Introduction

Currently, the modern world faces a series of worldwide energy security issues, starting with a growing population and an accompanying rise in energy consumption that results in environmental pollution and climate change [1, 2]. Among them, the ‘shortage of energy’ is a foremost issue because it is closely linked to the serious pollution induced by excessive use of conventional fossil fuels. Fossil fuel sources such as coal, natural gas, and petroleum are continuously being utilized to fulfill a substantial portion of the global energy demands [2, 3]. The escalating use of fossil fuels has resulted in a persistent rise in the levels of greenhouse gases (GHGs) including carbon dioxide (CO2) and methane (CH4), and atmospheric pollutants like sulfur oxides (SOx), nitrogen oxides (NOx) and so on. This has consequently given rise to significant environmental challenges [4, 5]. Climate change arises from global warming due to the extensive dependence on fossil fuels. This dependency contributes to elevated surface temperatures, sea level rise because of melting glaciers, and various other environmental concerns [6, 7]. The increasing awareness of environmental damage and GHG emissions, in conjunction with fluctuations in the price of petroleum products and gasoline, has opened opportunities to explore alternative and renewable energy sources [8]. Therefore, it is imperative to develop suitable, sustainable, and renewable alternatives to fossil fuels to mitigate GHG emissions and adverse environmental consequences that come with them.

In contrast to fossil fuels, biomass energy conversion processes are sustainable and renewable. Solid biomass, such as lignocellulosic and municipal waste, has been receiving growing interest as a reliable alternative to fossil fuels for the past several decades. Biomass energy conversion methods, including biochemical and thermochemical processes, have been recently growing owing to their carbon neutrality and eco-friendly advantages [9]. Biomass consists of around 38–50% of cellulose, 23–32% of hemicellulose, 15–25% of lignin, and 5–13% of other elements, including inorganic substances and extractives like sodium, potassium, and calcium [10,11,12]. These components can be converted into heat, electricity, fuels, and other value-added products through either biochemical or thermo-chemical processes.

Biochemical conversion processes such as anaerobic and syngas fermentation utilize bacteria, enzymes, and microorganisms to decompose biomass into fuels and other products such as biogas, biodiesel, bioethanol, and other specific constituents [13, 14]. However, the main drawback of these biochemical processes is their low reaction rate, which requires hours, days, or even weeks to complete the necessary reactions [7]. Thermo-chemical conversion of biomass is more commonly used than biochemical processes due to its several advantages and higher conversion efficiency. Unlike biochemical processes, thermochemical conversion does not rely on microorganisms. Instead, they utilize thermal energy and chemical reactions to decompose the biomass into various products [15,16,17]. They are relatively faster than biochemical processes, with reactions taking only a few minutes or seconds. Additionally, these processes can generate high yields and are adaptable to different types of feedstocks [18].

Thermochemical conversion processes are categorized into four classes: pyrolysis or torrefaction, gasification, hydrothermal treatment, and combustion or incineration [2, 15, 17]. Pyrolysis is a crucial technology used for transforming various types of biomasses, including lignocellulosic and waste biomass, into useful fuel or materials such as bio-oil, biochar, and gases such as H2, CH4, and so on. Pyrolysis is commonly classified into slow, fast, and flash depending on heating rate and solid residence time in pyrolizers [19, 20]. The efficiency and characteristics of pyrolysis products are significantly influenced by operation factors including reaction temperatures, solid residence time, heating rate, and water content of biomass [21]. Gasification is the process of transforming solid biomass into syngas. This can be classified into conventional and hydrothermal types based on factors such as reaction conditions, gasifying agents, and the syngas quality. [2, 15, 22]. Compared to other technologies that use dry biomass, hydrothermal treatment (HTT) technology transforms wet biomass into end products such as gaseous and solid fuels that are entirely different from those produced by dry conversion processes by utilizing hot pressurized water as a reactant, catalyst, and solvent [7]. HTT can be categorized as either hydrothermal carbonation or hydrothermal liquefaction, depending on factors such as temperature, pressure, and proportion of the intermediate products [23]. Combustion converts solid biomass into heat and electricity for industrial and household uses [17, 24]. The combustion process can be classified as single combustion or co-combustion, depending on the fuel type. Figure 1 shows the various types of thermochemical conversion processes. However, the modeling of these products, technologies, and systems presents various technical bottlenecks, most of which are due to their multiparametric and complex characteristics.

Fig. 1
figure 1

Classification of thermochemical conversion processes utilizing biomass [7]

Full size image

The optimization of typical thermochemical conversion processes at diverse research levels often requires numerous rigorous experiments to identify the operating conditions and the optimal design. To enhance the efficiency of optimization and lower operational costs, innovative modeling approaches like computational fluid dynamics (CFD), kinetic and thermodynamic models have been employed [25,26,27]. However, implementing these types of modeling using conventional techniques can be difficult and sometimes unrealistic owing to their assumptions, complexities, and limitations [28]. In addition, these models are difficult to incorporate with different variations in thermochemical conversion process yields and kinetics under operating conditions [27, 29]. For instance, CFD modeling is time-consuming and computationally costly due to its significant amount of calculation and predefined parameters, whereas the kinetic model requires complex reaction rate estimations since the reaction mechanism is either unidentified or not completely comprehended. Thermodynamic modeling frequently assumes the equilibrium state, despite the fact that the majority of experimental reactions rarely reach equilibrium [2]. Therefore, it is essential to devise a precise, time-efficient, economical, and resilient modeling method to establish correlations for the intricacies of thermochemical conversion processes.

