Highly effective adsorption of caffeine by a novel activated carbon prepared from coconut leaf

The disposal of coconut wastes is costly and damaging to the environment, but its uses are advantageous activated carbons production. Coconut leaves waste were used for activated carbon production by pyrolysis at 500° C and activation with potassium carbonate. The activated carbon was used for caffeine removal from aqueous solution. The coconut leaves activated carbon showed a predominantly amorphous structure from X-ray diffraction analysis and a pH at the zero charge point of 7.9. From the N2 adsorption/desorption method, the adsorbent showed a predominance of mesopores, with average pore size of 45.48 ηm and a surface area of 678.03 m2/g. From kinetic studies the data followed the pseudo-second order, where the intraparticle diffusion can be neglected. The adsorption isotherms were satisfactorily adjusted for the Redlich-Peterson model and a type curve L was identified. The thermodynamic parameters showed that adsorption occurred spontaneously, was exothermic and governed by physical adsorption. The artificial neural networks developed were capable of predicting both kinetics and equilibrium adsorption data under different operating conditions and was comparable to the traditional models available in literature in the training experiments, encouraging its use for data generalization when an efficient dataset is used. In conclusion, coconut leaves waste showed to be a promising feedstock to produce activated carbon aiming caffeine removal from water and wastewater.


Introduction
Coconut is a culture with several uses in people's daily lives and can be processed to obtain a wide variety of valueadded products. Coconut grows in more than 97 countries on most continents, Asia, Africa, and America (Kappil et al. 2021). In 2019, the coconut harvested area was 11,807,156 hectares, producing 62,455,08 tons. Indonesia is the world's largest producer of coconut, followed by the Philippines and India. Brazil is the 5 th largest coconut producer in the world with 2,330,949 tons (FAO 2020). The coconut wastes (shells, leaves and bunches) discarding is both costly and harmful to the environment and its structure, high carbon content, and very low ash content are beneficial to synthesize activated carbons (Sekhon et al. 2021).
Activated carbons are currently used as an adsorbent in different industries as food, pharmaceutical, chemical, oil, nuclear, automotive, and the treatment of water and wastewater (Bispo et al. 2018;Anastopoulos et al. 2020a, b). Researchers such as Anastopoulos et al. 2021, focus on the use of residual biomass as a precursor for the production of biochar and activated carbons, and their application as efficient and alternative adsorbents to remove potential toxic elements from wastewater. Efficiency and promising operation are the highlight of adsorption technique. This process has great technological maturity and is superior in terms of initial costs, flexibility and simplicity in projects, high efficiency, even at low concentration of pollutants (Rouquerol et al. 1999;Al-Degs et al. 2001). Adsorption methods are already well known, and the literature is full of studies focused on the phenomenological analysis of the process. However, due to the large application of adsorption technology, the search for new and unexplored adsorbents, as well, synthesis routes adsorbents, the precise adjustment of the operating parameters for a low-cost preparation, and the effective removal of specific contaminants are still subjects that attract the interest of researchers around the world . As an example, Alidokht et al. 2021, gave an overview of recently developed nanomaterials (2017)(2018)(2019)(2020)(2021) and their interaction mechanisms with contaminated water.
In recent years, a new class of contaminants called emerging pollutants (EP) has caused great attention in scientific community. EP are resistant to conventional treatments, toxic, and bio-accumulative and they are not comprised in schedules for monitoring water quality and are harmful to the environment and to humans. Their effects and comportment in the environment still demand studies Geissen et al. 2015;Patino et al. 2015). These pollutants are not fully removed in wastewater treatment units and drinking water stations because of their high complexity, low concentration, and the presence of other compounds that can compete or work as inhibitors (Patino et al. 2015;Patiño et al. 2015).
Thus, several systems have been studied for the removal of EP (Alakhras et al. 2021;Arfanis et al. 2016). Though, adsorption has been attracting the notice of various researchers due to the high efficiency and low cost . Amid emerging pollutants, the caffeine is considered a chemical indicator for surface water pollution because its regular consumption (Ferreira et al. 2015). Caffeine is an alkaloid used as a respiratory stimulant, as an additive in medications to increase analgesic effect and also as an appetite regulator. In addition to the natural occurrence of caffeine, it can be found in several widely consumed products such as coffee, chocolate, and tea (dos Santos Lins et al. 2019).
In the present work, we propose for the first time the production of the activated carbon from the coconut leaves to evaluate its potential for caffeine (CFN) removal from aqueous solution. For this purpose, the activated carbon was obtained by vacuum pyrolysis (500° C) and chemically activated with potassium carbonate (K 2 CO 3 ). The raw materials were characterized by thermogravimetric analysis (TGA) and the activated carbon by N 2 adsorption/desorption analysis, X-ray diffraction (XRD), the pH at the point of zero charge (pH PZC ). In order to explore the potentiality of the adsorbent to remove caffeine from aqueous solution, adsorption kinetic assessments were carried out and the data were evaluated using pseudo-first-order, pseudo-second-order, and intraparticle diffusion models. The effect of temperature was studied in the range 30-50° C and for the better understanding of adsorption phenomenon the Langmuir, Freundlich, Sips, and Redlich-Peterson models were fitted to the equilibrium data. The thermodynamics study was conducted to comprehensive the analysis of adsorption mechanism. And to complete the study proposed in this work, artificial neural networks were developed to compare with the conventional models used to foresee kinetic and adsorption equilibrium data.

