Modeling an experimental system often results in a number of alternative models that are all justified by the available experimental data. To discriminate among these models, additional experiments are needed. Existing methods for the selection of discriminatory experiments in statistics and in artificial intelligence are often based on an entropy criterion, the so-called information increment. A limitation of these methods is that they are not well-adapted to discriminating models of dynamical systems under conditions of limited measurability. Moreover, there are no generic procedures for computing the information increment of an experiment when the models are qualitative or semi-quantitative. This has motivated the development of a method for the selection of experiments to discriminate among semi-quantitative models of dynamical systems. The method has been implemented on top of existing implementations of the qualitative and semi-quantitative simulation techniques QSIM, Q2, and Q3. The applicability of the method to real-world problems is illustrated by means of an example in population biology: the discrimination of four competing models of the growth of phytoplankton in a bioreactor. The models have traditionally been considered equivalent for all practical purposes. Using our model discrimination approach and experimental data we show, however, that two of them are superior for describing phytoplankton growth under a wide range of experimental conditions.