Chemical reactions occurring in the atmosphere greatly control processes of regional air pollution and global climate change. A complete knowledge of all atmospheric chemical reactions would aid in how well we can understand – and thus predict – changes to these processes over time. This is, however, still out of reach, and necessitates active research in discovering unknown reactions to explain discrepancies between models and observation. Here, we develop a data-driven modeling approach that can handle more information than existing discovery methods in order to predict new reactions. Our model, a variant of graph neural networks known as Graph Autoencoders (GAE), accomplishes this by using chemical concentration data and an incomplete set of reactions as inputs. The GAE variant learns a low-dimensional representation for each chemical species, which it then uses to predict the probability of unseen reactions. We assess our model with a widely used chemical reaction mechanism and find it achieves higher performance results compared to other graph-based modeling algorithms. The model’s ability to recover reactions from a known mechanism shows promise in its ability to discover new reactions using observational chemistry data.