Application of machine learning and statistical approaches for optimization of heavy metals (Cd²⁺, Pb²⁺, Cu²⁺, and Zn²⁺) adsorption onto carbonized char prepared from PET plastic bottle waste

Rapid industrialization and urbanization development (e.g. diverse domestic and industrial activities) release an enormous amount of heavy metals (HMs)–containing wastewater, which finally discharges to our aquatic and terrestrial environment (Chakraborty et al. 2021a, 2023a; Chanda et al. 2021). Environmental pollution by metals and metalloids is considered a matter of global concern among researchers due to their persistent, non-biodegradability, potentially toxic character and is a severe disease creator to all living organisms via bioaccumulation and biomagnification (Al-Malack & Basaleh 2016; Ghosh et al. 2020a; Chakraborty et al. 2022a, 2022b, 2023b). Therefore, wastewater must be free from HMs before being released into the receiving environment to maintain environmental quality and develop an eco-friendly industry (Boulaiche et al. 2019; Chakraborty et al. 2021b, 2023c). Numerous treatment methods including membrane filtration, sedimentation, oxidation, nanofiltration, photo-catalytic degradation, ozonation, chemical precipitation, biological operations, electrochemical technologies, coagulation/flocculation, solid-phase extraction, ion exchange, and adsorption have been applied for elimination of diverse pollutants (e.g. organic and inorganic) from wastewater (Ghosh et al. 2018, 2020b; Bhattacharjee et al. 2020; Zaman et al. 2022). Furthermore, several treatment methods are either costly or have several drawbacks (Zaman et al. 2022). Barakat (2011) has extensively explored the benefits and limitations of various treatment methods. Adsorption is mostly using effluent treatment techniques with aspects such as easy to operate and design, cost-effectiveness, greater performance, source availability, lower sludge production, flexibility, and insensitivity to hazardous elements (Chakraborty et al. 2020, 2021b). The usability of marketable activated carbon is reducing due to high cost (Zaman et al. 2022). In this favor, many wastes and biomaterials were used as adsorbents for both inorganic and organic pollutants, which are broadly reviewed by Bhattacharjee et al. (2020). Nowadays numerous researchers are trying to develop novel, low-cost, and ecologically viable adsorbents with high performance from waste products for the elimination of hazardous substances from wastewater. In the modern age, plastic products are extensively used in our daily activities. Polyethylene terephthalate (PET) is a widely used polymer around the world for its particular properties such as being lightweight, cost-effective, clearness, a good insulator, exceptional flexibility, and easy to handle (Djahed et al. 2016; Rahmawati et al. 2019), so a huge amount of plastic waste is generated yearly that is a pressing concern for the environment (El Essawy et al. 2017). PET is widely produced from public and industrial discarded materials that have no potential beneficiary application. Therefore, PET waste management is considered a global burden, especially in developing countries (Mallakpour & Behranvand 2016), where burning and landfilling are the commonly used disposal methods, which are highly responsible for environmental pollution by emitting toxic gaseous pollutants and polluting landfilling surrounding the ecosystem (e.g. aquatic and terrestrial) (Saha & Ghoshal 2005). Though recycling is another method for PET waste management instead of burning and landfilling, this method is not effective due to a lack of technical difficulties and lower economic return; in this viewpoint, activated carbon/char preparation (e.g. thermal and chemical activation) from PET waste would be a substitute technique as compared with other techniques (e.g. incineration, pyrolysis) (Djahed et al. 2016). Providentially, PET discarded products contain high carbon and lower impurities, which act as a new window for developing activated carbon/char and pollutants adsorbing agents (Cuerda-Correa et al. 2016). In experimental studies, process modeling and optimization are very important to improve the system performance, but conventional methods could optimize a single parameter at a time, which increases experimental time and cost (Gadekar & Ahammed 2019). Correct optimum conditions could not be attained because of the incompetence of experimental results to reflect relations among factors affecting the process (Musa et al. 2023). Recently, many researchers have paid attention to applying artificial neural networks (ANN) and response surface methodology (RSM) for optimizing and modeling the diverse experimental parameters at a time that enhances the system performance (Khan et al. 2020). Those machine learning approaches are a reliable and powerful tool that help overwhelm the system limitations and assess actual results using experimental data. Consequently, the combination of a smaller number of experiments and the hidden performances of post-process parameters could attain a more actual solution. So, modeling using the RSM and ANN approaches could assess the real associations between the input parameters and the response parameters of a process by applying experimental data (Gadekar & Ahammed 2019; Musa et al. 2023). It is a soft computing technique where required results can be achieved via alternating network weights (Gadekar & Ahammed 2019). So, it does not need any particular understanding of the physical/chemical procedure that moves the system. Nowadays, ANN- and RSM-based approaches have been used in diverse areas of environmental engineering, especially for water and wastewater treatment processes (Nair & Ahammed 2014; Agarwal et al. 2016; Karri & Sahu 2018; Musa et al. 2023), but very few studies have been conducted for HM adsorption. So, the objectives of this study were (i) to investigate the removal performance of carbonized PET plastic bottles for HM (Cd²⁺, Pb²⁺, Cu²⁺, and Zn²⁺) removal from simulated wastewater using diverse operational conditions such as pH, time of contact, HM concentration, and adsorbent dose applying machine learning statistical analysis; and (ii) to explore the adsorption mechanism using diverse models, such as isotherm, kinetic, and thermodynamic models.

