Potential of artificial intelligence and response surface methodology to predict CO2 capture by KOH-modified activated alumina

Abstract

The present study focused on developing predictive neural networks and response surface methodology (RSM)-based model. In order to develop the predictive model, experimental data of CO2 capture by KOH-modified activated alumina was obtained through laboratory scale adsorption setup. Three independent input variables, including time (t: 0–1800 sec), initial temperature (Tin: 20–80 °C), and initial pressure (Pin:1.651–10.028 bar) of the reactor, were considered in the training process. Furthermore, CO2 adsorption capacity, and final pressure were considered as the output. The multilayer perceptron (MLP) and radial basis function (RBF) networks have been employed. The best corresponding optimized MLP network, out of all the 460 different structures, was chosen to be a structure trained with Levenberg Marquardt back propagation algorithm with four hidden layers, in which there were 25, 23, 7, and 20 neurons. The best corresponding transfer functions for the first, second, third, and fourth hidden layers plus the output layer were Purelin, Logsig, Tansing, Logsig, and Purelin, respectively. Finally, the performance of the RBF was explored on the experimental data. Out of the 385 built structures, the optimized corresponding RBF network was chosen to be the one with 2.5 as its spread value and 40 as its neuron numbers. Lastly, in the RSM design, the cubic model with square root transformation presented a comparably better performance. The coefficient of determination values for CO2 adsorption capacity in MLP, RBF, and RSM models were calculated as 0.999, 0.998, and 0.949, respectively.