چكيده به لاتين
Adsorption of CO₂ on porous carbon materials is an effective and cost-efficient method to reduce the emission of this gas. One promising approach for modeling CO₂ adsorption is the use of machine learning techniques, which, with their high capability in predicting nonlinear behaviors, can accurately estimate the CO₂ uptake by solid adsorbents. In this study, the effects of independent parameters such as specific surface area, micropore and mesopore volume, as well as operational conditions including temperature, pressure, and time on CO₂ adsorption capacity were investigated. For process modeling, machine learning methods and artificial neural networks (ANNs) were employed. Two types of neural networks, multilayer perceptron (MLP) and radial basis function (RBF), along with response surface methodology (RSM), were utilized, and the required data were extracted from literature sources. The dataset was randomly divided into training (70%), validation (15%), and testing (15%) subsets to accurately assess model performance. Initially, RSM was applied to evaluate interaction effects among variables and optimize adsorption conditions. Results indicated that specific surface area, micropore volume, and pressure significantly influenced adsorption capacity, and the optimum conditions for maximum CO₂ uptake were determined. The RSM model achieved a correlation coefficient of 0.9299, indicating good accuracy. In the simulation phase, the MLP architecture was tested with three backpropagation learning algorithms, and their performance was compared in terms of accuracy, computational speed, and processing cost. The Levenberg–Marquardt (LM) and Bayesian Regularization (BR) algorithms outperformed the Scaled Conjugate Gradient (SCG) algorithm, with BR achieving a minimum mean squared error (MSE) of 0.0005411 and a correlation coefficient of 0.9912, whereas SCG showed the lowest accuracy. Regarding the RBF network, the optimal structure consisted of 327 neurons in the hidden layer and a spread parameter of 0.01, with an MSE of 1.08691e-05. Overall, RBF exhibited higher accuracy compared to MLP. To evaluate model generalization, both networks were tested on 50 new data points, confirming that MLP with BR and RBF with 0.01 spread provided the highest accuracy. Finally, analysis revealed that at low pressures, CO₂ adsorption is more influenced by micropore volume, while at higher pressures, mesopore volume plays a more significant role. Additionally, specific surface area has a substantial effect on the overall adsorption capacity.
