Summary: In many practical situations, we know the dynamics of the analyzed system. Based on this known dynamics, we can build a computer-based model simulating the system’s behaviour. In principle, it is possible to directly use this model to predict the system’s behaviour, but in many cases, the model is very computationally complicated, and its direct use would require a large amount of computation time on high performance computers. A natural way to drastically reduce the computation time needed for the prediction of the system’s behaviour is to use machine learning to come up with a simpler faster-to-compute model. To train the corresponding machine learning algorithm, we can use the results of applying the existing (complex) computer-based model to different inputs – and these inputs can be selected by an appropriate experimental design technique. In this paper, we illustrate this methodology on a simple example of a pendulum; as machine learning techniques, we use support vector machine method. The results show that this methodology indeed drastically reduces the computation time and still provides reasonably accurate results.