Removal of Cr(VI) ion from aqueous solution using jute stick lignin and application of machine learning approach to determine optimal parameters

2.1 Preparation of adsorbent

The preparation of the adsorbent method for the characterization of adsorbent, adsorption experimental procedures, and batch experiments for the study of adsorption kinetics and isotherms have been described in the supplementary information.

2.2 Data preprocessing for ML algorithms

From the experiment of Cr(VI) adsorption onto Lignin, 145 data points were extracted and used for ML models. The parameters considered as input data for the models were (i) initial concentration of Cr(VI); (ii) adsorbent dosage; (iii) contact time; (iv) solution pH; and (v) temperature. At the same time, the percentage removal of Cr(VI) was selected as the output of the models. Linear dependence between any two input parameters or between any input parameter and the targeted output parameter (i.e., percentage removal of Cr(VI)) was inspected by Pearson’s correlation coefficient (PCC) (Benesty et al., 2009), given by Eq. (1).

Where, $r_{xy}$ is the PCC value between any two parameters $x$ and $y$. $\overline{x}$ or $\overline{y}$ are the mean value of the variable $x$ or $y$. PCC ranges between $1$ and $-1$. $r_{xy}=1$ denotes a strong linear dependence between the variables, while $r_{xy}=-1$ indicates a high negative correlation. $r_{xy}=0$ suggests there is no correlation between the variables.

Prior to training the ML models, the data of input variables undergo normalization using the Z-score standardization approach as described in Eq. (2).

Where, $x_i^{\text{*}}$ indicates the normalized value of the original variable $x_i$. $\mu$ and $\sigma$ are the mean value and the standard deviation of $x_i$, respectively. Some data points were observed that significantly deviate from the mean value of the dataset. Generally, these data points can exhibit unusual behavior or extreme observations. They were identified and eliminated using the Z-score. The data points with a z-score ≥3 or ≤ -3 are considered extreme and are flagged for removal.

2.3 ML models

For predicting Cr(IV) uptake by lignin under different conditions in terms of material characteristics and adsorption conditions, four supervised ML algorithms, i.e., RF, XGB, ANN, and KNN, were selected.

RF represents an ensemble methodology, integrating several DTs to enhance accuracy in prediction and reduce overfitting (Breiman, 2001). RF produces a collection of DTs, with each tree trained on a distinct, randomly drawn portion of the original dataset and a randomly sampled offset of input features. These trees independently train, and their predictions aggregate through averaging to yield the final output. By leveraging the diverse perspectives of multiple trees, RF effectively reduces overfitting and enhances prediction accuracy.

XGBoost is an advanced ensemble learning method based on the gradient boosting framework (Chen and Guestrin, 2016). This framework sequentially constructs DTs in a boosting approach, where each tree actively corrects errors from preceding trees by assigning greater weights to mispredicted samples. Gradient descent optimization techniques are utilized to minimize its loss function. XGBoost incorporates regularization techniques into the objective function, allowing it to control the complexity of individual trees and prevent overfitting. It is designed to be highly parallelizable, making it more efficient for multi-core processors (Mamudur and Kattamuri, 2020). It can build trees in parallel, which significantly speeds up training compared to traditional Gradient Boosting, which builds trees sequentially.

ANN draws inspiration from biological neural systems (Jain et al., 1996). ANNs are composed of multiple interconnected layers of artificial neurons, which include input, hidden, and output layers. Each connection between neurons has a weight that adjusts during training. Neurons apply activation functions (i.e., Sigmoid, ReLU, Leaky ReLU, Tanh, etc.) to their weighted inputs, enabling complex computations (Apicella et al., 2021). The process initiates with feedforward propagation, where input data flows through the network, producing an output. A loss function measures the prediction error compared to the actual target. Backpropagation adjusts weights by propagating error backward through the network, using techniques like gradient descent. This iterative process continues until the network’s performance is satisfactory. In our study, ANN consisted of an input layer with five independent variables according to the experimental data, followed by three hidden layers comprising 50, 350, and 50 neurons, and lastly, an output layer that predicted removal efficiency.

The KNN method forecasts outcomes for new data points using the K most proximate neighbors in the dataset. (Kramer, 2013). The process involves selecting an optimal K, computing distances between the new data point and all training data using metrics such as Euclidean or Manhattan distance, and choosing the K closest neighbors. Finally, the prediction is computed from the average of the KNN target output.