Chemicals
Potassium carbonate (K 2 CO 3 , 85%) and caffeine (purity ≥ 99%) were purchased from Sigma-Aldrich. All other chemicals presented analytical grade. CFN has a molecular weight of 194.19 g.mol −1 and a pKa of 8.30, which means it is a weak base (Barbas et al. 2000). Furthermore, CFN has high water solubility (Ks > 10,000 mg.L −1 ), so that hinders its removal in common water and wastewater treatment processes (Beltrame et al. 2018).

Raw material
The coconut leaves (CL) were collected in the city of São Miguel dos Milagres, Alagoas, Brazil, at latitude 9° 14′ 17.4′′ South and longitude 35° 21′ 16.3′′ West. The leaves were removed at different stages of life. The CL were cut into pieces and dried in an oven (Fanem, Orion 515, São Paulo, Brazil) at 100° C for 24 h (dos Santos et al. 2019a, b;Araújo et al. 2021). The proximate analysis was carried out in triplicate according to the standard methodology from American Society for Testing and Materials (ASTM)- D1762-1984. The dried sample was analyzed by thermogravimetric (TG) and derivative thermogravimetric (DTG) technique using a thermobalance (TA Instruments, SDT 650, USA). The sample (10 mg) was heated between 27 and 800° C, with a heating rate of 10° C.min −1 at an inert atmosphere (N 2 gas) with a flow of 40 mL.min −1 .

Preparation of activated carbon
Coconut leaves biochar (CLB) were produced by vacuum pyrolysis using a reactor composed of a tubular furnace (JUNG/LT6, 1 kW, 2010) heated by electric resistances. A temperature controller (NOVUS, 1043, Rio Grande do Sul, Brazil) was connected to keep the temperature at 500° C with a heating rate of 10° C.min −1 . The system was also composed of 6 condensers responsible for condensing the process gases, collected in three decantation funnels (CB1, CB2, and CB3). A thermostatic bath (Tecnal, TE-184, Brazil) was used to cool the liquid products down to 4° C. A vacuum pump (Fanem, 089/CAL, Brazil) was used to keep a negative pressure of 20 kPa within the system. Pyrolysis experiments were carried out in triplicate using 250 g of CL during a residence time of 2 h (Sandes et al.;Araújo et al. 2021). Biochar and bio-oil yields were calculated using Eq. 1.
where m p is the produced mass of biochar and bio-oil (g), m l is the initial mass of crude lignin (g), and Yield p is the product yield (%). Gas yield was calculated by difference based on the conservation of mass.
About 6 g CLB samples were chemically activated using 18 g anhydrous potassium carbonate (K 2 CO 3 ) at a weight ratio of 1:3 (biochar:K 2 CO 3 ). The mixture was homogenized using distilled water, inserted in a porcelain pot and filled to the reactor under a pressure of 20 kPa and heating rate of 10° C•min −1 achieving 600° C for 1 h (dos Santos et al. 2019a, b). The system was composed of a stainless steel reactor (height of 20.0 cm, diameter of 20.0 cm, and volume of 6.8 L) associated to a vacuum pump (Fanem, 089/CAL, São Paulo, Brazil) and a temperature controller (Novus, N1200, Campinas, Brazil). Steps involved in the activation of the biochar are biochar maceration, mixture homogenization, activation in the reactor, washing (neutralization), and drying. After the activation, the coconut leaves activated carbon (CLAC) samples were cooled and washed with distilled water to neutralize (pH 7) and to unblock the pores. After, biochar was decanted and separated using a paper filter. Lastly, the sample was dried in the oven (Fanem, Orion 515, São Paulo, Brazil) at 100° C overnight.

