1 s2.0 S2589014X22000330 Main

Bioresource Technology Reports 17 (2022) 100976
Contents lists available at ScienceDirect
Bioresource Technology Reports

journal homepage: www.sciencedirect.com/journal/bioresource-technology-reports
Machine learning methods for modeling conventional and hydrothermal

gasification of waste biomass: A review
Great C. Umenweke a, Inioluwa Christianah Afolabi b, Emmanuel I. Epelle c, Jude A. Okolie d, e, *
a
Department of Chemistry, University of Kentucky, Lexington, KY, USA
b
Department of Pure and Applied Chemistry, Ladoke Akintola University of Technology, P.M.B. 4000, Ogbomosho, Oyo State, Nigeria
c
School of Computing, Engineering & Physical Sciences, University of the West of Scotland, Paisley PA1 2BE, United Kingdom
d
Department of Chemical and Biological Engineering, University of Saskatchewan, Saskatoon, Saskatchewan, Canada
e
St. Peter's College, Muenster, Saskatchewan, Canada
A R T I C L E I N F O A B S T R A C T
Keywords: Conventional and hydrothermal gasification are promising thermochemical technologies for the production of
Gasification syngas from waste biomass. Both gasification processes are complex, with several intermediate reactions
Biomass occurring simultaneously and at different times. Therefore, traditional modeling approaches, including ther
Machine learning
modynamic and kinetic models, process modeling and computation fluid dynamics (CFD), are sometimes used to
Catalyst
describe the process and study the influence of process parameters on syngas yield. However, most traditional
Process optimization
models are impractical and often challenging to model the input-output relationship. Machine learning (ML)
methods provide a promising alternative to traditional modeling approaches. This study outlined the advance
ment in the application of ML to gasification research. Different ML algorithms, including artificial neural net
works, deep learning and support vector machines, are briefly described. Challenges and limitations of ML-based
methods for gasification research are also discussed. When implemented effectively, ML – approaches could
foster the development of novel heterogeneous catalysts for biomass gasification.
1. Introduction could eventually deplete.

The challenges of climate change, elevating energy demand and
The present world is faced with several challenges such as increasing environmental pollution have promoted a shift from a fossil-based
population and energy demand, climate change issues and environ economy to renewable energy resources. Renewable energy sources
mental pollution. Fossil fuel resources such as coal, natural gas and such as solar, nuclear, geothermal, and wind are promising from envi
petroleum are currently used to provide up to 80% of the worldwide ronmental perspectives. However, they are unsuitable for liquid trans
energy requirements (Mandegari et al., 2018). However, the increasing portation fuel production (Okolie et al., 2020). In contrast, waste
fossil fuel consumption has left humanity with some serious environ biomass such as woody biomass, agricultural residues, plastic wastes
mental concerns. The combustion of fossil fuels is associated with the (including COVID – 19 facemask), food and beverage wastes are plen
release of significant quantities of greenhouse gases such as CO2 (Nanda tiful for producing green fuels and chemicals. The conversion of waste
et al., 2016). Furthermore, a large deposit of the global petroleum re biomass to biofuels could help mitigate climate change and environ
serves is found in areas with economic and political uncertainty, ulti mental pollution challenges.
mately leading to fuel price instability (Okolie et al., 2020). It should be Waste biomass can be transformed into green fuels through ther
noted that with the current rate of fossil fuel consumption, the reserves mochemical (e.g., gasification, liquefaction, and pyrolysis) and
Abbreviations: ANN, Artificial Neural Network; AWRA, Adaptive Weighted Rank Aggregation; BTG, Biomass-to-gas; CFD, Computational Fluid Dynamics; CGE,
Carbon Gas Efficiency; DNN, Deep Neural Network; DTR, Decision Tree Regression; GBR, Gradient Boosting Regression; GPR, Gaussian Process Regression; HHV,
Higher Heating Value; LHV, Lower Heating Value; MAPE, Mean Absolute Percentage Error; MIMO, Multiple Input Multiple Output; MISO, Multiple Input Single
Output; ML, Machine Learning; MLP, Multilayer Perceptron; PCA, Particle Component Analysis; PSO, Particle Swarm Optimization; RAE, Relative Absolute Error;
RBF, Radial Basis Function; RF, Random Forest; RMSE, Root Mean Square Error; SCW, Supercritical Water; SVM, Support Vector Machine.
* Corresponding author at: Department of Chemical and Biological Engineering, University of Saskatchewan, Saskatoon, Saskatchewan, Canada.
E-mail address: [email protected] (J.A. Okolie).
https://doi.org/10.1016/j.biteb.2022.100976
Received 2 January 2022; Received in revised form 19 January 2022; Accepted 1 February 2022
Available online 5 February 2022
2589-014X/© 2022 Elsevier Ltd. All rights reserved.
G.C. Umenweke et al. Bioresource Technology Reports 17 (2022) 100976
biological (syngas fermentation and anaerobic digestion) processes. The implement their results (Wong et al., 2013). The CFD models require
latter uses bacteria, microorganisms, or enzymes to decompose waste sizeable computational power and predefined engine-specific parameter
biomass into green fuels and chemicals (Brethauer and Studer, 2015). to obtain a reliable simulation. The kinetic model is highly dependent on
On the other hand, thermochemical processes use thermal energy and the estimation of complex reaction rates. In addition, the complete re
chemical reactions to convert waste biomass into sustainable fuels and action mechanism is either unknown or not fully understood during
chemicals. Thermochemical processes are advantageous in terms of kinetic modeling. On the contrary, thermodynamic modeling often as
improved product yield and conversion efficiency. They do not need a sumes an equilibrium state even though most experimental reactions
microorganism culture solution. They are suitable for different types of hardly attain equilibrium. The process modeling approach usually re
feedstocks and require short residence time (Okolie et al., 2020). Fig. 1 quires specialized software such as ASPEN Plus, which is relatively
illustrates the different classifications of thermochemical conversion expensive in most cases. Therefore, developing a reliable, accurate,
processes. A detailed description of each biomass conversion process robust, and fast modeling approach to correlate the complexity of the
and their advantages and limitations are beyond the scope of the present gasification process is required.
study. However, the readers are referred to the recent review by Okolie Machine learning (ML) is a reliable alternative to conventional
et al. (2021b). models to correlate and study the complexity of waste biomass gasifi
Gasification is a promising biomass-to-gas (BTG) thermochemical cation. ML has been used to study the reaction mechanism of different
conversion process for producing hydrogen-rich syngas from waste biomass conversion processes (Aghbashlo et al., 2021; Okolie et al.,
biomass. Gasification could be further classified as hydrothermal and 2021a; Wang et al., 2022). Aghbashlo et al. (2021) recently reviewed the
conventional gasification (Okolie et al., 2021b). This classification is application of data-driven ML methods to biodiesel research. They
based on the type of gasifying agent, reaction conditions and the quality stated that data-driven ML methods have proven superior in modeling
of syngas produced. Although the main product obtained from the two complex chemical processes compared to traditional methods. In addi
types of gasification is similar (syngas), the process conditions and tion, the authors also suggested that future studies should focus on using
syngas composition are different. Several process factors such as tem ML techniques for real-time process monitoring and control of biodiesel
perature, pressure, reactor type, catalysts, residence time and heating systems to improve the product yield. However, their study is mainly
rate affect the syngas composition during hydrothermal and conven concerned with biodiesel production processes while other biomass
tional gasification (Abdoulmoumine et al., 2015). Moreover, the hy conversion processes were not considered. Ascher et al. (2021) reported
drothermal gasification process comprises several intermediate recent advances in the use of ML to promote pyrolysis and gasification
reactions, and its reaction kinetics and production pathway are least research. They summarized previous studies that used ML methods to
understood (Zhang et al., 2019). The complexity of the gasification optimize and predict product yield during pyrolysis and conventional
process makes the precise modeling more complicated and sophisti gasification. However, the authors did not consider hydrothermal gasi
cated. It should be emphasized that an accurate modeling approach is fication studies. In addition, they suggested that research emphasis
essential for efficient reactor design, process scale-up, techno-economic should be on model interpretability to present a sound understanding of
analysis and environmental assessment. Moreover, modeling of gasifi the process being studied (Ascher et al., 2021).
cation processes could provide detailed information and insights on the A recent study outlines the role of ML in improving biofuel and
reaction mechanisms. Modeling a process also helps to correlate the bioenergy research (Wang et al., 2022). The authors concluded that ML
input parameters with the desired output and attain a real-time process methods could effectively predict and optimize complex nonlinear bio
control (Aghbashlo et al., 2021). energy systems. In addition, future studies should focus on developing a
Traditional modeling techniques such as thermodynamic and kinetic high-quality database and testing new algorithms to improve model
models, process modeling and computation fluid dynamics (CFD) are predictability. However, to the best of the authors' knowledge, a
sometimes impractical and often challenging to model the input-output comprehensive review of ML applications for hydrothermal and con
relationship. Some of these models are developed based on numerous ventional gasification is not available in literature considering ML-
assumptions and oversimplifications, rendering it challenging to driven models' ascendancy in recent years. Such information could be
Fig. 1. Classification of different thermochemical conversion processes for the production of biofuels and chemicals. Adapted from Okolie et al. (2021b) with
permission from Elsevier.
2
valuable for process optimization and commercialization of the gasifi on the type of gasifying agent. For instance, the use of air as gasifying
cation process. Therefore, this study evaluates all possible areas where agents produces syngas with low LHV values ranging from 3.7–8.4 MJ/
data-driven ML could be implemented in gasification research, including Nm3 (Yang and Chen, 2015). On the contrary, pure steam and oxygen
catalysts screening and gaseous yield prediction. It is noteworthy that produce syngas with higher LHV of 12.2–13.8 MJ/Nm3 and 12–28 MJ/
the mathematical background inherent in developing various ML Nm3, respectively (Yang and Chen, 2015). However, air is relatively
models is not presented in this study. However, such information has cheap and often used in small or medium-sized gasification plants.
been meticulously described elsewhere (Burkov, 2019; Priddy and PK, Although steam and oxygen produce syngas with increased LHV, they
2005; Mohri et al., 2018). add to the overall capital cost of the plant.
Studies related to conventional gasification are focused on the opti
2. Conventional and hydrothermal gasification mization of process conditions (Yusup et al., 2014), techno-economic
analysis (Mohammed et al., 2011), thermodynamic and kinetic
Most of the papers involving hydrothermal gasification (also known modeling (Baruah and Baruah, 2014) and reactor design (Puig-Arnavat
as supercritical water gasification) and conventional gasification in et al., 2013b). Some recent studies also focus on tar reduction and the
cludes the differences, advantages, and limitations of each type of development of catalysts (Valderrama Rios et al., 2018). Research
gasification. Specifically, most review papers outline different factors emphasis on hydrothermal gasification is mainly concerned with reactor
affecting the gasification of waste biomass, the type of catalysts used, design (Pinkard et al., 2019), heat integration, techno-economic analysis
and operational challenges. Furthermore, only a few papers can be cited and life cycle assessment (Chen et al., 2019b; Özdenkçi et al., 2019),
that outlined different modeling and optimization strategies for hydro process modeling and parametric studies (Chen et al., 2019a), and the
thermal gasification (Okolie et al., 2021a) and conventional gasification integration of hydrothermal gasification with other biomass conversion
(Baruah and Baruah, 2014). In the following references, readers can processes (Okolie et al., 2021d). Table 2 summarizes previous research
learn about the characteristics and properties of different types of on conventional and hydrothermal gasification of waste biomass.
biomass gasifiers, the economics of biomass gasification (Ahmad et al.,
2016; Baruah and Baruah, 2014) and the unique properties of super 3. Overview of machine learning
critical water that makes hydrothermal gasification a promising tech
nology (Okolie et al., 2019). In addition, a recent study also discusses the ML comprises of an array of different techniques that are often
viability of hydrothermal gasification to produce biofuels and remarks applied to numerous fields. This engaging, fast-growing technology has
on future studies required to foster the advancement of the technology gained recognition for its innovative application in image classification,
towards commercialization (Ortiz, 2021). speech recognition, natural language processing, weather forecasting,
The most relevant distinction between hydrothermal and conven stock prediction, self-driving cars and so on (Selvaratnam and Koodali,
tional gasification are summarized in Table 1. 2021; Wang et al., 2021). Recently, ML has gained visibility in chemical
Hydrothermal gasification is advantageous because of its ability to research such as drug discovery (Yang et al., 2019), predictive toxi
handle high moisture content feedstock without drying requirement cology (Wang et al., 2021), cheminformatics, environmental sustain
(Kang et al., 2017). In contrast, drying is needed for waste biomass with ability, green chemistry (Afolabi et al., 2020a, 2020b), experimental
moisture content above 35 wt% during conventional gasification. Hy material chemistry, and catalysis, (Guan et al., 2022). However, to
drothermal gasification uses the unique features of supercritical water better discern how ML can be applied within conventional and hydro
(SCW) such as liquid-like density and gas- like- viscosity to promote the thermal gasification research, it is imperative to understand the under
decomposition of biomass into biofuels (Castello et al., 2017). Hydro lying theory. In addition, the advantages and limitations of different ML
thermal gasification is characterized by fast hydrolysis, reaction rate, algorithms should also be documented. In developing the ML model, the
and enhanced mass transfer. In addition, the produced syngas from machine learns useful patterns between a set of input data and its cor
hydrothermal gasification are often free from nitrogen and sulphur- responding output through training. The data used in training the ML
containing compounds (Okolie et al., 2021a). In contrast, conventional model is called the training dataset. Following the model's training is
gasification is beneficial due to the lower overall capital and operating validation, then repetitive training and validation of the model are done
cost and less risk of explosion due to decreased operating pressure. until an optimal model is achieved. The useful patterns learnt by the
Conventional gasification also requires gasifying agents such as air, optimal model during the training, generates a rule that help predict the
oxygen, CO2, steam, or their mixtures. It should be emphasized that the output for a new datasets called testing data.
product selectivity and lower heating value (LHV) of the syngas depends Machine learning algorithms can be classified as supervised, semi-
supervised, unsupervised and reinforced learning (Guan et al., 2022;
Table 1 Sahoo et al., 2020). In supervised learning, the input data (features) and
Comparison between conventional and hydrothermal gasification (Ahmad et al., corresponding output data (labels) are fed into the model. Then, the
2016; Baruah and Baruah, 2014; Okolie et al., 2021a, 2021b). model learns how to map input data to its target (Guan et al., 2022). The
Parameters Conventional gasification Hydrothermal gasification resulting model determines the output predictions for a specified input.
In addition, the predicted output results are compared with the actual
Moisture contents Requires drying when Suitable for feedstocks with
of the feedstock with more than moisture content up to 99% outcome. The model is optimized by minimizing the actual and pre
feedstocks 35 wt% moisture content is dicted output differences. It should be emphasized that the choice of a
used. desirable supervised learning method should be based on the size and
Gasifying agents Gasifying agents could be Gasifying agent used is feature of the available database (Guan et al., 2022).
oxygen, steam, air, or CO2 supercritical water (Water at
P > 22.1 MPa, and T > 374 ◦ C)
Supervised learning methods could be further divided into regression
The reaction The temperature range is Temperature and pressure is and classification methods, as shown in Fig. 2. Although both algorithms
temperature between 500 and 1400 ◦ C. above the critical point of deal with labelled datasets and are often used in ML predictions, they are
and pressure Most gasification studies are water (P > 22.1 MPa, T > pretty different in how they are applied to solve ML problems. While the
performed at atmospheric 374 ◦ C)
classification methods attempt to determine behavioural patterns be
pressure. However, there
are some cases of tween datasets, the regression methods aim to find the best fit correla
pressurized gasification. tion between the dependent and independent variables (Wang et al.,
Processing time 30–60 min for bench-scale 15–60 min for batch reactors. 2021).
studies. 30–90 s for continuous In the unsupervised ML, only input data is fed into the model; the
reactors
model learns by identifying similarities and differences in the input data
3
Table 2
Studies related to the conventional and hydrothermal gasification of waste biomass.
Feedstock Gasification reaction conditions Gasifying Key findings References
agent
Municipal solid waste • Temperature: 700–900 ◦ C Steam • Temperature had the most significant effect Luo et al.
(mixture of five different components • Steam/carbon ratio (S/C): 2.41 on catalytic steam gasification. (2012)
including kitchen garbage, paper, and • Catalyst: NiO/γ-Al2O3 and calcined
textile, wood, plastic) dolomite
• Reactor type: Fixed – bed. • NiO/γ -Al2O3 showed a better catalytic
performance than calcined dolomite.
• The addition of steam increased the gas yield

