1 Introduction

Rapid industrialization and economic growth have led to an energy consumption rate that far exceeds its production rate. Most current energy demands are met by fossil fuels, which has resulted in global climate change, energy crisis, and environmental pollution (Zhang et al. 2023a). Therefore, it is necessary to develop renewable energy. Biomass is considered the only carbon-based renewable energy source (Seo et al. 2022) that is abundant and has a high energy potential. Biomass (especially biowaste) is the most commonly used renewable energy source because it is carbon neutral, environmentally friendly, and economically viable.

Hydrothermal carbonization (HTC), which uses sub-critical water as the reaction solvent, is a promising biomass treatment and recycling method for converting wet biomass into hydrochar and chemicals without the need for pre-drying. The properties of hydrochar are affected by various factors, such as the composition of biomass feedstock and reaction parameters (Zhang et al. 2023b). Different biomass combinations have been studied to explore the effects of biomass types and reaction parameters (temperature, reaction time, and water content) on the hydrochar yield and fuel characteristics. The results have shown that temperature is the most important factor affecting hydrochar yield and energy density (Toptas Tag et al. 2018). Compared with temperature, the residence time and water content have similar but weaker effects on the yield of solid products (Xu et al. 2018). The introduction of a catalyst can directionally regulate the deoxygenation and denitrification reactions to improve the yield and quality of hydrochar (Marzbali et al. 2021; Xu et al. 2022), thereby generating functional carbon-based materials. Traditional approaches mainly measure the physical and chemical properties of hydrochar products and HTC conditions, but these experimental approaches are expensive and time-consuming.

Machine learning (ML) methods can address the above-mentioned problems by predicting and regulating the mining of experimental data to assist in the production and optimization of biofuels, chemicals, and materials in the HTC system (Li et al. 2023). Recently, several researchers have used ML to investigate the HTC process of biomass, and their current limitations are summarized in Table 1. For example, the yield and energy recovery of hydrochar can be predicted (average R2 = 0.90) by the random forest (RF) and artificial neural network (ANN) models (Li et al. 2020b; Gao et al. 2022), but these studies merely explored the impact of the biomass components on hydrochar. The fuel and structural properties of hydrochar have also been predicted for some specific types of biomass (e.g., sewage sludge, food waste, manure Mu et al. 2022; Zhu et al. 2023), or kitchen waste (Zhang et al. 2023c)). A support vector regression (SVR) model was used to construct relationships between biomass properties/HTC conditions and hydrochar composition (Shafizadeh et al. 2023). Most previous research only established one or several properties of hydrochar prediction models separately but did not carry out model verification, interpretation, and application, which were limited to the collected dataset. Thus, it is necessary to develop reliable ML models for predicting and optimizing multiple properties of hydrochar with clear mechanisms. Moreover, the biochemical composition was used to predict bio-oil properties during the hydrothermal liquefaction of biomass, which showed that lipids were the major component affecting the bio-oil yield and energy recovery (Sheng et al. 2018; Li et al. 2021a). The impacts of biochemical composition on synergism/antagonism of the bio-oil yield were more than that of reaction conditions (Yu et al. 2023), which may be suitable for HTC processes. The nitrogenous components were mainly derived from proteins (Zhang et al. 2021; Leng et al. 2022a). Generally, three primary biochemical components of biomass (i.e., proteins, lipids, and carbohydrates) were not given simultaneously with the elementary composition of HTC, even though they are principal factors in the formation of hydrochar (Li et al. 2019). However, few studies have considered the effects of the biochemical components on the hydrochar properties.

Table 1 Comparison of research designs with applied ML models between previous studies and this work in the HTC process
Full size table

In this study, biochemical components (i.e., proteins, lipids, and carbohydrates) of biomass were predicted first by elemental composition, which contributed to biomass characterization. The impacts of biochemical components on the hydrochar properties were explored by ML models, guiding the HTC process design. Then, ML multi-target models were constructed to predict and optimize the hydrochar properties (yield, elementary composition, elemental atomic ratios, and HHV). Single-biomass composition (elementary and biochemical), proximate analyses, and HTC conditions were used as inputs. The in-depth feature analysis through ML provides insights into the mechanisms of hydrochar formation. An intelligent system was also formed and shared based on a multi-target model, which was verified by predicting elemental atomic ratios (N/C, O/C, and H/C). It was not only used to guide the optimal production of hydrochar derived from HTC from single-biomass samples. Finally, hydrochars from the co-HTC of mixed biomass samples reported in previous references were compared.

