Optimal design of AI-based models integrated with Ensemble ML paradigms for crude-to-chemicals application

The dramatic increase in development of high throughput experiments (HTE) in the area of catalysis is attributed to the high amount of data collected within short time. Therefore, the current study utilizes the application of four stand-alone artificial intelligence techniques including; Regression tree (RT), least square BOOST (LSQ-BOOST), Gaussian process regression (GPR) and Robust linear regression (RLR) for modeling the yield of ethylene (C2=) and propylene (C3=) in the catalytic cracking of crude oil feeds using temperature, feed, conversion (%), total gas, dry gas and coke as the independent variables. Correlation matrix and non-linear sensitivity analysis were conducted to filter the input combinations into three different classes (M1, M2 and C3) based on the dependent-independent variables relationship. Four different performance metrics including; Co-efficient of determination (R²), Pearson correlation co-efficient (PCC), Mean squared error (MSE) and Root mean squared error (RMSE) were used in evaluating the prediction performance of the techniques in both calibration and validations stages. The results of the stand-alone models at certain instances showed weak performance accuracy. Hence, this led to the development of four Ensemble paradigms namely; Generalized Regression Neural Network Ensemble (GRNN-E), Support vector regression ensemble (SVR-E), Feed forward Neural Network Ensemble (NNE) and Adaptive neuro-fuzzy inference system Ensemble in order to improve the performance accuracy of the stand-alone paradigms. The performance efficiency of the Ensemble paradigms depicts that these techniques can improve the performance of the single models up to 58% and 62% in the calibration and validation phases respectively. Lastly, the findings of this study demonstrate the robust application and their possibilities in understanding different catalytic cracking phenomena.