and carbon gasification efficiency (CGE).
• Maximum gas yield (1.75 Nm3/kg) and

carbon gasification efficiency (84.87%) was
obtained with a S/C of 2.41
Raw and torrefied starchy food waste • Temperature: 600–1000 ◦ C Steam • Increasing reaction temperature beyond Huang et al.
• Steam/carbon ratio (S/C): 2 700 ◦ C led to a reduction in tar formation. (2019)
*Torrefaction was performed at • Reactor type: Vertical fixed – bed.
240–300 ◦ C for 30 min,
• CGE of torrefied starchy food waste rose from
16.7% at 600 ◦ C to 91.1% at 1000 ◦ C.
• CGE of torrefied food waste is higher than the

raw waste.
Wheat straw and lignite • Temperature: 950 ◦ C Air, oxygen • The addition of the desired amount of CaO Liu et al.
• steam/mixed feedstocks (S/MF) ratio: 4 could improve the chemical looping co- (2022)
*Co-gasification of wheat straw and lignite • CaO/mixed feedstocks (CaO/MF) ratio: gasification performance.
(1:1), 0.75
• Reactor type: chemical looping co-
gasification with iron-based oxygen • The addition of CaO also improved the
carrier in a fixed bed reactor. stability of the iron-based oxygen carrier.
• Chemical looping co-gasification helps to

reduce the net CO2 emission from coal and
biomass gasification.
Carbonaceous waste (Industrial sludge, • Temperature: 1490 K and 5 kW power. Steam • Gaseous product with molar ratios of H2/CO Piatkowski
Sewage sludge, Scrap tire powder, Fluff • Radiative flux concentrations up to = 1.5 and CO2/CO = 0.2 was produced. et al. (2009)
South, African coal, Beech charcoal) 2953 suns • The conversion efficiency of solar-to-chemical
• Reactor type: Packed bed solar reactor energy ranges from 17.3% to 29%.
Sawdust • Temperature: 700 ◦ C Steam, CO2 • The results showed that the water-gas shift Salaudeen
• Reactor type: fluidized bed reactor with reaction is the most influential gasification et al. (2020)
calcined eggshell (CES) as bed material reaction.
• The use of CES led to the depletion in the

concentration of CO2 and, at the same time, a
considerable rise in hydrogen yield.
Empty fruit bunch (EFB) • Temperature: 873–973 K Steam • The hydrogen yield also increases with Inayat et al.
• Steam to biomass ratio: 1.4–2.2 increased steam to biomass ratio and (2021)
• Sorbent to biomass ratio: 0.7–1.1 temperature.
• Reactor type: Steam gasification
fluidized bed reactor with in-situ CO2
capture. • The maximum H2 yield produced from the
gasifier was 127 g of H2/kg
Biomass (wood as saw dust, straw) and • Temperature: 600 C ◦
Supercritical • The addition of KOH and K2CO3 led to the Schmieder
wastes (sewage sludge and lignin) • Pressure 25 MPa. water complete gasification of all biomass et al. (2000)
• Catalyst: KOH and K2CO3 compounds.
• The addition of potassium decreases the CO
concentration and increases CO2 and H2 in the
product gas.
• The CO, CH4 and C2–C4 compounds are low in
the product gas.
Chicken manure • Temperature: 600 ◦ C Supercritical • Temperature plays a significant role in the Cao et al.
• Pressure 25 MPa. water formation of gaseous products. (2016)
• Catalyst: Activated carbon • Maximum H2 yield of 25.2 mol/kg was
• Reactor type: Fluidized bed attained at 600 ◦ C with activated carbon as
catalyst.
4
Fig. 2. Machine learning classification.
and classifying data accordingly. It should be stated that unsupervised 3.1. Brief overview of the common ML algorithms used for gasification
ML deals mainly with unlabeled data and is very suitable to identify research
hidden patterns. On the other hand, semi-supervised learning helps to
address the issues of supervised learning and can deal with systems Several ML algorithms have been employed in hydrothermal and
containing both unlabeled and labelled data (Guan et al., 2022). Often conventional gasification research. The widely explored algorithms in
the labelling of data for supervised learning is challenging and time- this research field include Decision trees, Random Forest, Support
consuming. However, semi-supervised learning can address the limita Vector Machines (SVM), Gradient boosting and neural network algo
tions by ensuring that the model combines a fraction or the overall rithms such as artificial neural networks (ANN) and deep neural net
unlabeled data into the supervised learning process (Guan et al., 2022; works (DNN). The subsequent section herein provides a brief description
Sahoo et al., 2020; Selvaratnam and Koodali, 2021). of these algorithms.
Reinforcement learning involves a learning system called an agent.
The agent keeps learning its environment better through a series of ac 3.1.1. Decision trees (DT)
tions that could give it reward or penalty. A positive action yields re These ML algorithms have been vastly used to solve real-world
wards, and negative action a penalty. The agent knows its environment problems, owing to their ability to handle complex datasets without
better without human intervention by taking the decision that maxi preprocessing. The unique ability of the tree-based model in solving
mizes profit and reduces penalties (Selvaratnam and Koodali, 2021; linear and nonlinear problems and easy understanding and interpreta
Taoufik et al., 2022). tion of the model makes them more relevant than other black box ML
Overall, employing the ML process involves five distinct steps out algorithms (Doerr et al., 2013). These methods employ recursive binary
lined below: splitting of input data (features) into several smaller sub-groups to
1. Defining the problem: This involves breaking down the ML task minimize the actual and predicted value variation. A tree structure that
into supervised or unsupervised learning based on the data at hand. could perform classification or regression tasks is then built. The depth
2. Building the dataset: This is an essential step in developing the of the tree structure can be tuned to improve the performance of DT.
ML model. An in-depth understanding of data will inform the choice of Moreover, the DT algorithm produces an optimized outcome dependent
model and algorithm. on the tree-structured selection process.
3. Model training: It is expedient to split the dataset before training Three main components are inherent in the DT algorithm; decision
the model. In most cases, the training dataset takes the majority of the nodes (helps with feature storage), decision links (helps in criterion
data, say 80%, while testing data take less (20%). Just as training data is selection) and decision leaves (helps in classification) (Guan et al.,
used to train the model, test data are used for model evaluation. 2022). Although the DT algorithm is preferred for large datasets, it tends
4. Model evaluation: Several evaluation metrics are tailored to the to develop complex tree structures as data size increases, resulting in
use case. Those metrics are used to determine the accuracy and reli overfitting (Kannangara et al., 2018; Taoufik et al., 2022). Another issue
ability of the ML models (Betiku et al., 2016). A detailed description of with basic DT is the difficulties in capturing noise in the dataset and the
different ML model evaluation metrics is presented in Table 3. lack of smoothness of their respective functions (Ascher et al., 2021).
5. Model deployment: After the evaluation of the model has given a Although, multivariate adaptive regression splines have addressed this
satisfactory result, the model is ready to solve real problems that it has limitation (Hastie et al., 2009b). Other methods such as ensemble and
not been trained with. The result yielded here is then compared with the boosting have been employed to enhance individual DT, and they will be
expected result to confirm the model's efficiency. briefly discussed in subsequent sections.
5
Table 3 Table 3 (continued )