2 Materials and methods

2.1 Dataset preparation and preprocessing

The related keywords (e.g., hydrothermal carbonization, biomass, algae, lignocellulos*, and hydrochar) were retrieved from literature databases such as Web of Science and Google Scholar for searching amounts of SCI papers. The  criterion for selecting suitable literature was that the data in the chosen articles must contain inputs (biomass composition and HTC conditions) and outputs (yield and elementary composition of the hydrochar). The hydrochar dataset was organized from the experiments of 87 selected articles among reviewed publications since 2010. All data points were collected from tables or read from figures with PlotDigitizer (a data extraction software). The detailed datasets #1–6 are provided in the Additional file 2. The collected data were divided into two datasets: hydrochar from single biomass (dataset #1) for model development, and hydrochar from mixed biomass (dataset #6) for model validation. In detail, the elementary composition (i.e., C, H, N, O, S, and ash), proximate composition (fixed carbon (Fc), volatiles (V), and ash), and operational conditions (residence time (RT), temperature (T), and water content of biomass (WC)) were treated as the input variables. The output variables of the HTC experiments  including the hydrochar yield (Yield_char) and elementary composition of hydrochar. Finally, 51 types of forestry leftovers, 43 types of domestic waste, 29 types of crop straw, and 17 types of organic waste were collected in dataset #1. To further explore the effect of biomass composition and HTC conditions on hydrochar quantity and quality, 199 kinds of biomass from 59 publications with listed elemental and biochemical composition were compiled in dataset #2. This consisted of 108 types of microalgae, 41 types of macroalgae, 20 types of sludge, 12 types of manure, 16 types of food waste, and 2 types of lignocellulose. The units of input and output variables in all datasets were standardized using international units to improve the consistency of units and better compare each feature in the datasets. The biochemical components were predicted from a model built based on dataset #2, and they were input into dataset #3 in combination with the HTC conditions to predict the hydrochar properties. Then, the predicted biochemical components were added to dataset #1 to form dataset #4. In addition, the atomic ratios were calculated according to the elemental compositions of hydrochar and then compiled to form dataset #5. The data of hydrochar related to the co-HTC of mixed biomass from references were integrated into dataset #6. The detailed information of all datasets is listed in Table 2.

Table 2 Detailed description of various compiled datasets used in this study
Full size table

The composition of biomass in datasets containing biochemical, proximate, and elemental composition was normalized using the following three equations (dry basis), respectively:

$$\text{Protein + Lipid + Carbohydrate + Ash}\text{ = 100}$$
(1)
$$\text{V +Fc + Ash }= \text{ 100}$$
(2)
$$\text{C + H + N + O + S + Ash }= \text{ 100}$$
(3)

where V, Fc, C, H, N, O, and S are the contents (wt%) of volatiles, fixed carbon, carbon, hydrogen, nitrogen, oxygen, and sulfur in biomass, respectively.

The elementary compositions of hydrochar in datasets were treated on a dry basis and unified with the following formula:

$$O\_char = 100 {-} C\_char {-} H\_char {-} N\_char {-} Ash\_char$$
(4)

where O_char, C_char, H_char, and N_char are the elemental contents (wt%) of oxygen, carbon, hydrogen, and nitrogen of hydrochar, respectively. The elementary composition of biomass and hydrochar in datasets were directly collected from the literature or calculated on a dry basis using Eqs. 3 and 4, respectively.

Through data preprocessing, the original data can be made more standardized, accurate, and usable, and the quality of the dataset can be improved, thereby improving the accuracy of data mining and prediction results (Kirchner et al. 2016).

Pearson correlation analysis was conducted to understand correlations between inputs and outputs (Leng et al. 2023), and the Pearson correlation coefficient (PCC, r value) was calculated using Eq. 5:

$${r}_{xy}=\frac{{\sum }_{i=1}^{n}({x}_{i}-\overline{x }){\sum }_{i=1}^{n}({y}_{i}-\overline{y })}{\sqrt{{\sum }_{i=1}^{n}{({x}_{i}-\overline{x })}^{2}}\sqrt{{\sum }_{i=1}^{n}{({y}_{i}-\overline{y })}^{2}}}$$
(5)

where \({r}_{xy}\) represents the PCC value of the two variables, and \(\overline{x }\) and \(\overline{y }\) are the mean values of the input feature x and another variable y, respectively. The value of \({r}_{xy}\) is between −1 and 1, where 0 indicates no linear correlation between the two variables. A higher PCC indicates a higher correlation between two variables, and a negative value indicates a negative correlation and vice versa.

2.2 Machine learning models establishment

The scope of experimental values of various features varied widely. To obtain a consistent range of orders of magnitude between different variables without changing the distribution relationships of each feature, all input features were normalized using the Z-score method before training using Eq. 6:

$${Z}_{i}^{*}=\frac{{Z}_{i}-\mu }{\sigma }$$
(6)

where \({Z}_{i}^{*}\) and \({Z}_{i}\) are the standardized and actual experimental values of each input variable, respectively;\(\sigma\) and \(\mu\) express the standard deviation and mean values of each \({Z}_{i}\), respectively. All input features were treated with similar scales of importance to avoid generating significant errors from range inconsistency or outlier points (Zhang et al. 2024).

The data points in the dataset were split into two parts (the training dataset and the testing dataset) during the ML model development process. Herein, 80% of data points in the whole dataset were randomly chosen and included in the training dataset, while the remaining 20% were used as the testing dataset to estimate ML predictive models.