In the entire experiment, every chemical and reagent was laboratory-grade, purchased from Sigma-Aldrich (Germany), namely, CuSO₄.5H₂O (≥98%), ZnCl₂ (≥98%), PbCl₂ (98%), and CdCl₂ (98%). The required amount of metal salts is used to prepare the stock solution (1,000 ppm), and the stock solution's pH value was less than 2.0 using HNO₃.

Adsorption isotherm studies were carried out into 250 mL HM solutions of varied HM concentrations (25–300 mg/L), at pH 7, where 1 g/L carbonized char dose was added, and the solution was stirred at 200 rpm at 25 °C temperature for 90 min. While kinetic experiments were run in 300 mL HM solution at a fixed concentration (100 mg/L) and the other condition kept constant, then samples were taken out after the following time intervals 1, 5, 7, 10, 15, 20, 30, 60, 90, 120, 150, and 180 min; filtered; acidified; and analyzed. This study applied Langmuir and Freundlich isotherms for equilibrium data modeling, while pseudo-first-order and pseudo-second-order models were used for kinetic modeling (detailed in Table S1).

The adsorption peak near 1,600 cm⁻¹ matching the in-plane C = C vibration indicates the graphite properties, which are the inborn properties of sp2 graphitic ingredients (Hu et al. 2017). The raw adsorbent shows four major peaks at 3,446, 2,930, 2,363, and 1,583 cm⁻¹ (Figure S1a), indicating the stretching vibration of OH, C–H_, C–N, and C = C, respectively (El Essawy et al. 2017). Slight alterations were observed after HM adsorption (Cd²⁺, Pb²⁺, Cu²⁺, and Zn²⁺) (Figure S1a); the above-mentioned adsorption peaks can link with the HM ions, while, more specifically, the hydroxyl group (OH⁻) highly regulates the linkage between the adsorbent and the adsorbate. A similar result was reported by Chanda et al. (2021) for Ni adsorption using chemically treated mahogany sawdust. The surface feature of carbonized char (before and after adsorption) was estimated through SEM analysis; the results attained are presented in Figure S1. The EDX spectroscopy indicates that after the adsorption process, each HM is adsorbed onto the carbonized char surface (Figure S2b–e).

For adsorption science, pH is considered a vital factor because it influences the metal salt ionization rate, the modification of adsorbent inner and outer binding sites, and shifting of the adsorbent surface charge (Chakraborty et al. 2023a). Figure 2(b) shows the adsorption performance of HMs using carbonized adsorbent at varied solution pH of 3–7. The removal of HM increase (Cd²⁺ = 64–72%; Pb²⁺ = 7–35%; Cu²⁺ = 37–64% and Zn²⁺ = 22–32%) with increasing solution pH (3–7) (Figure 2(b)), because the surfaces of adsorbents were deprotonating, so electrostatic attraction between the absorbate and the adsorbent also increased at an increasing pH. The adsorbent surface becomes positive in a highly acidic environment due to the influence of oxygen-linked functional groups, where high proton clouds enrich electrostatic repulsion between adsorbate and adsorbent, resulting in lower HM removal. Though high pH increases the adsorption capacity, it also assists to precipitate the HMs by influencing the –OH group present in the solution confirmed by Wang & Qin (2005), which makes the adsorption study more complex. Therefore, a pH of 7 was selected as the optimum pH to avoid the precipitation of HMs. Similar results for the adsorption of HMs onto Carnauba straw powder were reported by Pereira et al. (2020).

