Researchers at Delft University of Technology are working on the CompleteRxn project to address incompleteness in open chemical reaction data, which limits the development of accurate AI models for chemistry. The project began with a sustainability focus, as evaluating environmental impact requires a reliable mass balance of inputs and outputs, but open databases often miss core reactants or by-products. Open reaction datasets, commonly derived from chemistry-related literature and patents in the US Patent & Trademark Office (USPTO) database, are heavily used in machine learning to predict reaction outcomes or synthesis routes, yet many entries are incomplete, missing reactants, ignoring by-products, or containing misidentified molecules from automated extraction tools. Researcher Jana notes, “We don’t know yet how much this incompleteness affects predictive models, but it might be a limiting factor in developing accurate AI for chemistry.” The USPTO dataset is technically open and FAIR (Findable, Accessible, Interoperable, and Reusable), but molecules are encoded as text strings, making them difficult for experimentalists or sustainability practitioners to use directly. Multiple research groups have created modified subsets of the USPTO data, filtering and cleaning in different ways, but these transformations are not always traceable.