulation. ML helps in the research of combustion phenomena, such as detecting thermoacoustic
combustion oscillation, portioning regimes of ignition and detonation, and reconstructing cellular surface of gaseous
detonation. ML has been also applied to study physicochemical properties of fuels and to design the next generation fuels.
In the industrial research with respect to combustion, ML is mainly applied to produce electricity and power by power plants
or engines, and less to other fields. ML could figure out problems of combustion in various kinds of furnaces and post-
combustion emissions in power plants. In addition, ML plays important roles in biodiesel engine, Homogenous Charge
Compression Ignition (HCCI), and operation control or monitoring in the engines. Moreover, ML can also be applied to other
industrial studies related to combustion, mainly to particulate matters. The methods of the mentioned studies are
summarized in details and the potential applications of ML in combustion community are proposed.

Keywords: Combustion; Machine learning; Industry; Combustion theory.

Received: 01 September 2020; Accepted: 18 September 2020.
Article type: Review article.

1. Introduction the total installed power capacity.[2] Nonetheless, as a main

As the main approach to getting energy, the combustion of fuel method to supply the most energy that people desire,
is commonly applied to the industry, power generation, combustion of fossil fuels is responsible for pollutions, which
heating supply, waste disposal and related fields. About 80 % greatly endanger ecological and people’s living environments.
of world energy comes from combustion of fossil fuels.[1] Amount of CO2 causes environmental problems that threaten
Combustion contributes a lot to the electric power. For human survival, such as the greenhouse effect, sea-level rise,
instance, the installed thermal power capacity of China in global warming and etc. The incomplete combustion of fossil
2019 is about 1.19 billion kilowatts, accounting for 59 % of fuels will produce pollutants like nitrogen oxides (NOX),
sulfur oxides (SOX), volatile organic compounds (VOCs),
polycyclic aromatic hydrocarbons (PAHs) and soot. Efficient
1Institute of Engineering Thermophysics, Chinese Academy of and clean combustion plays an important role in relieving
Sciences, Beijing 100190, China energy pressure, ensuring energy security, lowering emissions
2University of Chinese Academy of Sciences, Beijing 100049, China
and keeping sustainable development. The goals of the
3School of Software Engineering, Beijing Jiaotong University,
combustion study are to improve combustion efficiency and
Beijing100044, China reduce emissions. Despite vast applications of combustion in
*E-mail: [email protected] (E. Bao), [email protected] (H. Zhang),
industrial production, the combustion study still remains a lot
[email protected] (Z. Tian). of gaps. As combustion is a kind of flowing oxidation
† Authors contributed equally.
reactions, it needs to couple the fluid dynamics, chemical

© Engineered Science Publisher LLC 2020 ES Energy Environ., 2020, 9, 1-15 | 1

Review article ES Energy & Environment

kinetics, heat and mass transfer, it also requires powerful give feedback on those classification behaviors. Common
methods to investigate. algorithms include clustering, hierarchical clustering, k-means,
Machine learning (ML), which is subject to constructing principal component analysis (PCA), local principal
models from existing data, is a branch of Artificial Intelligence component analysis (LPCA) and association rules.
(AI) and has been applied in many other research fields. For Semi-supervised learning is a two-stage or multi-stage
example, many scholars have used it in the design of process that uses supervised (or unsupervised) learning as pre-
functional thermal materials and achieved good results.[3] With processing for unsupervised (or supervised) learning.[17] Its
ML, researchers are able to predict the thermal conductivity or main goal is to use a small number of labeled samples and a
interface thermal conductivity of various materials,[4-6] large number of unlabeled samples for training and
discover the abnormal thermal properties of different classification, so as to reduce the cost of labeling and improve
nanomaterials[7,8] and modulate the thermal conductivity of learning ability. Semi-supervised learning is applied in solving
aperiodic superlattices.[9] Mitchell et al.[10] defined ML as: “A both classification and regression problems. Its algorithm
computer program is said to learn from experience E with includes some extension forms of commonly-used supervised
respect to some class of tasks T and performance measure P, if learning algorithms, like graph inference and Laplacian SVM.
its performance at tasks in T, as measured by P, improves with RL, different from supervised learning and unsupervised
experience E.” The definition has been accepted by the learning, does not require any data in advance. RL obtains
academic world. The dependence on experience may suggest learning information and updates model parameters by
a shortage of ML. For example, systematic biases in society receiving feedback that includes strengthening effective rules
will lead to biases in data, which will raise questions and make and weakening invalid or bad rules from interactions between
unfair judgements[11] when using ML in policy decision- the RL system and the environment. Dynamic systems and
making.[12] Domingos et al.[13] believed that learning = robot control are the common application fields of RL. Typical
representation + evaluation + optimization. The purpose of algorithms of RL include Q-Learning and temporal difference
representation is to express the model with the language that learning.
the computer can handle. More than one classifier will be In particular, as a new research direction in the field of ML,
represented, so the evaluation and optimization functions are deep learning (DL), one of the fastest-growing learning
needed to score the models and search for high-scoring method,[18] has received a lot of research attention due to the
classifiers.[13] ML is shifting our power[11] in social and substantial improvement in computing power in recent years.
economic life, whose influence is considered in policy- It is closer to the original goal of ML-AI, and has been widely
making.[14] used in related fields including data mining and natural
According to the learning methods, ML can be categorized language processing (NPL). Here is a brief introduction to
into supervised learning, unsupervised learning, semi- several common models of DL. Convolutional neural network
supervised learning, reinforcement learning (RL), and etc.[15] (CNN) is one of the most widely-used models for DL, with
The respective characteristics of these methods are:1) applications in image recognition, recommender systems and
supervised learning needs to provide labeled sample sets; 2) NPL. CNN is a type of feedforward neural network (NN) that
unsupervised learning does not need to provide labeled sample includes convolution calculation and has a deep structure. Its
sets; 3) semi-supervised learning needs to provide a small weight-sharing network structure makes it more similar to a
number of labeled samples; and 4) RL requires a feedback biological NN, reducing the complexity of the network model
mechanism. Supervised learning[15] is to use a finite set of and reducing the number of weights. This advantage allows
labeled training data to build a model through a certain the image to be directly used as the input of the network,
learning strategy to achieve labeling of new instances. avoiding the complex feature extraction and data
Supervised learning requires that the classification labels of reconstruction processes in traditional recognition algorithms,
the training samples are known. The higher the precision of which is conducive to the rapid learning of NNs and reduces
the classification labels and the more representative the the memory overhead during calculation.[19]
samples are, the higher the accuracy of the learning model In order to meet the needs of modeling in changes of time
predictive for previously unseen data[16] we can get. sequence during NPL and Automatic Speech Recognition
Supervised learning, a powerful set of tools that have (ASR), Recurrent Neural Network (RNN) appeared. RNN is a
successfully been used in applications spanning the themes of kind of recursive NN that takes sequence data as input and
automation, modeling and discovery,[16] is often used for performs recursion in the evolution direction of the sequence
classification and regression, with typical algorithms like and all nodes (recycling units) are connected in a chain.[20]
Decision Tree, random forest, Support Vector Machine RNN has memory, parameter sharing and turing completeness,
(SVM), Naïve Bayes, Artificial Neural Network (ANN), and so it has certain advantages when learning the nonlinear
so on.[15] Unsupervised learning uses unlabeled finite data to characteristics of the sequence. The general RNN has a long-
describe the inherent laws of unlabeled data. Instead of telling term dependence problem, and the ML field has developed a
the computer how to classify data set, unsupervised learning long-short term memory unit (LSTM), which is a time-
makes the computer learn to classify data by itself, and then recursive NN that realizes the memory function in time