Characterizations
The N 2 adsorption/desorption analysis was conducted at − 195.91° C using a Micromeritics ASAP 2020 equipment 16. This analysis aimed to determine CLAC textural characteristics such as surface areas and pore diameter. The sample (0.221 g) was pre-treated at 100° C under vacuum for 12 h. The specific area and the adsorption/desorption isotherms were estimated based on the Brunauer-Emmett-Teller (BET) method until a relative pressure of P/P 0 = 0.97 and by the average pore width in the adsorption. The X-ray diffraction (XRD) and patterns were recorded using a diffractometer (Shimadzu, DRX 6000, Japan), operating with a Cu Kα (λ = 0.1540598 Å) radiation source in a 2θ range of 2-70° with step of 0.02°. In this case, 0.1 g of the powder sample was analyzed using a voltage of 40 kV and a current of 30 mA. The analysis of pH at the point of zero charge (pH PZC ) was determined based on the methodology proposed by El-Sayed et al. (El-Sayed et al. 2014). In this case, 50 mg of the sample was added in a 50-mL container of solution with the initial pH values ranging from 1 to 12, these values were achieved adding 0.1 mol.L −1 of HCl or NaOH. The samples were mixed at 140 rpm, 30° C for 24 h, after that centrifuged to measure the pH. Finally, inlet and outlet pH values were plotted to obtain the equilibrium pH, which corresponds to the pH PZC of the CLAC (Beltrame et al. 2018).

Kinetic study
The kinetic study, performed in a finite bath at 27 ± 2° C, with Erlenmeyers (250 mL) was filled with 30 mL of CFN solution, in two different concentrations, 75 and 125 mg. L −1 , and 50 mg of activated biochar (d p < 0.212 mm). The solutions were kept under constant agitation of 140 rpm (Shaker SL 222 incubator) under time intervals between 2 and 120 min (Araújo et al. 2021). For each time, the samples were obtained and separated by centrifugation (Solab, SL-700) for 5 min at 3000 rpm. The concentration of CFN in the sample was obtained by a UV-Vis spectrophotometer (Shimadzu, UV 1800 Spectrophotometer, Japan), at a wavelength of 272 ηm (Araújo et al. 2021). The adsorption capacity ( q t , mg•g −1 ) was calculated by Eq. 2.
where C 0 is the initial CFN concentration, V is the solution volume, and M is the mass of adsorbent.
To assess the kinetic comportment, the pseudo-first-order (Lagergren 1898), the pseudo-second-order (Ho and McKay 1999), and Weber-Morris (Weber and Morris 1963) models, presented in Eqs. 3, 4, and 5, were used to fit the data.
where q t (mg.g −1 ) is the adsorption capacity at t, t (min) is the time, q e (mg.g −1 ) is the adsorption capacity at equilibrium, k 1 (min −1 ) is the constant of the pseudo-first-order model, k 2 (g. −1 mg.min −1 ) is the constant for the pseudosecond order model; k d (mg.g −1 .min −0.5 ) is the intraparticle diffusion rate constant and C (mg.g −1 ) signifies the thickness of the boundary layer (Özer et al. 2006;Yang et al. 2014;Santos et al. 2019a, b, c).