Machine learning (ML) techniques, in response, have been increasingly utilized as a reliable alternative to conventional modeling techniques for comprehending complex biomass thermochemical conversion processes. Unlike conventional models, ML models can effectively describe intricate relationships between input and output variables, offering a more accurate and dependable approach to modeling complex problems. Several studies have employed ML in various aspects of biomass thermochemical conversion, including bioenergy and biofuel conversion processes, emissions prediction of coal-fired boilers, biomass pyrolysis in fluidized bed reactors, and steam methane reforming (SMR) [30,31,32,33,34,35]. Most pioneering review studies on thermochemical conversion processes have focused on the application of conventional modeling techniques to product yields and kinetics [36,37,38]. However, recent studies have demonstrated the potential of ML models for thermochemical conversion processes, particularly pyrolysis and gasification. Nevertheless, there is a shortage of comprehensive reviews on the application of ML models in HTT and combustion processes [2, 31].

In a recent review of the application of ML in thermo-chemical conversion processes, Ascher et al. [31] highlighted the recent advances in pyrolysis and gasification product yields, distributions, and kinetics optimization and prediction. However, other thermochemical conversion processes, including HTT and combustion, were not considered. Umenweke et al. [2] investigated and summarized biomass gasification and its recent advances in the application of ML to product yields and distribution during conventional and hydro-thermal gasification. However, this review also did not consider other thermochemical conversion processes, such as pyrolysis, HTT, and combustion. Recently, Jeon et al. [39] reviewed the detailed procedures of ML development for specific biomass conversion processes, including pyrolysis, gasification, and hydrothermal treatment. They compared the relative importance of input variables in predicting output variables and discussed the application of ML in techno-economic analysis. However, their focus primarily centered on the detailed ML development procedure to enhance model performance, overlooking crucial insights into improving the efficiency of biomass conversion processes, such as variations in product yields under different operation conditions.

Information gathering through a comprehensive review of all thermochemical conversion processes could help determine the most important parameters and optimal operating conditions that affect product yields and kinetics, which can aid in the optimization and large-scale deployment of the processes. Therefore, this study aims to comprehensively cover all areas in which ML has been employed in this field, highlight the challenges encountered in existing ML techniques, and suggest potential solutions and future research directions. The application of ML in biomass thermochemical conversion processes, encompassing the combustion process, is reviewed to determine optimal operational conditions for a given biomass. Given that the successful commercialization of biomass thermochemical conversion relies on reactor and process designs [40], this review also investigates the types of reactors employed in the literature. Ultimately, this study elicits recent advances in the application of ML to solid biomass thermochemical conversion processes and proposes strategies to enhance the efficiency of biomass thermochemical conversion process under specific conditions.

General Machine Learning Approaches

Machine learning (ML) methods have extensively been applied in addressing societal challenges across diverse fields, including biomass thermochemical conversion processes. These approaches offer the benefit of being efficient and yielding accurate results that closely align with experimental findings. ML algorithms can be categorized as supervised, unsupervised, and semi-supervised.

Supervised ML involves feeding labelled input and corresponding output data to the model, which then learns to map input data to its targeted output [41]. This type of learning can be divided into regression and classification techniques. Both approaches employ similarly labelled datasets for predictions distinctively. While regression methods aim to find the best match between dependent and independent variables in continuous data, classification methods aim to find that in discrete data or identify behavioral patterns across datasets [42]. Unsupervised ML, also known as self-organization, differs from supervised learning in that it works well with unlabeled data and excels at identifying hidden patterns in the data. When input data is fed into an unsupervised ML model, the model searches and sorts the similarities and differences among the data. Unsupervised ML methods encompass principal component analysis (PCA), dimensionality reduction, and clustering.

In contrast, semi-supervised ML integrates the benefits of both supervised and unsupervised learning. It can handle systems with both labeled and unlabeled input and helps address challenges in supervised learning [41]. Labeling data for supervised learning can be difficult and time-consuming, but semi-supervised learning can overcome this limitation by integrating a portion of the unlabeled input data into the supervised learning process [41, 43]. An overview of the classes of ML techniques applicable in the fields of science and engineering, including biomass applications is illustrated in Fig. 2. A wide variety of ML models have been employed for various aspects of biomass thermochemical conversion processes, with the artificial neural network (ANN) being mostly explored model in this research area. A brief discussion of the ANN and other models that have been applied in diverse aspects of this field are discussed below.

Fig. 2
figure 2

Classification of general ML algorithms

Full size image

Artificial Neural Network

Artificial neural network (ANN) is an ML algorithm that imitates the human brain's ability to extract, process, and interpret information. It follows a mathematical model based on the functioning of biological neurons in solving complex problems [44]. It offers the benefit of straightforward implementation and does not necessitate a predetermined or recognized connection between the inputs (features) and outputs (labels). The ANN is comprised of interconnected nodes, and its capacity to process information is encoded in the weights that connect these nodes. The arrangements of connections among the nodes in the ANN are referred to as the ANN network architecture. Figure 3 illustrates a standard depiction of the ANN network architecture.

Fig. 3
figure 3

A typical network architecture of ANN [47]

Full size image

The ANN network architecture comprises three essential layers: an input layer containing input features, a hidden layer comprising hidden neurons, and an output layer that provides the predicted target. Also, ANN can be divided into two main network architectures: feed-forward and feedback or recurrent network [45]. The feed-forward networks lack loops, as seen in a multilayer perceptron where input and output neurons are layered with one-way direction between them. In general, the feed-forward networks generate a set of outputs rather than a sequence in response to a given input. In contrast, feedback or recurrent networks have variability [46]. Regardless of the mentioned network architectures, the crucial aspect is the learning process.

