The integration of prior knowledge into the training of machine learning (ML) models can improve their inter- and extrapolation capabilities and increases the trust of domain experts in model predictions. Shape-constrained regression is one category of ML algorithms capable of integrating knowledge about the shape of the model. Such knowledge is represented by boundary information of partial derivatives of different orders. However, the translation or formulation of (intrinsic) domain expert knowledge into such constraints is challenging and requires experience. Sometimes, this knowledge may even be unavailable for certain domains. We propose an approach that can automatically infer such knowledge from observational data. We envision this approach as an additional tool in the data analysis toolbox that provides suggestions, which can be incorporated into the training of prediction models. In this work, we describe our approach for automated knowledge inference from data. Additionally, we show the applicability of our approach by testing it on synthetic data generated from a set of physics equations.