This paper presents a deep kernel learning (DKL) model for predicting chemical reaction outcomes and optimizing reaction conditions. The model combines the feature learning capabilities of neural networks with the uncertainty quantification of Gaussian processes (GPs). It demonstrates superior performance across various input representations (molecular descriptors, fingerprints, and molecular graphs) compared to standard GPs and achieves comparable results to graph neural networks (GNNs) while providing uncertainty estimates. The model's uncertainty estimations facilitate its integration with Bayesian optimization (BO) for efficient reaction optimization.

Sukriti Singh, José Miguel Hernández-Lobato