The use of data-driven methods to model constitutive laws in mechanics has gained significant popularity in recent years due to their ability to accurately capture complex material behavior. In particular, Physics-Augmented Neural Networks (PANNs) combine flexibility with built-in compliance with fundamental physical principles, making them especially attractive. However, while considerable effort has been devoted to the design of neural networks, there is still a lack of well-defined protocols for identifying such models from training data obtained through real experiments. In particular, key questions remain unanswered: which specimens should be selected, and which multi-axial loading paths should be applied? This work initiates a reflection on these issues by drawing inspiration from the field of Design of Experiments. One of the simplest and most commonly used criteria is space-filling, i.e., a geometric criterion that evaluates how well the data cover a given domain. Unlike traditional constitutive models, which aim to describe global regimes (e.g., elasticity, plasticity), the present approach deliberately restricts the modeling effort to a controlled target subspace, defined by the operating conditions of a specific structure. Within this subspace, not all points are of equal importance, both for model accuracy and for the targeted application, especially when performance is evaluated with respect to specific engineering criteria. To support experimental design decisions, several data-quality metrics based solely on the geometric dispersion of the data are proposed and analyzed, with the objective of identifying those that correlate with model performance. Finally, a preliminary study explores the possibility of directly learning such a quality indicator from the data.