Comparative predicting study of heterogeneous catalysis using support vector regression and neural networks with chaotic particle swarm optimization algorithm

Abstract

This paper presents a comparative study of two artificial intelligence based heterogeneous catalysis modeling strategies, namely ANN-CPSO and SVR-CPSO, for modeling dimethyl ether (DME) in direct synthesis from syngas process (called STD process), for reducing both high temporal costs and financial costs, and accelerating the process of industrialization synthesis of DME. In the two hybrid approaches, catalyst compositional models and catalytic reaction mechanism models are simultaneously constructed for correlating process data comprising values of input variables of catalyst compositional, operating conditions and output variables of performance of catalyst (catalyst selectivity and conversion). Next, in order to improve predictive accuracy and generalization ability of models, the structure and connection weights for ANN and three hyper-parameters for SVR are automatically optimized using chaos particle swarm optimization(CPSO). The CPSO possesses certain unique advantages over the commonly used GA and PSO optimization algorithms, which has high search efficiency, high precision and not the local optimum. The major advantage of the two hybrid strategies are that modeling can be conducted exclusively from the historic data wherein the detailed knowledge of process phenomenology (reaction mechanism, rate constants, etc.) is not required and difficult to get. Another advantage is that they have avoided the blindness and contingency of the traditional catalyst "trial and error" method. Finally, ANN-CPSO model and SVR-CPSO model were verified experimentally to be feasible, the forecasting values based on SVR-CPSO model were more close to the measured ones. The SVR-CPSO is a new soft computing methodology for heterogeneous catalysis modeling, which is more suitable for the non-linearity uncertainty, higher-dimension, chaos and small quantity of training data. The obtained results indicate that the SVR-CPSO model can be used as a promising alternative method for chemical process modeling.