Two classical tree-based ensemble learning ML algorithms, random forest (RF) and gradient boosting decision tree (GBDT), were used to train the predictive models (Deng and Lin 2022; Xiong et al. 2023), which could provide explainable prediction process. During RF learning, bagging or bootstrap aggregation was applied to develop hundreds of decision trees with the input dataset, where each tree was a regressor (Fang et al. 2022). The final model prediction result for RF was the average output value of all individual decision trees. The GBDT algorithm was developed using the boosting method (Natekin and Knoll 2013), in which a leaf (an initial predicted value for obtaining the initial errors) was obtained from the mean of the observed output values. The gradient boost strategy involves building the first regressor tree to predict errors according to changes between the actual and original values. Subsequently, each tree minimizes the errors from the prior tree to obtain a highly accurate and efficient GBDT model (Yuan et al. 2021). The optimal hyperparameters of these two algorithms were determined through fivefold cross-validation combined with a simulated annealing algorithm (SSA) based on maximal mean regression coefficient (R2) during model training for fast optimization (Leng et al. 2022b). The training dataset was randomly split into five parts, of which four parts were used to determine the optimal hyperparameters of the ML model during training, and the remaining part was used to evaluate the training performance. For SSA, the hyperparameters of the RF or GBDT models were used as inputs, and the average mean R2 of the fivefold cross-validation was the output (Additional file 1: Fig. S1). The best hyperparameters of these two were finally confirmed by the values corresponding to maximized mean R2 (Additional file 1: Fig. S1). In this present study, the tuned hyperparameters of the RF were the number of trees (n_estimators) and the maximum depth of trees (max_depth), while those for the GBDT were n_estimators, max_depth, learning rate, and loss function. The model and optimization algorithm employed the scikit-learn library (scikit-learn.org/stable) and Hyperopt library (Putatunda and Rama 2018) in the Python programming platform, respectively.

2.3 Model evaluation

To comprehensively predict and evaluate the output characteristics, multi-target models were applied to predict the biochemical components of biomass, the biomass yield, and the properties of hydrochar in this work. A total loss function (\({L}_{M}\)) of the multi-target predictive model was specified as follows (Liu et al. 2017):

$${L}_{M}=\sum_{t=1}^{M}{\alpha }_{t}{L}_{t}({y}_{n}^{t},{\widehat{y}}_{n}^{t})$$
(7)

where \(M\) is the number of targets; \({\alpha }_{t}\) represents the weight factor for each target t, which was used to underline the degree of relative importance for this task, where all targets in this work were given equal weight; \({L}_{M}\) indicates the loss function for each t; \({y}_{n}^{t}\) is the actual value for each t; \({\widehat{y}}_{n}^{t}\) is the predicted value for each t. All outputs were treated equally in the multi-target predictions (i.e., with the same weight).

The determination coefficient (R2), a statistical index, was employed to evaluate the model, which generally represents the fitting degree of the relationship between input and output variables.

$${R}^{2}=1-\frac{{\sum }_{i=1}^{N}{({y}_{i}^{*}-{\widehat{y}}_{i})}^{2}}{{\sum }_{i=1}^{N}{({y}_{i}^{*}-\overline{y })}^{2}}$$
(8)

where N indicates the number of test samples; \({y}_{i}^{*}\) and \({\widehat{y}}_{i}\) indicate the ith experimental value and corresponding predicted value, respectively; \(\overline{y }\) indicates the mean of N experimental values.

In addition, the root mean square error (RMSE) was calculated to evaluate the model’s prediction performance by representing the degree of dispersion between the actual and predicted values using Eq. 9.

$$RMSE=\sqrt{\frac{1}{N}({\sum }_{i=1}^{N}{({y}_{i}^{*}-{\widehat{y}}_{i})}^{2})}$$
(9)

Since biochemical components were both outputs and inputs for ML models, residue analysis was conducted on biochemical composition predictive models. The Shapiro–Wilk (S–W) test (Shapiro and Wilk 1965; Yap and Sim 2011; Rodrigues et al. 2023) (confidence interval = 95%, median residues value = 0) was analyzed based on the obtained residues to calculate significant values (p_S–W) of predicted and actual values. If p_S–W > 0.05, the residues conform to the normal distribution (Shapiro and Wilk 1965), which shows high prediction feasibility and reliability of ML models (Leng et al. 2023).

2.4 Model interpretation

After the ML models were built and their performance was assessed, the impact of input features on targets was analyzed to understand relationships between crucial input variables and output variables. Embedded feature analysis [e.g., feature importance and partial dependence plot (PDP)] and Shapley additive explanation (SHAP) (Ribeiro et al. 2016) for input features on targets were used to understand the ML predictions to understand the underlying ML model. The formula of the SHAP value is as follows:

$${\phi }_{i}=\sum \frac{\left|S\right|!\left(M-\left|S\right|-1\right)!}{M!}[{f}_{x}(S\cup \left\{i\right\}-{f}_{x}(S)]$$
(10)

where M and S are the dimensions of the set of all features and a subset of M, respectively; (S ∈ M); \({f}_{x}(S)\) and \({f}_{x}(S\cup \left\{i\right\}\) are the mean sample values calculated by the models trained with and without output variables, respectively (Zhu et al. 2023). The PDP of ML models can also display visual relationships (e.g., linear, monotonic, or more complex correlations) between one or two feature input and output variables (Palansooriya et al. 2022).

2.5 Model applications and verification

2.5.1 Graphical user interface

An intelligent system based on an online graphical user interface (GUI) website was developed to provide a convenient platform for other researchers to use. Herein, the predictive ML model was uploaded to the GUI website by using hypertext mark-up language (HTML), cascading style sheets (CSS), Python program (version 3.8), and Flask framework (version 1.1) (Leng et al. 2022a).

2.5.2 Prediction of elemental atomic ratios (N/C, O/C, and H/C)

The elemental atomic ratios (N/C, O/C, and H/C) provide indicators for measuring the fuel properties of hydrochar. After the system was developed, we aimed to evaluate whether the elemental composition of the hydrochar were precise or not. Hence, the elemental atomic ratios were calculated by the values predicted by the multi-target ML model, and the results were compared with the actual results.

