Comparison of single- and multiobjective design of experiment in combinatorial chemistry for the selective dehydrogenation of propane.

Two genetic algorithms for the single- and multiobjective design of combinatorial experiments were applied to the optimization of a solid catalyst system active in the selective catalytic oxidation of propane to propylene. The two different optimization strategies, namely, the single objective optimization of the yield and the multiobjective optimization of the conversion and selectivity were implemented and compared. It was observed that the multiobjective approach optimized the yield in a similar way compared to the single objective approach. With respect to the selectivity, however, the multiobjective outperformed the single objective approach. It was also found that by applying the multiobjective optimization more interesting possible combinations were discovered.

On the suitability of different representations of solid catalysts for combinatorial library design by genetic algorithms.

Genetic algorithms are widely used to solve and optimize combinatorial problems and are more often applied for library design in combinatorial chemistry. Because of their flexibility, however, their implementation can be challenging. In this study, the influence of the representation of solid catalysts on the performance of genetic algorithms was systematically investigated on the basis of a new, constrained, multiobjective, combinatorial test problem with properties common to problems in combinatorial materials science. Constraints were satisfied by penalty functions, repair algorithms, or special representations. The tests were performed using three state-of-the-art evolutionary multiobjective algorithms by performing 100 optimization runs for each algorithm and test case. Experimental data obtained during the optimization of a noble metal-free solid catalyst system active in the selective catalytic reduction of nitric oxide with propene was used to build up a predictive model to validate the results of the theoretical test problem. A significant influence of the representation on the optimization performance was observed. Binary encodings were found to be the preferred encoding in most of the cases, and depending on the experimental test unit, repair algorithms or penalty functions performed best.