2 | ES Energy Environ., 2020, 9, 1-14 © Engineered Science Publisher LLC 2020

ES Energy & Environment Review article

through the switch of the gate. Due to its unique design governed by highly non-linear equations and difficult to be
structure, LSTM is suitable for processing and predicting predicted. Some scholars have used ML algorithms to solve
important events with very long intervals and delays in time problems involving combustion in CFD simulation.[23-25]
series, and has been well applied in the fields of hand writing Detailed kinetic model is the key factor to simulate
recognition (HWR) and ASR. laminar and turbulent flames. The computing cost of coupling
detailed mechanism with CFD is too high, because detailed
models contain too many species and elementary reactions,
which recall the simplification of kinetic models. However, to
simulate the entire flow field with a sole simplified mechanism
may get a result that is significantly different from the one
obtained from a detailed mechanism. D'Alessio et al.[23]
applied ML to simulate flames, and solve the conflict with
Sample-Partitioning Adaptive Reduced Chemistry (SPARC)
method. The core of such method is to partition the data-space
obtained from detailed simulations in different clusters and to
generate a special reduced mechanism for each cluster
respectively through Directed Relation Graph with Error
Fig. 1 The variation of paper number about combustion and ML.
Propagation. The last step of SPARC is the CFD simulation of
ML has been used in several studies related to combustion. a reacting system. LPCA was used in the second step to
This work reviewed the current state of ML in combustion partition the datasets into several clusters. The feasibility of
research, especially the recent five years. As can be seen in Fig. the method was proven by the successful simulation of co-
1, the application of ML in combustion field is attracting more flow diffusion methane flame under both steady and dynamic
and more attention. It should be noted that the paper number conditions. The CPU time was significantly reduced while the
published in 2020 was taken into account up to August. The precision of the result almost remained. Compared to the
studies of ML in combustion will be divided into theoretical calculation cost for transport step, chemistry was not the
studies and industrial ones. The future development of ML in biggest computational burden in the flame simulation due to
combustion will be prospected in the end. the reduction.
PDF is a common method to simulate turbulent
2. Theoretical research combustion, which needs to tabulate parameters like
As mentioned above, ML has the capacity to find patterns in temperature, density and mass fraction. The traditional way is
large sets of data that are incompressible to experimental to perform linear interpolation on a table involving such
practitioner,[21] especially incomplete or uncertain data sets.[22] parameters. However, it will consume a lot of memory despite
ML seems to be able to solve problems where the of its simplicity. ANN can correlate input with output, as
dependencies are nonlinear relationships between inputs and shown in Fig. 2. Nonetheless, previous ANN tabulation
outputs or insufficiently known to be programmed in a studies need too much training time and have too many
traditional and rigid manner.[22] The advantages make ML play nonlinear activation functions to solve, which enlarge the
an important role in studies related to combustion theory, such computational costs and time, offsetting the advantage of
as Computational Fluid Dynamics (CFD) simulation of memory saving. In combination with multi-layer perception
combustion,[23-25] prediction of combustion phenomenon[26-29] (MLP) neural networks for regression and self-organizing
and fuels.[30-35] ML can help to reduce mechanism scales,[23] maps (SOM) for clustering data, Ranade et al.[24] developed a
optimize probability density function (PDF)[24] and Large small and compact ANN, which is more accurate and cost-
Eddy Simulation (LES)[25] method in combustion simulation. saving in training and computation compared to previous
When investigating combustion phenomenon, ML algorithms ANN. The advantage of MLP-SOM is to find optimum
were used to detect thermoacoustic combustion number of neurons and subsequently construct a simpler
oscillations,[26,27] distinguish donation and ignition[28] and network that can reduce training and interpolation time. In
reconstruct the detonation surface.[29] In various fuel studies, order to compare MLP-SOM approach with traditional
ML also played an important role in the investigations about interpolation in solution accuracy and computation time, the
fuel properties,[30-32] the bio-[33,34] and next generation[35] fuels. Reynolds-Averaged Navier-Stokes Models and LES are used
to simulate a turbulent jet diffusion flame which is called
2.1 CFD simulation DLR-A.[36] A great agreement is achieved between the linear
With the rapid increase of computational resources and data PDF interpolation and MLP-SOM based ANN according to
storage capabilities, ML has had a great impact on various the simulation result. Compared to linear PDF interpolation
fields such as speech, image and text recognition, robotics, approach, MLP-SOM method owns the advantage of easily
health-care and many others. Its ability to learn from data can accommodating more parameterized quantities due to 1000
be used to simulate turbulent and reacting flow, which is times smaller memory requirements.

© Engineered Science Publisher LLC 2020 ES Energy. Environ., 2020, 9, 1-14 | 3

Review article ES Energy & Environment

chemistry and ML are delivering impacts in new catalysis

design[22] and reaction prediction.[39] Applying ML to chemical
kinetics may construct a perfect mechanism that can
accurately predict all species profiles.

2.2 Prediction of combustion phenomenon

ML has severely impacted almost all scientific and
engineering subjects, and has great potentials to build a new
combustion control system. Therefore, it has been used to
study combustion phenomena.

Fig. 2 Schematic diagram of MLP-SOM based ANN (adapted

from Ranade et al.[24]).
Seltz et al.[25] used a direct numerical simulation (DNS)
database of a turbulent premixed stoichiometric methane/air
jet flame to train CNN to predict both the filtered progress
variable source term and the unresolved scalar transport terms.
Then LES method was used to numerically simulate turbulent
Fig. 3 first-second components feature space borrowed from
premixed stoichiometric methane/air flame. The simulation
Hachijo et al..[26] (The reuse permission was granted through
results proved that the CNN could provide quantitatively
Rights Link)
accurate predictions of both the source and flux terms. The
predictive capabilities networks are weakly sensitive to Kobayashi et al.[27] and Hachijo et al.[26] combined
variations in filter width, which is an important attribute for statistical complexity and ML for detecting thermoacoustic
any sub-grid scale model. combustion oscillations to avoid it. By combining PCA to
Via using CNN, researchers[25] are allowed to establish a estimate the principal component plane and SVM as a binary
unified modelling framework for all unresolved terms classifier, a two-dimensional feature space was made to
including source and scalar transport terms in the filtered predict thermoacoustic combustion oscillations, as shown in
progress variable transport equation in LES simulations of Fig. 3.[26] The complexity-entropy causality plane (CECP) is
turbulent premixed flames. Using a database from DNS of a another promising way to evaluate complexity in time series
turbulent premixed stoichiometric methane/air jet flame, CNN data, potentially useful in detecting a precursor of combustion
is well trained and shows quantitatively good predictions of all oscillations. The training data samples were divided into three
unresolved terms in a priori study, which is in the context of clusters by k-means clustering, which can minimize the
flame tabulated chemistry (premixed flamelet). distance between noncentral data points and the centroid of
Combustion involved into the coupling of many subjects each cluster with an SVM classifier.[27] The result turned out
like fluid dynamics, chemical kinetics, heat and mass transfer that the CECP feature space constructed by k-means clustering
and so on. Chemical kinetics plays an important role in deeply and SVM was an effective way to capture the subtle changes
understanding the mechanism of combustion, even the during a transition from low-amplitude combustion noise state
formation mechanism of astrochemical species like interstellar to large-amplitude combustion oscillations state, as shown in
enols.[37] So far, the application of ML in combustion theory Fig. 4.
study has only involved the reduction of chemical kinetic Grogan et al.[28] used ML algorithm to estimate the
model and optimization of CFD simulation, and has not been predictive capability of different combustion parameters like
used to develop chemical kinetic models. Only constructing an ignition delay time, heat release, chemical sensitivity and etc.
accurate mechanism, can we correctly simulate combustion under different facilities, fuels and configurations. By means
phenomenon in combination with CFD? The process of of experimental and simulation data, an ML method which is
constructing a chemical kinetic model needs to adjust the logistic regression was utilized to demarcate regime transition
number of chemical parameters, which can be gotten by in shock tubes and Rapid Compression Machines (RCMs) to
solving Schrodinger equation. However, the size of chemical distinguish detonation and ignition. The decision boundary
space is still overwhelming to model molecules and smart between ignition and detonation in acoustic residence-
navigation is required, therefore ML can play an important chemical resonance plane is obtained. Moreover, it is found
role because the computational strategies automatically that the increased chemical sensitivity has a regularizing effect
improve through experience.[38] Advances in computational when plotted against the heat release rate. The classification