2.5.3 Forward optimization of hydrochar from specified biomass and experimental verification

To further verify the generality of the model and system, the input HTC conditions were adjusted by taking soybean straw from our laboratory as the raw material (C = 41.05%, H = 5.52%, N = 2.90%, O = 41.39%, S = 0.28%, ash = 8.87%, FC = 74.38%, and V = 16.75%). Among them, the temperature was set to 5 °C, from 140 °C to 310 °C; the residence time was 5 min, from 5 to 240 min; the water content step was 5%, from 75% to 99%. The optimal hydrochar objective function was set to maximize yield, C_char, and H_char and to minimize N_char and O_char, belonging to multi-objective optimization (Liu et al. 2022). Experimental validation was conducted using the optimal iteration outputs to guide the experiments to generate the desired hydrochar. Soybean straw was collected from a farm and was smashed into < 50 μm particles during HTC experiment validation. Then, it was dried to constant weight at 105 °C and stored in a glass container. The detailed operating procedures (e.g., reactor setup, reaction program, procurement of products, and analysis) are provided in our previous studies (Zhang et al. 2021; Leng et al. 2022c). Briefly, the reactor was heated to a set temperature at a rate of 4–5 °C min−1 , and HTC was reacted at a given residence time. The solid product was filtered from the mixture after the reactor was cooled to room temperature, and the solid product was dried at 105 °C for 24 h before characterization.

2.5.4 Prediction of hydrochar from co-HTC

The system was established using hydrochar derived from single biomass HTC without considering a biomass mixture. To explore the feasibility of using this system for mixed biomass, the validation dataset of hydrochar characteristics from different biomass mixtures in the literature was collected. The values predicted by the system were compared with the actual experimental values from previous studies using the same input variables.

3 Results and discussion

3.1 Dataset integration, filling, and analysis

The box plots show the minimum, average, median, maximum, and quartile values of the input and output features. Figure 2a–c show the distribution of each feature in dataset #1. Elements C and O were the two major components of the elementary composition of biomass, while V was the major component in proximate composition (Fig. 2a). The sludge generally contained low amounts of C and V and high contents of N, S, and ash, which are shown as outliers in Fig. 1a. For the HTC operating conditions, the most common values of RT ranged from 5 to 240 min, T ranged from 140 °C to 310 °C, and WC ranged from 75.00% to 99.90%, respectively (Fig. 2b). For the output variables, the median yield, C, and O were 55.26%, 50.60%, and 23.46% (Fig. 2c). Moreover, the Van Krevelen diagrams of H/C and O/C showed that three reactions (i.e., decarboxylation, dehydration, and demethylation) could occur in the HTC (Fig. 2d), of which dehydration and decarboxylation were dominant. The overall distributions of all six output variables were approximately normal (Additional file 1: Fig. S2), which expressed that data randomly selected from the literature satisfied the laws of probability. The normal distributions of outputs also examined the reliability of hydrochar samples, which validated the assumption that normality is of fundamental concern for machine learning modeling and analysis (Yap and Sim 2011). A total of 27 hydrochar data from the HTC of mixed biomass were compiled, and the range of input variables is shown in Fig. 2e and f, where the distribution ranges of input variables of mixed biomass were within the data range of single biomass (Fig. 2). The range of mixed biomass was more concentrated without obvious outliers. In addition to elemental and proximate composition, the biochemical components were used as another indicator to identify the type of biomass. However, biochemical components (i.e., proteins, lipids, and carbohydrates) are often ignored or unknowable during the HTC process. To fill the missing data in dataset #1, the new dataset #2 was compiled using the elemental and biochemical composition to develop ML models (details provided in Sect. 3.2.1). Figure 2g represents the numerical intervals of elemental and biochemical composition. The data points covered nearly the whole three-phase graph space, which indicated that the collected dataset #2 covered most biomass types (Fig. 2h). Then, the biochemical components from dataset #1 were inserted using the above prediction model to compile a new dataset #4.

Fig. 1
figure 1

Method flowchart of this study

Full size image
Fig. 2
figure 2

Statistical data visualization of elementary and proximate composition  (a), operating conditions of the HTC (b), hydrochar characteristics (c) and Van Krevelen diagram (d) in dataset #1, input variables (e, f) in dataset #6, and elementary composition (g) and biochemical components (h) of biomass in dataset #2. The three components (proteins, lipids, and carbohydrates) were unified to ensure that they added up to 100% (data unified with a dry-ash-free basis)