Description of different statistical criteria used to determine the performance, Statistical criteria Description
accuracy, and reliability of ML models (Afolabi et al., 2020a; Aghbashlo et al.,
2021; Betiku et al., 2016; Gal and Ghahramani, 2016; RPubs, 2015; Smritis, √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
2019). 1 ∑n
RMSE = (actual − predicted values)2
n i=1
Statistical criteria Description Standard deviation Describes the variability of the dataset with respect to the
(SD) mean. SDE values close to zero means that the calculated
Absolute error (AE) Shows the absolute difference between the actual and
value is closer to the mean of the dataset.
predicted output.
Note: Absolute means that a result with a negative sign is √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
∑
ignored. (actual − mean)2
SD =
n
Standard error (SE) SE is used to determine the precision of the calculated
AE = |actual − predicted values|
sample mean and therefore it is dependent on the sample
Coefficient of A statistical parameter that describes the proportion of
size. Increasing samples size leads to a decline in SE.
determination (R2) variance in the dependent variable that the independent
variable can explain. A value closer to 1 indicates a better
SD
model performance. SE = √̅̅̅
n
R2=1 - (sums of squares error/sums of squares total)
Adjusted coefficient of An R2 value is used to account for a model's goodness of fit
determination (Adj by quantifying the variability of a parameter relative to 3.1.2. Random forest (RF) and gradient boosting (GB)
R2) another. A value closer to 1 indicates a better model RF may be considered an evolved version of the DT algorithm. RF is
performance. an ensemble learning algorithm that overcomes the overfitting limita
[ ] tion associated with DT by averaging the smaller decision trees rather
( n− 1
)
Adj R2 = 1 − 1 − R2 ×
n− k− 1
than using a single massive DT (Auret and Aldrich, 2012; Scornet et al.,
Mean absolute It shows the average distance between an observed value 2015). The algorithm exploits random feature bagging and selection
deviation (MAD) and the mean. Larger values of MAD show that the data strategy to ensemble DT and provides average prediction. It learns
spread out from the mean.
simultaneously on several constructed DTs trained on subsets of original
1 ∑n data, then average predictions when data is quantitative or use it as a
MAD = |Actual output − mean actual output|
n i=1 vote for qualitative data. Depending on the problem, the number of
Mean absolute error Expressed as the absolute sum of the differences between
decision trees that run in parallel is an important parameter that in
(MAE) the actual and predicted output divided by the total
number of outcomes. Not sensitive to outliers and also not fluences the algorithm's performance (Nguyen et al., 2021). The
suitable for models with significant errors. Lower value maximum number of variables in each tree (Nvar) and the number of
indicated better model performance. decision trees (n) are the most predominantly applied parameters out of
several RF parameters. The major drawback of DT in handling nonlinear
1 ∑n
MAE =
n i=1
|actual − predicted values| input features and the high risk of overfitting has been enhanced in the
Mean absolute Measures the accuracy of a model in percentage. Performs RF algorithm (Vigneau et al., 2018). Fig. 3 illustrates the DT and RF
percentage error better if the data has no extremes and no zeros. methodology.
(MAPE)
⃒ ⃒ Gradient Boosting (GB) is an ensemble ML technique that trains
1 ∑n ⃒⃒Actual − predicted value ⃒⃒
MAPE = ⃒ × 100 using boosting strategy (Noori et al., 2009). A series of weak prediction
n i=1 ⃒ Actual value
Mean error (ME) Refers to the differences between the actual and predicted models based on chain rule is integrated with a final robust model with
output divided by the total number of outcomes.
good prediction ability. GB leaf is obtained by finding the average of
1 ∑n observed values for output as an initial value to calculate initial errors.
ME = (Actual − Predicted values)
n i− 1 GB then builds the first tree to predict this error. The new tree with a
Mean Squared Error This is defined as the average of the square of the
learning rate then calculates the new prediction value from the previous
(MSE) difference between the actual and predicted values. Very
useful when the dataset contains outliers or unexpected prediction value and predicted error. The next tree is developed based
values. Lower value indicated better model performance. on the new residual prediction value from the previous tree and
observed value until there is no noticeable decrease in the residue. The
1 ∑n essential hyperparameter needed to be tunned during the modeling
MSE = (actual − predicted values)2
n i=1
Normalized root mean Relates the RMSE to the recorded variable range. Often process includes many trees or the number of boost stages and the
squared error used to compare models with different scales. maximum depth of the tree (Cai et al., 2020; Rzychoń et al., 2020).
(NRMSE) Overall, DT methods appear easier to implement when compared
RMSE
NRMSE = with other ML methods. The algorithm's nature ensures that missing
Average of observation values
Relative absolute error Often used to measure the performance of a predictive predictors are handled easily while data pre-processing steps can be
(RAE) model. Expressed as the ratio of the residuals to the error performed easily. In addition, the categorical variables can be used
obtained from the model.
directly without transforming into numerical variables (Auret and
[
∑n
]1/2 Aldrich, 2012).
I=1 (Predicted − actual values)2
RAE = [
∑n
]1/2 3.1.3. Support vector machine (SVM)
Actual value2
i=1 This robust machine learning algorithm can perform both classifi
Relative percentage It is defined as the mean absolute error ratio to the
cation and regression in linear and nonlinear problems with minimal
error (RPE) variable's mean value expressed in percentage.
risk (Bhagat et al., 2020; Hlihor et al., 2015). The algorithm finds the
Mean absolute error best boundary separating input features into two categories, called bi
RPE = × 100
Mean value of observation nary classification. A subgroup of the original dataset forms the
Root mean square error Same as the MSE but with a square root. It is beneficial for
(RMSE) models with significant errors that could drastically affect
boundary called support vectors, maximizing the distance between the
their performance. It does not consider the absolute value two classes (Fig. 4).
of the error. Therefore it is a very useful parameter. Lower SVM algorithm is used for handling nonlinear, multidimensional,
value indicated better model performance. small dataset sizes with local minima. It employs different nonlinear
mapping functions such as sigmoid, polynomials, radial basis functions
6
Fig. 3. An overview of the decision tree (DT) and random forest (RF) methodology. Individual decision trees are represented with the green box, while the black
border represents the RF methodology. The highlighted branches indicate the path that provides a tree's prediction for a specified output. Adapted from Ascher et al.
(2021) with permission from Elsevier. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
Fig. 4. Sample representation of the support vector machine.
(RBF) for solving problems. In support vector regression (SVR), the 3.1.4. Artificial neural network (ANN)
machine seeks to reduce the separating distance between the two classes Artificial neural network (ANN) is one of the most sophisticated ML
(maximum-margin hyperplane) to find the regression line (Noori et al., algorithms that identify complex relationships between input and
2009). Fig. 4 describes the SVM wherein support vectors are used to output data. It is a mathematical model inspired by the principle of
develop a hyperplane. It should be noted the support vectors indicates biological neurons in solving complex problems (Vasseghian et al.,
data points from either class that is nearest to the hyperplane. 2020). ANN structure comprises a densely interconnected adaptive
7
processing unit known as artificial neurons or nodes that could perform variable is multiplied with a random weight value added to a bias. Then
massive parallel computing for data processing and knowledge repre the output is passed into an activated function that transforms each
sentation. Several unique features of ANN have made it an attractive ML node's output and turns them into input features for the next layer. The
algorithm for predictive tasks. ANN is a self-adaptive data-driven model output of this layer is repeatedly turned into the input feature of the
that does not work based on assumption. In addition to this, ANN has a successive layers; this continues until the final layer is reached. At the
good generalization ability in predicting the output of the new dataset end of the training process, the final weight and bias computed in the
closely to the actual value, given the model quality data and optimal trained model are then used to predict the output of new data fed into
parameters. Also, ANN has been a universal function approximator that the neural network. The advantages and limitation of the described ML
enhances continuous approximation on any complex or nonlinear algorithms are outlined in Table 4.
problem to a higher degree of accuracy (Shenbagaraj et al., 2021).
ANN architecture comprises three essential layers: the input layer, 4. Application of ML technology in conventional and
which contains input features, the hidden layer, which includes hidden hydrothermal gasification
neurons and the output layer, which has the predicted target. Since no
exact formula decides the optimal ANN architecture and algorithm to ML has been applied in gasification for syngas yield prediction,
solve a particular problem, this decision is often made via the trial-and- parametric studies, process modeling, quantifying the char and tar for
error method. The search for optimal architecture could start from a mation as a function of process parameters, and catalysts development.
simple network structure to a more complex structure until the gap This section reviewed and meticulously outlined studies in this area to
between actual and predicted values, i.e., the error, is minimized (Afo clarify the application of the ML approach to advance the field.
labi et al., 2020b).
3.1.5. Deep learning 4.1. Machine learning applications in conventional gasification