4 | ES Energy Environ., 2020, 9, 1-14 © Engineered Science Publisher LLC 2020

ES Energy & Environment Review article

considered the influence of combustion devices and the error for property prediction, seeking to learn a function that maps
is reduced to 11.3 %. However, the technique perhaps cannot a molecular material to the property of choice.[38] Wang et al.[30]
be used to guide further theoretical ignition analysis for lack built the correlations between mid-Infrared Fourier Transform
of the most crucial physics. Infrared (mid-IR FTIR) absorption spectra and physical and
chemical properties for 64 hydrocarbon fuels. The least
absolute shrinkage and selection operator (LASSO)
regularized linear model, which takes FTIR spectra as input
parameters, was developed to accurately predict the physical
and chemical properties of fuel. Saldana et al.[31] used the ML
method to predict melting points and net heat of combustion
of fuels like hydrocarbons, alcohols and esters. The consensus
model was developed on the basis of molecular descriptors
and functional group count descriptors with the use of various
statistical algorithms. The algorithms include Genetic
Function Approximation (GFA), Partial Least Squares (PLS)
regressions, Feed Forward Artificial Neural Networks
Fig. 4 CECP feature space borrowed from Kobayashi et al..[27] (FFANN), general regression neural networks (GRNN), SVM
(The reuse permission was granted by American Physical Society) and Graph Machines (GM), as shown in Fig. 5. The predicted
In order to extract the main features of the flow field and values of the model agree well with experimental ones. ML
eventually reconstruct the detonation surface, Zhang et al.[29] could be a facile and efficient way to design the next
combined the proper orthogonal decomposition (POD) with generation fuels, which was utilized by Li et al.[32] to conduct
ANN to achieve the goals, instead of inconvenient a proof-of-concept study for discovering high energy density
experimental measuring. The POD method decomposed the hydrocarbon (HEDH) fuels. The model based on ML, as
detonation flow fields were into several basic features along shown in Fig. 6, identified 28 hydrocarbon molecules with
with time-varying coefficients, which were subsequently high density, high net heat of combustion, high specific
reduced to release the designing and training process of ANN. impulse and low melting point from 319, 895 hydrocarbon
ANN was key to reach the eventual goal by establishing the molecules. The result turned out that this high-throughput
mapping relationship between the reduced coefficients screening strategy was a promising technique to search or
corresponding to full flow field and those of the post-surface design new fuels.
subfield, whose feasibility was verified by successfully
reconstructing the cellular surface of both normal and unstable
oblique detonations.
ML can be used to other studies of combustion
phenomenon in the future, such as the negative temperature
coefficient (NTC) effect, cool flame, and production of soot.
NTC effect is a phenomenon that the fuel and ignition delay
time increase while immediate species, products and burning
velocity decrease with increasing the temperature during a
specific temperature span, which emerges in the oxidation of
linear alkanes. A well-behaved kinetic model can accurately
predict the NTC effect. Cool flame is a kind of flame emerged
at a temperature of several hundred K lower than auto-ignition
temperature. The formation of soot is believed to be relevant
to PAHs, whose profile is hard to be accurately predicted by
the existing mechanism. The aforementioned phenomena can
be investigated if ML could be used to establish chemical
kinetic models.
Fig. 5 Schematic representation of the methodology followed for
the development of the consensus models (adapted from Saldana
2.3 Fuel
et al.[31]).
In the theoretical studies of fuels, there are cases of studying
fuels properties and designing the next generation fuels. The In the theoretical studies of biofuels, ML was used to
compositions of jet fuels are too complex and it’s difficult to discriminate the biomass pellets[33] and predict activation
study physical and chemical properties of fuels by energy.[34] Liu et al.[33] developed models to discriminate
constructing physical models. In the context of materials, ML biomass pellets in combination with Laser-induced
techniques, a new tool in synthetic chemistry,[40] are often used breakdown spectroscopy (LIBS) technique and several ML

© Engineered Science Publisher LLC 2020 ES Energy. Environ., 2020, 9, 1-14 | 5

Review article ES Energy & Environment

algorithms including PCA, SVM, extreme learning machine chemical reaction kinetics processes and other features. ML,
(ELM) and so on, and tested the new technique by has been successfully used to predict fuel properties[30,31] and
discriminating ice husk, redwood, pine wood and rubber wood. design the next generation fuels,[32] perhaps will be applied to
The accurate discrimination indicated that the newly built construct surrogate fuels.
approach of LIBS combined with chemometrics methods was
reliable. Compared to the five chosen algorithms, the radial
basis function neural network model was the most reliable one
and used to develop a visual pseudo-color image to
discriminate biomass pellets intuitively with different colors.
The study would benefit for screening out biomass pellets with
good combustion performance to improve the energy
utilization. Jiang et al.[34] studied combustion activation energy
of alkali-catalyzed straw pyrolytic carbon. The kinds of straw
utilized in reference[34] are wheat straw, corn straw and
sorghum straw. The catalysts are NaOH, KOH and their blends.
The algorithms include Linear Regression (LR), Support
Fig. 7 The flow chart of the programming (adapted from Jiang et
Vector Regression (SVR) and Random Forest Regression
al.[34]).
(RFR). Fig. 7 is the flow chart of programming. The realistic
activation energy was analyzed by thermogravimetric analysis.
3. Industrial research
The result indicated that it is feasible to apply ML to predict
Compared to theoretical studies, there are more industrial ones
the combustion activation energy. The RFR model was the
using ML algorithms. This section will introduce application
most suitable model to predict combustion activation energy
of ML in industrial studies related to combustion, including
of alkali-catalyzed straw pyrolytic carbon among the models
power plant,[41-45] engine,[46-54] fuel[55-58] and other fields.[59-64] To
in reference.[34] Besides the application in biofuels, ML also
reach the goal of efficient and clean combustion in electric
plays an important role in the combustion study of biowaste.
generation, ML was used to classify coals,[41] monitor
Wen et al.[35] experimentally quantified the co-combustion/
combustion process[42] and treat emissions[43-45] in power plants.
pyrolysis performances of textile dyeing sludge and incense
Biodiesel engine[46-48] HCCI combustion[49,50] and improving
sticks in combination with thermogravimetric analysis and
operational methods[51-54] are three branches of engine design
regression ML model under different atmosphere, blend ratio,
where ML made contributions. Industrial studies about fuel,
heating rate and temperature. The operational conditions were
such as coal fuel[55,56] and predictions of cetane number (CN)[57]
optimized by using Box Behnken design and general linear
and octane number,[58] also utilized ML algorithms. Besides,
model to identify the significant interactions among blend
ML was also applied to other studies, including gas turbine,[59]
ratio, heating rate, atmosphere and temperature for different
particulate matters,[60-62] organic Rankine cycle[63] and
thermogravimetric measurements. The two most important
hydrogen array sensor detection.[64]
predictors for every measuring method were found based on
Random Forest model.
3.1 Power plant
The boiler is one of the most crucial devices in power plants,
with a responsibility to transfer heat from flame to working
fluid. ML was mainly applied to the combustion of furnace
and pollutant treatment. The previous studies include fuel
classification[41] based on flame characteristics and control of
combustion processes,[42] while the latter ones focus on the
treatment of CO2[43] and NOX.[44,45]
Annadurai et al.[41] utilized ML such as Trees, Random
Forest, Naïve Bayesian, SVM, ANN and Existing algorithm
Fig. 6 Schematic diagram for discovering new fuels via ML
adopted by Bharat Heavy Electrics Limited, to distinguish fuel
enabled high-throughput screening (adapted from Li et al.[32]).
types by flicker frequency range. It was challenging to
Besides the mentioned above, ML may also be used to correctly predict fuels due to the ambiguity and uncertainty
construct surrogate fuels. Real fuels are composed of several existing in the algorithms. Annadurai developed a new model
hundred compounds. In order to investigate combustion that is Time Series Analysis with Fuzzy Support Vector
chemical kinetics, researchers usually choose to blend some Machine (TSA-FSVM), as shown in Fig. 8. The performance
representative compounds to get a surrogate fuel similar to real criteria like accuracy, Micro-Precision, Micro-Recall and
one. When constructing surrogate fuels, researchers need to Micro-F-Measure were compared among the algorithms, and
consider physical and chemical properties of real fuels, the values of TSA-FSVM are the highest. TSA-FSVM