Full size image

3.2 Model predictions and feature analysis

3.2.1 Predictions and impact analysis of biochemical components of biomass

Based on dataset #2, RF and GBDT single-target prediction models were used to predict the protein, lipid, and carbohydrate contents using biomass elementary composition as the inputs. The hyperparameters and training results of the three prediction models are shown in Additional file 1: Table S1, Fig. S3, and Table 3. A comparison of the training results showed that both RF and GBDT had better predictive performance for proteins and lipids, and the prediction effect (R2 > 0.90) of RF was slightly better than that of GBDT (R2 = 0.88). The R2 of RF and GBDT models for carbohydrates was only about 0.80, possibly because values of carbohydrates may have come from various sources, including experimentally measured values and calculated values (Eq. 1). The measurement methods vary from one carbohydrate to another, including anion-exchange chromatography (Hoekman et al. 2011), nuclear magnetic resonance spectroscopy (Yuan et al. 2011), and the phenol–sulfuric acid method (DuBois et al. 1956). The use of different methods led to a sizeable systematic error in the carbohydrate data, ultimately resulting in poor predictive accuracy and efficiency of the model for carbohydrates. It could be seen that most residues of the data points were concentrated between –5 and 5, and the residual distributions of protein and lipid prediction models were closer to zero than that of the carbohydrate prediction model (Fig. 4a–c). According to the S–W test results, the residues of the training and testing datasets obeyed the normal distribution for all RF models for predicting protein, lipid, and carbohydrate (p_S–W > 0.05, Fig. 4). However, the residues of the testing data in the carbohydrate prediction model were relatively large and diffused (Fig. 4c), indicating poor generalization ability. To more accurately determine the carbohydrate content and reduce prediction errors, the protein and lipid contents predicted by the RF model and those experimentally determined from ash were used to calculate the carbohydrate content by difference (Eq. 1). The calculated carbohydrate contents fitted well with the actual values in dataset #2, with R2 = 0.92 and RMSE = 5.77 (Fig. 3f), showing the model’s good predictive performance for proteins and lipids (Fig. 3a and b). Moreover, the residual distribution of calculated carbohydrate contents was similar to that of the training dataset in the prediction model (Fig. 4c and d, p_S–W > 0.05), validating that the calculated carbohydrate contents could reduce errors compared with direct prediction. Various ML models were compared for their ability to predict carbohydrate contents (i.e., cellulose, hemicelluloses, and lignin) of syngas from biomass fluidized gasifier as the input. However, the predictive performance was mediocre, and the highest R2 was 0.77 among all models (Kim et al. 2023). Carbohydrates from previous studies were obtained mainly by difference (Duan et al. 2013; Liu et al. 2021). Hence, it is reasonable to employ a similar method to obtain carbohydrates in this study.

Table 3 Result of ML prediction model performance under different HTC datasets
Full size table
Fig. 3
figure 3

Model performance and feature importance analysis for predicting (a, d) protein, (b, c) lipid, and carbohydrate (c) training and (f) validating based on RF models (dataset #2)

Full size image
Fig. 4
figure 4

The residue analysis for predicted values and actual values of (a) protein, (b) lipid, (c) carbohydrate, and (d) calculated carbohydrate based on RF models (dataset #2)

Full size image

To accurately establish the relationship between elemental components and biochemical composition, the RF model was selected for subsequent model evaluation and prediction. Figure 3 shows the predictions of the RF model for proteins and lipids based on dataset #2 and the ranking of their respective feature importance. The protein content was concentrated between 20% and 60%, and the lipid content was mainly below 30%. The ML model showed more accurate predictions for data ranges, while the confidence interval widened and the validated prediction accuracy decreased for intervals with relatively little data (Fig. 3a–c). The feature analysis exhibited that the N content of biomass significantly affected the protein content (Fig. 3c), while other elements did not (relative importance < 0.05) because proteins are nitrogen-containing compounds (Leng et al. 2020). In addition, the C content of biomass had the greatest influence on the lipid content, followed by the H content (Fig. 3d) because lipids are mainly composed of C and H atoms.

The elemental composition, proximate analysis, and HTC conditions in dataset #1 were applied as inputs for the above optimal RF models to predict proteins and lipids, thereby also determining carbohydrates by difference.

3.2.2 Multi-target predictions and impact analysis of hydrochar properties by biochemical composition and HTC conditions

After obtaining 648 data points of biochemical composition, the corresponding HTC conditions (T, RT, and WC) and hydrochar properties were compiled into dataset #3 to predict the properties of hydrochar. The ternary phase diagram displays the various data distribution relationships between biochemical composition and hydrochar properties (Additional file 1: Fig. S4) caused by interactions between biochemical components (Duan et al. 2018). The collected biomass types in dataset #1 were limited, so the ranges of predicted biochemical components (Additional file 1: Fig. S4) were narrower than those in dataset #2 (Fig. 2h). Table 3 shows the performance of the ML model in terms of using biochemical components and HTC conditions as inputs. The optimal hyperparameters are determined and listed in Additional file 1: Table S1, in which all hydrochar properties were predicted with relative accuracy (training R2 > 0.97 and test R2 > 0.82 of each output). The biochemical components showed good linear correlation with hydrochar properties (|r|> 0.40) except Yield_char and T was correlated with Yield_char among HTC conditions (Fig. 5), resulting in high predictive performance of hydrochar properties (average test R2 ~ 0.90). The overall feature importance of hydrochar and the individual PDP of each biochemical component to various hydrochar properties are shown in Fig. 5a, Additional file 1: Fig. S5, S6. In particular, ash was the dominant feature of hydrochar, followed by T, while RT and WC had only slight impacts on hydrochar (Additional file 1: Fig. S6a). One reason for the major impact of ash on Yield_char was that the ash contents were used to calculate Yield_char in the literature. A higher carbohydrate content promoted O_char, C_char, H_char, and HHV_char. When the carbohydrate content rose from 10% to ~ 40%, Yield_char gradually increased (Additional file 1: Fig. S5), indicating that carbohydrate hydrolysis may have occurred in this interval to form hydrochar. Upon further increasing the carbohydrate content, the average C_char and O_char also increased from 20–25% to 45–52%, respectively, resulting in the polymerization of precursors (e.g., α-carbonyl aldehyde) from carbohydrates (Shi et al. 2019). Moreover, proteins are N-containing components and could participate in the Maillard reaction with carbohydrates, which affected the Yield_char and N_char (Fig. 5). Lipids exhibited a steady trend to Yield_char, N_char, and O_char, indicating that they had little contribution (Additional file 1: Fig. S4). However, lipids preferentially generated C_char, H_char, and HHV_char, with nearly linear relationships (Fig. 5 and Additional file 1: Fig. S5). Although the ML model here performed well, the biochemical components and HTC conditions (six inputs) were insufficient to describe the HTC process (six outputs), especially Yield_char. Hence, it will be necessary to consider more input descriptors to train more accurate ML models.

