Feed Forward Neural Network for Predicting Enantioselectivity of the Asymmetric Negishi Reaction

Density functional theory (DFT) has become a popular method to model transition state (TS) energies to predict enantioselectivity, but the associated errors present challenges. Machine learning has emerged as a powerful tool to model enantioselectivity but generally requires large datasets for training. Herein, we describe the development of a feed forward neural network for predicting enantioselectivity of the Negishi cross-coupling reaction with Boehringer Ingelheim (BI)-type phosphines. The selectivity predicted from DFT TS energies is upgraded through the neural network based on input features including geometries, electron population, and dispersive interactions. This new approach to modeling enantioselectivity is compared to conventional approaches, including exclusive use of DFT energies, and data science approaches using features from ligands or ground states with simple neural network architectures.