6 | ES Energy Environ., 2020, 9, 1-14 © Engineered Science Publisher LLC 2020

ES Energy & Environment Review article

overcome ambiguity and will play an important role in LSTM, RNN and Least Squares Support Vector Machine
generating electricity more efficiently and conserving energy. (LSSVM) to build and compare three NOX emission models.
The optimization algorithms like Stochastic Gradient Descent,
Root Mean Square prop and Adaptive Moment Estimation
were used in the LSTM model to improve the network
performance. Comparison results of the three NOX emission
models proved the superiority of LSTM model in solving the
long term memory problem. Selective catalyst reaction (SCR)
system is a common and expensive approach to NOX emission
control in coal-fire boiler, which is negative to economic
benefit of the power plant. To balance the unit thermal
efficiency, SCR reagent consumption and NOX emissions, Si[45]
implemented an integrated combustion optimization scheme
by genetic algorithm and accurate on-line SVR to achieve the
Fig. 8 Schematic diagram of TSA-FSVM (adapted from Xu et goal of multi-objective optimization. The relationships
al.[41]). between process controllable variables and optimization
targets and targets were investigated by means of the scheme
via the field case study performed at a 160 MW coal-fired unit.
The validity of the integrated combustion optimization
approach was proven by the field case study, which obtained
the feasible optimal solutions in the first stage and achieved
the lowest operation cost in the second stage.
The boiler is a device that transfers the chemical energy of
fuel into internal energy of the working fluid. There may be
catastrophe like heat transfer deterioration during the transfer
of heat in flame into steam. An ML model can be trained by
the combination of combustion parameters and steam
parameters to predict, monitor and control the operation of
Fig. 9 Bootstrap aggregated ELM (adapted from Bai et al.[43]).
boiler. The relevant research is still lacking up to now. Apart
To control the combustion processes in the furnace and from NOX and CO2, there are other pollutants like CO, SOx
minimize the incomplete combustion heat loss. Rahmanda et VOCs and soot, which can be abated by the appliance of ML
al.[42] optimized the control and monitoring system in Furnace in the future. For instance, as an effective way to abate CO and
Boiler Prototype at Industrial Steam Power Plant by means of VOCs,[66-69] catalyst can be facilely synthesized with the help
ELM and NN. The result indicated that ELM is better than NN, of ML.
the former one has less training time and average errors.
ML was also applied to the studies of post-combustion 3.2 Engine
pollutants. As shown in Fig. 9, Bai et al.[43] built a bootstrap Engine is a kind of important power equipment. In order to
aggregated ELM model consisting of many ELM models, solve the problems caused by consumption of fossil fuels, ML
which were developed on bootstrap re-sampling replication of has been used to many studies about engine like diesel engine
the original training data to predict CO2 capture rate and level. consuming biodiesel,[46-48] Homogeneous Charge Compression
Choosing the mean square error of the prediction results of Ignition (HCCI)[49,50] and the operation of engine.[51-54]
different hidden neurons as a comparing criterion, the Biodiesel is an effective way to reduce pollution and
bootstrap aggregated ELM model excelled a single ELM enlarge energy source. ML has been used to study engines
model in providing more accurate and reliable predictions and consuming biodiesel. Aghbashlo et al.[46] designed a method
model prediction confidence bounds. The training data (50 %), based on Extreme Learning Machine with Wavelet Transform
testing data (25 %) and unseen validation data (25 %) came (ELM-WT) algorithm to estimate the exergetic performance
from operation data of thermal power plant, so the bootstrap of a DI diesel engine running on diesel/ biodiesel blends,
aggregated ELM would be a promising way to optimize and including fuel exergy rate, exergy rate of exhaust gas, exergy
control post-combustion CO2 capture processes. transfer rate to cooling water and ambient, exergy destruction
NOX emissions mainly come from the combustion of coal rate, exergy efficiency and sustainability index. Besides ELM-
and coal-derived fuels (like pyridine)[65] in thermal power plant, WT, other algorithms like ELM, Genetic Programming and
and recall an effective way to reduce. Yang et al.[44] managed ANN were also used to predict seven parameters. The
to model the relationship between operational parameters and outperformance of ELM-WT was proved by comparing error
NOX emission by ML and applied to a 660 WM boiler. After statistical indicators of the four models: root mean square error,
processed by PCA, the input data were respectively put into Pearson correlation coefficient and coefficient of

© Engineered Science Publisher LLC 2020 ES Energy. Environ., 2020, 9, 1-14 | 7

Review article ES Energy & Environment

determination. ELM is a single layer feed forward NN, as SG-ELM and online sequential extreme learning machines
shown in Fig. 10. Silitonga et al.[47] used Kernel-based (OS-ELM) respectively, to get HCCI engine state variable and
Extreme Learning Machine (KELM) to analyze the dynamic operating envelope. The results prove that good
performance, emission and combustion characteristics of a generalization performance can be achieved by both means
turbocharged diesel engine fueled with Jatropha curcas and the SG-ELM might have an advantage in terms of stability,
biodiesel-diesel blends. The coefficient of determination of the crucial for designing robust control systems.
parameters is within a range of 0.9805–0.9991 for both the
KELM model and the experimental data, while the mean
absolute percentage error (MAPE) is within a range of
0.1259–2.3838. Silitonga et al.[48] used the same model to
evaluate the engine performance exhaust emissions of a
single-cylinder diesel engine fueled with biodiesel-
bioethanol-diesel blends. Besides, results showed that this
algorithm is of a learn speed thousands times faster than the
conventional popular learning algorithms. The engine
performance parameters evaluated are the brake specific fuel
consumption and brake thermal efficiency whereas the
exhaust emission parameters evaluated are carbon monoxide,
nitrogen oxide, and smoke opacity. The predictive results
agree well with experimental ones, which suggests that the
KELM model is a reliable technique to predict the engine
performance and exhaust emission parameters of biodiesel-
bioethanol-diesel blends.
Fig. 11 Process of the energy management strategy based on
DDQL algorithm (adapted from Han et al.[51]).

Han et al.[51] proposed an energy development strategy

based on double deep Q-learning (DDQL) algorithms for a
dual-motor driven hybrid electric tracked-vehicle, as shown in
Fig. 11. It is easy to lead to an overoptimistic evaluation of the
results if the same network is used for both selecting action
and evaluating value function. Han et al. utilized two deep
NNs naming action selecting network Q and target updating
network Q’ to decouple action selection and evaluation. Han
et al. also used conventional deep Q-learning (DQL) and
dynamic programming (DP) to get driving schedules. The
Fig. 10 Single layer feed forward NN (adapted from Aghbashloet
simulation results were compared between DQL and DDQL
al.[46]).
based on the results gotten from DP. The comparison
HCCI is a new combustion method that the fuel is burned suggested that the fuel economy of DDQL achieved 7.1 %
by autoignition instead of spark. The primary challenge of better than DQL and reached 93.2 % level of DP. The results
HCCI autoignition is to ensure that the burn timing is substantiate the effectivity and robustness of proposed
synchronized against the motion of the piston, to avoid noisy methodology in terms of SOC convergence performance and
combustion and unstable combustion oscillation. HCCI fuel consumption level.
involved nonlinear chemistry, nonlinear physics, period- Li et al.[52] developed an air-fuel ratio (AFR) control
doubling bifurcation, turbulent mixing and so on, which result method combined with fuzzy feed-forward, intelligent PI
in high cyclic variability and difficult prediction of feedback and self-learning control, as shown in Fig. 12. The
combustion timing. Vaughan et al.[49] proposed a Weighted- conclusion was drawn from the simulation on Matlab and
Ring ELM, an online adaptive ML model to offer an accurate, engine bench test that this technique could perfectly correct
causal combustion timing prediction. The algorithm can also the AFR deviation during the compressed natural gas engine
enable a new class of real-time model predictive strategies for stationary and transient operations. The method could rapidly
HCCI, emitting less NOX and CO2. Janakiraman et al.[50] respond to the dynamic changes of engines, improve the AFR
developed an ELM model based on stochastic gradients (SG- control accuracy required by three-way catalyst converter
ELM). Janakiraman et al. built the online models by means of purification rate and reduce exhaust emissions.