To investigate the performance of carbonized adsorbents for HM adsorption with diverse concentrations (25–300 mg/L), an experiment was conducted where other factors were kept constant (t = 25 °C, contact time = 90 min, pH = 7, and dose = 1 g/L). Figure 2(c) shows that with increasing HM concentration (25–300 mg/L), the elimination performance of HMs (Cd²⁺ = 87–63%, Pb²⁺ = 77–25%, Cu²⁺ = 72–45%, and Zn²⁺ = 57–23%) (Figure 2(c)) decreases due to saturating carbonized adsorbent external surface at a fixed dose, where HM ions cover the vacant space. However, the adsorption capacity of the carbonized adsorbent increased (Cd²⁺ = 23–190 mg/g, Pb²⁺ = 19–74 mg/g, Cu²⁺ = 18–136 mg/g, and Zn²⁺ = 14–69 mg/g) due to a higher interface between HM ions and adsorbent active sites with increasing HM concentration, where high driving forces led to overcome the mass transfer between liquid and solid phases (Zaman et al. 2021).

Pollutant elimination from wastewater is significantly influenced by adsorbent dosage (Chakraborty et al. 2020). To assess the suitable adsorbent dose, batch adsorption experiments of HMs onto carbonized adsorbent were conducted at 25 ± 2 °C, with different adsorbent doses (1–6 g/L) and keeping the other conditions constant (pH = 7, HM concentration = 100 mg/L, CT = 90 min). Figure 2(d) reveals that the HM removal percentage improved with rising carbonized adsorbent doses (0.5–6 g/L) for HMs (Cd²⁺ = 64–89%, Pb²⁺ = 24–67%, Cu²⁺ = 49–78%, and Zn²⁺ = 24–43%), respectively, due to lots of active exchangeable sites on the carbonized adsorbent surface, resulting in greater adsorption (Chakraborty et al. 2023a). While the adsorption capacity gradually reduces (Cd²⁺ = 129–15 mg/g, Pb²⁺ = 47–11 mg/g, Cu²⁺ = 49–15 mg/g, and Zn²⁺ = 48–7 mg/g) with increasing adsorption dose (0.5–6 g/L) might be the aggregation of HM ions onto the carbonized adsorbent. A related elucidation was given by Al-Malack & Basaleh (2016).

Adsorption kinetics is needed to design an effluent treatment plant unit, where the adsorbent could remove pollutants at a certain rate. In the present study, two commonly used kinetics models such as Lagergren's pseudo-first-order and Ho's pseudo-second-order, were applied for HM (Cd²⁺, Pb²⁺, Cu²⁺, and Zn²⁺) adsorption onto carbonized adsorbents. The model accuracy depends on the high correlation coefficient (R²) and lower error values (RSS, chi-square (χ²), and RMSE) of the model, and the applied model parameters are shown in Table 1. The pseudo-second-order kinetic model shows higher R² and lower RSS, χ², and RMSE values for HM adsorption as compared with pseudo-first-order kinetic; in addition, the calculated (q_e,cal) value from the pseudo-second-order kinetic model also corresponds to the experimental value (q_e,exp), indicating that pseudo-second-order was the best-fitted kinetic model for adsorption data. Therefore, chemical adsorption controlled the total adsorption process (i.e. electrostatic interactions between the absorbate and adsorbent), where rapid adsorption occurred at the first phase and gradually slowed at time progress due to the lack of adsorbent site that is supported by Figure 2(a); previous studies also reported a similar outcome (Chakraborty et al. 2023d; Lemessa et al. 2023).

The experimental data were also evaluated by intraparticle (IP) diffusion to determine the diffusion mechanisms. The IP plot did not pass through the origin (Figure S3) and high intercept values (Table 2), representing that exterior diffusion was the rate-limiting step as interior diffusion for HM adsorption using carbonized adsorbent, which may happen at the same time. Futalan et al. (2019), Chanda et al. (2021), and Chen et al. (2022) observed related findings in their adsorption study.

