Modern high-resolution numerical weather prediction generates vast amounts of data; however, much of the data is just turbulence and not really meaningful to both experts and end-users. At the same time, numerical weather forecasts can be viewed as (sequences of) multidimensional images, so that tools from computer vision can be applied to identify patterns. Specifically, autoencoders are often used to derive lower-dimensional representations of complex high-dimensional data. These so-called embeddings have useful properties that enable search and clustering, which in turn can be used to characterize and summarize weather patterns in forecasts. This is not only useful for downstream applications and end-users, but also for research and analysis, as embeddings are often associated with particular physical processes. In this presentation we will see how embeddings are created for the High Resolution Rapid Refresh forecasts of the National Weather Service (US), and discuss their properties and applications.