Parameter Optimization of Group Contribution Methods in High Dimensional Solution Spaces

Abstract

The prediction of certain thermodynamic properties of pure substances and mixtures with calculation methods is a frequent task during the process design in chemical engineering. Group contribution methods divide the molecules into functional groups and if the model parameters for theses groups are known, predictions of thermodynamic properties of compounds that comprise these groups are possible. Their model parameters have to be fitted to experimental data, which usually leads to a multi-parameter multi-modal optimization problem. In this paper, different approaches for the parameter optimization are tested for a certain class of substances. One way to carry out the optimization is to fit only one group interaction at a time, which results in six parameters, that have to be fitted. The downside of this procedure is, that incompatibilities between different parameter sets might occur. The other way is to fit more than one group interaction at a time. This further increases the variable dimension but prevents incompatibilities and leads to thermodynamic more consistent parameters because of a greater data base for their optimization. Therefore, investigations on those different optimization procedures with the help of encapsulated Evolution Strategies are made.

Member of the Collaborative Research Center SFB 531: “Design and Management of Technical Processes and Systems by Using Methods of Computational Intelligence”, supported by the Deutsche Forschungsgemeinschaft DFG