<p>Results of numerous variants of numerical simulations of casting processes serve as training and test data to teach and test artificial neural networks. The data base is generated with a commercial simulation program, MAGMASOFT. The optimization function of this software can be used to automatically generate the data needed for training and testing. In addition to suitable standard modules for data preparation, the widely used open-source programming language Python also has modules for the generation and optimization of artificial neural networks (e.g., TensorFlow). This makes it possible to build a network that makes sufficiently accurate predictions of target process variables when the relevant input parameters are fed in. With a trained network, the target variables can be obtained not only for a particular but for any variation of all input process parameters in their respective ranges. To be able to train a network in a meaningful way, the essential input process parameters affecting the target variables must be known. This limitation results from the exponential increase of variations for generating training data by means of simulation. In this work, examples of trained neural networks with up to eight input variables are presented, resulting in about ten thousand needed simulation variants. The purpose of this work is to establish methods to save computation time in numerical simulations by substituting a process phase (e.g., shot chamber filling or piston moving in high-pressure die casting) by a surrogate model. Another shown application is a porosity prediction model for gravity die-casting case, which is also shown as example.</p>