Material models describing the relationship between strains and stresses are of great importance for the quality of FE-simulations. Recently data-driven models based on machine learning (ML) methods such as artificial neural networks have been shown to possess the potential to substitute the classic analytical models, promising fast computation, a high level of flexibility and thus the adaptability to new materials. We present a method for training artificial neural networks using only data available in experiments by resorting to physical equations for training the ML material models in order to avoid the need for a classic analytic material model for generating the training data.