There is a real danger of assuming that large numbers of compounds enable ''several hundreds or thousands of experiments to be performed at once'' if only very shallow gradients over the whole parameter space are applied. The information extracted from these experiments may, owing to larger experimental noise in the simpler set-up, be even less than in a two-step experimental run with a much larger ''grid width'' in a superficial first run followed by a second run ''zooming'' in on the area of interest.
Tools that configure the loop as a whole as efficiently as possible are given by different experimental design approaches. Besides the classical way of varying one parameter at a time, and the simple way of scanning the whole parameter space by an exorbitant number of experiments, there are intelligent designs such as statistical experimental design and genetic algorithms.
The statistical design of experiments is a common tool that is used to examine critical factors at several levels in order to characterize a chosen parameter space. Statistical analysis of the data enables a so-called response surface model to be set up, which allows the interactions between the factors to be analyzed. This analysis can then serve as the basis for further optimization and a good understanding of the process. Typical designs are factorial or reduced factorial designs, central composite designs, or fractional factorial Plackett-Burman designs. Once a design table for one loop has been set up, the experiments can be run in parallel. Over recent years, statistical experimental design has been applied in many case studies [31-33].
Not as popular as statistically based design is the use of genetic algorithms, which are copied from nature. For example, in the first step of optimizing a heterogeneous catalyst, the number of components is defined, then a first generation of catalysts is initialized and tested in parallel . The results are evaluated, leading to a selection of catalysts from the population to create the next generation. The next generation is created by application ofevolutionary operators such as crossover, or qualitative or quantitative mutation. The loop of testing, selection, and mutation is repeated until the convergence criterion is satisfied. In heterogeneous catalyst development, sometimes the question of whether a quaternary will show activity if all binary and ternary combinations of a given set of elements are completely inactive must be considered. Here, the application of genetic algorithms seems to be a promising strategy . A similar approach in which a genetic algorithm is used increasingly is in directed evolution for the creation of new enzymes displaying improved enantioselectivity in a given reaction .
Was this article helpful?