8 | ES Energy Environ., 2020, 9, 1-14 © Engineered Science Publisher LLC 2020

ES Energy & Environment Review article

prototype. The simulation and test results suggested that the

good control of the stable reciprocating motion was obtained
by the strategy.
There are many kinds of engines. The ignition modes of
engine include spark ignition, compression ignition, HCCI,
Partially Premixed Combustion (PPC) and Reactivity
Controlled Compression Ignition (RCCI). Every ignition
method has its flaws. The spark ignition will produce CO and
HC along with severe detonation. More NOX and soot will be
formed by compression ignition. HCCI cannot control the
combustion timing and exothermic rate. PPC and RCCI have
Fig. 12 Structure of AFR control method (adapted from Li et low burning efficiency and more HC and CO. ML can be used
al.[52]). to solve the problems.

Namigtle-Jimenez et al.[53] proposed a faulty detection and 3.3 Fuel

diagnosis scheme that extracted the oil pressure as the only Besides theoretical studies described in Section 2.3, ML was
signal for isolating the damaged injector of an internal also used to overcome difficulties about fuel in industrial
combustion engine. The arrangement of scheme consisted of research.
three coupled ANNs shown in Fig. 13. ANNs were trained by
Levenberg-Marquardt (LM), BFGS quasi-Newton, and
Gradient Descent (GD). The offline comparison of the four
training algorithms was carried out to select the best one to
function online. The results demonstrated that the LM was the
best option that reached a 100 % classification offline and near
100 % classification online.

Fig. 14 Diagram of the multilayer NN (adapted from

Abdurakipov et al.[55]).

Coal is the most widely used fuel in the world. It is

necessary to develop new, economically viable and
environmentally friendly technologies for coal combustion.
The treatment of coal fuel in high-stress pulverizers on newly
formed surfaces forms active centers that can accelerate the
progress of physicochemical reactions. Abdurakipov et al.[55]
experimentally investigated the influence of high-stress
pulverization on the degree of burnout of coal fuel and
analyzed the applicability of a neural network-based
mathematical model for predicting the degree of burnout of
coal fuel treated in high-stress pulverizers. The model of four-
Fig. 13 Neuronal networks final structure (adapted from
layer ANN was developed as shown in Fig. 14, which
Namigtle-Jimenez et al.[53]).
consisted of an input layer (17 input parameters), three hidden
Yan et al.[54] obtained a reciprocating motion control layers (neurons are 64, 32 and 8 respectively) and an output
strategy of free piston by ML. The control problem was layer (one neuron). Learning data consisted of temperature,
decomposed to the stroke control, the top turning center heating rate, coal type, and a degree of pulverization,
control and the bottom turning center control, corresponding including training (80 %) and testing parts (20 %). A mean
to a finite-state machine, iterative learning controller and absolute percentage error is used as a model error function that
combustion state estimation as control method respectively. A is minimized as the NN learns. The results showed that the
combined simulation model was presented based on the model of the four-layer ANN quite accurately reproduced the
combination of the three methods, including combustion main effects of the influencing parameters and allowed one to
fluctuations and the validity of the model was proved by a highly accurately determine the coal mass loss coefficient

© Engineered Science Publisher LLC 2020 ES Energy. Environ., 2020, 9, 1-14 | 9

Review article ES Energy & Environment

using the given parameters: coal type, pulverization method, 3.4 Other industrial studies about combustion
temperature, and heating rate. There are amount of other fields about combustion or post-
Coal and its combustion products contain rare earth combustion disposal where ML was applied, such as studies
elements like Eu. The existence of Ba will interfere with the of gas turbine,[59] particulate matter (PM),[60-62] organic Rankine
measuring of Eu, there is a threshold value for Ba/Eu beyond cycle[63] and detection of hydrogen sensor.[64]
which the Eu concentration gotten by quadrupole-based Yan et al.[59] proposed a deep Semi-Supervised Anomaly
inductively-coupled plasma mass spectroscopy (ICP-Q-MS) Detection (deepSSAD) combining DL and semi-supervised
is not reliable. Xu et al.[56] implemented LR, regression tree to learning algorithms to detect gas turbine combustor anomalies.
predict the threshold value of Ba/Eu. The established model The model was composed of two steps: using DL to learn
predicted that the interference of Ba to the measurement of Eu representations or features from multivariate and time-series
could be neglected when Ba/Eu values were less than 363. sensor measurements and using one-class classification to
model normality in the learned feature space. The approach
achieved a good performance comparing the data collected
from a real-world gas turbine combustion system. This method
outperformed others too.

Fig. 15 Model architecture including inputs, two hidden layers of

32neurons, and an output (adapted from Kessler etal.[57]).

The method to predict the CN which is the measure of Fig. 16 Endpoint forecasting network of composite regeneration
ignition quality can accelerate the research process of the next- due to coupling cerium-based additive and microwave for diesel
generation alternative fuels. Derived cetane number (DCN) particulate filter (adapted from E et al.[60]).
could be an alternative of CN and obtained by ignition quality
tester (IQT). ANN could be used to establish an effective With the help of Fuzzy Least Squares Support Vector
model to predict CN of a fuel,[70] the structure and key learning Machines (FLS-SVM), E et al.[60] developed a mathematical
variables of which were optimized by Kessler et al.[57] to model which is shown in Fig. 16 to investigate the major
incorporate more kinds of fuels, especially furans and furan factors affecting the time of composite regeneration of diesel
derivatives. The optimized structure consisted of an input particulate filter. The regeneration was affected by the
layer containing 15 molecular descriptors, two hidden layers coupling cerium-based additive and microwave. The
of 32 neurons each and an output neuron, as shown in Fig. 15. simulation results with the relative error less than 3.5 % for
The optimization effect was validated from the more accurate endpoint forecasting indicated that main factors affecting the
CN predictions and slightly larger Root-Mean-Square-Error regeneration time are the oxygen concentration of exhaust gas,
(RMSE) than the previous. the mass flow rate of exhaust gas, the microwave power, the
Wang et al.[58] applied the ELM, OS-ELM, Self-adaptive temperature of exhaust gas and the amount of cerium-based
evolutionary Extreme Learning Machine (SaDE-ELM) additive.
respectively to predict the octane number and compared the Siegel et al.[61] built a classification model by decision tree,
three models by means of the statistical parameters like RMSE, which can detect filter contaminant loading from audio data
Correlation Coefficient, R2 and the execution time. The collected by a smartphone or stand microphone. The
prediction accuracy, generalization ability and stability of the robustness of the model was verified by changing vehicle type
ELM were improved by OS- and SaDE-ELM in different (Hondo and Mazda), loading states (clean, dirty and filthy) and
levels. Besides, many other ML algorithms have been used to microphone type (iPhone and TASCAM). The results showed
analyze and predict gasoline components like ANN, RBFNN, that 80 % of accuracy can be achieved.
Adaptive Neuro-Fuzzy Interface Systems and SVM. Rutherford et al.[62] combined excitation emission matrix
The incomplete combustion will result in the formation of (EEM) fluorescence spectroscopy and CNN to identify PM
HC and CO which is responsible for pollution. ML has not source. PM came from wood burning, diesel exhaust and
been used to study incomplete combustion yet. cigarettes using filters, with the limits of detection being 0.7,