To achieve accurate predictions and ensure robust generalization capabilities, it is essential to carefully tune the hyperparameters for each ML model. Hyperparameters, parameters fixed prior to training and not learned from training data (Probst et al., 2019), were optimized using grid search (Bergstra and Bengio, 2012) combined with 5-fold cross-validation (Refaeilzadeh et al., 2009). This method systematically explores a predefined hyperparameter grid by randomly splitting the dataset into five folds. For each hyperparameter set, models are trained on four folds and evaluated(test) on the remaining fold. Repeating this process five times, the average performance identifies the best hyperparameters, ensuring models generalize well and avoid overfitting.

2.4 Performance evaluation

The performance of ML algorithms in this research was measured using four different metrics i.e., mean absolute error (MAE), mean squared error (MSE), root mean square error (RMSE), and coefficient of determination (R²). MAE, MSE, RMSE, and R² can be expressed mathematically using Eqs. (3-6), respectively.

Where, $y_i$ and ${\hat{y}}_i$ are actual and predicted values of the target parameter, respectively. $\overline{y}$ denotes the average of all instances of the target variable in the dataset. $N$ denotes the number of instances. A ML model’s predictive performance is considered better when it exhibits a higher R² value and lower values for MAE, MSE, and RMSE(N. Moriasi et al., 2007; Shahmansouri et al., 2019).

2.5 Feature importance and partial dependence

Achieving optimal Cr(VI) adsorption by lignin requires a thorough analysis of each input parameter’s impact on the adsorption process. This analysis reveals optimal material characteristics and experimental conditions. To quantify each parameter’s contribution to removal efficiency, feature importance analysis was performed. Among four ML models, XGBoost and KNN were chosen due to their superior performance. Permutation feature importance (Altmann et al., 2010) was used to calculate feature importance for these models. Permutation feature importance measures the effect on the model’s prediction error when the values of a single feature are permuted (Molnar et al., 2020). By disrupting the relationship between the input data and the target feature, this approach indicates the extent to which the algorithm relies on the feature. Although permutation importance identifies the individual contribution of features to overall prediction, it does not reveal how these variables affect adsorption efficiency. To address this, partial dependence plots (PDPs) were applied, illustrating the effect of one or two features on adsorption efficiency in the selected models (Friedman, 2001). Results from both permutation importance and PDP analyses have been detailed in the results section, providing comprehensive insight into the roles of individual input parameters in optimizing Cr(VI) adsorption by lignin.

3.1 Characterization of the adsorbent

To characterize the jute stick lignin, fourier transform infrared (FTIR) data were recorded over 400–4000 cm^-1. A broad peak arises in the wavelength range of 2800–3800 cm^-1 (Fig. 1), which can be assigned as the stretching vibration of hydroxyl (–OH) groups of lignin. The bands observed at 1669 cm^-1, 1631 cm^-1, and 1459 cm^-1 might be due to aromatic skeletal vibration of lignin (Choudhary et al., 2020; Keiluweit et al., 2010). The band at 1050 cm^-1 is attributed to C–O stretching vibrations of alcohol groups or in-plane C–H aromatic deformations. The band at 685 cm^-1 describes out-of-plane C–H stretching of aromatic skeletal, while the peak at 560 cm^-1 is assigned to the out-of-plane bending vibration of O–H groups. The FTIR spectrum of lignin exhibited analogous absorption peaks, with shifting or reduced intensities in several bands, stipulating the uptake of chromium ions due to the adsorption of Cr(VI) onto the lignin surface. After adsorption of Cr(VI), a fresh peak appeared at 2945 cm^-1 due to C–H stretching of –CH₂ groups, while additional peaks at 1243 cm^-1 and 1103 cm^-1 might be assigned to C–H stretching of the tertiary alcoholic groups (Papaevangelou et al., 2023). The experimental FTIR data for the jute sticks lignin are well consistent with the recently published literature on lignin derived from agricultural waste (Md Salim et al., 2021; Nguyen-Thi et al., 2024).

Close

Fig. 1.

FTIR spectrum of lignin (a) before adsorption, (b) after Adsorption of Cr(VI) from aqueous solution by using jute stick lignin.