Equilibrium study
The equilibrium experiments were carried out in a finite bath system, using 50 mg of activated carbon (d p < 0.212 mm) which was placed in a 250-ml conical flask containing 30 ml of CFN solution in concentrations 75,85,95,105,115,125,135, and 145 mg.L −1 . The solutions, in duplicate, were left in contact under agitation for 40 min (time defined in the adsorption kinetics experiments) at 140 rpm, 30, 40, and 50° C. After the contact time, the samples were centrifuged for 7 min at 3000 rpm for separation between adsorbent and aqueous solution. The CFN final concentration was obtained using a UV-Vis spectrophotometer. The equilibrium data were fitted using the nonlinear models, Langmuir (Langmuir 1918), Freundlich (Freundlich 1906), Sips (Sips 1948), and Redlich-Peterson (Redlich and Peterson 1959), presented in the Eqs. 5, 6, 7, and 8, respectively.
where q max (mg.g −1 ) is the adsorption capacity capacity; k L (L.mg −1 ) is the constant of Langmuir model; k F (mg.g −1 ) (mg.L −1 ) −1/n is the constant of Freundlich model; 1/n ( −) is the factor of heterogeneity; q S is the maximum adsorption capacity for the Sips model (mg.g −1 ); K S (L.mg −1 ) is the constant of Sips model; and m S ( −) is the exponent of the Sips model; k R (L.mg −1 ), a R (L.mg −1 ) β , and β are constants for Redlich-Peterson.
and y cal i are the experimental and calculated (model) values from dependent variables q t ; n is the number of measured points; and y n is used to calculate the mean of the experimental values, n p is the number of model parameters .

Thermodynamics
Thermodynamics is an important used tool to understand the influences of temperature on the adsorption mechanisms and their nature. The Gibbs-free energy change (ΔG°), enthalpy change (ΔH°), and entropy change (ΔS°) were determined using data from the isotherms and by eqs. 13, 14, and 15 (Milonjić 2007;Piccin et al. 2017).
R is the universal constant of gases, T (K) is temperature, K e is the thermodynamic equilibrium constant, which was determined by the constant of the isotherm model that best fit the experimental data (Milonjić 2007;Piccin et al. 2017).

Ionic strength and influence of aqueous matrices
Ionic strength tests were conducted at NaCl concentrations of 0.5, 10, 15, and 30% (m/v). The experiments were assessed at 30 °C with 75 mg.L −1 of caffeine solution in contact with 50 mg of CLAC, at 140 rpm for 40 min of contact. Subsequently, the liquid solution was separated from solids by centrifugation for 7 min at 3000 rpm. The final concentration of CFN was obtained using a UV-Vis spectrophotometer .
With the intention of assess the effect of different aqueous mediums on the caffeine removal, adsorption experiments were achieved using lagoon (Mundaú Lagoon, Maceió/AL, Brazil), tap, and mineral water. The initial caffeine concentration was 75 mg.L −1 , 50 mg of CLAC, 30° C and 40 min contact time were used for the tests (Melo et al. 2020).

Artificial neural networks
In this work, neural networks were developed to compare their results with the traditional models in the literature used to predict kinetic and adsorption equilibrium data. The purpose of the neural model was also to assess its ability to generalize data. To do so, the adsorption kinetics data for the condition of 125 mg.L -1 and equilibrium data at the temperature of 40 °C were not included in the network database, being used in the network test and validation experiments. The other data referring to the other conditions studied were included in the database and consisted of network training experiments. The metrics used to assess the performance of the network developed and provide its comparison with theoretical models were the coefficient of determination and the average relative deviation.
The feed-forward neural model was used, with the Levenberg-Marquardt technique included into the backpropagation optimization process to calculate the weights and Bayesian regularization to prevent data overfitting. For the neurons in the hidden and output layers, hyperbolic tangent sigmoid and linear activation functions were employed, respectively. To increase the performance of the artificial neural network, the database was first normalized for the training stage, in order to remove any difficulties connected to the varied magnitudes of the data. The data normalization involved setting the rows of each data matrix to 0 and the deviations to 1. Optimization tests were done to establish the network's ideal structure once the database used for training trials was defined (number of neurons and number of intermediate layers). Thus, starting with the addition of one neuron, the size of the intermediate layer was changed, and changes in the R 2 and ARD were continually monitored. This method was repeated until an optimal number of neurons and hidden layers were found, resulting in excellent output prediction outcomes and, at the same time, physically consistent values to a new pattern that was not addressed in the tests and has yet to be addressed.