Decision Tree and Random Forest

The decision tree (DT) model possesses a distinctive ability to address both linear and nonlinear problems, rendering it well-suited for intricate non-linear models. The tree models are more pertinent than other black box ML models due to their straightforward comprehension and interpretability [48, 49]. The primary drawback linked to DT models is the risk of overfitting with the increased number of datasets as the model tends to develop complex tree structures [50]. Another constraint of conventional DT models is their incapacity to account for noise in the dataset and the absence of smoothness in their corresponding functions [31]. The DT models, however, have benefits in terms of simplicity and ease of interpretation coupled with the low computational cost [51]. Therefore, other models such as random forest (RF) have been concurrently employed to improve the veracity of each DT models.

RF is a collection of DT designed to address the overfitting limitation. To mitigate overfitting, RF employs an averaging approach across smaller decision trees rather than relying on a single massive DT [52, 53]. RF stands out as a user-friendly ML model suitable for both supervised and unsupervised learning techniques because it has fewer hyper-parameters. Even with sub-optimal hyperparameters, it exhibits commendable performance and accuracy. The two most commonly used parameters among the various options in RF are the total number of individual tree parameters (Nvar) and decision trees (n). The performance of model is significantly influenced by the number of decision trees running in parallel, depending on the desired output [54]. Figure 4 illustrated a standard graphical representation of both DT and RF techniques.

Fig. 4
figure 4

A schematic diagram of decision tree and random forest pathway (modified from [55])

Full size image

Support Vector Machine

Support vector machine (SVM) employs a kernel approach, integrating both classifications, known as support vector classifier (SVC), and regression, known as support vector regression (SVR). This is applicable to linear, non-linear, and multi-dimensional problems, guided by the Vapnik–Chervonenkis (VC) theory [56, 57]. A kernel method is an analysis technique merely based on the dot-products of available datasets. The SVM encompasses the combination of four fundamental concepts: the hyperplane of separation, the hyperplane with maximum margin, flexible margin, and kernel function. The operation of the SVMs depends on identifying the hyperplane which minimizes the distance between examples [58]. A typical structure of the SVM is depicted in Fig. 5.

Fig. 5
figure 5

A typical structure of support vector machine in two-dimension (modified from [59])

Full size image

Deep Learning

Deep learning (DL) is characterized as the distinct type of neural network architecture featuring multiple layers. A graphical representation of the distinction between shallow learning and deep learning is shown in Fig. 6. Different from shallow learning that only has one hidden layer, DL is characterized by multiple hidden layers and activation functions. DL employs a linear regression framework built upon numerous neural nodes or networks. The capacity of DL models to formulate intricate hypotheses renders them potent for comprehending complex, nonlinear, and multi-dimensional relationships. Therefore, the DL can solve problems that shallow learning models could hardly solve [60]. Examples of DL algorithms include recurrent neural network (RNN), long short-term memory (LSTM), convolutional neural network (CNN) and deep neural network (DNN). RNN is a kind of network combinations in a loop through which information persistence occurs [61]. The LSTM are a form of RNN that is explicitly designed to avoid the problem of RNN networks’ long-term dependency. The standard LSTM module has four neural network layers known as the repeated module interacting. The DNN is one of feed-forward neural networks containing multiple layers of hidden units situated between its input and output layers.

Fig. 6
figure 6

Distinction between simple and deep neural network architecture (modified from [2])

Full size image

Machine Learning Applications in Biomass Thermochemical Conversion

The implementation of ML techniques has been expanded in the thermo-chemical conversion processes. Such implementation allows prediction of yield and kinetics of pyrolysis process, syngas or producer gas yield prediction for a gasification process, process modeling, and quantifying char and tar formation among others. The following section and sub-sections are aimed at summarizing existing literature in this field and highlighting the application of ML techniques for operation optimization and research advancement.

Machine Learning Applications in Pyrolysis Processes

Pyrolysis is usually implemented in an oxygen-free condition at elevated temperatures of about 300–700 ℃ [62]. ML techniques have been applied to biomass pyrolysis thermo-chemical conversion processes including the prediction of pyrolysis yield and kinetics. Some of the commonly employed input variables include particle characteristics and operating conditions. The particle scale parameters comprise the particle size and chemical compositions. The operating conditions comprise temperature, gas velocity, heating rate, production capability, and sample weight [63]. The application of ML techniques in biomass pyrolysis prediction is discussed in the sub-sections that follow.

Yield Prediction of Biomass Pyrolysis

The prediction of biomass pyrolysis yields has been previously reported in the literature using the ANN and some other ML algorithms, in terms of biochar, bio-oil, and biogas and their constituents inclusive of fixed carbon, volatile matter, ash, and water content. The quality of end products of pyrolysis has been known to depend strongly on various operation conditions and particle characteristics [21]. These variables are the features usually used to develop the pyrolysis yield prediction ML models.

Earlier studies employed the ANN model to estimate the quality and quantity of liquid and gaseous pyrolysis products generated from three biomasses waste [64]. Catalyst type, amount, temperature, and biomass diversity were input variables for the pyrolysis process while the output of the model is the hydrogen-rich gas (H-rG) ratio. A total of 168 datasets consisting of 102 training, 33 testing, and 33 validation datasets were used to progress the ANN model. The forecasting performance and accuracy of the ANN model were superior with the regression coefficient (R) of 0.975, 0.955, and 0.902 and mean square error (MSE) of 3.25, 6.97, and 9.20 for training, testing, and validation datasets, respectively. The developed model can be applied to similar experimental programs provided the range of model parameters are within the range used in this model.