Export to PPT

Scanning Electron Microscopy (SEM) images were used to observe the structural morphology of lignin, see Fig. 2. The SEM images reveal lignin consists of a 3D layered arrangement of sheets of of varying sizes and shapes. The lignin sheets display a very flat and smooth surface with a porous microstructure. SEM images of lignin after the adsorption of Cr(VI) also demonstrate a 3D layer arrangement of sheets, but surface roughness is more pronounced compared to the lignin surface before the adsorption of chromium due to the uptake of Cr(VI) onto the surface of lignin.

Close

Fig. 2.

SEM image of lignin (a) before adsorption, (b) after adsorption of Cr(VI) ions.

Export to PPT

Fig. 3(a) shows the UV-visible spectra of lignin prepared from jute sticks. The spectra showed a wide adsorption peak from 250–300 nm and a shoulder peak at 300–350 nm. The characteristic absorption band between 250–300 nm ascribes the π → π* transition of unsaturated benzene rings and the aromatic hydroxy groups of lignin. The associated shoulder band observed at 300–350 nm is assigned to the n → π* transition of α-β unsaturated and α-carbonyl functionalities of lignin molecules (Mohamad Ibrahim et al., 2011). These characteristic and prominent absorption peaks indicate that lignin is structurally composed of different aromatic rings and alcohol groups such as synapyl alcohol, p-coumaryl alcohol, and coniferyl alcohol (Kumar and Dwivedi, 2021). To further characterize lignin, powder X-ray diffraction was carried out, and Fig. 3(b) demonstrates the X-ray diffractogram of lignin. A broad, intense reflection band is observed over a wide 2θ range of 20°–30°. This suggests that the lignin surface is amorphous in nature, where 2D graphene sheets are packed together into small domains of carbon nanosheets. Furthermore, the maximum values of 2θ arise at 20.42°; this stipulates that type-I cellulose predominantly forms the lignin structure (Dutta et al., 2022).

Close

Fig. 3.

(a) UV-visible spectra, (b) X-ray diffraction pattern, and (c) TGA thermogram of lignin.

Export to PPT

Fig. 3(c) represents the thermogravimetric analysis (TGA) of jute stick lignin. The thermogram shows that the degradation of lignin occurred in four different stages within temperature ranges of 50–600°C. Stage-I (50–130°C) is the volatilization region; water and moisture content evaporate and resulting in about a 10% weight loss of lignin. Stage-II (130–290°C) is less steep than stage-I; the different polymeric binders, such as hemicellulose content, degrade slowly over this region. In stage-III (300–480°C), decomposition of lignin occurred quite harshly in this area due to the complex linkages between fundamental structural molecules like phenolic, carbonyl, and benzylic hydroxyl groups and the initiation of carbonization in lignin molecules (Nguyen-Thi et al., 2024). In stage-IV (>480°C), charcoal or coke carbon is formed due to extensive cross-linking and the formation of a highly dense aromatic structure.

3.2 Comparison of different ML models

In this study, four ML models (RF, XGBoost, ANN, and KNN) were used to predict Cr(VI) adsorption by lignin, utilizing the input parameters outlined in the data collection section. During model training, hyperparameter tuning was accomplished using a grid search technique and five-fold cross-validation. The hyperparameters of the four models were optimized to get superlative predictive performance. Table S2 of the Supplementary information lists the hyperparameters and their tuned values for all ML models. After the development of the ML models and the optimization of their hyperparameters, an evaluation was conducted to assess their prediction accuracy on the dataset. The joint scatter plots of the actual values vs. the predicted values of Cr(IV) uptake onto lignin by different models have been shown in Fig. 4. In each plot, a precise match between the predicted removal and actual removal is demonstrated by a diagonal straight line. Validation parameters encompassing R², RMSE, MSE, and MAE underpinned the selection of the most accurate model for chromium removal from water. The values of different metrics for various models have been presented in Table 1. It is evident from Fig. 4 that all the models demonstrate a strong alignment with the ideal fit represented by the diagonal straight line. This is also supported by the selected evaluation metrics for this study, indicating that our predictions closely match the actual values. All models exhibited high R² values (0.9905 for KNN, followed by 0.9522 for XGBoost, 0.9084 for RF, and 0.8802 for the ANN model). Furthermore, in terms of other metrics (MSE, MAE, and RMSE), the models demonstrated minimal values, indicating minimal error in their predictions, as shown in Table 1. Consequently, the higher R² values and lower RMSE, MSE, and MAE values for this study imply that all four models could be applied efficiently for predicting the chromium ions by lignin. Although the performances of all the models were comparable, KNN produced the most accurate predictions, followed by XGBoost, according to the joint scatter plots in Fig. 4 and the results shown in Table 1. These results stipulate that KNN and XGBoost models outperform more competently and accurately to predict the removal efficiency of chromium by the adsorption method, although all the ML models are well acceptable for predicting the adsorption of Cr(VI) onto lignin. Herein, KNN and XGBoost models were selected as the best ML algorithms among the four models to develop a chromium adsorption model for further analysis. These results align with the literature on the removal of antibiotics using carbon-based adsorbents (Sigmund et al., 2020).