ML can also be grouped as shallow learning and deep learning.
Shallow learning occurs in a machine when a nonlinear conversion layer 4.1.1. Parametric studies and syngas yield prediction
(i.e., the hidden layer) does not exceed two layers. Some of the shallow ML studies are often applied to predict the yield and composition of
learning structures include SVM (support vector machines) (Ingo, 2008), syngas and to study the impact of different process parameters (tem
logistic regression (Kleinbaum and MK, 2006), Gaussian mixture model perature, Steam/carbon ratio, reaction time and feed particle size) on
(Verbeek et al., 2003) and more. The limit in the number of conversion the yield. ANN is one of the most common ML methods used to model
layers in shallow learning makes solving problems with multiple re biomass gasification due to its flexibility in understanding the non-
strictions efficient. Still, it limits its capacity to handle complex real- linear relationships between input and output variables during gasifi
world challenges such as voice recognition. Therefore, shallow cation (Ascher et al., 2021). Experimental data are often used to train the
learning cannot be handled like the brain in information processing. ANNs, while some of the metrics outlined in Table 3 are used to verify
Deep learning is a logistic regression-based model containing mul the model's accuracy.
tiple conversion layers and an activation function. Even though its Earlier studies applied two ANN architectures to model and predict
principle is based on linear regression, it differs from traditional statis syngas composition, including CO, CO2, H2, and CH4, during biomass
tical regression with its multiple neural nodes, also called neural net gasification (circulating and bubbling fluidized bed gasifier) (Puig-
works. Each of the classifying nodes is called perceptron or neural unit. Arnavat et al., 2013a). Although the data used for the models were small
Deep learning architecture is built on interconnecting hidden layers (18–36 experimental runs), the model achieved high predictive accuracy
between input and output layers. The nodes in the hidden layer are with R2 > 97 (Puig-Arnavat et al., 2013a). In addition, the model was
referred to as hidden nodes. The unique feature of deep learning in also used to study the relative effect of input variables (biomass
generating complex hypotheses automatically makes it a powerful tool composition and equivalent ratio) on the producer gas composition.
for learning complex nonlinear relationships. With this unique feature, Both variables were found to affect the producer gas composition
deep learning could solve problems that shallow learning models could significantly.
not solve (Verbeek et al., 2003). In another study, the ANN model showed high accuracy in predicting
The deep neural network (DNN) is an extension of an artificial neural syngas yield and composition during biomass gasification in a fixed–bed
network with multiple hidden layers. A shallow network and a simple gasifier (Mikulandrić et al., 2014). The model was also used to study the
four-layered DNN architecture are shown in Fig. 5. From the left, in effect of reaction time, air flow and temperature on syngas yield and
dependent variables (input features) are fed into the network, and each composition (Mikulandrić et al., 2014). Although the described ANNs
models showed promising results, they cannot be applied to feedstocks
Fig. 5. Comparison between the simple and deep neural network architecture.
8
Table 4
Advantages and limitations of selected ML algorithms used in conventional and hydrothermal gasification studies.
Machine learning Advantages Limitations References
algorithm
Decision Tree • They are suitable for both classification and regression problems. • As the tree structure in DT grows larger, it is prone to Hastie et al.
(DT) overfitting. (2009a)
Schmidhuber
• They are simple to implement than other ML algorithms. (2015)
• The hierarchical nature of this algorithm leads to
high variance.
• Data requires minimal preprocessing as missing predictors are easily
handled.
• A larger number of trees increases computational
demands.
• It performed well on categorical predictors as they need not be turned into
numerical values.
• The algorithm lacks smoothness in its underlying
function and finds it difficult to capture data in
• It allows users to extract feature importance with relative ease compared to datasets.
ANN, which is considered a black box.
Random Forest • It improves on individual decision tree by creating an ensemble of decision • The larger the number of trees used in the RF Breiman (2001)
(RF) trees algorithm, the computational demands increases. Hastie et al.
(2009a)
• It trains each tree on fractions of randomly chosen data • A small change in the data has considerable changes
on the algorithm
• It is an ensemble method that uses a vast number of trees to overcome

overfitting problems in decision trees • It is prone to overfitting, hyperparameter tuning is
required
• It is robust to overfitting as the ensemble methods have in-built regulari

zation mechanisms • The final model is challenging to interpret.
• RF has fewer hyper parameters and performs well even for sub-optimal
hyper-parameter choices.
Gradient • GB is an ensemble method that overcomes the overfitting problems in • GB tuning is challenging Hastie et al.
Boosting (GB) decision trees by using shallow trees, which are small trees with a low depth (2009a)
called weak learners, instead of a large number of trees used by random
forest • It is prone to overfitting
• When well-tuned, it performs better than random forest. • Models are computationally expensive as they take a
longer time to train
• It computes second-order gradients of the loss function to minimize loss and

advanced regularization (L1 and L2) • Final models are challenging to interpret.
• It takes more detailed approximations into account when determining the

best model
• They train faster on large datasets and fast to interpret

Support Vector • SVM is well known for binary classification but well suited for classification • It is not suitable for large datasets Noori et al. (2009)
Machine and regression tasks. Serrano and
(SVM) Castelló (2020)
• It does not perform well with noisy datasets
• They show high prediction accuracy for data set of limited size in real-world
applications.
• SVM will not perform well When the number of
features for each data point exceeds the number of
• It is a good option for a high-dimensional dataset. training sample
• They have low sensitivity to outliers resulting from the role of support
vector in constructing optimal plane hyperplane.
• It is relatively memory efficient

Neural Networks • This is a powerful tool for modeling complex nonlinear processes. • The black box nature and lack of interpretability Noori et al. (2009)
have been significant drawbacks in using this Serrano and
algorithm. Castelló (2020)
• It requires less statistical training to develop.
(continued on next page)
9
Table 4 (continued )
Machine learning Advantages Limitations References
algorithm
• It can easily map all possible interactions between predictor variables • If not carefully observed, it is susceptible to
overfitting.
• It requires more significant computational resources

Deep learning • Compared with other traditional machine learning methods, it performs • Hyper parameter tuning required when using this Sain (1996)
better in several cases, particularly learning from large datasets. model makes it computationally expensive Schmidhuber
(2015)
• They could perform automatic feature extraction with the help of their
multiple hidden layers.
and experimental conditions outside the study range. step. A summary of previous studies that used ML to predict syngas yield
Baruah et al. (2017) developed ANN model by using the reduction and composition during conventional gasification is presented in
zone temperature and C, H, O, ash, moisture content data from different Table 5.
woody biomass as input for the ANN to address the limitations of the
feedstock types model (Baruah et al., 2017). The developed model was 4.1.2. Prediction of solid residues yield
able to predict syngas composition as a function of the input parameters. The formation of solid residues, including ash, char and tar during
The output showed good agreement with experimental results (R2 range biomass gasification, has been a long-standing issue that restricts its
from 0.98–0.99) (Baruah et al., 2017). A similar study used the multi- large-scale implementation. Char and tar contain hydrocarbon mixtures
layer feed-forward neural network with Levenberg–Marquardt back- that could potentially create severe problems such as clogging, surface
propagation algorithm to predict syngas yield and study the influence deposition and catalysts deactivation (Klinghoffer et al., 2015). More
of temperature and equivalence ratio on syngas composition (George over, the accurate prediction of char yield could help develop an
et al., 2018). Data from five different waste biomass were used for the effective char mitigation strategy. ML methods have been applied for
ANN model to address the lack of generalization in ANN models re modeling and predicting char yield during gasification by some authors
ported in previous studies. The model showed high predictive accuracy (Pandey et al., 2016; Serrano and Castelló, 2020). It should be empha
with an R2 value of 0.9. sized that ML modeling of char is scarce due to the limited data avail
Some other supervised ML algorithms have also been applied to ability and the heterogeneity of the available data in the literature
study biomass gasification. Elmaz et al. (2020) compared the perfor (Ascher et al., 2021).
mance of four regression algorithms in predicting syngas yield and ANN was used to model and predict char deposition during woody
composition during biomass gasification in a downdraft gasifier (Elmaz biomass gasification in a bubbling fluidized bed reactor (Serrano and
et al., 2020). Multilayer perceptron (MLP) and DT regression methods Castelló, 2020). Different experimental tar sampling, collection, analysis
performed significantly better than other regression methods (SVR and methods, and data from literature were used to train the model. The
Polynomial Regression - PR) in predicting CO, CO2, CH4, H2 and higher developed tar model showed good predictive capability (R2 > 0.97) for
heating value (HHV) outputs of the downdraft biomass gasification. the test and validation. In addition, the relative errors were less than
Elmaz and Yücel (2020) performed a comparative study on the perfor 20% for most of the tested samples (Serrano and Castelló, 2020).
mance of a simpler polynomial regression model and non-linear autor Parametric studies with the proposed ANN model reveal that trend of tar
egressive with exogenous neural networks (NARXNN) in predicting yield with temperature and equivalent ratio (ER) is similar to experi
syngas composition. The NARXNN performed better with R2 > 0.98 mental values reported in the literature.
compared to the polynomial model (R2 > 0.80) (Elmaz and Yücel, In another study, multiple-input and single-output (MISO) and
2020). Another study compares the predictive performance of several multiple-input and multiple-output (MIMO) ANN models, trained with
regression models such as K Nearest Neighbours (KNN) Regression, the Levenberg-Marquardt backpropagation algorithm was used to
Linear Regression (LR), Support Vector Machine Regression (SVMR) and develop a gasification predictive model. The model was based on several
Decision Tree Regression (DTR) in predicting syngas composition and its input parameters such as C, H, N, S, O, moisture, ash content, temper
heating value (Ozbas et al., 2019). All the models showed good agree ature, and ER (Pandey et al., 2016). The output parameters include the
ment with R2 values >0.99. LHV of syngas and tars. The predictive performance of the ANN-based
ANN methods are often considered a “black box” by some re models showed good agreement with experimental values, and the
searchers. Therefore its data interpretability is limited (Ozbas et al., MSE is very low. Moreover, the models demonstrated R2 > 0.99 for tar
2019; Pandey et al., 2016; Serrano and Castelló, 2020). This means that and syngas LHV value prediction (Pandey et al., 2016).
it is often unclear to explain the fundamental mechanism underlying the
relationships between inputs and output. As a result, some studies
integrating thermodynamics and ANN has been performed (Guo et al., 4.2. Application of ML technology in hydrothermal gasification
2001; Safarian et al., 2020). Thermodynamic modeling combined with
ANN allows data generation flexibility and the study of the effect of 4.2.1. Prediction of hydrogen yield, parametric studies, and process
different parameters on the syngas yield (Umenweke et al., 2021). optimization
Moreover, the integrated approach is advantageous because of its Hydrothermal gasification is a relatively new technology compared
robustness, simplicity, and convenience. to conventional gasification. Therefore studies related to the application
Predicting the evolution of syngas yield with time is also very of ML for product yield prediction, parametric studies and optimization
important for process scale-up and techno-economic studies. The are limited, as outlined in Table 5. ML models such as ANN, SCM
NARXNN has been used to study and predict syngas composition with regression methods and gradient boosting regression (GBR) have been
time (Elmaz and Yücel, 2020). The model uses the exogenous time series used to predict H2 yield and studies the effect of several parameters such
and incorporates a feedback mechanism. The latter permits the learning as temperature, pressure, feed concentration and reaction time on H2
of time – series by employing the input and output pairs in a repetitive yield and selectivity during hydrothermal gasification (Li et al., 2020;
Zhao et al., 2021).
10
Table 5 Table 5 (continued )

