Machine-learning potentials enable molecular simulations of water over system sizes and timescales that are inaccessible to direct electronic-structure calculations. Our recent work, enabled by NAISS 2025/22-909, has, however, demonstrated an important challenge in constructing such models: accurately reproducing training data does not necessarily guarantee accurate dynamical properties. In this project, we will continue the work initiated during NAISS 2025/22-909 to train and systematically evaluate NEP and qNEP machine-learning potentials for water based on several widely used levels of theory. The GPU resources will be used extensively for model training and active learning, followed by long molecular dynamics simulations that provide stringent tests of the resulting models. In particular, we will evaluate structural, vibrational, dielectric, and transport properties over a range of temperatures and pressures and compare with available experimental reference data. The outcome of these simulations will in turn guide refinement of the training data and models. The project addresses both the AI/ML challenge of constructing and validating machine-learning potentials for complex molecular systems and the underlying physics question of how accurately different models describe the dynamics of liquid water. The resulting models already include the ions relevant to our longer-term work and will provide the foundation for studying ionic conductivity at experimentally relevant conditions and, subsequently, its possible modification under vibrational strong coupling.