Raw material characterization
The thermogravimetric analysis (TGA) and the differential thermal analysis (DTG) curves for coconut leaf are presented in Fig. 1(a). This analysis is important to obtain information regarding the ideal temperatures for pyrolysis process. The degradation of biomass basically presents three zones. The first zone is up to 200 °C, it is observed approximately 25% of weight loss. From this point, it is observed the degradation of hemicellulose, cellulose, and lignin begins. From 150 to 500 °C is the second zone, with a high mass loss rate, directly linked to the biomass carbonization process. In the third stage, there is a stabilization of the mass loss rate, which also indicates a decrease in the conversion process. Soon after, on the DTG curve, it was possible to observe the presence of two peaks: the first started between 200 and 235 °C, with a 12% of weight loss of hemicellulose degradation and the second, between 235 and 330 °C related to the cellulose degradation (Kim et al. 2012;Iberahim et al. 2018). Lignin is a macromolecule that is difficult to decompose and presents showing a prominent degradation from 330 °C until the stabilization at approximately 500 °C. After that, the mass losses were minimized and defining this to be the best point for pyrolysis in order to obtain biochar. This behavior was similar to that described by Almeida et al. (2013) who studied coconut fiber pyrolysis and Tsai et al. (2006) who pyrolyzed the coconut fiber, sugarcane bagasse, and rice husk.

Pyrolysis yield
Pyrolysis experiments were performed in triplicate, with an initial mass of 250 g each. With the results obtained, it was possible to calculate the process yield using Eq. 2 as shown in Table 1

Characterizations
Through N 2 adsorption/desorption, analysis was determined the CLAC surface area of 678.03 m 2 /g, pore volume of 0.03 cm 3 /g, and pore diameter of 45.48 nm. Materials with pores between 20 and 50 ηm are classified as mesoporous; thus, the activated carbon of the coconut leaf is within this range (Thommes et al. 2015). The CFN molecule presents the dimensions 0.78 nm/0.61 nm/0.21 nm; this means that the pore diameter of the biochar, 45.48 nm, is much larger than the dimensions of the CFN molecule, which allows the molecule to access the available active sites inside the pores for adsorption (Pendolino 2014). Regarding the surface area, the value of 678.03 m 2 /g −1 is acceptable for the adsorption process when related to the materials used for comparison (Ahmedna et al. 2000;Liu et al. 2007;Foo and Hameed 2012;Ferreira et al. 2016;Paixão et al. 2018;Alves et al. 2019). The N 2 adsorption/desorption isotherms are shown in Fig. 1(b), they are classified as type IV, characteristic of mesoporous material. The hysteresis also characterizes mesoporous materials and is suitable for the type IV isotherm model. The shape of hysteresis indicates a limitation. Despite easily adsorbing N 2 , it is not easily release it ).
The CLAC XRD diffractogram pattern is presented in Fig. 1(c) suggesting a typical amorphous characteristic. The diffraction peaks indicate an amorphous material. The broad peak at 2θ = 23.32° can be related to the typical silica characteristic, and the peak 2θ = 42.96° implies the formation of graphite with a turbostratic structure of amorphous carbon (Kalapathy et al. 2002;Xiao et al. 2014;Sayʇili et al. 2015;Shen and Fu 2018). The point of zero charge for the activated carbon was 8.0. Thus, when the activated carbon is in contact with a liquid solution in which the pH is below pH pzc , it indicates that this surface is positively charged. However, when the pH is above pH pzc , the surface is negatively charged. The chemical compounds are transformed into cations and anions depending on the pKa and pH values  of the solution (Nam et al. 2014). Considering the CFN pKa = 8.3, the adsorption at pH values close to 8.0 occurs in the neutral form of the CFN molecule, confirmed by the experiments presented in this work (Ferreira et al. 2015).