Previous studies on the applications of machine learning in conventional and Feedstock ML method Predicted References
hydrothermal gasification. parameters
Feedstock ML method Predicted References Multi-layered feed- Syngas yield and Shenbagaraj
parameters forward back- composition et al. (2021)
Woody biomass ANN (R2 > 0.99 for Syngas Baruah et al. propagation
CH4 yield and CO; composition (CO, (2017) algorithm-based
R2 > 0.98 for CO2 CO2, H2, CH4) artificial neural
and H2 yield.) network (FFBPNN)
Simulated ANN and Syngas Safarian et al. models(R2 > 0.98)
lignocellulosic thermodynamic composition (CO, (2020)
biomass models (R2 > 0.99) CO2, H2, CH4)
Woody biomass ANN, R2 > 0.97 Syngas and Tar, Serrano and Li et al. (2020) applied the gradient boosting regression (GBR) model
LHV Castelló with optimal hyper-parameters for a non-catalyst aided hydrothermal
(2020) gasification process. The model was used to predict syngas yield and
Woody biomass ANN (R2 > 0.930) Syngas yield, Yucel et al.
and Nonlinear composition (CO, (2019)
composition, including H2, CH4, CO2, and CO (Li et al., 2020). The
autoregressive CO2, H2, CH4,) model showed good prediction capability with R2 values of 0.92, 0.90,
exogenous (NARX) and heating value. 0.95, and 0.92 for H2, CH4, CO2, and CO prediction, respectively (Li
(R2 > 0.999) et al., 2020). Although promising results were attained, the effect of
Woody biomass SVR, polynomial Syngas yield, Elmaz et al.
biomass properties on the gaseous yields was not studied.
regression |(PR), composition (CO, (2020)
decision tree CO2, H2, CH4,) The limitation of biomass properties was addressed in another study.
regression (DTR) and and heating value. The H2 yield predictive performance of four ML methods (ANN, GPR,
multilayer SVM and RF) during hydrothermal gasification was compared (Zhao
perceptron (MLP) et al., 2021). The model was also used to evaluate the influence of
Biomass (Olive pit) Linear Regression Syngas yield, Ozbas et al.
(LR), K Nearest composition (CO, (2019)
biomass properties (C, H, O contents) on H2 yield and the exergy effi
Neighbours (KNN) CO2, H2, CH4,) ciency. All the assessed models displayed promising predictive capa
Regression, Support and heating value. bility (R2 > 0.98). The proposed ML models could be used for future
Vector Machine design and optimization of hydrothermal gasification process.
Regression (SVMR)
Gopirajan et al. (2021) proposed a ML method based on the Tunable
and Decision Tree
Regression (DTR) Decision Support System (TDSS) and Tunable Recommendation System
(R2 > 0.999 for all (TRS) for the optimization of reaction conditions during the gasification
models) of waste biomass in SCW. The TDSS uses the adaptive Multivariate
Palm kernel shell ANN (R2 > 0.998) Syngas yield and Shahbaz et al. Random Forest (AMVRF) and Adaptive Weighted Rank Aggregation
(using CaO composition (CO, (2019)
adsorbent and CO2, H2, CH4)
(AWRA) module for classification and feature selection, respectively
coal bottom ash (Gopirajan et al., 2021). The ML methods showed >94% accuracy for
as a catalyst) optimizing the process parameters (e.g., temperature, pressure, catalysts
Coal and biomass Gaussian process Total gas yield Mutlu et al. loading, reaction time, and solvent to biomass ratio) during hydro
regression (GPR), and carbon (2019)
thermal gasification. In another study, the multi-layered feed-forward
DT, ANN, linear AND conversion
PR. efficiency (CGE) back-propagation algorithm-based artificial neural network (FFBPNN)
Woody and Feed and cascade Syngas yield and Serrano et al. model was developed and used to study the effect of different factors
herbaceous forward back composition (CO, (2020) affecting syngas composition during the hydrothermal gasification of
biomass, sewage propagation ANN CO2, H2, CH4) food waste (Shenbagaraj et al., 2021). The model achieved a prediction
sludge,
municipal solid
accuracy greater than 98% and an MSE value less than 0.3 in predicting
waste, plastics the influence of feed concentration, reaction temperature and time on
Municipal solid Optimized ensemble Syngas yield, Kardani et al. syngas composition (Shenbagaraj et al., 2021). Similar studies by Zhang
waste methods based on composition and (2021) et al. (2021) compared the predictive performance of the traditional
decision tree (DT), LHV and LHV of
linear regression with the single-layer neural network and two-layer
extreme gradient other gasification
boosting (XGB), products. neural network during the hydrothermal gasification of waste biomass
random forest (RF), in SCW (Zhang et al., 2021). The single-layer neural network performed
multilayer better than the linear regression and two-layer neural network.
perceptron (MLP) Since hydrothermal gasification is a complex process, as noted
and support vector
earlier, the black-box nature of most ML methods makes it difficult to
regression (SVR).
Co – gasification Radial basis function H2 yield Ayodele et al. explain the fundamental mechanisms and the relationships between
(Plastics and (RBF) and multilayer (2021) input and output. To address the limitations, a novel – hybrid data-
rubber waste) perceptron (MLP) driven ML method was integrated with mechanistic modeling to study
MLP has an R2 of
the influence of process parameters on gas yield during hydrothermal
0.990
Several Gradient boosting Syngas Li et al. gasification (Li et al., 2021b). The GBR was first developed and used to
combinations of regression (GBR) composition (CO, (2020) predict H2 yield with an R2 > 0.90 and then combined with Particle
waste biomasses model for non- CO2, H2, CH4) Swarm Optimization (PSO). The results were compared and validated
catalytic systems with mechanistic modeling from Aspen plus simulation (Fig. 6). Com
Microalgae ANN (MSE = 0.568 Bio methanol Fózer et al.
parison of ML with mechanistic models fosters the user confidence of the
biomass and and R2 ≥ 0.965) production from (2021)
wastes hydrothermal GBR – PSO models.
gasification
Lignocellulosic RF, Gaussian, ANN H2 yield Zhao et al. 4.2.2. Catalysts design, screening, and selection
biomass and SVR (2021)
Catalysts are often applied during hydrothermal gasification to
Food Wastes
improve the H2 yield and selectivity and decrease the heating re
quirements. Hydrothermal gasification catalysts could be homogeneous
11
or heterogeneous catalysts. While homogenous catalysts such as alkali interpretability. This way, the interconnection between input and target
oxide and carbonates are soluble in water and cannot be recycled or variables could be easily learnt. Comprehensive predictive models
reused, they significantly increase H2 yield (Okolie et al., 2021c). On the capable of predicting large dimensions of relevant factors should be
contrary, the heterogeneous catalyst can be recycled and reused. The developed to create more holistic models.
application of ML to the development of homogeneous and heteroge Data availability for ML-based gasification studies is also another
neous catalysts during hydrothermal gasification has been scarcely major issue. Moreover, performing hydrothermal and conventional
documented in the literature. Moreover, the recent advancement in the gasification experiments is very expensive and time-consuming due to
generation of catalysis–related experimental data and theoretical cal the severe temperature and pressure conditions (Okolie et al., 2020).
culations data has promoted the use of ML methods for catalysts selec Some studies have also relied on literature data that are characterized by
tion, design and screening for several chemical applications (Guan et al., inconsistencies. It should be emphasized that the quality of the ML
2022). Specifically, ML is often used for heterogeneous catalysts model depends on the data's quality and learning algorithm. Dataset size
screening (Gopirajan et al., 2021; Li et al., 2021b). of high quality, with relevant features that represent the problem to be
A recent study proposed a ML model for the screening and selecting modelled in sufficient dataset size, will yield a high-performing model.
catalysts during hydrothermal gasification of waste biomass. The au Therefore, accurate data collection strategies should be devised during
thors divided the total dataset into three subcategories of non-catalyst, ML studies.
alkali-metal catalyst, and transition-metal catalyst with the principal As detailed previously, studies related to developing efficient char
component analysis (PCA) (Li et al., 2021a). The developed model prediction ML-based models during hydrothermal and conventional
showed promising results in identifying and screening catalysts to gasification of waste biomass are relatively scarce in the literature. This
maximize H2 yield and decrease CO2 yield during hydrothermal gasifi is primarily due to the scarcity of char formation data. However, it is
cation of waste biomass. A similar study also used the ML methods based recommended that future studies adapt thermodynamic/process
on the TDSS to optimize heterogeneous catalyst loading during hydro modeling to generate a large char formation dataset. These datasets
thermal gasification (Gopirajan et al., 2021). The model showed that could be used to evaluate the mechanism of char formation and develop
catalysts loading positively affected the H2 yield. appropriate guidelines for minimizing char deposition.
The ANN method based on the Levenberg-Marquardt and Bayesian ML application in catalytic gasification is an area that requires
Regularization algorithms were used to model the NaOH catalyzed hy further studies. ML could be used to study the relationship between
drothermal gasification of waste biomass and to study the environ catalytic properties and performance (e.g., activity, selectivity, and
mental impact of the process. The ML-based model aided in the stability). ML-based studies, when implemented effectively, could foster
optimization and prediction of catalysts composition with R2 ≥ 0.965 the development of novel heterogeneous catalysts for biomass gasifica
(Fózer et al., 2021). Their studies showed that the application of alkali tion. Moreover, catalysts deactivation studies based on ML have not
catalyst increased the global warming potential of the process. There been performed. Catalyst deactivation during gasification is often
fore, it was suggested that future studies should consider the environ influenced by several catalytic properties as well as the reaction con
mental implications and performance of the catalysts during ditions. Modeling such systems with ML could be challenging. More
hydrothermal gasification. over, the dataset on catalyst deactivation during gasification is relatively
scarce.
5. Challenges and future directions Finally, ML models could be coupled to the process sensors to ach
ieve a fully integrated digital twin system. A digital twin is a simulation
As outlined in previous sections, most ML methods are referred to as or virtual representation of a process/system, but with a real-time
“black box.” Therefore, applying such methods in studying gasification functionality that enables a user or operator to optimally interact with
mechanisms and pathways is very difficult. Future studies should focus the process/system without being physically present (Yu et al., 2020).
on combining ML models with traditional modeling approaches such as In summary, ML should not be considered a universal panacea to all
kinetics studies to provide detailed information on reaction mechanisms challenges in gasification research. However, the technology could
and pathways. Some authors have suggested the use of feature permu complement the available technologies. Moreover, the ML approach
tations to evaluate how different input variables behave as well as their provides an excellent medium for controlling, monitoring, and opti
relevance (Ascher et al., 2021). Future work should be aimed at illu mizing syngas yield during gasification.
minating the black-box nature of the ML algorithm to enhance its
Fig. 6. Comparison of syngas yield and composition between the GBR- PSO method and the mechanistic modeling based on mechanistic modeling from Aspen plus
(Li et al., 2021b).
12
6. Conclusions Cai, J., Xu, K., Zhu, Y., Hu, F., Li, L., 2020. Prediction and analysis of net ecosystem
carbon exchange based on gradient boosting regression and random forest. Appl.
Energy 262, 114566. https://doi.org/10.1016/j.apenergy.2020.114566.
