Integrating Mechanistic Knowledge into Machine Learning Enables Improved Reaction Prediction in Low-Data Regimes

One of the limitations of routinely using machine learning tools for practical chemistry is the requirement for large volumes of data. The current direction in this field is towards low-abundance data models supported by chemical understanding. In this article, we show that embedding chemical expertise directly into the model via mechanism-based descriptors, creates an inductive bias that reflects how organic chemists reason about reactivity and considerably improves predictive performance in a low data regime with significant gains in recall and F1-scores compared to standard representations to identify low and high yield reactions.