Fig. 5
figure 5

Pearson correlation coefficients (r) matrix of inputs features and targets (dataset #4)

Full size image

3.2.3 Multi-target predictions and impact analysis of hydrochar properties by biomass components and HTC conditions

After filling dataset #1 with the predicted biochemical components, the GBDT model was applied to predict the yield, elemental composition, and HHV of hydrochar according to 14 input features. These features included the elemental composition, proximate analysis, biochemical components of biomass, and HTC conditions (Additional file 1: Fig. S6).

The PCC matrix of input and output based on dataset #4 (Fig. 5) shows that C_char, H_char, and HHV_char of hydrochar were positively correlated with the C, H, and V contents of biomass (r > 0.50), and strongly negatively correlated with ash content (r < −0.60). The N and protein contents of biomass were strongly positively correlated with N_char and N/C (r > 0.75), which indicates that the N content of the raw material was the main factor affecting the N content of hydrochar. The N/C ratio of hydrochar was negatively correlated with the O content and carbohydrate content (r <  −0.70). Temperature showed a weak negative correlation with hydrochar, excluding C_char and HHV_char. The output variables such as Yield_char, H/C, and O/C had no significant linear correlation with the input characteristics. The target features also showed correlations, e.g., the atomic ratios of hydrochar were positively or negatively correlated with the corresponding elemental composition. Therefore, the elemental atomic ratios (H/C, O/C, and N/C) were classified as a separate category to be predicted to avoid similar information with the elemental composition of hydrochar. HHV_char was positively correlated with C_char and H_char (r > 0.70), and Yield_char was negatively correlated with C_char (r = −0.39) and HHV_char (r = −0.42). In summary, the input variables were mostly correlated with the outputs, with PCC values mostly between –0.5 and 0.5 (except C_char, H_char, and HHV_char), showing that these descriptors were adequately representative. Thus, the feasibility of using multi-target ML prediction was demonstrated, showing it could be used to draw generalized conclusions about typical correlations.

Selected optimal hyperparameters are shown in Additional file 1: Table S2, and the comprehensive prediction performance is presented in Table 3. In Fig. 6, the horizontal coordinate represents the original experimental data, and the vertical coordinate represents the data predicted by the model. The closer the scatter point to the bisect line (function y = x), the more accurate the model; otherwise, the model error will be larger. After adding the elemental composition and proximate analysis into dataset #3, the training accuracy of hydrochar properties was improved, with both training R2 values reaching 0.99 and the average testing R2 increasing from 0.87 to 0.93 (Table 3 and Fig. 6). This value was superior to that (0.86–0.91) of a previous study that did not consider biochemical components (Li et al. 2020a). The results showed that adding biochemical components improved the prediction performance of hydrochar properties to a certain extent, providing a means to characterize biomass raw materials. Therefore, considering more key features of biomass components (containing elemental composition, proximate analysis, and biochemical components) could comprehensively represent biomass and further improve the prediction accuracy of ML models for hydrochar properties, due to correlations between inputs and outputs variables (Fig. 5).

Fig. 6
figure 6

Multi-target predicted data versus the actual data based on the optimized GBDT for predicting (a) yield, (b) N content, (c) O content, (d) C content, (e) H content, and (f) HHV of hydrochar (dataset #4)

Full size image

As illustrated in Additional file 1: Fig. S7, the input features had various effects on the different targets of hydrochar. SHAP assigned an importance value to each feature variable for each predicted data point, where a positive SHAP value indicates that the input features are positive to the predicted target. A red or blue point indicates a high or low original value of the input features, respectively. It can be assumed that the input feature is positively related to the output target when the data points of the input feature change from blue to red upon increasing the SHAP value. Moreover, the mean absolute SHAP (|SHAP|) values of each input feature were aggregated for the six characteristics of hydrochar. The importance ranking of input features during the multi-task prediction process was conducted according to the sum of |SHAP| values for the six outputs. The one-way PDP provided more specific functional information of the input variables to each output (Additional file 1: Figs. S8–12).