Close

Fig. 4.

(a-d) Comparison of removal predicted by different machine learning models and actual removal in the experimental data for the adsorption of Cr(VI) onto lignin.

Export to PPT

3.3. Effect of parameters on the adsorption of Cr(VI) by lignin

The well-trained KNN and XGBoost algorithms serve as the reliable predictive models for the removal of chromium by lignin. Thus, these two models are applied to assess the impact of various variables on chromium removal efficiency. The effect of adsorbent dose (experimental black dotted line) and partial dependence plot of a single parameter, like load of adsorbent (red line and blue line obtained from XGBoost and KNN, respectively), has been presented in Fig. 5(a). The removal efficiency demonstrates an upward trend as the lignin dose increases, reaching its peak at 50 mg/L for experimental results, and more than 90% of the dye is removed. However, as the lignin content increases after 50 mg/L, the removal efficiency gradually decreases. This observation suggests the paramount importance of adsorption capacity for eliminating chromium. The observed relationship between removal efficiency and load of lignin can be attributed to the presence of more surface areas and an increased number of active sites for Cr(VI) adsorption. On the contrary, a slight decrease in the removal efficiency with the increase of lignin load over 50 mg/L could be attributed to a reduction in the number of active adsorbed sites, which is caused by the formation of lignin aggregates through π–π interactions between the aromatic rings of adjacent lignin molecules (Ma et al., 2020; Zhao et al., 2017). The predicted results of KNN and XGBoost models also demonstrated similar trends as the experimental results, except that the maximum removal is found to occur over a range of 40–60 mg/L adsorbent load, and the maximum removal is found to be 78%. The experimental results for removal efficiency as a function of initial concentration of Cr(VI) solution, along with a partial dependence plot for the same input parameter, have been presented in Fig. 5(b). The Experimental results closely match the predicted results obtained from the KNN and XGBoost models. The removal efficiency of chromium is found to increase with the initial concentration of Cr(VI) solution, reaching the maximum value at an initial concentration of 50 ppm for experimental results. However, the maximum adsorption is found over the range of 40–60 ppm for both KNN and XGBoost models. As the concentration of Cr(VI) increases, more Cr(VI) ions are available for colliding with the active adsorption sites on the lignin surface, and there is a higher chance for the uptake of Cr(VI) by lignin (Di et al., 2022). Moreover, after a certain concentration (above 50 ppm for experimental and over 60 ppm for predicted results) removal rate is found to decline again gradually with the rise of Cr(VI) concentration. For a definite amount of lignin dose, a certain number of active sorption sites are found. Once Cr(VI) ions saturate the active sites, any rise in metal ion concentration does not produce extra adsorption. This is because the quantity of chromium ions exceeds the number of available active adsorption sites. In addition, as the concentration of metal ions increases, mass transfer resistance increases for the adsorbate to become adsorbed, resulting in a decrease in adsorption. Similar results were also reported for the adsorptive removal of Pb²⁺ ion from aqueous solution by lignin (Liu et al., 2019).

Close

Fig. 5.

Comparison between experimental results and partial dependence plots obtained from machine learning models for (a) initial concentration of Cr(VI), (b) lignin dosage, (c) time of adsorption, (d) pH of the medium, (e) adsorption mechanism (f) temperature. Blue and red lines represent the results obtained from KNN and XGBoost models, respectively.

Export to PPT

Adsorption duration serves as a pivotal factor indicating the kinetics of adsorption under a specific metal ion concentration. The results show that the slope of removal efficiency is high at the initial stage of adsorption, and later it reaches a steady value at about 80 min (Fig. 5c). The faster rate of adsorption might be ascribed to the availability of higher active sites at the surface of the adsorbent for the adsorption of chromium. The predicted results exhibit trends similar to the experimental results, but the equilibrium appears to be reached earlier than in the experiment. It is also noticed that more than 90% chromium is removed at equilibrium time (experimental), whereas the predicted results show a maximum removal efficiency of about 70% at equilibrium time.