1 0 | ES Energy Environ., 2020, 9, 1-14 © Engineered Science Publisher LLC 2020

ES Energy & Environment Review article

2.6 and 0.9 μg/m3, respectively. The overall classification which are relevant to fuels,[30-35] part of which are targeted to
accuracy for all three sources was 89%. study combustion phenomenon,[26-29] and only a few of studies
Xu et al.[63] applied ML to optimize organic Rankine cycle concentrate on CFD simulation.[23-25] In industrial studies, ML
waste heat recovery system. Three real-time implementable has been used to study power plant boiler,[41-45] engine,[46-54]
power optimization methods respectively based on fuel[55-58] and other fields.[59-64]
Proportional-Integral-Derivative (PID), Nonlinear Model ML can be used to reduce mechanism, save PDF table
Predictive Control (NMPC) and Dynamic Programming memory and predict terms filtered by LES, in order to balance
Random Forest (DP-RF) were compared. Xu et al. adopted computation costs and accuracy of CFD combustion
random forest model to extract rules for DP, as shown in Figs.simulation. In the combustion phenomenon studies, ML is
17 and 18. The comparison results showed that the DP-RF used to construct feature space, to predict thermoacoustic
method had similar performance to NMPC method and combustion oscillations, distinguish detonation and ignition in
outperformed the PID method while the computation costs of RCMs and shock tubes, and reconstruct the detonation surface.
DP-RF were less than NMPC. ML also plays an important role in studying chemical and
physical properties, designing the next generation fuel and
studying biofuels.
Compared with theoretical studies, ML is more widely
used in industrial ones. ML in power plant studies focused on
monitoring combustion in furnace and disposing post-
combustion pollutants. In the engine studies, ML was used to
study the performance of diesel, combustion timing and
dynamic operating envelope of HCCI and operation or control
of engine. In the industrial studies about fuels, ML was utilized
to investigate the influence of high-stress pulverization on the
degree of burnout of coal fuel, obtain the threshold value of
Ba/Eu in coals and combustion products, predict the CN of
Fig. 17 Diagram of DP-RF optimization method training process biofuels and octane number of gasoline. Apart from those
(adapted from Xu et al.[63]). fields, ML can also be used in PM, monitoring gas turbine
combustor anomalies, fault diagnosis and reconfiguration of
hydrogen sensor array and optimization of organic Rankine
cycle waste heat recovery system.

5. Perspectives
As a powerful instrument to build model, ML is attracting
more and more attention in combustion investigations recently.
This paper reviewed the applications of ML in theoretical and
industrial combustion studies, and some recommendations
will be given here.
1) From the present results, ML applies more to industrial
studies of combustion and less on theoretical ones. Theoretical
studies need more assistance from ML, especially combustion
chemical kinetics. Combustion kinetic study will help us
Fig. 18 Schematic of RF model in DP rule extraction (adapted
deeply understand combustion process, and give guidance to
from Xu et al.[63]).
efficient combustion and low pollution emissions. The most
Song et al.[64] utilized moving window principle important thing in kinetic research is to construct the kinetic
component analysis and extreme learning machine (MWPCA- models. A perfect mechanism can accurately predict profiles
ELM) to propose a novel fault diagnosis and reconfiguration of reactants, products and intermediate species. A chemical
strategy for hydrogen sensor array. The effectiveness and kinetic model usually contains hundreds of species and
feasibility of the strategy were illustrated by the decreased thousands of reactions, which means a large number of
average relative error from 1.18% to 0.82% of hydrogen parameters need to be adjusted. Determination of parameters
concentration estimated from a monitoring experimental is a task depending on experience. ML is good at building a
system. model or function from data, which may help to build an
efficient mechanism if applied in chemical kinetic studies.
4. Summary 2) NTC effect and cool flame will emerge in combustion
The appliances of ML to combustion studies were reviewed in of linear alkanes. It’s an important criterion to judge kinetic
the paper. Among the combustion theory studies, most of models that whether the NTC effect can be accurately

© Engineered Science Publisher LLC 2020 ES Energy. Environ., 2020, 9, 1-14 | 11

Review article ES Energy & Environment

predicted. The current chemical kinetic models can hardly get CNN Convolutional Neural Network
an accurate prediction of PAHs, which are precursors of soot,
DCN Derived Cetane Number
therefore the mechanism can’t predict soot satisfactorily. If a
well-behaved chemical kinetic model is developed based on DDQL Double Deep Q-Learning
ML, then the phenomena like NTC effect, cool flame and soot DL Deep Learning
will be successfully investigated. DLR DeutschesZentrumfürLuft-und Raumfahrt
3) The realistic fuels like gasoline, kerosene, diesel and so DNS Direct Numerical Simulation
on are usually composed of hundreds of components, which
DP Dynamic Programming
makes it impossible to simulate real fuels. The popular method
is to simulate surrogate fuels having similar properties like DP-RF Dynamic Programming Random Forest
carbon-hydrogen ratio, molar mass, combustion heat and so DQL Deep Q-Learning
on. ML has been used to predict fuels properties and design EEM Excitation Emission Matrix
the next generation fuels, which means it will play important
ELM Extreme Learning Machine
roles in constructing surrogate fuels in the future.
4) The effects of furnace are to function as a combustion Extreme Learning Machine with Wavelet
ELM-WT
chamber and transfer heat from flame to working fluid. Heat Transform
transfer deterioration will cause catastrophes, damage boiler FFANN Feed Forward Artificial Neural Networks
and hurt people. ML can be used to develop a strategy to FLS-SVM Fuzzy Least Squares Support Vector Machines
monitor boiler parameters and furnace parameters meanwhile
GD Gradient Descent
to avoid tragedies like heat transfer deterioration.
5) The combustion methods of engine include spark GFA Genetic Function Approximation
ignition, compression ignition, HCCI, PPC and RCCI. Low- GM Graph Machines
temperature combustion exists the problems like low GRNN General Regression Neural Networks
efficiency and high emissions of HC and CO. The high
HCCI Homogenous Charge Compression Ignition
temperature will produce more NOX and soot. HCCI cannot
control combustion timing and exothermic rate. ML has been HEDH High Energy Density Hydrocarbon
widely used in engine research, but more efforts will be paid HWR Handwriting Recognition
to solve the mentioned problems. ICP-Q-MS Inductively-Coupled Plasma Mass Spectroscopy
6) As for study with respect to the post-combustion
IQT Ignition Quality Tester
pollutants, ML has only been applied to deal with CO2, NOX
and particulates up to now. However, the emissions formed in KELM Kernel-based Extreme Learning Machine
combustion are far more than the 3 species. CO and PAH will Least Absolute Shrinkage and Selection
LASSO
emerge during incomplete combustion of fuels, harming the Operator
environment and the human body. ML will be helpful in the LES Large Eddy Simulation
disposal of other pollutants. LIBS Laser-Induced Breakdown Spectroscopy
Acknowledgements LM Levenberg-Marquardt
The authors are thankful for the financial support from the LPCA Local Principal Component Analysis
NSFC (No. 51888103/51976216/51606192), MOST
LR Linear Regression
(2017YFA0402800), Recruitment Program of Global Youth
Experts and the CAS Pioneer Hundred Talents Program. LSSVM Least Squares Support Vector Machine

Supporting information LSTM Long-Short Term Memory

Not applicable. MAPE Mean Absolute Percentage Error

Conflict of interest mid-IR FTIR mid-Infrared Fourier Transform Infrared

There are no conflicts to declare. ML Machine Learning

Nomenclature MLP Multi-layer Perception Neural Networks

AFR Air-Fuel Ratio Moving Window Principle Component Analysis

MWPCA-ELM
and Extreme Learning Machine
ANN Artificial Neural Network
NLP Natural Language Processing
ASR Automatic Speech Recognition
NMPC Nonlinear Model Predictive Control
CECP Complexity-Entropy Causality Plane
NN Neural Network
CFD Computational Fluid Dynamics
NOX Nitrogen Oxides
CN Cetane Number
NTC Negative Temperature Coefficient