Kinetic study
Adsorption kinetic reveals the significant characteristic of adsorption, the transport rate of the adsorbate to the adsorbent surface (Hiew et al. 2019). The kinetic curves are presented in the Fig. 2(a), showing a typical behavior. At the beginning, a high adsorption rate was observed, reaching the equilibrium at 15 min (44.0 mg.g −1 ) and 30 min (64.0 mg.g −1 ) for 75 and 125 mg.g −1 , respectively. Such a behavior, the fast initial adsorption may be associated with the large availability of surface area and active sites. The concentration difference between the solution bulk and the adsorbent surface can act as a driving force to overcome mass transfer barrier. After this initial step, the surface area and sites are gradually blocked and, instead, the adsorption rate decreases. The quantity of desorbing adsorbate on the adsorbent surface and the amount of adsorbent adsorbate on the surface are in dynamic equilibrium at this moment.
Another issue to be addressed is that at higher concentrations the adsorption capacity is higher. This is because at high concentrations there are more molecules occupying the available active sites (El Haddad et al. 2014;Hiew et al. 2019;Georgin et al. 2019;Khan et al. 2021).
The pseudo-first-order (PFO), pseudo-second-order (PSO), and intraparticle diffusion were adjusted to the experimental data and the parameters are presented in Table 2. The higher coefficients of determination (R 2 > 0.89) and lower average relative deviation showed a better fit of the experimental data to the pseudo-secon-order model. Despite the fact that neither PFO nor PSO models describe the adsorption mechanism, the difference between the mean solid-phase concentration and the equilibrium concentration is assumed to represent the driving factor for adsorption by both models. In this case, the adsorption rate for the PFO model would be proportional to the driving force, while for the PSO model, it would be proportional to the square of the driving force (Chang and Juang 2004;Yang and Al-Duri 2005;Meili et al. 2019). Figure 2(b) shows the adjustment of the experimental data to the intraparticle diffusion model. When plotting q t as a function of t 0.5 , the intraparticle diffusion step will govern adsorption if the resulting curve is linear and passes through the origin. If two phases can be distinguished, the first curve represents diffusion in the particle's outer layer, while the second curve represents diffusion in the particle's inner layer. As a result, the two stages of external diffusion and intraparticle diffusion regulate caffeine adsorption by coconut leaves activated carbon (CLAC) (Yang et al. 2014;Wang and Wang 2018;Henrique et al. 2020).

Equilibrium study
In order to comprehend the interaction amid the caffeine molecules and the coconut leaves activated carbon, the adsorption isotherms were obtained, shown in Fig. 2 (c-e). Giles et al. (Giles et al. 1960) describes a classification of solid-liquid isotherm and suggests that its shape can be used to define the adsorption mechanism, physical nature of the solute, substrate surface, and specific product surface area. Thus, the obtained isotherms can be considered type L2, which presents an initial curvature facing downwards, indicating a reduction in the availability of active sites with the increase of adsorbate concentration. As more places in the adsorbent surface are occupied, it becomes more difficult for the CFN molecules to find an empty spot. The plateaus suggest that all available locations on the original adsorbent surface were filled Georgin et al. 2019). Besides, the isotherms revealed an endothermic comportment since there is a small intensification in the adsorption capacity with the increase of the temperature. This dependence can be associated to the consequence of thermal collisions or modifications in the adsorbent arrangement as function of temperature. Nonlinear isothermal models were used to fit the experimental data: four models were tested: Langmuir, Freundlich, Redlich-Peterson, and Sips. The results are presented in Table 2. Considering R 2 values, all models have values above 0.98, which makes this criterion not decisive. The second statistical criterion analyzed was the R 2 adj , where the best values were for the Redlich-Peterson and Sips models. The smallest ARE values was obtained for the Redlich-Peterson model. However, according to the AIC criterion, the lowest values were for the Freundlich model. Yet, the Redlich-Peterson model has components of the Langmuir and Freundlich equations. For the constant β values close to 1, the Redlich-Peterson model presents Freundlich characteristics, and for β values close to 0, it presents Langmuir characteristics, so in the Redlich-Peterson model can represent the equilibrium data caffeine adsorption onto coconut leaves activated carbon.
Finally, Table 3 compares the adsorption performances of alternative adsorbents in order to cover the adsorbent's future applicability to various contaminants. It can be noted that the adsorption capacity findings obtained are on the same order of magnitude as those obtained in the current work, checking the positive result achieved.
The negative values of ΔG° determine that the adsorption is spontaneous and favorable. In addition, the positive ΔS° value indicates that there was a change in the structure of the adsorbent and an increase in randomness during CFN fixation in active sites. ΔH° negative value indicates the adsorption is exothermic. The negative and low ΔH° values indicate the weak adsorption forces and physical and exothermic nature (Khan et al. 2010;Wang and Li 2013;Lins et al. 2019;Quintela et al. 2020).

