Catalyst Design by Machine Learning and Multiobjective Optimization

Abstract

The computer technologies of machine learning and multiobjective optimization were introduced to develop the catalyst for fluid catalytic cracking (FCC). Response surface methodology was applied for a training set consisting of 1000 data points with varied catalyst compositions which consist of a variety of catalysts compositions, feedstock properties, pseudo-equilibrium conditions, cracking performance test conditions as input parameters and the cracking test results as outputs. At first, response surface model (RSM) was obtained with four approximation methods, among which the radial basis function (RBF) method was found to give the highest score accurate RSM with the smallest average error and the highest coefficient of determination among them. Then the virtual experiments were carried out with the RSM applied with multiobjective genetic algorithm (MOGA) to optimize the catalyst design considering the multiobjective; to yield less bottoms, less coke, more gasoline and less gas. After 5000 virtual experiments with RSM were carried out, we found that the pareto front was obtained. Finally, the optimum catalyst design was selected from the designs on the pareto front. As a result, the selected catalyst design showed 2.7 % higher gasoline yield and was confirmed to show the excellent performance over conventional FCC catalyst.