Artificial neural network modeling of phenol adsorption onto barley husks activated carbon in an airlift reactor

摘要

Purpose and method of the study: The production of activated carbon fromudbarley husks (BH) by chemical activation with zinc chloride was optimized by using a 23udfactorial design with replicates at the central point, followed by a central compositeuddesign with two responses (the yield and iodine number) and three factors (theudactivation temperature, activation time, and impregnation ratio). Both responses wereudsimultaneously optimized by using the desirability functions approach to determine theudoptimal conditions of this process. The experimental data from the batch phenoludadsorption onto barley husks activated carbon (BHAC) was represented by adsorptionudisotherms (Langmuir and Freundlich) and kinetic models (pseudo-first and pseudosecondudorder, and intraparticle diffusion models), besides the regeneration of phenolloadedudBHAC with different solvents was evaluated. Experimental data confirmed thatudthe breakthrough curves were dependent on BHAC dosage, phenol initialudconcentration, air flow rate, and influent flow rate. Adaline and feed-forward backpropagationudArtificial Neural Networks (ANNs) were developed to predict theudbreakthrough curves for the adsorption of phenol onto barley husks activated carbonud(BHAC) in an airlift reactor. Feed-forward back-propagation networks were tested withuddifferent quantity of neurons at the hidden layer to determine the optimal number ofudneurons in the ANN architecture to represent the breakthrough curves performed atuddifferent operational conditions for the airlift reactor.udContributions and conclusions: After the simultaneous dual optimization ofudBHAC production, the maximal response values were obtained at an activationudtemperature of 436 °C, an activation time of 20 min, and an impregnation ratio of 1.1 gudZnCl2 g BH-1ud, although the results after the single optimization of each response wereudquite different. At these conditions, the predicted values for the iodine number and yieldudwere 829.58 ± 78.30 mg g-1 and 46.82 ± 2.64%, respectively, whereas experimentaludtests produced values of 901.86 mg g-1 and 48.48%, respectively. Moreover, activatedudcarbons from BH obtained at the optimal conditions mainly developed a porousudvstructure (mesopores > 71% and micropores > 28%), achieving a high surface areaud(811.44 m2 g-1ud) that is similar to commercial activated carbons and lignocellulosic-basedudactivated carbons. These results imply that the pore width and surface area are largeudenough to allow the diffusion and adsorption of pollutants inside the adsorbent particles.udFreundlich isotherm model satisfactorily predicted the equilibrium data at 25 andud35 °C, whereas the Langmuir isotherm model well represented the equilibrium data atud45 °C. The maximum phenol adsorption capacity onto BHAC was 98.83 mg g-1 at 25 °Cudand pH 7, similar to phenol adsorption onto commercial activated carbons. The kineticuddata were adequately predicted by both the pseudo-first order and intraparticle diffusionudmodels. The external mass transfer was minimized at stirring speeds greater than 400udmin-1ud, and the adsorption kinetics are affected by both initial phenol concentration andudtemperature. Adsorption equilibrium was reached within 40 and 200 min at initial phenoludconcentration of 1000 mg L-1 at 35 °C and 30 °C, respectively. Ethanol/water solutionsudat 10% V/V were the most effective regenerating agent, with desorption capacity ofud47.79 mg g-1 after five adsorption-desorption cycles.udThe breakthrough curves of phenol adsorption onto BHAC in an airlift reactor inudcontinuous operation were adequately predicted with feed-forward back-propagationudANN architecture with 2 neurons in the hidden layer for the single-input single-outputudproblem. Correlation coefficients higher than 0.95 were observed between theudbreakthrough curves predicted by the developed Adaline network and those obtainedudexperimentally for the multiple-input single-output problem. Further improvements andudgeneralization of the developed predictive Adaline network are discussed.