A novel approach for constructing a machine-learned potential energy surface (MLP) from unlabeled training data is presented. Utilizing neural networks augmented with a pool-based active learning sampling method, a potential energy surface (PES) is developed for the accurate modeling of interfaces of hematite iron oxide and water, fitting the much more expensive density functional theory (DFT). Molecular dynamics simulations were performed using this DFT-based PES to characterize the structural and energetic properties of the system. By utilizing the developed machine learning potential (MLP), it was possible to simulate much larger systems for extended periods of time, which will be important for leveraging machine learning potentials as accurate and pragmatic simulation-led molecular design and prototyping tools whilst preserving the ab initio accuracy.