Physics-informed neural networks (PINNs) for numerical model error approximation and superresolution