Machine learning has recently been used to generate models of chemical data with impressive levels of predictive power. This statistical success is largely quantitative, but insufficient to provide the chemical insights that drive the field of chemistry. We begin with an examination and deconstruction of black box machine learning models of a computationally generated chemical reactivity dataset. We experiment with chemical representations and modeling techniques which provides insights into the nature of how statistical accuracy can arise, even when the model is devoid of generalizable physical principles. By peeling back the layers of these complicated models we arrive at a minimal, chemically intuitive model (and no machine involved). This model is based on systematic reaction type classification and Evans-Polanyi relationships within reaction types which are easily visualized and interpreted. Through exploring such a model and interactively adding incremental complexity, we gain deeper understanding of the dataset and produce models more relatable to chemical intuition.

Understanding the limits of machines in learning chemical reactivity

2019 Spring National Meeting

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