The surface characteristics of the adsorbent, ionization probability, the charge on the adsorbent surface, and the amount of metal ions in the aqueous medium are strongly governed by pH. The dependency of removal efficiency on pH (experimental) and partial dependence plots for chromium removal by the KNN and XGBoost models have been shown in Fig. 5(d). Higher removal efficiency is observed under acidic conditions and progressively downward with the increase of pH for experimental results. It was observed that uptake of Cr(VI) ions by modified lignin principally results from ionic interactions of Cr(VI) ions and protonated functional groups of lignin (Demirbaş, 2005). At lower pH (2), the different functional groups of lignin, such as phenolic, hydroxyl, carbonyl groups, etc., are protonated and become positively charged. On the other hand, chromium exists predominantly as HCrO₄^–. As a result, electrostatic interaction between these two oppositely charged particles favors the uptake of Cr(VI) onto lignin (Albadarin et al., 2011). It has been reported that HCrO₄^– ions can easily be diffused toward the lignin surface and get adsorbed (Tazerouti and Amrani, 2009). The removal efficiency is lower at higher values of Tazeroutiinterfere with HCrO₄⁻ adsorption by occupying active sites on the adsorbent. (Albadarin et al., 2011). Moreover, some chromium ions are deposited as chromium hydroxide by the formation of a precipitate at higher pH (Afzaal et al., 2022). It has been reported that conversion of HCrO₄^– to CrO₄^2– occurs at higher pH; on the contrary, at higher pH surface charge of lignin is reduced due to deprotonation and hence forms a neutral or a negative surface (Kwak et al., 2016). However, it is observed that more than 40% chromium is removed at higher pH. At higher pH, lignin surface becomes negatively charged and Cr(VI) is reduced to Cr(III) when it comes in contact to lignin via different functional groups (–OH, –O–, C=C, C–O, –OMe), which act as an electron donor, and adsorbs on the lignin surface through H-bonding (Elrayyes et al., 2025; Kwak et al., 2016), as shown in Fig. 5(e). Predicted results of pH also resemble the experimental data. The predicted results of XGBoost demonstrate that removal efficiency is almost steady up to pH 6 and then decreases in a similar way as that of experimental results. The predicted results of KNN demonstrated little deviation from the experimental results, and it is shown that the removal of chromium is not dependent on the change of pH of the suspension. Moreover, the predicted results reveal a little bit lower removal efficiency (72%) than that of the experimental results (85%). The above results suggest the possible mechanism of chromium (VI) ions by lignin and electrostatic interactions, and redox reaction plays the pivotal role for the uptake of Cr(VI) ions by lignin, as represented by Fig. 5(f).

The effect of temperature on the adsorption efficiency has been shown in Fig. 5(f). The figure shows that the percentage of chromium removal initially shows a slightly decreasing trend and then increases sharply as the temperature rises, and a maximum of 98% chromium is removed at 60°C. The decrease in adsorption observed at elevated temperature implies a physical adsorption mechanism. As the temperature rises, the kinetic energy of adsorbate molecules increases, and adsorbate molecules may be released from the active sites of lignin, following a drop in Cr(VI) uptake up to 40°C. When the temperature increases beyond 40°C, the adsorption efficiency is found to increase, which indicates that the process of adsorption is endothermic in nature. The high temperature might alter the adsorbent, its active sites, and activity, and augment the chemical interaction between Cr(VI) and lignin. Furthermore, high temperature might enhance the diffusion rate of chromium from the bulk to the surface of lignin, which results in an increase in Cr(VI) adsorption. The predicted results also demonstrate similar behavior as experimental results, but the percentage of removal is found to be lower than experimental results.