Ionic strength and influence of aqueous matrices
The effect of ionic strength on the adsorption of caffeine onto coconut leaf activated carbon is depicted in Fig. 3(a). Despite changing the ionic strength of the solution, the adsorption capacity of the CLAC remained nearly constant, according to the experimental results. As a result, electrostatic interactions have no effect on the adsorption of caffeine onto CLAC.
The use of coconut leaf activated carbon for caffeine removal from real matrices is an important complementary step in this study. The collected data of this step is presented in Fig. 3(b). The removal was 44.7, 40.8, and 65.9% for lagoon, tap, and mineral water, respectively. The delineated

Artificial neural networks
The results of the neural network predictions for the adsorption kinetic data and equilibrium isotherms are presented in Figs. 4 and 5, respectively. As it can be seen in the results presented, the developed neural network was adequate to predict the kinetics adsorption data, but it had difficulty in predicting the adsorption equilibrium data for values of C e lower than 5 mg.L −1 for the data not used in the training step (Fig. 6b), but there is room for improvement of the network. It must be taken into account, however, that C e values are different for each temperature condition tested, that is, the amount adsorbed at equilibrium is dependent on the concentration values for each temperature condition. As a neural model is a model that interpolates values, the global adjustment is compromised, given the network's difficulty in structuring its learning in the face of the non-uniformity of the independent variable's data. As presented in Fig. 4, this was not observed for the results referring to the adsorption kinetics, in which, for the same time interval, the network performance was satisfactory, even considering only the adsorption kinetic data for the dataset of the 125 mg.L −1 condition. These results do not detract from the neural network, which, in addition to presenting a good performance for adsorption kinetics, proved to be an excellent tool to predict equilibrium data for the other analyzed temperature conditions used in its database (Fig. 5 a and c). In fact, with a more efficient database, the neural network could be more accurate. Even so, for data not used in the training step, the ARD value was below 5%, encouraging its use for future studies on the optimization of effluent adsorption processes using the solid waste of an organic nature, both regarding kinetics and isotherms of adsorption.
It is worth mentioning that the developed neural network presented a very simple structure and low computational cost, with no tendencies towards the overfitting problem. The resulting mathematical model is simple and can be used with good accuracy to predict caffeine adsorption data using coconut leaf residues within the range of operating conditions studied. The coefficient of determination was greater than 0.99 for the training experiments and equal to the unity for the kinetic experiments. Little data was used in the training of the network, which reinforces its good performance. The ARD value was below 5% in all the cases analyzed, showing a very little deviation from the experimental values.
In comparison with the classical models presented in the literature, such as the Langmuir, Freundlich, Redlich-Peterson, and Sips models, the accuracy of the neural network was comparable to these models in predicting the data used in the training experiments. It should be noted that these models do not have the ability for data generalization.
Similar results were also found for the adsorption kinetics experiments.

Conclusion
From the results obtained in this study, the coconut leaf activated carbon achieved good results of caffeine adsorption reaching a maximum theoretical adsorption capacity of 73.83 mg.g −1 at 30° C in 40 min of adsorption. The pseudosecond-order model adequately described the kinetic data and the Redlich-Peterson model well fitted the equilibrium isotherms. The adsorption was exothermic, spontaneous, and favorable. It was determined a change in the structure of the adsorbent and an increase in the randomness during CFN fixation in active sites. The ionic strength of the solution has no effect on the adsorption of caffeine onto CLAC. The studied adsorbent demonstrated is suitable for caffeine removal in different water mediums. The artificial neural networks developed in this work were capable of predicting both kinetics and equilibrium adsorption data under different operating conditions. Even for data not used in the network database in the training experiments, the highest mean absolute percentage error obtained was less than 5%, showing that the neural model can be considered an interesting tool for future studies on optimization of adsorption processes and data generalization. The neural model also proved to be comparable to the traditional theoretical models presented in the literature for predicting isotherm and adsorption kinetics data. Finally, the coconut leaf-activated carbon, unexplored for this use, appears to be an exceptional alternative adsorbent, as it is an abundant biomass available at low cost.