1 2 | ES Energy Environ., 2020, 9, 1-14 © Engineered Science Publisher LLC 2020

ES Energy & Environment Review article

OS-ELM Online Sequential Extreme Learning Machine [7] H. Wei, H. Bao and X. Ruan, Nano Energy, 2020, 71,
104619, doi: 10.1016/j.nanoen.2020.104619.
PAHs Polycyclic Aromatic Hydrocarbons
[8] P. R. Chowdhury, C. Reynolds, A. Garrett, T. Feng, S. P.
PCA Principal Component Analysis
Adiga and X. Ruan, Nano Energy, 2020, 69, 104428, doi:
PDF Probability Density Function 10.1016/j.nanoen.2019.104428.
PID Proportional-Integral-Derivative [9] R. Hu, S. Iwamoto, L. Feng, S. Ju, S. Hu, M. Ohnishi, N.
PLS Partial Least Squares Nagai, K. Hirakawa and J. Shiomi, Phys. Rev. X, 2020, 10,
021050, doi: 10.1103/PhysRevX.10.021050.
PM Particulate Matter
[10] T. M. Mitchell, Machine learning, New York, the USA:
POD Proper Orthogonal Decomposition McGraw-Hill Education, 1997.
PPC Partially Premixed Combustion [11] P. Kalluri, Nature, 2020, 583, 169, doi: 10.1038/d41586-
RBFNN Radial Basis Function Neural Network 020-02003-2.
[12] D. Coyle and A. Weller, Science, 2020, 368, 1433-1434, doi:
RCCI Reactivity Controlled Compression Ignition
10.1126/science.aba9647.
RCMs Rapid Compression Machines [13] P. Domingos, Commun. Acm., 2012, 55, 78-87, doi:
RDR Random Forest Regression 10.1145/2347736.2347755.
RL Reinforcement Learning
[14] L. Gokhberg, Nature, 2020, 583, 360-360, doi:
10.1038/d41586-020-02086-x.
RMSE Root-Mean-Square-Error
[15] Z. Li, Int. J. Heat Mass Transf., 2016, 5, 198.
RNN Recurrent Neural Network [16] K. J. Bergen, P. A. Johnson, M. V. De Hoop and G. C.
Self-adaptive evolutionary Extreme Learning Beroza, Science, 2019, 363, 0323, doi: 10.1126/science.aau0323.
SaDE-ELM
Machine [17] I. Lee and Y. J. Shin, Bus. Horizons, 2020, 63, 157-170, doi:
SCR Selective Catalyst Reaction 10.1016/j.bushor.2019.10.005.
[18] X. Lin, Y. Rivenson, N. T. Yardimei, M. Veli, Y. Luo, M.
SG-ELM Stochastic Gradients Extreme Learning Machine
Jarrahi and A. Ozcan, Science, 2018, 361, 1004-1008, doi:
SOM Self-Organizing Maps 10.1126/science.aat8084.
Sample-Partitioning Adaptive Reduced [19] I. Goodfellow, Y. Bengio and A. Courville, Deep Learning,
SPARC
Chemistry Cambridge：MIT Press, 2016, 1, 326-366, doi:10.1007/s10710-
SSAD Semi-Supervised Anomaly Detection 017-9314-z.
SVM Support Vector Machine [20] I. Goodfellow, Y. Bengio and A. Courville, Deep Learning,
Cambridge：MIT Press, 2016, 1, 367-415,
SVR Support Vector Regression
https://doi.org/10.1007/s10710-017-9314-z.
Time Series Analysis with Fuzzy Support Vector
TSA-FSVM [21] A. Zahrt, J. Henle, B. Rose, Y. Wang, W. Darrow and S. Den-
Machine
markt, Science, 2019, 363, 5631, doi: 10.1126/science.aau5631.
VOCs Volatile Organic Compounds [22] R. Blaga, A. Sabadus, N. Stefu, C. Dughir, M. Paulescu and
References V. Badescu, Prog. Energy Combust. Sci., 2019, 70, 119-144, doi:
[1] Y. Liu, The reaction kinetic stud ies of C9 surrogate fuels of 10.1016/j.pecs.2018.10.003.
kerosene. In Insititute of Engineering Thermophysics, Chinese [23] G. D'alessio, A. Parente, A. Stagni and A. Cuoci, Combust.
Academy of Science: Beijing, China, 2019. Flame, 2020, 211, 68-82, doi: 10.1016/j.com-
[2] N. B. O. Statistics, China Academic Journal Electronic Pub- bustflame.2019.09.010.
lishing House, 2020, 01, 16-18. [24] R. Ranade, G. Li, S. Li and T. Echekki, Combust. Sci. Tech-
[3] X. Wan, W. Feng, Y. Wang, H. Wang, X. Zhang, C. Deng nol., 2019, 1-21, doi: 10.1080/00102202.2019.1686702.
and N. Yang, Nano Lett., 2019, 19, 3387-3395, doi: [25] A. Seltz, P. Domingo, L. Vervisch and Z. M. Nikolaou, Com-
10.1021/acs.nanolett.8b05196. bust. Flame, 2019, 210, 71-82, doi: 10.1016/j.com-
[4] H. Zhang, K. Hippalgaonkar, T. Buonassisi, O. Løvvik, E. bustflame.2019.08.014.
Sagvoldenand and D. Ding, ES Energy Environ.,2018, 2, 1-8, doi: [26] T. Hachijo, S. Masuda, T. Kurosaka and H. Gotoda, Chaos,
10.30919/esee8c209. 2019, 29, 103123, doi: 10.1063/1.5120815.
[5] H. Wei, S. Zhao, Q. Rong and H. Bao, Int. J. Heat Mass [27] T. Kobayashi, S. Murayama, T. Hachijo and H. Gotoda,
Transf., 2018, 127, 908-916, doi: 10.1016/j.ijheatmasstrans- Phys. Rev. Appl., 2019, 11, 064034, doi: 10.1103/PhysRevAp-
fer.2018.08.082. plied.11.064034.
[6] S. Wu, Y. Kondo, M.A. Kakimoto, B. Yang, H. Yamada, I. [28] K. P. Grogan and M. Ihme, Shock Waves, 2018, 28, 941-954,
Kuwajima, G. Lambard, K. Hongo, Y. Xu, J. Shiomi, C. Schick, doi: 10.1007/s00193-018-0852-y.
J. Morikawa and R. Yoshida, NPJ Comput. Mater., 2019, 5,5, doi: [29] Y. Zhang, L. Zhou, H. Meng and H. Teng, Combust. Flame,
10.1038/s41524-019-0203-2. 2020, 212, 156-164, doi: 10.1016/j.combustflame.2019.10.031.
[30] Y. Wang, Y. Ding, W. Wei, Y. Cao, D. Davidson and R.