3.4 Correlation analysis

Fig. 6 displays the relative importance of various input parameters based on two best-performing models (KNN and XGBoost). It has been found from Fig. 6(a) that the removal efficacy of Cr(VI) ions has a strong correlation with adsorbent dose, temperature, and initial concentration, but has a weak correlation with time and pH of the medium. These results suggest that the common vital independent parameters are adsorbent (lignin) dose, initial concentration of chromium, and temperature. For the KNN model, the adsorbent dose ranked highest among these independent variables (31%), followed by initial concentration (25%) and temperature (22%). For the XGBoost model, the most influential parameters are temperature (29%), initial concentration (27%), and adsorbent dose (23%), respectively. This result reveals that the adsorption capacity of chromium by lignin is primarily affected by the lignin dose, initial concentration of chromium ions, and temperature. Similar findings have been reported for dye adsorption by agricultural waste using several machine learning algorithms in previous studies (Moosavi et al., 2021). Moreover, the removal efficiency of chromium ions is found to be positively correlated with the initial concentration of chromium ions, lignin dose, adsorption time, temperature, and pH of the medium, as shown by the Pearson correlation matrix in Fig. 6(b). Initial concentrations of chromium have a positive correlation with adsorbent dose because as the dose of adsorbent and initial concentration of chromium ions increase, more fresh adsorbed sites become available for adsorption of more chromium ions. Again, the lignin dose negatively correlates with the time. This result indicates that once all adsorption sites are occupied and equilibrium is reached, no significant adsorption occurs beyond this point due to a lack of available sites after a definite time interval.

Close

Fig. 6.

(a) Relative importance of input features based on KNN and XGBoost models, and (b) Pearson correlation matrix between any two variables and removal efficiency.

Export to PPT

3.5 Double factor analysis

The two-factor PDP analysis was utilized to evaluate the synergistic relationship between adsorption conditions. Figs. 7(a and b) for XGBoost and KNN, respectively) represents the combined effect of two variables (initial concentration of Cr(VI) and lignin dose) on the removal efficiency of chromium. It is found that the dependence of removal efficiency is higher at lower concentrations of chromium ions and lower doses of adsorbent. The maximum uptake was observed at chromium ion concentration of 30–50 ppm and an adsorbent dose of 40–60 mg/L for both models, followed by a gradual decline at higher concentration and dose. The overall maximum removal efficiency was 75%, which is lower than the experimental results mentioned earlier. With increasing chromium concentration and adsorbent dose, more metal ions and more active adsorption sites become available, leading to improved efficiency (Kabir et al., 2025; Miah et al., 2017). However, at higher concentrations, mass transfer is restricted (due to a reduced effective diffusion coefficient of ions (Bilbao et al., 2016; Miah et al., 2016). Additionally, some lignin molecules (at a higher dosage) might aggregate due to π–π interactions between adjacent aromatic layers, reducing the number of available active sites for adsorption (Ma et al., 2020).

Close

Fig. 7.

Two-factor partial dependence (a and b) effect of initial concentration of chromium ion vs. dose of adsorbent, (c and d) effect of initial concentration vs. temperature, (e and f) effect of contact time vs adsorbent dose according to XGBoost and KNN model, respectively, for the removal of chromium by lignin.

Export to PPT

The combined effect of two variables, the initial concentration of chromium ions and temperature, has been presented in Figs. 7(c and d) for XGBoost and KNN, respectively. It is noticed that the temperature had a comparatively greater impact on the removal efficiency than the initial concentration of chromium ions for both models. The results show that removal efficiency is better at higher temperatures and moderate initial concentrations of chromium ions. In contrast, it has been shown that the percentage of removal decreases at lower temperatures, both for lower concentrations and higher concentrations of metal ions, likely owing to the combined influence of these factors. It is noteworthy that the removal efficiency exhibits an increasing trend with an initial increase in concentration, followed by minimal fluctuation, and thereafter a declining trend with further increases in concentration. The maximum removal efficiency is found within the concentration range of 40–50 ppm. However, an increasing trend of removal efficiency is observed with the rise of temperature, and the maximum efficiency is found at 60°C for both KNN and XGBoost models. At elevated temperatures, mass transfer kinetics are enhanced because of the increase in kinetic energy of adsorbates, and more adsorption sites are activated through physical and chemical interactions, resulting in higher removal efficiency (Bilbao et al., 2016).

The mutual interaction of two variables, adsorbent dose and adsorption time, has been presented in Figs. 7(e and f) for XGBoost and KNN, respectively. The combined effect of contact time and adsorbent dose manifests in the gradual increase in removal efficiency with contact time for both models. An initial increase of removal efficiency followed by a decreasing trend with adsorbent dose is found for both models, and the maximum removal efficiency is observed over the time frame of 80–120 min and adsorbent dose of 40–60 mg/L. An abundant number of functional groups (alcoholic –OH, phenolic –OH, –CO–, –COOH, methoxy group, and sulfonic groups) of lignin are available to interact with chromium ions. Thus, higher removal is observed at moderate adsorbent load, but at higher load, some adsorbent molecules might form aggregates or coagulate, resulting in a decline in removal efficiency (Kabir et al., 2025).

