Investigating ground or thermal states of quantum phases of matter is of great interest in condensed matter physics. However, obtaining observables of interesting states is expensive, due to the necessity of physical experiments and the limitations in capabilities and accessibility of current quantum devices. Furthermore, there are no efficient ways of simulating these properties on classical computers. It is however possible to interpolate between experiments from relatively small amounts of data. Recent studies have proposed algorithms to learn to predict properties of the aforementioned states and provided theoretical guarantees. In this work, we aim to build on both the theoretical and the practical results. We aim to show better scaling with training data for Gaussian Processes, which we want to verify empirically. Moreover, we study their performance compared to the performance of Graph Neural Networks.