During the recent COVID-19 crisis, mathematical modeling has been one of the principal forms to provide evidence on the effectiveness of public health and social interventions [1]. Due to local outbreaks and human contact patterns, infectious disease dynamics are often heterogeneous on a spatial or demographic scale and for efficient mitigation, local transmission dynamics should be integrated in a mathematical model. Over the last years, a large number of authors has made contributions to predict the development of SARS-CoV-2. Simple ordinary differential equation based models have been used for their efficiency in time-critical moments and through metapopulation modeling, various authors addressed the spatially heterogeneous spread of SARS-CoV-2 very efficiently; see, e.g., [2]. % On one hand, these models are very efficient and integrate a meaningful level of complexity, on the other hand, they can not be executed on-the-fly. While a HPC infrastructure can be used to run a parallelized version or to simulate thousands to millions of runs for parameter estimation or scenario considerations in short time, we need to ensure low-barrier access to reliable and up-to-date results for public health experts and decision makers. This is even more important as time is a critical factor in pandemics and evaluations of mitigation and reaction strategies have to be conducted in narrow time windows, with constantly changing initial conditions. Here, we suggest to combine expert mechanistic modeling with AI-based approaches to allow for an on-the-fly execution of reliable and accurate infectious disease models. We build upon an already validated spatially and demographically resolved metapopulation model [2] whose predictions we use as data to train a machine learning model. As the output is already in the form of a graph, a graph-based machine learning model is a natural choice and we therefore propose to employ graph neural networks (GNNs). In combining advanced mechanistic modeling and machine learning approaches, we significantly advance pandemic preparedness. Providing predictions on spatially resolved disease dynamics in a fraction of seconds enables a novel level of evidence-based decision making. [1] European Centre for Disease Prevention and Control, Tech. rep. (2024). doi:10.2900/253991. [2] M. J. Kühn et al, Mathematical Biosciences (2021).