In a recent study [65], ANN network architecture was used to calculate the basic yields of liquid, gas, and solid product from a pyrolizer. Eighteen datasets with eight input parameters consisting of particle scale parameters and operating conditions, particularly temperature for four biomasses waste, were used to develop the ANN model. The model developed showed high predictive accuracy with an R value of 0.9999 and 0.9941 and MSE of 0.0176 and 5.1714 for both the training and testing dataset, respectively. The major yields or products of pyrolysis (char, gas, and tar) using seven different ML algorithms including Linear regression (LR), ANN, K-nearest neighbor (KNN), SVR, DT, RF, and DNN were predicted in a spouted bed [66]. The labeled dataset for the training of the ML models was generated by the computational particle fluid dynamics (CPFD) simulation. The input parameters used for the simulation include temperature and gas residence time. The developed ML models showed a better agreement with the product yields of the CPFD than those of the lumped process models (Fig. 7). Finally, the results of the ML applications were evaluated with those of the CPFD, and process analysis (PA) lumped models for all products. The comparison study indicates that the developed ML models produce highly accurate predictions compared to lumped process models. A summary of other recently developed ML models for pyrolysis yield prediction is highlighted in Table 1.

Fig. 7
figure 7

Comparisons of the yields of fast pyrolysis products for various prediction models and CPFD with respect to the reaction temperature and inlet gas velocity (modified from [66])

Full size image
Table 1 Summary of pioneering investigation on machine learning application for biomass pyrolysis
Full size table

The review of constructed models on pyrolysis product yields showed that the biomass pyrolysis condition is the most sensitive feature for pyrolysis yield prediction. Tang et al. [77] found that the pyrolysis conditions predominantly have a larger effect on the process outputs than the biomass composition. However, in the other two studies by Cheng et al. [80] and Zhu et al. [72], pyrolysis temperatures were observed to have a powerful effect on the production of biochar from a range of biomass feedstocks.

Machine Learning Applications in Biomass Pyrolysis Kinetic Parameters Prediction

The ANN has been widely applied to predict the kinetic parameters or constituent distributions of a biomass pyrolysis process. Zhong et al. [81] developed an ANN model to build CFD-based reduced order models (ROMs) using pyrolysis temperature and coordinates in the reactor as input variables to predict the distributions or kinetics of gas, liquid, and solid fractions in a bubbling fluidized bed pyrolyzer. The developed ANNs have high predictive accuracy with an average R2 of 0.9940 and were able to show good results for tar, gas, and biomass distributions and kinetics simultaneously.

Sasithorn et al. [82] employed a total of 150 datasets from the thermogravimetric experiments of various biomass resources to develop an ANN model to correlate biomass constituents with the kinetic parameters of the pyrolysis process, in terms of pre-exponential constant (k0), activation energy (Ea), and reaction order (n). Although non-linear relationships were depicted between the biomass components and the output variables, the developed ANN model shows a predictive accuracy, R2 of more than 0.9. The results obtained from the developed ANN models showed minor deviation compared to the experimental results from thermogravimetric analyses (TGA).

In another study, the RF method was successfully applied to develop a prediction model for the pyrolysis activation energy of 5 different biomass resources [83]. For this study, 281 datasets comprising 10 features (types, ultimate analysis results (C, H, O, N, S), H/C ratio, O/C ratio, ash, and model-free method) were used to develop the model. The developed model showed a high predictive performance with the coefficient of determination reaching a value as high as 0.9964. Based on the findings, the reliability of the RF model for the estimation of pyrolysis kinetic has been verified. A summary of other recently developed ML models for biomass pyrolysis kinetic parameters prediction is highlighted in Table 2.

Table 2 Summary of machine learning applications for biomass pyrolysis kinetic parameters prediction
Full size table

Machine Learning Applications in Gasification Processes

Gasification is an important thermo-chemical conversion process that has been vastly researched to produce hydrogen-rich syngas or producer gas from biomass. The gasification process usually involves four distinct processes including drying, pyrolysis, partial combustion and tar cracking, and reduction. The application of novel ML techniques in gasification studies has been tailored towards the prediction of syngas yield and solid residue (char and tar) yield, and catalysis screening and selection. The application of ML techniques in the gasification process has helped in designing a cost-effective process with higher carbon conversion efficiency, as well as reducing the processing time and expenses of complex and time-consuming practices [84]. The various applications of ML techniques in the prediction of biomass gasification processes are elaborated in the subsequent "Machine learning applications in conventional gasification process" and "Machine learning applications in hydrothermal gasification process".

Machine Learning Applications in Conventional Gasification Process

Conventional gasification has a lot of merits, such as low capital cost, low operating cost and less risk of explosion with decreased operating pressure. However, the process requires drying of waste biomass with moisture content above 35wt% which is different from the hydrothermal gasification. Pioneer studies with the implementation of ML algorithms in the conventional gasification process have demonstrated some promising results. Two separate ANN network architectures were employed to predict the syngas yields including CO, H2, CO2, and CH4, during the gasification of biomass in fluidized bed gasifiers [93]. The datasets used for the ANN model development were obtained from 18 and 36 experimental runs. Biomass composition and equivalent ratios were used as input conditions. The developed ANN models displayed robust and accurate predictive performance with R2 of > 0.97. However, the model is constrained to only one type of feedstock and the specific range of the experimental conditions used in model development. The sensitivity analysis was also carried out to figure out the relation between inputs and predicted outputs and showed that two inputs have greater effects on the syngas yields [93]. The results and limitations depicted in the previous study [93] coincided with the results obtained by Mikulandrić et al. [94]. They used gasification temperature, gas flow, and residence time to estimate the yield and compositions of syngas in a fixed-bed gasifier.

In their attempt to solve the lack of model generalization in the pioneer studies, Baruah et al. [95] developed an ANN model by utilizing experimental datasets from different types of woody biomass to predict syngas yield and composition. The inputs included the zone temperature, ultimate, and proximate analysis of different woody biomass utilized for the ANN model. The model performance was high, and the predicted output showed adequate agreement with experimental results with an R2 value between 0.98 and 0.99 [95]. A similar study by George et al. [96], that utilized datasets from five different waste biomass for ANN model training, also showed a high predictive accuracy with an R2 value of 0.9. These studies suggested that expanding the used dataset is desirable to improve the generalization of the model. Similar studies that utilized ANN for model development and prediction were reported by Ascher et al. [31].