3.6 Adsorption isotherm analysis

Adsorption isotherms were studied by using chromium solutions of 3 mg/L to 40 mg/L, and the resulting data were applied to the Langmuir isotherm and the Freundlich isotherm models, represented by Eqs. (7) and (8), respectively.

Where C_e is the concentration (mg/L) of Cr(VI) at equilibrium, Q_e is the amount of chromium (mg/g) at equilibrium, q_max is the adsorption capacity (mg/g) of lignin at monolayer coverage, K_L is the Langmuir adsorption constant, K_F is the Freundlich adsorption constant (mg^1-1/n L^-1/n g^-1), n is a constant related to adsorption intensity.

Fig. 8(a) shows that the correlation coefficient for the adsorption of chromium by lignin is 0.987 for the Langmuir model and 0.930 for the Freundlich model. These results indicate that experimental data of the adsorption of chromium by lignin is more well-fitted to the Langmuir model than the Freundlich model. Fitting the experimental data to the Langmuir isotherm provided that the amount of chromium for monolayer coverage, i.e., the maximum adsorption capacity of chromium by lignin, was 97.06 mg/g. Fitting of experimental data to the Freundlich isotherm provided the value of $1/n$ = 0.887, indicating that the adsorption of chromium onto the surface of lignin is favorable (Sume et al., 2024).

Close

Fig. 8.

Study of (a) adsorption isotherm, (b) adsorption kinetics for the removal of Cr(VI) by lignin, and (c) recycling test of lignin.

Export to PPT

3.7 Adsorption kinetic analysis

To elucidate the adsorption kinetics, several experiments were conducted at various time spans, and the resulting data were applied to the pseudo-first-order and second-order kinetic models, as described by Eqs. (9) and (10).

Where $t$ is the time of adsorption, $Q_t$ is the adsorption capacity (mg/g) at time $t$, $Q_e$ is the equilibrium adsorption capacity (mg/L), and $K_1$ (min^-1) and $K_2$ (g/mg min) are the adsorption rate constants for pseudo-first-order and second-order models, respectively.

The experimental data for the adsorption of chromium by lignin were fitted to the two kinetic models, and the fitted graphs have been presented in Fig. 8(b). It has been found that the correlation coefficient for the pseudo-first-order kinetic model is 0.978, and the pseudo-second-order kinetic model is 0.899. This indicates that the pseudo-first-order kinetic model is more appropriate to describe the adsorption behavior of chromium by lignin (Kayes et al., 2016; Sume et al., 2024).

3.8 Recycling test and comparison with lignin-based adsorbent

To evaluate the regeneration efficiency of lignin, after completion of adsorption, Cr(VI) loaded lignin was immersed in 0.1M NaOH solution and stirred continuously for 120 min to achieve complete desorption of chromium. The adsorption–desorption process was repeated for up to six cycles, and the regeneration efficiency of lignin has been shown in Fig. 8(c). From the figure, it is observed that chromium removal efficiency declines gradually, reaching about 80% after six consecutive cycles. It was reported that lignin molecules dissolved to some extent at higher pH. During regeneration of lignin with NaOH treatment, the pH of the solution approaches 13, which contributes to lignin dissolution and the reduction of Cr(VI) (Albadarin et al., 2011). Although the regeneration efficiency of lignin decreases over repeated cycles, it still demonstrates a considerable potential (∼80%) for eliminating Cr(VI) ions from the water environment. This suggests that lignin serves as an efficient, cost-effective, and widely abundant biosorbent for eliminating heavy metals from wastewater.

The maximum adsorption capacities of different bio-based adsorbents have been presented in Table 2. Lignin demonstrates a comparable adsorption capacity for the removal of Cr(VI) ions from aqueous solution, establishing itself as a strong contender for contaminant removal. Considering prime factors such as the low cost, wide availability of raw materials, and the ability to regenerate and reuse with minimal performance loss, strengthens its practical value, making it well-suited for large-scale applications. Even though its capacity is slightly lower compared to some modified biochar/composite adsorbents, it still performs competitively against high-end adsorbents like functionalized activated carbons.