Most chemical reactions “fail”—and while these experiments are carefully recorded in laboratory notebooks, they are not published in the open literature. But these unreported “dark” reactions contain valuable information that can be exploited by machine learning algorithms to improve the synthesis process, by clearly defining the boundary between “successful” and “unsuccessful” reaction conditions. In this talk, I will describe our work on the “Dark Reactions Project”—an effort to develop an open-source public repository for storing and utilizing hydrothermal synthesis reaction data. We have collected &gt;9000 historical syntheses of organically-templated metal oxides. We convert the reaction description (reactant names, masses, pH, etc.) into a set of physicochemical properties using cheminformatics, and then build machine-learning models to predict reaction “success”. A recent experimental test of 500 new syntheses with 34 novel amine precursors (which were absent from our database and are under-represented in the Cambridge Crystal Structure Database) showed that our machine learning model can successfully predict successful reaction conditions 89%, beating human experiments. Moreover, the model can be “inverted” to yield novel, human-readable hypotheses for compound formation. In addition to presenting an overview of the project, I will discuss our approach to solving the informatics challenges associated with representing and capturing solid-state materials synthesis data, as well as our new directions for improving the reaction and structural outcome descriptors.

Dark reactions project: Machine learning-assisted materials discovery using failed experiments

253rd National Meeting (2017)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.