Recent research efforts have aimed to explore alternative ML algorithms for predicting syngas or producer gas yields and composition, driven by the desire to overcome the black-box nature of ANN and to assess the predictive capabilities of different ML approaches. Elmaz et al. [97] conducted a comparative study involving four ML algorithms: decision trees (DT), support vector regression (SVR), polynomial regression (PR), and multilayer perceptron ANN (MLP-ANN). These models were utilized to estimate the yield and compositions of syngas in a downdraft gasifier. The developed models exhibited high predictive performance, with both MLP-ANN and DT models outperforming the other two in predicting the levels of CO, CO2, CH4, H2, and HHV. Similarly, Ozbas et al. [98] also predicted the syngas composition and its HHV using four different ML models, including the KNN, Linear Regression (LR), SVR, and DT. Their predictive performances are also compared. The developed models demonstrated high accuracy, with R-squared (R2) values exceeding 0.99. Similar studies were also conducted by Fang et al. [99] and Li et al. [100].

An attempt has been made recently by Serrano and Castelló [101] to include the reactor’s bed material as one of the input variables for the prediction of syngas compositions and gas yield in a bubbling fluidized bed gasifier. Nine input variables were chosen for the model training in this study. Apart from the bed materials, the remaining input variables include C, H, O, N, S, MC, ash, equivalence ratio (ER), temperature, and steam/biomass ratio. It was noted that the developed models successfully forecasted the compositions and yield of syngas with good accuracy (R2 of > 0.94 and MSE of < 1.7 × 10−3) [101].

Prediction of outputs other than syngas yield and composition such as solid residues (char, tar, and ash) yield during gasification are believed to be equally important. Solid residue generation can be one of the limiting factors hindering commercial scale utilization. Therefore, the accurate prediction of these solid residue yields could help in establishing an effective solid residue reduction strategy. To this end, Serrano and Castelló [101] developed an ANN model to predict char deposition in a bubbling fluidized bed gasifier. Datasets used for the model training were obtained from experimental tar sampling, collection, analysis methods, and literature. The developed model showed good predictive ability with R2 of > 0.97 for both testing and validation datasets. In addition, the calculated relative errors were less than 20% for most of the tested samples. Parametric studies conducted with the developed ANN model reveal that the tar yield profile with both temperature and equivalent ratio (ER), exhibits an excellent fit to experimental values reported in the literature, outperforming previous models. A similar study employed various inputs including C, H, N, S, O, moisture, ash content, temperature, and ER to predict LHV of syngas and tars [102]. In the investigation, two ANN models were employed: one with multiple-input and single-output (MISO) and another with multiple-input and multiple-output (MIMO). These were trained using the Levenberg–Marquardt backpropagation algorithm. They showed a high estimation performance with R2 of > 0.99 for tar and syngas LHV value prediction [102]. A summary of other recently developed ML models for syngas yield and composition, and residue yield during conventional gasification are highlighted in Table 3 (Fig. 8).

Table 3 Summary of machine learning applications in conventional gasifiers for syngas and residue yield
Full size table
Fig. 8
figure 8

Comparisons of experimental results with simulation results in the literature [(times symbol) Stark et al. [103]; (white square), (diamond suit) Sridhar [104]; (black down pointing small triangle) Wojnicka et al. [105]; (black right-pointing pointer) Hejazi et al. [106]; (white circle) Rameshkumar and Mayilsamy [107])

Full size image

Machine Learning Applications in Hydrothermal Gasification Process

Hydro-thermal or supercritical water gasification is preferable to conventional gasification because of its capability to handle feedstock of high moisture content without requiring drying [108]. In addition, the produced syngas from this method are often free from nitrogen and sulfur-containing compounds [109]. Different ML algorithms have been applied to hydrothermal gasification processes, particularly for H2 yield prediction and optimization, and catalyst selection and screening. ANN, GPR, SVM, and RF models were developed to estimate H2 yield for supercritical water gasification (SCWG) of agricultural waste and municipal solid waste biomass.

The developed models utilized eight input variables that includes both biomass characteristics: C, H, O, ash; and the SCWG conditions: biomass concentration (BC), gasification temperature (T), pressure (P) and residence time (RT). All the developed ML models showed promising predictive capability with R2 of > 0.98. They also assessed the effect of biomass properties (C, H, O contents) on H2 yield and the exergy efficiency. The proposed ML models showed good prospects to be used for future design and optimization of hydro-thermal gasification process [113].

In another study, Shenbagaraj et al. [114] developed a multi-layered feed-forward back-propagation algorithm-aided artificial neural network (FFBPNN) model to evaluate the effect of different factors affecting syngas composition during the SCWG of food waste. The model displayed a high prediction accuracy greater than 98% and an MSE value less than 0.3. The effect of feed concentration, reaction temperature and time on syngas composition was also studied [114]. A similar study was reported by Zhang et al. [115] on SCWG of waste biomass using linear regression with the single- and two-hidden layer neural networks. Predictive performance of the single-hidden layer neural network has shown superior performance compared to a two-hidden layer neural network. Li et al. [116] proposed the use of Gradient Boosting Regression (GBR) and hybrid GBR-Particle Swarm Optimization (PSO) algorithms to predict the hydrogen (H2) yield in a hydro-thermal gasification process. The aim was to address the black-box nature of ML algorithms and ensure the interpretability of the developed models. The GBR model was constructed based on the composition of the feedstock (C, H, O, N, ash) and operation parameters (temperature, pressure, and time), achieving an R2 value exceeding 0.90. Subsequently, the model was enhanced by integrating Particle Swarm Optimization (PSO). In order to validate the ML models, a comparison was made with mechanistic modeling using Aspen Plus simulation. This comparative analysis between ML algorithms and mechanistic models enhances user confidence in the GBR-PSO models.

