Not yet reviewed by Pith; the record is open.

Machine-learned interatomic potentials (MILPs) are rapidly gaining interest for molecular modeling, as they provide a balance between quantum-mechanical level descriptions of atomic interactions and reasonable computational efficiency. However, questions remain regarding the stability of simulations using these potentials, as well as the extent to which the learned potential energy function can be extrapolated safely. Past studies have reported challenges encountered when MILPs are applied to classical benchmark systems. In this work, we show that some of these challenges are related to the characteristics of the training datasets, particularly the inclusion of rigid constraints. We demonstrate that long stability in simulations with MILPs can be achieved by generating unconstrained datasets using unbiased classical simulations if the fast modes are correctly sampled. Additionally, we emphasize that in order to achieve precise energy predictions, it is important to resort to enhanced sampling techniques for dataset generation, and we demonstrate that safe extrapolation of MILPs depends on judicious choices related to the system's underlying free energy landscape and the symmetry features embedded within the machine learning models.