The present study comprehensively reviewed and outlined critical Cao, W., Cao, C., Guo, L., Jin, H., Dargusch, M., Bernhardt, D., Yao, X., 2016. Hydrogen
discussions related to the application of ML technology in conventional production from supercritical water gasification of chicken manure. Int. J. Hydrog.
and hydrothermal gasification research. The distinction between BTG Energy 41, 22722–22731. https://doi.org/10.1016/j.ijhydene.2016.09.031.
Castello, D., Rolli, B., Kruse, A., Fiori, L., Castello, D., Rolli, B., Kruse, A., Fiori, L., 2017.
technologies is first outlined. Additionally, different ML algorithms used Supercritical water gasification of biomass in a ceramic reactor: long-time batch
for gasification studies are briefly discussed. Specifically, the ANN, SVM, experiments. Energies 10, 1734. https://doi.org/10.3390/en10111734.
DT, and deep neural networks. The advantages and limitations of each Chen, J., Xu, W., Zhang, F., Zuo, H., EJ, Wei, K., Liao, G., Fan, Y., 2019. Thermodynamic
and environmental analysis of integrated supercritical water gasification of coal for
algorithm are also outlined. ML-based gasification research primarily power and hydrogen production. Energy Convers. Manag. 198, 111927. https://doi.
focuses on process optimization, parametric studies, prediction of char org/10.1016/j.enconman.2019.111927.
formation during conventional gasification and catalyst screening for Chen, J., Xu, W., Zuo, H., Wu, X., Jiaqiang, E., Wang, T., Zhang, F., Lu, N., 2019b. System
development and environmental performance analysis of a solar-driven supercritical
hydrothermal gasification. In the last section, challenges and sugges water gasification pilot plant for hydrogen production using life cycle assessment
tions were outlined to help advance ML–models' application to gasifi approach. Energy Convers. Manag. 184, 60–73. https://doi.org/10.1016/j.
cation research. enconman.2019.01.041.
Doerr, B., Kötzing, T., Lengler, J., Winzen, C., 2013. Black-box complexities of
combinatorial problems. Theor. Comput. Sci. 471, 84–106. https://doi.org/
10.1016/j.tcs.2012.10.039.
Declaration of competing interest Elmaz, F., Yücel, Ö., 2020. Data-driven identification and model predictive control of
biomass gasification process for maximum energy production. Energy 195, 117037.
The authors declare that they have no known competing financial https://doi.org/10.1016/j.energy.2020.117037.
Elmaz, F., Yücel, Ö., Mutlu, A.Y., 2020. Predictive modeling of biomass gasification with
interests or personal relationships that could have appeared to influence machine learning-based regression methods. Energy 191, 116541. https://doi.org/
the work reported in this paper. 10.1016/j.energy.2019.116541.
Fózer, D., Tóth, A.J., Varbanov, P.S., Klemeš, J.J., Mizsey, P., 2021. Sustainability
assessment of biomethanol production via hydrothermal gasification supported by
artificial neural network. J. Clean. Prod. 318, 128606 https://doi.org/10.1016/j.
Acknowledgement
jclepro.2021.128606.
Gal, Y., Ghahramani, Z., 2016. Dropout as a Bayesian approximation: representing model
The authors would like to thank everyone who contributed to the uncertainty in deep learning. In: 33rd Int. Conf. Mach. Learn. ICML.
manuscript's success, including the editors and reviewers. George, J., Arun, P., Muraleedharan, C., 2018. Assessment of producer gas composition
in air gasification of biomass using artificial neural network model. Int. J. Hydrog.
Energy 43, 9558–9568. https://doi.org/10.1016/j.ijhydene.2018.04.007.
References Gopirajan, P.V., Gopinath, K.P., Sivaranjani, G., Arun, J., 2021. Optimization of
hydrothermal gasification process through machine learning approach:
Experimental conditions, product yield and pollution. J. Clean. Prod. 306, 127302
Abdoulmoumine, N., Adhikari, S., Kulkarni, A., Chattanathan, S., 2015. A review on
https://doi.org/10.1016/j.jclepro.2021.127302.
biomass gasification syngas cleanup. Appl. Energy. https://doi.org/10.1016/j.
Guan, Y., Chaffart, D., Liu, G., Tan, Z., Zhang, D., Wang, Y., Li, J., Ricardez-Sandoval, L.,
apenergy.2015.05.095.
2022. Machine learning in solid heterogeneous catalysis: recent developments,
Afolabi, I.C., Popoola, S.I., Bello, O.S., 2020a. Modeling pseudo-second-order kinetics of
challenges and perspectives. Chem. Eng. Sci. 248, 117224 https://doi.org/10.1016/
orange peel-paracetamol adsorption process using artificial neural network.
j.ces.2021.117224.
Chemom. Intell. Lab. Syst. 203, 104053 https://doi.org/10.1016/j.
Guo, B., Li, D., Cheng, C., Lü, Z.A., Shen, Y., 2001. Simulation of biomass gasification
chemolab.2020.104053.
with a hybrid neural network model. Bioresour. Technol. 76, 77–83. https://doi.org/
Afolabi, I.C., Popoola, S.I., Bello, O.S., 2020b. Machine learning approach for prediction
10.1016/S0960-8524(00)00106-1.
of paracetamol adsorption efficiency on chemically modified orange peel.
Hastie, T., et al., 2009. Springer series in statistics the elements of statistical learning.
Spectrochim. Acta A Mol. Biomol. Spectrosc. 243, 118769 https://doi.org/10.1016/
Math. Intell. 27, 83–85.
j.saa.2020.118769.
Hastie, T., Tibshirani, R., JF, 2009. The Elements of Statistical Learning.
Aghbashlo, M., Peng, W., Tabatabaei, M., Kalogirou, S.A., Soltanian, S., Hosseinzadeh-
Hlihor, R.M., Diaconu, M., Leon, F., Curteanu, S., Tavares, T., Gavrilescu, M., 2015.
Bandbafha, H., Mahian, O., Lam, S.S., 2021. Machine learning technology in
Experimental analysis and mathematical prediction of Cd(II) removal by biosorption
biodiesel research: a review. Prog. Energy Combust. Sci. 85, 100904 https://doi.org/
using support vector machines and genetic algorithms. New Biotechnol. 32,
10.1016/j.pecs.2021.100904.
358–368. https://doi.org/10.1016/j.nbt.2014.08.003.
Ahmad, A.A., Zawawi, N.A., Kasim, F.H., Inayat, A., Khasri, A., 2016. Assessing the
Huang, J., Qiao, Y., Wei, X., Zhou, J., Yu, Y., Xu, M., 2019. Effect of torrefaction on steam
gasification performance of biomass: a review on biomass gasification process
gasification of starchy food waste. Fuel 253, 1556–1564. https://doi.org/10.1016/j.
conditions, optimization and economic evaluation. Renew. Sust. Energ. Rev. 53,
fuel.2019.05.142.
1333–1347. https://doi.org/10.1016/j.rser.2015.09.030.
Inayat, A., Khan, Z., Aslam, M., Shahbaz, M., Ahmad, M.M., Abdul Mutalib, M.I.,
Ascher, S., Watson, I., You, S., 2021. Machine learning methods for modelling the
Yusup, S., 2021. Integrated adsorption steam gasification for enhanced hydrogen
gasification and pyrolysis of biomass and waste. Renew. Sust. Energ. Rev. 111902
production from palm waste at bench scale plant. Int. J. Hydrog. Energy 46,
https://doi.org/10.1016/j.rser.2021.111902.
30581–30591. https://doi.org/10.1016/j.ijhydene.2020.12.008.
Auret, L., Aldrich, C., 2012. Interpretation of nonlinear relationships between process
Ingo, Steinwart, 2008. Support Vector Machines [WWW Document], 2008 Ed.
variables by use of random forests. Miner. Eng. 35, 27–42. https://doi.org/10.1016/
(September 15, 2008). Springer (accessed 12.31.21). https://books.google.com.mt
j.mineng.2012.05.008.
/books?hl=en&lr=&id=HUnqnrpYt4IC&oi=fnd&pg=PP7&dq=Support+vector+
Ayodele, B.V., Mustapa, S.I., Kanthasamy, R., Zwawi, M., Cheng, C.K., 2021. Modeling
machines&ots=g9fLEv_rQa&sig=afLocfpUCTEvzN-2Jwh7v5ZEPLQ&redir_esc=y#v
the prediction of hydrogen production by co-gasification of plastic and rubber wastes
=onepage&q=Supportvectormachines&f=false.
using machine learning algorithms. Int. J. Energy Res. 45, 9580–9594. https://doi.
Kang, K., Azargohar, R., Dalai, A.K., Wang, H., 2017. Hydrogen generation via
org/10.1002/er.6483.
supercritical water gasification of lignin using Ni-Co/Mg-Al catalysts. Int. J. Energy
Baruah, D., Baruah, D.C., 2014. Modeling of biomass gasification: a review. Renew. Sust.
Res. 41, 1835–1846. https://doi.org/10.1002/er.3739.
Energ. Rev. 39, 806–815. https://doi.org/10.1016/j.rser.2014.07.129.
Kannangara, M., Dua, R., Ahmadi, L., Bensebaa, F., 2018. Modeling and prediction of
Baruah, D., Baruah, D.C., Hazarika, M.K., 2017. Artificial neural network based modeling
regional municipal solid waste generation and diversion in Canada using machine
of biomass gasification in fixed bed downdraft gasifiers. Biomass Bioenergy 98,
learning approaches. Waste Manag. 74, 3–15. https://doi.org/10.1016/j.
264–271. https://doi.org/10.1016/j.biombioe.2017.01.029.
wasman.2017.11.057.
Betiku, E., Odude, V.O., Ishola, N.B., Bamimore, A., Osunleke, A.S., Okeleye, A.A., 2016.
Kardani, N., Zhou, A., Nazem, M., Lin, X., 2021. Modelling of municipal solid waste
Predictive capability evaluation of RSM, ANFIS and ANN: a case of reduction of high
gasification using an optimised ensemble soft computing model. Fuel 289, 119903.
free fatty acid of palm kernel oil via esterification process. Energy Convers. Manag.
https://doi.org/10.1016/j.fuel.2020.119903.
124, 219–230. https://doi.org/10.1016/j.enconman.2016.07.030.
Kleinbaum, D.G., MK, 2006. Analysis of matched data using logistic regression. Logist.
Bhagat, S.K., Tiyasha, T., Tung, T.M., Mostafa, R.R., Yaseen, Z.M., 2020. Manganese
Regres. 227–265. https://doi.org/10.1007/0-387-21647-2_8.
(Mn) removal prediction using extreme gradient model. Ecotoxicol. Environ. Saf.
Klinghoffer, N.B., Castaldi, M.J., Nzihou, A., 2015. Influence of char composition and
204, 111059 https://doi.org/10.1016/j.ecoenv.2020.111059.
inorganics on catalytic activity of char from biomass gasification. Fuel 157, 37–47.
Breiman, L., 2001. Random forests. Mach. Learn. 451 (45), 5–32. https://doi.org/
https://doi.org/10.1016/j.fuel.2015.04.036.
10.1023/A:1010933404324.
Li, J., Pan, L., Suvarna, M., Tong, Y.W., Wang, X., 2020. Machine learning prediction of
Brethauer, S., Studer, M.H., 2015. Biochemical conversion processes of lignocellulosic
syngas composition of hydrothermal gasification from wet organic wastes. Energy-
biomass to fuels and chemicals - a review. Chimia (Aarau) 69, 572–581. https://doi.
proceedings. 2020.
org/10.2533/chimia.2015.572.
Burkov, A., 2019. The Hundred-Page Machine Learning Book [WWW Document]. http://
www.mlebook.com/wiki/doku.php (accessed 12.29.21).
13
Li, J., Pan, L., Suvarna, M., Wang, X., 2021a. Machine learning aided supercritical water Pandey, D.S., Das, S., Pan, I., Leahy, J.J., Kwapinski, W., 2016. Artificial neural network
gasification for H2-rich syngas production with process optimization and catalyst based modelling approach for municipal solid waste gasification in a fluidized bed
screening. Chem. Eng. J. 426, 131285 https://doi.org/10.1016/j.cej.2021.131285. reactor. Waste Manag. 58, 202–213. https://doi.org/10.1016/j.
Li, J., Suvarna, M., Pan, L., Zhao, Y., Wang, X., 2021b. A hybrid data-driven and wasman.2016.08.023.
mechanistic modelling approach for hydrothermal gasification. Appl. Energy 304, Piatkowski, N., Wieckert, C., Steinfeld, A., 2009. Experimental investigation of a packed-
117674. https://doi.org/10.1016/j.apenergy.2021.117674. bed solar reactor for the steam-gasification of carbonaceous feedstocks. Fuel Process.
Liu, X., Hu, J., Zhao, S., Wang, W., Zhang, Q., Yan, X., 2022. Chemical looping co- Technol. 90, 360–366. https://doi.org/10.1016/j.fuproc.2008.10.007.
gasification of wheat straw and lignite with calcium-enhanced iron-based oxygen Pinkard, B.R., Gorman, D.J., Tiwari, K., Rasmussen, E.G., Kramlich, J.C., Reinhall, P.G.,
carrier for syngas production. Fuel Process. Technol. 227, 107108 https://doi.org/ Novosselov, I.V., 2019. Supercritical water gasification: practical design strategies
10.1016/j.fuproc.2021.107108. and operational challenges for lab-scale, continuous flow reactors. Heliyon 5,
Luo, S., Zhou, Y., Yi, C., 2012. Syngas production by catalytic steam gasification of e01269. https://doi.org/10.1016/j.heliyon.2019.e01269.
municipal solid waste in fixed-bed reactor. Energy 44, 391–395. https://doi.org/ Priddy, K.L., PK, 2005. Artificial neural networks: an introduction [WWW Document].
10.1016/j.energy.2012.06.016. https://scholar.google.com/scholar_lookup?title=Artificialneuralnetworksinche
Mandegari, M., Farzad, S., Görgens, J.F., 2018. A new insight into sugarcane micalengineering&publication_year=2017&author=ABasile&author=CurcioStefano
biorefineries with fossil fuel co-combustion: Techno-economic analysis and life cycle (accessed 12.29.21).
assessment. Energy Convers. Manag. 165, 76–91. https://doi.org/10.1016/j. Puig-Arnavat, M., Tora, E.A., Bruno, J.C., Coronas, A., 2013. State of the art on reactor