An attempt was made by Li et al. [128] for screening and selection of catalysts during the hydro-thermal gasification of waste biomass using ML method of principal component analysis (PCA). Applied datasets were divided into three subcategories of non-catalyst, alkali-metal catalyst, and transition-metal catalyst with the aid of the PCA. The developed model displayed high prediction accuracy in identifying and screening materials for increasing H2 and decreasing CO2 during SCWG of waste biomass. Similar studies on catalyst screening and selection for H2 yield during SCWG of biomass using ML algorithms are reported by Gopirajan et al. [117], Guan et al. [41], and Fózer et al. [118]. A summary of other recently developed ML models for syngas yield and composition, and residue yield during hydro-thermal gasification are highlighted in Table 4.

Table 4 Summary of machine learning applications in SCWG for syngas and residue yield prediction
Full size table

Machine Learning Applications in Hydrothermal Processes

In HTT, wet biomass is converted into gaseous or solid fuels and other constituents under elevated pressure and temperature [7]. As mentioned above, HTT is classified as hydrothermal carbonation (HTC) and hydrothermal liquefaction (HTL) depending on temperature, pressure, and the proportions of the converted intermediate products [23]. The HTC is usually conducted at relatively low temperatures of between 150 and 260 ℃ and pressure of 1.5–5 MPa to obtain a solid product (hydro-char), while HTL is mostly undertaken at moderate temperature of between 250 and 350 ℃ and pressure 5 ~ 20 MPa in subcritical water to yield liquid fuel (bio-oil) [120]. ML techniques have been applied to various aspects of wet biomass hydro-thermal treatment prediction, particularly for the prediction of biofuel and bio-char.

Machine Learning Applications in Hydrothermal Carbonation

ML algorithms have demonstrated their effectiveness in the field of hydrothermal carbonization (HTC) and the pre-treatment of biomass that can transform wet biomass into uniform, carbon-rich hydro-char with a higher heating value (HHV). The ML algorithms have been successfully utilized to predict diverse aspects of the HTC process, including the composition and properties of the final product (hydro-char), process kinetics, nitrogen, and phosphorus content in hydro-char, process optimization, and synergistic catalysis.

Vardiambasis et al. [121] employed four ANN models, developed using 144 datasets, to predict hydro-char yield and HHV during hydro-thermal carbonization of sewage sludge and food waste biomass. The developed models utilized elemental content (C, H, O), HTC temperature, and time as the input variables. The developed ANN models showed promising predictive capability with R2 of > 0.917. The sensitivity analysis results confirmed that carbon content (C) had the greatest influence on hydro-char yield and HHV. The proposed ML models confirmed good perspectives to be used for future design and optimization of the hydro-thermal gasification process [121]. Zhu et al. [72] developed an RF model to predict the yield, HHV, and carbon in char of municipal waste during HTC operation. The predictive results showed a high accuracy with R2 of 0.80, 0.91, and 0.95 for hydro-char yield, HHV, and C_ char, respectively. Similar studies that utilized ML models for HTC process prediction can be found in Li et al. [71] , Ismail et al. [122], Kapetanakis et al. [123], and Mu et al. [124].

The prediction of process kinetics during HTC of cellulose, poplar, and wheat straw biomass was reported by Aghaaminiha et al. [125]. Four different ML models of KNN, ANN, SVR, and DF employed 132 datasets with input variables of experimental type (isothermal, dynamic), temperature, time, nitrogen content, sulfur content, and hydrogen content. The developed models showed high predictive accuracy in all the scenarios investigated. In addition, the performance of the developed ML methods in interpolating kinetics results was evaluated, considering situations where experimental data is limited to only a few time-points. An ‘extrapolation model’ was additionally developed using kinetics data from the first three time-points as input. The kinetic data for subsequent time steps was predicted to evaluate its capability in extrapolating kinetics when data is available from only a small number of initial time-points. Djandja et al. [126] developed an RF model for HTC processes to access the phosphorous concentration of char based on 109 data points. Particle properties such as FC, VM, ash, moisture, C, H, O, N, S, and phosphorous concentration, and operation conditions including temperature, residence time, pH of feedwater, and dry matter content (DM), were considered as features for designing the model. The model predictive performance was high with R2 of > 0.92–0.95. In addition, operation conditions exhibit a positive effect on the phosphorous concentration in char [126]. A similar study by Djandja et al. [127] was developed to predict the nitrogen content in hydro-char. A summary of recently developed ML models for HTC is highlighted in Table 5 (Fig. 9).

Table 5 Summary of the existing research on machine learning applications for hydrothermal carbonization of wet biomass
Full size table
Fig. 9
figure 9

Comparison of the predictive performance of the machine learning models [125]

Full size image

Machine Learning Applications in Hydro-thermal Liquefaction

Hydro-thermal liquefaction (HTL) is considered as a viable option for producing bio-crude oil from wet biomass with varying moisture content, which can be upgraded to be used as a transportation fuel and for chemicals. HTL also produces solid biochar and a gaseous fraction. It has some advantages over the classical process, as it has a lower processing time and a higher yield [130]. The process is mainly dependent on feed quality, reaction time, temperature, catalyst, and mixing ratio. The application of ML in HTL is mainly for estimating the yield and composition of bio-oil.