In the Yield_char prediction, the HTC temperature, N content, and O content were identified as the three most significant variables, among which the biomass N content and temperature negatively affected the Yield_char (Additional file 1: Fig. S8). The low O content (< 20%) increased Yield_char. When the O content was between 20% and 26%, the relationship between O and Yield_char was linear negative, and further increasing the O content did not significantly change Yield_char (Additional file 1: Fig. S8). The N and protein contents were the two most critical contributors to N_char, showing that the N atoms in hydrochar were mainly derived from N-containing components of biomass (Fig. 7b). The positive correlation between N and N_char was stronger than that between protein and N_char because protein was mainly predicted and affected by the N content (Fig. 3d and Additional file 1: Fig. S9). In O_char prediction, the O content, T, and RT were the most critical input features, in which O_char was negatively correlated with T and RT and positively correlated with the O content (Additional file 1: Fig. S10). The cumulative O content indicated a reduction in the organic matter content and O_char aggravation. A higher T and RT accelerated deoxidation and dehydrogenation reactions (Additional file 1: Fig. S10) during the HTC process. The biomass ash was slightly linearly negatively correlated with C_char and H_char (Additional file 1: Figs. S11–12) due to an increase in inorganic matter (Eqs. 1–3). Therefore, the ash content was the key characteristic affecting C_char (Fig. 7d), which might have been conducive to decarboxylation and aromatization reactions during HTC (Buss et al. 2019). For HHV_char prediction, the H content, ash content, and temperature had the greatest influence, where the H content (6–8%) and temperature both positively impacted HHV_char, while ash and HHV_char were negatively correlated.

Fig. 7
figure 7

Average impact of the top three important features by one-way PDP for (a) Yield_char, (b) N_char, (c) O_char, (d) C_char, (e) H_char, and (f) HHV_char with GBDT model developed based on dataset #4

Full size image

Furthermore, ash, T, and N and C contents had the greatest overall impact on the six prediction objectives (Fig. 7b). In general, insoluble substances in ash decreased the yield and quality of hydrochar. However, the release of alkaline earth and alkaline earth metals (e.g., Ca) from the ash to the hydrochar surface increased the yield by catalyzing the hydrolysis of biomass to hydrochar (He et al. 2022a, 2022b). Thus, the impact of ash on Yield_char might have been neutralized, showing a steady trend (Additional file 1: Fig. S7). Therefore, temperature and C content could be identified as significant features because temperature affected the degree of thermal degradation of biomass, and C was the main energy source of hydrochar. However, the N content of biomass was another key factor that affected the properties of hydrochar, especially during the prediction of Yield_char.

The importance of elemental composition, biochemical components, proximate analysis, and operating conditions to the six properties of hydrochar was 35%, 25%, 21%, and 19%, respectively (Fig. 7b). These results indicate that elemental composition and biochemical components were fundamental characteristics that affected the properties of the resulting hydrochar (the total relative importance was about 60%). The relative importance of biomass components (elemental composition, biochemical components, and proximate analysis, > 80%) was much higher than that of HTC conditions (< 20%), showing hydrochar properties were more dependent on the composition of feedstocks. Because biochemical composition was predicted, some of the information about the biochemical components was similar to that of the elementary composition. Therefore, it was reasonable that the elemental composition was more important than the biochemical components. Furthermore, the feature importance ranking and analysis were almost the same as the models based on dataset #3 in which the ash and T were still the top two features (Additional file 1: Fig. S6). This was different from the previous results (Additional file 1: Fig. S6a), which showed that among biochemical components, carbohydrates were the main contributors to hydrochar formation (Additional file 1: Fig. S6b), consistent with previous research results (Li et al. 2019). N_char increased with the carbohydrate and protein contents (Fig. 7 and Additional file 1: Fig. S9) due to the Maillard reaction between proteins and carbohydrates during the HTC of biomass.

3.3 Practical validation for the online intelligent system

An online intelligent system (website link: http://43.139.222.20:8000/) for predicting and optimizing hydrochar’s properties based on the above multi-target ML models was proposed and made accessible to other researchers. The homepage of the website is shown in Additional file 1: Fig. S13. More critically, the reliability and generality of the proposed system should be verified by practical data other than the training and testing dataset.

3.3.1 Verification of the atomic ratios (N/C, O/C, and H/C) with the predicted elementary composition of hydrochar from the model

A multi-objective prediction model of elemental atomic ratios was established using dataset #5, which used biomass components and the HTC conditions as the inputs. The optimal hyperparameter and predictive performance results are shown in Table 3 and Additional file 1: Table S2, respectively. The overall predicted results were accurate (the average training R2 was 0.98, and the average test R2 was 0.86). Still, there was a significant difference between the test results of the three atomic ratios. Only the test R2 of N/C reached 0.94, while those of O/C and N/C were lower than 0.85 (Table 3), which might be because dozens of the H/C and O/C values of hydrochar were dispersed outside the concentrated area in the Van Krevelen diagram (Fig. 2).

Compared with the simultaneous prediction of all three atomic ratios, the predictive performance of the multi-target models for predicting six properties of hydrochar (dataset #4) was slightly improved by using the training dataset (R2 from 0.98 to 0.99; RMSE from 0.027 to 0.022). On the test dataset, the predictive accuracy of N/C and O/C was improved, with the R2 of O/C increased significantly from 0.78 to 0.97 (Table 3). The predictive accuracy of H/C slightly decreased (Table 3), possibly because the predictive performance of H_char (test R2 = 0.89) was worse than that of N_char and O_char (test R2 > 0.90) in the multi-target prediction model for hydrochar (Fig. 6). The direct prediction of elemental composition from the multi-target prediction model could reduce transfer errors (Qureshi et al. 2017) between H & O and C during calculations. Overall, the accuracy of predicting the elemental composition and then calculating the atomic ratios using the multi-objective prediction model for hydrochar properties was much better than that of predicting the atomic ratios directly. Therefore, the multi-target model based on dataset #4 accurately predicted the hydrochar elemental composition and provided good generalization performance for elemental composition. The system can be used to accurately calculate and predict the atomic ratios of hydrochar samples.