The link between an adsorbate and an adsorbent in an adsorption study is well explained by adsorption isotherms. In this present study, Langmuir and Freundlich isotherms were used, which are usually beneficial in the solid/liquid system, presented in Table 2. Langmuir isotherm was the best-fitted isotherm for Cu²⁺ adsorption onto a carbonized adsorbent due to its higher correlation coefficient value (R² = 0.972), lower error (RSS = 0.689), chi-square value (χ² = 0.044), and RMSE (0.262) as compared to the Freundlich isotherm (R² = 0.937, RSS = 4.610; RSME = 0.679, and χ² = 0.313) (Table 1), revealing that Cu²⁺ molecules produce a single continuous layer with similar dispersal to carbonized char, while Cd²⁺, Pb²⁺, and Zn²⁺ follow the Freundlich isotherm than Langmuir concerning correlation coefficients and error values, showing multi-layer adsorption onto carbonized adsorbents.

The monolayer maximum adsorption capacity of HMs onto carbonized adsorbent was 263.157, 78.740, 196.078, and 84.745 mg/g for Cd²⁺, Pb²⁺, Cu²⁺, and Zn²⁺, respectively (Table 2). The R_L values of carbonized char were between 0 and 1, representing that HM adsorption onto carbonized adsorbents was appropriate under the studied experimental conditions. Conversely, the value of adsorption intensity (n) was higher than 1, and higher K_F demonstrated that the adsorption process was promising for HM adsorption from aqueous solutions using carbonized adsorbents (Table 2). This study outcome is also in correspondence with the findings of El Essawy et al. (2017). In addition, the performance of HM adsorption using carbonized adsorbents is comparable with other adsorbents, as presented in Table S3.

A thermodynamic study represents the role of temperature in adsorption, the nature of the linkage between the adsorbate and the adsorbent, and the direction and mechanism of reaction with changing the experimental temperature (Chakraborty et al. 2023d). These study results are presented in Table S4 and Figure S5a. A Van't Hoff plot of ln K_d vs 1/T made a straight line (not shown in the figure) with R² values of 0.798, 0.996, 0.931, and 0.954 for Cd²⁺, Pb²⁺, Cu²⁺, and Zn²⁺, respectively. Table S4 shows that negative values of ΔG⁰ demonstrate that HM adsorption processes are thermodynamically spontaneous and useful (El Essawy et al. 2017). The exothermic type of adsorption is confirmed by the negative result of ΔH⁰, HM ion adsorption decreases with rising temperature, and no remarkable deviations were found after 298 K (250 °C). Therefore, the next experiments were conducted at this temperature. The positive ΔS⁰ also suggests the decreasing chance between the carbonized adsorbent and metal ion interface, which mainly occurred through the chemical adsorption mechanism through the ion exchange process. Sahmoune (2019) detected the exothermic reaction for HM adsorption using green adsorbents.

Analysis of variance (ANOVA) is an analytical technique that is applied to detect the validity and suitability of Fisher's, F-test, and Student's t-test models. The input effective variables (A = pH, B = HMs concentration, and C = adsorbent dose) were considered for statistical evaluation of the empirical models, presented in Tables S6–S9, respectively, where pH (A) and adsorbent dose (C) show a significant effect for HM removal from aqueous solution. The p-values of all quadratic models (Cd²⁺ = 0.0001, Pb²⁺ = 0.0001, Cu²⁺ = 0.0020, and Zn²⁺ = 0.0002) were less than 0.05, suggesting that the quadratic model fits the HM (Cd²⁺, Pb²⁺, Cu²⁺, and Zn²⁺) adsorption data on the carbonized adsorbent. Lack of fit is more significant than the pure error in terms of the p-value (p < 0.0001), revealing that the design model precisely explains the data performance for the experiments and the factors examined have very significant effects on removal efficiency (Chakraborty et al. 2023d). In addition, A (pH), C (adsorbent dose), AC (pH–adsorbent dose), BC (HM concentration–adsorbent dose), A² (pH), and B² (HM concentration) are significant model (p < 0.05) terms for HM adsorption using carbonized adsorbents. Moreover, the R² value of the HM adsorption quadratic model (Cd²⁺ = 0.99, Pb²⁺ = 0.97, Cu²⁺ = 0.94, and Zn²⁺ = 0.98) showed that 99, 97, 94, and 98% of the whole variability of the result was explained by this model. A good agreement was attained between experimental (R²) and projected (R²adj) results for all HMs (Figures 3 and 4), showing that the model is a good fit, where most of the data point near the straight line. On the other hand, the F-values of Cd²⁺, Pb²⁺, Cu²⁺, and Zn²⁺ were 118.05, 39.47, 11.42, and, 25.50, respectively, suggesting that the models are significant. There is only a 0.01, 0.01, 0.20, and 0.02% chance of creating noise for Cd²⁺, Pb²⁺, Cu²⁺, and Zn²⁺ adsorption onto carbonized adsorbent due to higher F-value, respectively. The signal-to-noise ratio for Cd²⁺ (34.99), Pb²⁺ (21.81), Cu²⁺ (12.01), and Zn²⁺ (17.26) was larger than 4 representing a satisfactory signal. So, these quadratic models are suitable for explaining the adsorption of HMs onto CPETPW. Renu et al. (2018) and Rouniasi et al. (2018) used this model for HM adsorption from wastewater using activated carbon oxide nanosheets and modified wheat bran, respectively.