enconman.2018.03.057. designs for solar gasification of carbonaceous feedstock. Sol. Energy 97, 67–84.
Mikulandrić, R., Lončar, D., Böhning, D., Böhme, R., Beckmann, M., 2014. Artificial https://doi.org/10.1016/j.solener.2013.08.001.
neural network modelling approach for a biomass gasification process in fixed bed Puig-Arnavat, Maria, Hernández, J.A., Bruno, J.C., Coronas, A., 2013. Artificial neural
gasifiers. Energy Convers. Manag. 87, 1210–1223. https://doi.org/10.1016/j. network models for biomass gasification in fluidized bed gasifiers. Biomass
enconman.2014.03.036. Bioenergy 49, 279–289. https://doi.org/10.1016/j.biombioe.2012.12.012.
Mohammed, M.A.A., Salmiaton, A., Wan Azlina, W.A.K.G., Mohammad Amran, M.S., RPubs, 2015. Predictive R-squared according to Tom Hopper [WWW Document]. https
Fakhru’L-Razi, A., 2011. Air gasification of empty fruit bunch for hydrogen-rich gas ://rpubs.com/RatherBit/102428 (accessed 12.30.21).
production in a fluidized-bed reactor. Energy Convers. Manag. 52, 1555–1561. Rzychoń, M., Zogała, A., Róg, L., 2020. An interpretable extreme gradient boosting
https://doi.org/10.1016/j.enconman.2010.10.023. model to predict ash fusion temperatures. Minerals 10, 487. https://doi.org/
Mohri, M., Rostamizadeh, A., Talwalkar, A., 2018. https://books.google.com.au/books? 10.3390/min10060487.
hl=de&lr=&id=dWB9DwAAQBAJ&oi=fnd&pg=PR5&dq=. In: Mohri, Mehryar, Safarian, S., Ebrahimi Saryazdi, S.M., Unnthorsson, R., Richter, C., 2020. Artificial neural
Rostamizadeh, Afshin, Talwalkar, Ameet (Eds.), Foundations of Machine Learning, network integrated with thermodynamic equilibrium modeling of downdraft
second edition. Google Books [WWW Document]. MIT Press (accessed 12.29.21). biomass gasification-power production plant. Energy 213, 118800. https://doi.org/
https://books.google.nl/books?hl=en&l 10.1016/j.energy.2020.118800.
r=&id=dWB9DwAAQBAJ&oi=fnd&pg=PR5&dq=machine+lea Sahoo, P., Roy, I., Wang, Z., Mi, F., Yu, L., Balasubramani, P., Khan, L., Stoddart, J.F.,
rning&ots=AypSUStZl5&sig=NrllsejTu1MKz8lhi7zlwg-2GaM#v=onepage&q=ma 2020. MultiCon: a semi-supervised approach for predicting drug function from
chinelearning&f=false%250A. chemical structure analysis. J. Chem. Inf. Model. 60, 5995–6006. https://doi.org/
Mutlu, A.Y., Yücel, Ö., Elmaz, F., 2019. Evaluating the effect of blending ratio on the co- 10.1021/acs.jcim.0c00801.
gasification of high ash coal and biomass in a fluidized bed gasifier using machine Sain, S.R., 1996. The nature of statistical learning theory [WWW Document].
learning. Mugla J. Sci. Technol. 5, 1–15. https://doi.org/10.22531/ Technometrics. https://doi.org/10.1080/00401706.1996.10484565.
muglajsci.471538. Salaudeen, S.A., Acharya, B., Heidari, M., Al-Salem, S.M., Dutta, A., 2020. Hydrogen-rich
Nanda, S., Reddy, S.N., Mitra, S.K., Kozinski, J.A., 2016. The progressive routes for gas stream from steam gasification of biomass: eggshell as a CO2 sorbent. Energy
carbon capture and sequestration. Energy Sci. Eng. 4, 99–122. https://doi.org/ Fuels 34, 4828–4836. https://doi.org/10.1021/acs.energyfuels.9b03719.
10.1002/ese3.117. Schmidhuber, J., 2015. Deep learning in neural networks: an overview. Neural Netw. 61,
Nguyen, X.C., Ly, Q.V., Peng, W., Nguyen, V.H., Nguyen, D.D., Tran, Q.B., Huyen 85–117. https://doi.org/10.1016/J.NEUNET.2014.09.003.
Nguyen, T.T., Sonne, C., Lam, S.S., Ngo, H.H., Goethals, P., Le, Q.Van, 2021. Vertical Schmieder, H., Abeln, J., Boukis, N., Dinjus, E., Kruse, A., Kluth, M., Petrich, G., Sadri, E.,
flow constructed wetlands using expanded clay and biochar for wastewater Schacht, M., 2000. Hydrothermal gasification of biomass and organic wastes.
remediation: a comparative study and prediction of effluents using machine J. Supercrit. Fluids 17, 145–153. https://doi.org/10.1016/S0896-8446(99)00051-0.
learning. J. Hazard. Mater. 413, 125426 https://doi.org/10.1016/j. Scornet, E., Biau, G., Vert, J.P., 2015. Consistency of random forests. Ann. Stat. 43,
jhazmat.2021.125426. 1716–1741. https://doi.org/10.1214/15-AOS1321.
Noori, R., Abdoli, M.A., Ameri Ghasrodashti, A., Jalili Ghazizade, M., 2009. Prediction of Selvaratnam, B., Koodali, R.T., 2021. Machine learning in experimental materials
municipal solid waste generation with combination of support vector machine and chemistry. Catal. Today 371, 77–84. https://doi.org/10.1016/j.cattod.2020.07.074.
principal component analysis: a case study of Mashhad. Environ. Prog. Sustain. Serrano, D., Castelló, D., 2020. Tar prediction in bubbling fluidized bed gasification
Energy 28, 249–258. https://doi.org/10.1002/ep.10317. through artificial neural networks. Chem. Eng. J. 402, 126229 https://doi.org/
Okolie, J.A., Rana, R., Nanda, S., Dalai, A.K., Kozinski, J.A., 2019. Supercritical water 10.1016/j.cej.2020.126229.
gasification of biomass: a state-of-the-art review of process parameters, reaction Serrano, D., Golpour, I., Sánchez-Delgado, S., 2020. Predicting the effect of bed materials
mechanisms and catalysis. Sustain. Energy Fuels. https://doi.org/10.1039/ in bubbling fluidized bed gasification using artificial neural networks (ANNs)
c8se00565f. modeling approach. Fuel 266, 117021. https://doi.org/10.1016/J.
Okolie, J.A., Nanda, S., Dalai, A.K., Berruti, F., Kozinski, J.A., 2020. A review on FUEL.2020.117021.
subcritical and supercritical water gasification of biogenic, polymeric and petroleum Shahbaz, M., Taqvi, S.A., Minh Loy, A.C., Inayat, A., Uddin, F., Bokhari, A., Naqvi, S.R.,
wastes to hydrogen-rich synthesis gas. Renew. Sust. Energ. Rev. 119, 109546 2019. Artificial neural network approach for the steam gasification of palm oil waste
https://doi.org/10.1016/J.RSER.2019.109546. using bottom ash and CaO. Renew. Energy 132, 243–254. https://doi.org/10.1016/
Okolie, J.A., Epelle, E.I., Nanda, S., Castello, D., Dalai, A.K., Kozinski, J.A., 2021a. j.renene.2018.07.142.
Modeling and process optimization of hydrothermal gasification for hydrogen Shenbagaraj, S., Sharma, P.K., Sharma, A.K., Raghav, G., Kota, K.B., Ashokkumar, V.,
production: a comprehensive review. J. Supercrit. Fluids 173, 105199. https://doi. 2021. Gasification of food waste in supercritical water: an innovative synthesis gas
org/10.1016/j.supflu.2021.105199. composition prediction model based on Artificial Neural Networks. Int. J. Hydrog.
Okolie, J.A., Epelle, E.I., Tabat, M.E., Orivri, U., Amenaghawon, A.N., Okoye, P.U., Energy 46, 12739–12757. https://doi.org/10.1016/j.ijhydene.2021.01.122.
Gunes, B., 2021b. Waste biomass valorization for the production of biofuels and Smritis, 2019. What is Mean Squared Error, Mean Absolute Error, Root Mean Squared
value-added products: a comprehensive review of thermochemical, biological and Error and R Squared? [WWW Document]. Study Tonight. https://www.studytonigh
integrated processes. Process Saf. Environ. Prot. https://doi.org/10.1016/j. t.com/post/what-is-mean-squared-error-mean-absolute-error-root-mean-squared-
psep.2021.12.049. error-and-r-squared (accessed 12.30.21).
Okolie, J.A., Mukherjee, A., Nanda, S., Dalai, A.K., Kozinski, J.A., 2021c. Catalytic Taoufik, N., Boumya, W., Achak, M., Chennouk, H., Dewil, R., Barka, N., 2022. The state
Supercritical Water Gasification of soybean Straw: Effects of Catalyst Supports and of art on the prediction of efficiency and modeling of the processes of pollutants
Promoters. Ind. Eng. Chem. Res. 60, 5770–5782. https://doi.org/10.1021/acs. removal based on machine learning. Sci. Total Environ. 807, 150554 https://doi.
iecr.0c06177. org/10.1016/j.scitotenv.2021.150554.
Okolie, J.A., Tabat, M.E., Gunes, B., Epelle, E.I., Mukherjee, A., Nanda, S., Dalai, A.K., Umenweke, G., Adesanya, Z., Onyeaka, H., Miri, T., 2021. Modular bio-refinery
2021d. A techno-economic assessment of biomethane and bioethanol production simulation of Nesogordonia papaverifera by fast pyrolysis (FP): a focus on bio-oil
from crude glycerol through integrated hydrothermal gasification, syngas enhancement. Biomass Convers. Biorefinery 1–11. https://doi.org/10.1007/s13399-
fermentation and biomethanation. Energy Convers. Manag. X 12, 100131. https:// 021-01430-z.
doi.org/10.1016/j.ecmx.2021.100131. Valderrama Rios, M.L., González, A.M., Lora, E.E.S., Almazán del Olmo, O.A., 2018.
Ortiz, F.J.G., 2021. Biofuel production from supercritical water gasification of Reduction of tar generated during biomass gasification: a review. Biomass Bioenergy
sustainable biomass. Energy Convers. Manag. X 100164. https://doi.org/10.1016/j. 108, 345–370. https://doi.org/10.1016/j.biombioe.2017.12.002.
ecmx.2021.100164. Vasseghian, Y., Bahadori, A., Khataee, A., Dragoi, E.N., Moradi, M., 2020. Modeling the
Ozbas, E.E., Aksu, D., Ongen, A., Aydin, M.A., Ozcan, H.K., 2019. Hydrogen production interfacial tension of water-based binary and ternary systems at high pressures using
via biomass gasification, and modeling by supervised machine learning algorithms. a neuro-evolutive technique. ACS Omega 5, 781–790. https://doi.org/10.1021/
Int. J. Hydrog. Energy 44, 17260–17268. https://doi.org/10.1016/j. acsomega.9b03518.
ijhydene.2019.02.108. Verbeek, J.J., Vlassis, N., Kröse, B., 2003. Efficient greedy learning of gaussian mixture
Özdenkçi, K., De Blasio, C., Sarwar, G., Melin, K., Koskinen, J., Alopaeus, V., 2019. models. Neural Comput. 15, 469–485. https://doi.org/10.1162/
Techno-economic feasibility of supercritical water gasification of black liquor. 089976603762553004.
Energy 189, 116284. https://doi.org/10.1016/j.energy.2019.116284.
14
Vigneau, E., Courcoux, P., Symoneaux, R., Guérin, L., Villière, A., 2018. Random forests: Yu, J., Liu, P., Li, Z., 2020. A Hybrid modelling approach to developing digital twins of
a machine learning methodology to highlight the volatile organic compounds an ultra-supercritical steam turbine control stage. Comput. Aided Chem. Eng. 48,
involved in olfactory perception. Food Qual. Prefer. 68, 135–145. https://doi.org/ 115–120. https://doi.org/10.1016/B978-0-12-823377-1.50020-3.
10.1016/j.foodqual.2018.02.008. Yucel, O., Aydin, E.S., Sadikoglu, H., 2019. Comparison of the different artificial neural
Wang, M.W.H., Goodman, J.M., Allen, T.E.H., 2021. Machine learning in predictive networks in prediction of biomass gasification products. Int. J. Energy Res. 43,
toxicology: recent applications and future directions for classification models. Chem. 5992–6003. https://doi.org/10.1002/er.4682.
Res. Toxicol. 34, 217–239. https://doi.org/10.1021/acs.chemrestox.0c00316. Yusup, S., Khan, Z., Ahmad, M.M., Rashidi, N.A., 2014. Optimization of hydrogen
Wang, Z., Peng, X., Xia, A., Shah, A.A., Huang, Y., Zhu, Xianqing, Zhu, Xun, Liao, Q., production in in-situ catalytic adsorption (ICA) steam gasification based on response
2022. The role of machine learning to boost the bioenergy and biofuels conversion. surface methodology. Biomass Bioenergy 60, 98–107. https://doi.org/10.1016/j.
Bioresour. Technol. 343, 126099 https://doi.org/10.1016/j.biortech.2021.126099. biombioe.2013.11.007.
Wong, K.I., Wong, P.K., Cheung, C.S., Vong, C.M., 2013. Modelling of diesel engine Zhang, B., Guo, S., Jin, H., 2021. Production forecast analysis of BP neural network based
performance using advanced machine learning methods under scarce and on yimin lignite supercritical water gasification experiment results. SSRN Electron.
exponential data set. Appl. Soft Comput. J. 13, 4428–4441. https://doi.org/ J. https://doi.org/10.2139/ssrn.3906523.
10.1016/j.asoc.2013.06.006. Zhang, Y., Li, L., Xu, P., Liu, B., Shuai, Y., Li, B., 2019. Hydrogen production through
Yang, H., Chen, H., 2015. Biomass gasification for synthetic liquid fuel production. Gasif. biomass gasification in supercritical water: a review from exergy aspect. Int. J.
Synth. Fuel Prod. Fundam. Process. Appl, 241–275. https://doi.org/10.1016/B978- Hydrog. Energy 44, 15727–15736. https://doi.org/10.1016/j.
0-85709-802-3.00011-4. ijhydene.2019.01.151.
Yang, X., Wang, Y., Byrne, R., Schneider, G., Yang, S., 2019. Concepts of artificial Zhao, S., Li, J., Chen, C., Yan, B., Tao, J., Chen, G., 2021. Interpretable machine learning
intelligence for computer-assisted drug discovery. Chem. Rev. 119, 10520–10594. for predicting and evaluating hydrogen production via supercritical water
https://doi.org/10.1021/acs.chemrev.8b00728. gasification of biomass. J. Clean. Prod. 316, 128244 https://doi.org/10.1016/j.
jclepro.2021.128244.
15