3.3.2 Forward optimization for hydrochar with specified biomass and experimental verification

From many prediction results from HTC of soybean straw with the system (12,420 iteration results, Additional file 1: Fig. S14), eight optimal solutions (maximized \({\text{optima}}{\text{l}}_{\text{char}}\)) with different operating conditions were screened for experimental verification (Table 4). Herein, \({\text{optima}}{\text{l}}_{\text{char}}\) was determined as follows:

$$Optimal_{char} = Maximum\left( {Yield\_char + C\_char + H\_char{-} O\_char {-} N\_char} \right)$$
(11)
Table 4 Forward optimization and experimental verification of hydrochar properties
Full size table

The WC was set at 75% or 90% in the forward experiments because the feature impact of WC was less significant for hydrochar than T and RT (Additional file 1: Fig. S6). In all cases, the absolute error of Yield_char between experiment and predicted values was < 5%. The elementary composition of hydrochar (N_char, C_char, and H_char) was experimentally verified, with an absolute error of < 1%. The absolute values of relative errors for all cases were lower than 15% (except one of 21.89%, negligible). In fact, most cases were within 10%, and one-third of cases were < 5% (Table 4), indicating satisfactory validating performance. However, the predicted values of O_char were higher than some experimental values (Runs 2, 3, and 6), in which the O_char calculated via difference (Eq. 4) might also be a possible reason for the deviation. Overall, the relative errors between the predicted and experimental values of hydrochar properties were small, and the generality of the model was verified by optimization experiments. This confirms that the intelligent system based on multi-target modeling prediction can be used for the aforementioned single biomass.

3.3.3 Verification of hydrochar from co-HTC

During actual production, it is necessary to choose the appropriate ratio according to the biomass raw materials to produce high-yield and high-quality hydrochar under optimal HTC conditions. To reduce the number of experiments, the system can be further applied to predict and adjust the HTC process of biomass mixtures.

Figure 8 shows the predictions of six hydrochar properties using mixed biomass-related input variables as inputs into the multi-target model. The validated R2 values of the GBDT multi-target model for the Yield_char, N_char, O_char, C_char, H_char, and HHV_char from mixed biomass were 0.70, 0.76, 0.72, 0.93, 0.83, and 0.88, respectively, with an average R2 of 0.79 (Fig. 8). The RMSE values were 5.79, 0.70, 4.81, 4.19, 0.54, and 2.24 for these six targets, respectively (Fig. 8). The validation result of Yield_char was the worst among all prediction targets, which resulted in the predictive performance of Yield_char (test R2 of 0.88) being lower than other hydrochar properties, with all test R2 ≥ 0.91 (Fig. 6). Moreover, the synergistic effects of different biomass could either promote or inhibit Yield_char during co-HTC process, which may have caused the Yield_char predicted by the multi-target model for single biomass to be lower or higher than the actual values (Fig. 8a). The validating performance of O_char was also unsatisfactory, similar to the forward experiments of single biomass (Table 4), due to the different acquisition methods of O_char values. The C_char, H_char, and HHV_char derived from biomass mixtures were predicted with validating R2 > 0.83 because T and ash were the same important features for all of them (Additional file 1: Fig. S6), which are generally provided accurately. In summary, the online intelligent system could be employed to preliminarily predict hydrochar properties in real industrial applications only considering the single-biomass composition without considering the synergistic effect from co-HTC.

Fig. 8
figure 8

Validation performance comparison of predicted values by the multi-target model (dataset #4) and experimental values of hydrochar obtained from a biomass mixture

Full size image

3.4 Future perspective and outlook

Although the system developed in this study was satisfactory, there is still room for improvement. More detailed input features could be considered to describe the HTC process, such as catalysts, solvents, ash compounds, pressure, and water recirculation. More descriptors for hydrochar, such as structural composition (functional groups), fuel properties (density), and material properties (e.g., surface area, surface morphology, pore volume, and cation exchange capacity), could be measured to more comprehensively assess the value of hydrochar for certain applications. The above descriptors are not considered in the proposed model, which is one limitation of this intelligent system based on the model.

The generality of this system for predicting hydrochar properties from biomass mixtures was inferior to that for single biomass, possibly because of the synergistic effects of biomass mixtures. Hence, more data related to hydrochar from biomass mixtures could be added to obtain datasets to improve the generality of our system. Additionally, more intelligent functions of the developed system could be exploited and upgraded, for example, the automated online reading of input datasets, the prediction of batches of data, reverse optimization, and error calculation.

4 Conclusions

In this study, biochemical components (i.e., proteins, lipids, and carbohydrates) of biomass were predicted first by their elementary composition. A novel intelligent system based on a multi-target ML prediction model (average test R2 of 0.93; RMSE of 2.36) was established to predict and optimize the yield, elementary composition (C_char, H_char, O_char, and N_char), and HHV of hydrochar by using single biomass components and HTC operating conditions as inputs. Interpretation of the model’s results showed that the ash content, temperature, and N and C contents of biomass were the most important factors affecting the six properties of hydrochar. Finally, an intelligent system was constructed based on a multi-target model and verified by applying it to predict the atomic ratios (N/C, O/C, and H/C), optimize hydrochar yield from the HTC of single biomass samples, and predict hydrochar properties from the co-HTC of mixed biomass samples. The proposed intelligent system provided accurate predictions and has practical generality, showing its great potential for industrial applications.