Cheng et al. [131] developed ML models by utilizing 570 experimental datasets for HTL of different biomass feedstocks in literature (algae, lignocellulosic biomass, food waste, manure, sludge, bioethanol residue, municipal solid waste, and seed plants) to predict bio-crude yield. Eight ML models including MLR, Ridge regression (RR), Lasso regression (LR), SVR, DT, MLP-ANN, RF, and XG Boost were developed for this purpose, and their performance was compared. The performance of the above models was relatively high, and results showed good agreement with experiments. This RF model was adjudged the optimal model with an RMSE of 8.07. The accuracy of the optimal RF model and a probabilistic economic analysis are found to be enough to arrange the order of resources delayed on the basis of the estimated minimum fuel price [131]. Katongtung et al. [132] predicted bio-crude yields and HHV of biomass resources in HTL processes by using 17 inputs in 325 datasets. Novel ML algorithms, including SVR, Kernel ridge regression (KRR), RF, and extreme GB (XGB) based on tenfold cross-validation were adopted for the prediction purpose. The developed models showed a high predictive accuracy with an R2 value of > 0.9. Similar studies that utilized ML algorithms for bio-crude and composition prediction from HTL of wet biomass can be found in the literature [133, 134].

In another study, two ML models of DT and RF with 257 datasets were developed to predict pH, TOC, TP, and TN in the aqueous phase (AP) produced in HTL processes by using algae, food waste, sludge, and manure [135]. Both developed models presented high predictive performance, with the DT showing better performance than RF for both single and multi-target predictions. In addition, the ML-based feature importance and partial dependence analysis showed that temperature, solid, and nitrogen content were mostly important factors for pH, TN, and TP. A similar study used both GBR and RF to predict bio-oil yield, content of oxygen (O_ oil), and nitrogen (N_ oil) from HTL of wet biomass [115]. The results indicated that the GBR, with an averaged R2 of > 0.90 for the test set, displayed better predictive capability than the RF model for both single and multi-target task prediction. A summary of other recently developed ML models for bio-crude and its composition prediction from HTL of biomass are highlighted in Table 6.

Table 6 Summary of pioneering investigation on machine learning application for hydrothermal liquefaction of wet biomass
Full size table

Machine Learning Application in Combustion Processes

In the combustion process, organic matters are thermally converted into CO2 and H2O with the help of an oxidant, generally oxygen. The combustion process generates heat and electricity for home and industrial use. The process can be classified into single combustion or co-combustion of two fuels depending on the fuel type [7]. In the combustion process, ML can be employed for modeling boiler wall at different operating conditions, fault diagnosis, automation in the generation process, and the prediction of gaseous pollutants, thermal properties of biomass, and output energy [136]. The thermal combustion characteristics, which include combustion reactivity [137], heat capacity [138, 139], oxidation kinetic parameters [140,141,142,143,144], and co-combustion of biomass are also included [145,146,147].

The presence of moisture content in biomass significantly affects its properties and combustion characteristics. Taking this into account, Rico-Contreras et al. [148] devised a prediction model based on FL and ANN to estimate the thermal properties of poultry litter. They utilized various input variables including density, temperature, duration of storage, and feedstock handling, which affect the moisture content of the feedstock. A total of 108 samples were tested with this artificial framework and an average moisture content of 30.16% was reported. In addition, the developed FL-ANN system showed an accuracy of 92.88% when modelled with 20 different farm study results. Importantly, this proposed model holds applicability across a wide range of bioenergy generation systems, encompassing not only the poultry sector but also other combustion or alternative processes. Krzywanski et al. [33] developed an FL model to predict emission contents from both wood chips biomass and bituminous coal in a lab-scale 5 KW dual fluidized chemical looping combustor. This model estimated various emissions including CO2, CO, NOx, and SOx. These pollutants were made across conventional and novel combustion processes including air-firing, oxy-fuel combustion, chemical looping combustion, and so on. The developed FL model displayed a high predictive capability with a maximum relative error between target and prediction lower than 8%. Prediction by the FL models for gaseous pollutant emissions was in good agreement with experimental results. A comparable study was conducted by Li et al. [149] and Li et al. [150] to investigate NOx emitted from biomass combustors using the DNN. In both studies, the predicted NOx emissions exhibit good agreement with the measurement results. Further studies on GHG emissions prediction from biomass-fired combustion process can be found in [151,152,153].

To become a sustainable society, carbon-neutral fuels like biomass, hydrogen, and ammonia should be utilized for heat and electricity and their usage consequentially increases in most countries. Therefore, HHV is an important factor in assessing the economic feasibility of biomass utilization. Noushabadi et al. [154] developed and proposed an ML methods framework to evaluate the HHV of biomass by using C, H, N, O, and S weight fractions. A total of 535 datasets were utilized to construct five ML models. Different types of biomasses, such as fruits, agricultural wastes, grasses/leaves/fibrous materials, wood chips/tree species, and various organic wastes including municipal solid wastes, were utilized in combustion processes. The developed models demonstrated excellent predictive performance, as evidenced by the low average absolute relative deviation (AARD) values.

Specifically, the MNR and GA-RBF algorithms achieved an AARD of 3.5% and 3.4%, respectively, indicating their suitability for estimating the HHV of biomass. You et al. [155] estimated the low heating value of biomass (municipal solid waste, MSW) using MLP-ANN, ANFIS, SVM, and RF models. The result indicated that a high-accuracy ML model could improve the CFB operation and contribute stable energy supply [155]. Further studies on the prediction of HHV from biomass-fired combustion processes can be found in [47, 86, 156,157,158,159]. A summary of other recently developed ML models for thermo-chemical combustion of biomass is highlighted in Table 7.