© Engineered Science Publisher LLC 2020 ES Energy. Environ., 2020, 9, 1-14 | 13

Review article ES Energy & Environment

Hanson, Fuel, 2019, 255, 115715, doi: 254, 113708, doi: 10.1016/j.apenergy.2019.113708.
10.1016/j.fuel.2019.115715. [52] P. Li, J. Wang and W. Cai, Proceedings of the 2013 Fourth
[31] D. Saldana, L. Starck, P. Mougin, B. Rousseau and B. Cre- International Conference on Intelligent Control and Information
ton, Sar. Qsar. Environ. Res., 2013, 24,259-277, doi: Processing, 2013, 684-688, doi: 10.1109/ICICIP.2013.6568160.
10.1080/1062936X.2013.766634. [53] A. Namigtle-Jimenez, R. F. Escobar-Jimenez, J. F. Gomez-
[32] G. Li, Z. Hu, F. Hou, X. Li, L. Wang and X. Zhang, Fuel, Aguilar, C. D. Garcia-Beltran and A. C. Tellez-Anguiano, ISA. T.,
2020, 265, 116968, doi: 10.1016/j.fuel.2019.116968. 2019, 100, 358-372, doi: 10.1016/j.isatra.2019.11.003.
[33] X. Liu, X. Feng and Y. He, Renew. Energ., 2019, 143, 176- [54] H. Yan, Z. Xu, J. Lu, D. Liu and X. Jiang, Electronics, 2020,
182, doi: 10.1016/j.renene.2019.04.137. 9, 245, doi: 10.3390/electronics9020245.
[34] W. Jiang, X. Xing, X. Zhang and M. Mi, Renew. Energ., [55] S. S. Abdurakipov, E. B. Butakov, A. P. Burdukov, A. V.
2019, 130, 1216-1225, doi: 10.1016/j.renene.2018.08.089. Kuznetsov and G. V. Chernova, Combust. Explo. Shock., 2019,
[35] S. Wen, M. Buyukada, F. Evrendilek and J. Liu, Renew. En- 55, 697-701, doi: 10.1134/S0010508219060108.
erg., 2020, 151, 463-474, doi: 10.1016/j.renene.2019.11.038. [56] N. Xu and Q. Li, Minerals, 2019, 9, 259, doi:
[36] W. Meier, R. Barlow, Y. Chen and J. Chen, Combust. Flame, 10.3390/min9050259.
2000, 123, 326-343, doi: 10.1016/S0010-2180(00)00171-1. [57] T. Kessler, E. R. Sacia, A. T. Bell and J. H. Mack, Fuel, 2017,
[37] J. Wang, Y. Li, T. Zhang, Z. Tian, B. Yang, K. Zhang, F. Qi, 206, 171-179, doi: 10.1016/j.fuel.2017.06.015.
A. Zhu, Z. Cui and C.Y. Ng, Astrophys. J., 2008, 676, 416-419, [58] X. Wang, K. Yang and J. H. Kalivas, Optik, 2020, 200,
doi: 10.1086/528732. 163325, doi: 10.1016/j.ijleo.2019.163325
[38] B. Sanchez-Lengeling and A. Aspuru-Guzik, Science, 2018, [59] W. Yan, Cogn. Comput., 2020, 12, 398-411, doi:
361, 360-365, doi: 10.1126/science.aat2663. 10.1007/s12559-019-09710-7.
[39] D. Ahneman, J. Estrada, S. Lin, S. Dreher and A. Doyle, Sci- [60] J. E, Q. Zuo, H. Liu, Y. Li and J. Gong, J. Cent. South. Univ.,
ence, 2018, 360, 186-190, doi: 10.1126/science.aar5169. 2016, 23, 2118-2128, doi: 10.1007/s11771-016-3268-9.
[40] K. Campos, P. Coleman, J. Alvarez, S. Dreher, R. Garbaccio, [61] J. E. Siegel, R. Bhattacharyya, S. Kumar and S. E. Sarma,
N. Terrett, R. Tillyer, M. Truppo and E. Parmee, Science, 2019, Eng. App. Artif. Intel, 2017, 66, 104-112, doi: 10.1016/j.en-
363, 0805, doi: 10.1126/science.aat0805. gappai.2017.09.015.
[41] S. Annadurai and M. Arock, Asia.-Pac. J. Chem. Eng., 2020, [62] J. W. Rutherford, N. Dawson-Elli, A. M. Manicone, G. V.
e2395, doi: 10.1002/apj.2395. Korshin, I. V. Novosselov, E. Seto and J. D. Posner, Atmos. Envi-
[42] A. A. Rahmanda, A. Soeprijanto, A. Muhammad, M. Syaiin, ron., 2020, 220, 117065, doi: 10.1016/j.atmosenv.2019.117065.
R. Y. Adhitya, B. Herijono, J. Endrasmono, A. Singgih, E. A. Zu- [63] B. Xu, D. Rathod, A. Yebi and Z. Filipi, Appl. Therm. Eng.,
liari, S. I. Haryudo, F. Afandi and B. S. Kaloko, 2017 Interna- 2020, 164, 114442, doi: 10.1016/j.applthermaleng.2019.114442.
tional Symposium on Electronics and Smart Devices, 2017, 2018, [64] K. Song, P. Xu, Y. Chen, T. Zhang, G. Wei and Q. Wang,
123-128, doi: 10.1109/ISESD.2017.8253317. IEEE ACCESS, 2019, 7, 115075-115092, doi: 10.1109/AC-
[43] Z. Bai, F. Li, J. Zhang, E. Oko, M. Wang, Z. Xiong and D. CESS.2019.2936128.
Huang, Computer Aided Chemical Engineering, 2016, 38, 2007- [65] Z. Tian, Y. Li, T. Zhang, A. Zhu and F. Qi, J. Phys. Chem. A,
2012, doi: 10.1016/B978-0-444-63428-3.50339-8. 2008, 112, 13549-13555, doi: 10.1021/jp8066537.
[44] G. Yang, Y. Wang and X. Li, Energy, 2020, 192, 116597, doi: [66] Z.Y. Tian, P. M. Kouotou, A. El Kasmi, P. H. T. Ngamou, K.
10.1016/j.energy.2019.116597. Kohse-Hoeinghaus, H. Vieker, A. Beyer and A. Goelzhaeuser, P.
[45] F. Si, C. Romero, Z. Yao, E. Schuster, Z. Xu, R. Morey and Combust. Inst., 2015, 35, 2207-2214, doi:
B. Liebowitz, Fuel, 2009, 88, 806-816, doi: 10.1016/j.proci.2014.06.111.
10.1016/j.fuel.2008.10.038. [67] S.B. Fan, P. M. Kouotou, J.J. Weng, G.F. Pan and Z.Y. Tian,
[46] M. Aghbashlo, S. Shamshirband, M. Tabatabaei, P. L. Yee P. Combust. Inst., 2017, 36, 4375-4382, doi:
and Y. N. Larimi, Energy, 2016, 94, 443-456, doi: 10.1016/j.en- 10.1016/j.proci.2016.07.087.
ergy.2015.11.008. [68] P. M. Kouotou and Z.Y. Tian, P. Combust. Inst., 2019, 37,
[47] A. Silitonga, M. Hassan, H. Ong and F. Kusumo, Environ. 5445-5453, doi: 10.1016/j.proci.2018.05.172.
Sci. Pollut. R., 2017, 24, 25383-25405, doi: 10.1007/s11356-017- [69] P. Kouotou, Z. Tian, U. Mundloch, N. Bahlawane and K.
0141-9. Kohse-Hoeinghaus, RSC Adv., 2012, 2, 10809-10812, doi:
[48] A. S. Silitonga, H. H. Masjuki, H. C. Ong, A. H. Sebayang, 10.1039/c2ra21277c.
S. Dharma, F. Kusumo, J. Siswantoro, J. Milano, K. Daud, T. M. [70] B. Creton, C. Dartiguelongue, T. De Bruin and H. Toulhoat,
I. Mahlia, W.-H. Chen and B. Sugiyanto, Energy, 2018, 159, Energ. Fuel., 2010, 24, 5396-5403, doi: 10.1021/ef1008456.
1075-1087, doi: 10.1016/j.energy.2018.06.202.
[49] A. Vaughan and S. V. Bohac, NPJ Comput. Mater., 2015, 70, Publisher’s Note: Engineered Science Publisher remains
18-26, doi: 10.1016/j.neunet.2015.04.007. neutral with regard to jurisdictional claims in published maps
[50] V. M. Janakiraman, X. Nguyen and D. Assanis, Neurocom- and institutional affiliations.
puting, 2016, 177, 304-316, doi: 10.1016/j.neucom.2015.11.024.
[51] X. Han, H. He, J. Wu, J. Peng and Y. Li, Appl. Energ., 2019,

1 4 | ES Energy Environ., 2020, 9, 1-14 © Engineered Science Publisher LLC 2020