1 s2.0 S2589014X22000330 Main

Uploaded by

Document Informationclick to expand document information

Document Informationclick to expand document information

Copyright:

Available Formats

1 s2.0 S2589014X22000330 Main

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

1 s2.0 S2589014X22000330 Main

Uploaded by

Copyright:

Available Formats

Bioresource Technology Reports 17 (2022) 100976

Contents lists available at ScienceDirect

Bioresource Technology Reports

Machine learning methods for modeling conventional and hydrothermal

1. Introduction could eventually deplete.

• The addition of steam increased the gas yield

• Maximum gas yield (1.75 Nm3/kg) and

• CGE of torrefied food waste is higher than the

• Chemical looping co-gasification helps to

• The use of CES led to the depletion in the

Fig. 2. Machine learning classification.

Table 3 Table 3 (continued )

Fig. 4. Sample representation of the support vector machine.

3.1.5. Deep learning 4.1. Machine learning applications in conventional gasification

• It is an ensemble method that uses a vast number of trees to overcome

• It is robust to overfitting as the ensemble methods have in-built regulari­

• It computes second-order gradients of the loss function to minimize loss and

• It takes more detailed approximations into account when determining the

• They train faster on large datasets and fast to interpret

• It is relatively memory efficient

(continued on next page)

• It requires more significant computational resources

Table 5 Table 5 (continued )

You might also like

• It is robust to overfitting as the ensemble methods have in-built regulari