Abstract:
The aim of this thesis is to extract knowledge for carbon dioxide adsorption over amine modified adsorbents using two data mining techniques that are decision trees (DT) and artificial neural networks (ANN). The experimental data were collected from 30 papers published between 2002 and 2016. The data set consisted of 1356 data points with 26 attributes; the adsorption capacity (mmol CO2/g adsorbent) and amino efficiency (CO2/N ratio) were chosen as performance (output) variables. In DT analysis, the adsorption capacity and amino efficiency were classified in four groups, and the conditions leading to high adsorption capacity and amino efficiency were identified. Training and testing accuracies for adsorption capacity were 83.4% and 82%, respectively; while these values were 80.8 % and 77.3 % for the amino efficiency; from which the analysis for both output variables could be considered as successful. ANN analysis was also used to model the same data to develop predictive models; tansig function was used as the activation function and trainlm and trainbr were used as the training algorithms for training and testing, respectively. The optimal network topology was determined as 26-10-1 (10 neurons in one hidden layer) for adsorption capacity and 26-9-1 (nine neurons in hidden layer) for amino efficiency. The R2 and RMSE values for adsorption capacity were respectively found to be 0.97 and 0.118 for training, and 0.90 and 0.250 for testing. Similar results were obtained for the amino efficiency. The R2 and RMSE were 0.97 and 0.0230 respectively for training while R2 was 0.83 and RMSE was 0.053 for testing. These results indicate that ANN models were also successful. The relative significances of input variables were also calculated by using the optimal neural network topology and change of mean square error method. Operational variables had greater significance on both amino efficiency and adsorption capacity relative to the adsorbent properties.