ANNs are widely used for recording the non-linear relation between independent and dependent variables and are suitable to apply to any condition (Gadekar & Ahammed 2019). This study applies a multi-level feed-forward neural network, which is directed in the following order: input–hidden–output. The applied ANN has 60% training, 20% validation, and 20% testing networks. The input parameters (pH, adsorbent dose, and initial HM concentration) were selected for ANN, while the percentage of HM removal was selected as the output layer. Trial and error techniques were applied to achieve model accuracy, and validation and testing were carried out using MATLAB (R2020a).

This study utilizes carbonized adsorbents for exploring the performance of removing HMs from real wastewater (RWW) experiments, where wastewater collected from an industrial area and optimum experiment conditions (adsorbent dose = 1 g/L, optimum pH = 7, temperature = 25 °C, and equilibrium contact time = 90 min) were applied for capitalizing the performance of the adsorbent in RWW. The properties of the collected wastewater were as follows: pH = 5.2, conductivity = 8,840 μS/cm, and total dissolved solids = 4,231 mg/L. The initial concentration of Cd²⁺, Pb²⁺, Cu²⁺, and Zn²⁺ in RWW was 0.501, 19.260, 0.683, and 0.355 mg/L, respectively. After the treatment at pH 7, the concentration was 0.004, 7.280, 0.012, and 0.126 mg/L, respectively. The removal rates of HMs (Cd²⁺, Pb²⁺, Cu²⁺, and Zn²⁺) were 99.2, 62.2, 97.6, and 64.4%, respectively (Figure S4), suggesting that a carbonized adsorbent will be suitable for eliminating HMs (Cd²⁺, Pb²⁺, Cu²⁺, and Zn²⁺) from industrial wastewater.

This process is applied to assess the possibility of further contamination when the treated adsorbent comes into the environment. The desorption rate of the adsorbent is significantly influenced by the nature of bonding (ionic bonds, Van der Waals forces, or covalent) between adsorbate and adsorbent (Chakraborty et al. 2023a). Here, a desorption study was conducted with diverse pH values (pH 5–11), where carbonized chars were taken from isotherm studies, dried, and the required amount was used for experiments. Figure S5b shows that the desorption percentage of HMs (Cd²⁺, Pb²⁺, Cu²⁺, and Zn²⁺) was not increasing significantly instead of increasing pH, which might be the possibility of the existing strong chemical bonding between HM ions and carbonized adsorbent (Figure S7), confirming the eco-friendly properties. Mouni et al. (2018) and Chakraborty et al. (2021b) provide a similar explanation in their study.

The presence of HMs in the environment from industrial sources is a matter of great concern. So, this study tried to develop a new, inexpensive, and environmentally friendly wastewater treatment method where carbonized char is produced from PET bottle waste for removing HMs from wastewater. The study finding shows that carbonized adsorbent shows high performance in removing HMs from wastewater. The elimination efficiency of HMs decreases with increasing HMs concentration and temperature while enhancing with increasing adsorbent dose. The equilibrium contact time of all HMs was 90 min. BBD and ANN models showed a good agreement between the predicted and experimental results. Pseudo-second-order was the best-matching kinetic model for HM adsorption data. The equilibrium data were well explained by the Langmuir isotherm for Cu²⁺, while the Freundlich isotherm was used for Cd²⁺, Pb²⁺, and Zn²⁺. Thermodynamic study shows that the adsorption is exothermic and spontaneous for HM removal. According to the study performance, carbonized char could be a suitable and effective adsorbent for removing HMs (Cd²⁺, Pb²⁺, Cu²⁺, and Zn²⁺) from wastewater, where a centralized wastewater treatment system is not accessible.

The authors would like to thank the Department of Environmental Science and Technology, Jashore University of Science and Technology, Jashore 7408, Bangladesh, for providing access to laboratory facilities.

All relevant data are included in the paper or its Supplementary Information.

The authors declare there is no conflict.