Table 7 Summary of machine learning applications for biomass combustion
Full size table

Futuristic Prospects

As discussed, ML is widely applied to biomass thermochemical conversion processes. However, there needs to be a more detailed description of the procedure of ML development to improve further. For example, in many instances, the ‘number of epochs’ is used as the termination criterion when the desired correlation (R2) and minimum RMSE are not achieved [160]. In this chapter, some challenges, and prospects in the application of ML to biomass thermochemical conversion are introduced with respect to the datasets and training and testing of the ML model. Furthermore, relying on a single statistical index is inadequate as it only provides a limited perspective on model errors, focusing on specific aspects of the error characteristics. To comprehensively assess model performance, a combination of statistical indices, including variance accounted for (VAF), mean absolute percentage error (MAPE), and mean squared deviation (MSD), should be employed. By considering multiple indices, a more thorough evaluation of the model's accuracy and performance can be achieved.

Previous research has revealed a scarcity of data collected directly from biomass thermochemical experiments because of the costly and time-consuming nature of the data acquisition process. To overcome this challenge, researchers often depend on importing experimental data from various sources. Therefore, employing these sets of experimental data often results in models with good predictive accuracy but low extrapolability since these models are only compatible with a specific dataset. The developed models are only applicable to experimental data within the range of the data they were trained on. Integrating ML models with theoretical modeling approaches, such as kinetics, thermodynamics, and CFD studies, to simulate the experimental conditions should provide a high-quality dataset. These datasets provide extensive information regarding fundamental reaction mechanisms and pathways, facilitating the development of optimal models during the training process and improving the predictive capabilities of the models.

Most ML algorithms are called "black box" because the relationships between the input and output variables are, most of the time, complex and are not always depicted in the form of tractable mathematical equations that can be easily understood. On the other hand, the advantages of developing a reactor model with detailed reaction kinetics and hydrodynamics for biomass thermochemical conversion processes lie in its better extrapolability and interpretability compared to ML models. This becomes particularly important when dealing with the inherent variability in biomass feedstock. While a physically and chemically consistent feedstock is required for the robust operation of the process, the inconsistency within a single resource due to varying growth and harvesting conditions [161], especially during climate crises, can present challenges in black-box modeling. Thus, elucidating the reaction pathways of biomass decomposition under various conditions is necessary for the robust operation of biomass thermochemical conversion processes. To enhance the extrapolability of ML models, it has been suggested to develop ML models guided by theoretical models. The methodology for such a combination was described elsewhere [39]. Future work should be directed towards unboxing the black-box nature to ease the implementation and increase the reproducibility and interpretability of these ML models. These models should be converted into tractable mathematical equations, allowing quantitative interpretation and accurate prediction. Additionally, researchers have proposed the use of feature permutations to examine the behavior and relevance of various input variables. This approach enabled a comprehensive evaluation of the significance of each input variable in the ML model [31], increasing the interpretability of the model. Moreover, it is advisable to incorporate sensitivity and techno-economic analyses to assess the influence of each predictor or feature on the predicted thermochemical conversion process and the economic importance, respectively.

Several ML algorithms employed to study biomass thermochemical conversion processes have shown good relevance. Nevertheless, even though ML algorithms offer several advantages over conventional modeling techniques, such as CFD, it is crucial to remark that no single ML algorithm can effectively address all engineering problems related to conversion processes, as stated by the No-Free-Lunch (NFL) theorem [162]. Advancements in conversion technologies, such as carbon capture and sequestration processes and liquid fuel transportation, which require additional inputs and increase the complexity of the models. At present, accurate predictive models for these conversion technologies have been extensively developed. Future studies should focus on applying other existing ML algorithms that are yet to be used to biomass thermochemical conversion processes while also developing new algorithms for these advanced technologies. Thus, researchers will be able to develop ML models with higher predictive accuracy and interpretability.

Assessing the predictive performance of ML models is as crucial as training them. During the performance evaluation, the predicted outputs of the models were compared with relevant observations or experimental data using various statistical measures or error metrics, commonly known as statistical indices. These statistical metrics illustrate the disparities between the anticipated and observed values of the outputs, specifically indicating the extent of the residuals' dispersion. The statistical indices used to appraise the performance of ML models have several limitations. The two primary error metrics, namely the mean square error (MSE) and root mean square error (RMSE) are regarded as less reliable owing to their sensitivity to different data fractions and their susceptibility to outliers [163]. In particular, the RMSE is affected by extreme scores, which can hinder the convergence of errors within specified tolerance limits.

Conclusions

This study provides a comprehensive review of published articles that focus on the application of ML techniques in biomass thermochemical conversion processes. Compared to conventional modeling techniques such as computational fluid dynamics (CFD), thermodynamic and kinetic models, process models, and ML techniques offer several advantages, including accuracy, efficiency, simplicity, and robustness in modeling the complex nature of these processes. Various ML algorithms have been successfully used to predict yields and kinetics and optimize pyrolysis, gasification, HTT, and combustion processes, demonstrating high predictive performance and accuracy. Among these techniques, ANNs have been widely utilized by researchers owing to their capacity to manage intricate associations between input and output variables without prior knowledge of their interconnections. However, a limitation of ANN and other ML models is their lack of interpretability because the relationships between inputs and outputs can be intricate. To improve the understanding and interpretability of these processes, it is crucial to develop models that are easy to interpret. Ensuring the interpretability of the developed ML models and improving their predictive accuracy is imperative for large-scale and industrial deployment of these processes. Fine-tuning the hyperparameters of the developed process models using metaheuristic algorithms could help ensure higher predictive accuracy of the processes. Establishing broad input and output variable datasets could aid in improving the model performance because it has been established that the model performance increases with the increasing number of datasets.

Consequently, there is a notable need for the development of additional ML algorithms and the inclusion of a wider range of operating conditions and parameters in the datasets used for model development. This approach can help to create universal models with high predictive accuracy. By using novel algorithms and incorporating comprehensive data, researchers can enhance the capabilities of ML models and ensure their applicability to different biomass thermochemical conversion processes. This will ultimately contribute to the advancement of the field and facilitate more accurate predictions in various operational scenarios.