پرديس مهرمحمدي

Carbon dioxide (CO2) adsorption on porous carbon materials offers a cost-effective and high-capacity approach to reduce CO2 emissions. The textural properties of these porous carbons, especially the presence of micropores, significantly affect the CO2 adsorption capacity. However, the effect of these textural properties on the adsorption mechanism is unclear. In this research, a MLP and RBF neural network are trained as models for simulating the amount of CO2 adsorption by operating conditions (temperature and pressure) and adsorbent textural properties (specific surface area, micropore volume, and mesopore volume). The required data was collected from international authoritative articles and 80%, 10% and 10% of the input data were randomly allocated for training, validation and testing of the network, respectively. In order to find the best model, MLP deep networks with different architectures and activation functions were optimized for 12 backpropagation learning algorithms. Optimal networks with different algorithms were compared with each other in terms of accuracy, time and cost of calculations. In MLP neural network, LM and GDM learning algorithms showed the highest and lowest accuracy with MSE values of 2.6293E-5 and 4.4192E-4 and correlation coefficients of 0.9954 and 0.9138, respectively. SCG and BFG algorithms took the least and the most run time with 0.7640 and 56.6110 seconds, respectively. Also, the lowest number of epochs was obtained for the LM algorithm and the highest for the GDM and GD algorithms. The RBF network with a spread of 9 and the number of 207 neurons in its hidden layer with an MSE value equal to 9.8401E-5 was selected as the optimal RBF network. In order to eva‎luate the efficiency of the created models, the networks were implemented with 50 new data, and the MLP network with LM and BR algorithm showed the highest accuracy among the models with %AARD equal to 2.80 and 4.27 and correlation coefficient of 0.9993 and 0.9988 respectively. Finally, the MLP deep neural network with LM and BR learning algorithm, due to having a lower MSE value and a higher correlation coefficient than the RBF network, is a more suitable model for process simulation. Also, the results obtained from the simulation showed that at lower pressures, the amount of carbon dioxide adsorption is more dependent on the volume of micropores, and at higher pressures, it is more affected by the volume of mesopores.

شبكه‌ عصبي پرسپترون چندلايه , شبكه عصبي پايه شعاعي , الگوريتم يادگيري , جذب سطحي CO2 , خواص بافتي جاذب

MLP , RBF , Training Algorithm , CO2 Adsorption , Textural Properties

pardis mehrmohammadi

ahad ghaemi