Direction-dependent calibration and imaging is a vital part of producing radio images that are deep and have a high fidelity and highly dynamic range with a wide-field low-frequency array such as LOFAR. Currently, dedicated facet-based direction-dependent calibration algorithms rely on the assumption that the size of the isoplanatic patch is much larger than the separation between bright in-field calibrators. This assumption is often violated owing to the dynamic nature of the ionosphere, and as a result, direction-dependent errors affect image quality between calibrators. In this paper we propose a probabilistic physics-informed model for inferring ionospheric phase screens, providing a calibration for all sources in the field of view. We apply our method to a randomly selected observation from the LOFAR Two-Metre Sky Survey archive, and show that almost all direction-dependent effects between bright calibrators are corrected and that the root-mean-squared residuals around bright sources are reduced by 32% on average.