An optimized deep convolutional neural network for yield prediction of Buchwald-Hartwig amination

Abstract Nowadays deep convolutional neural networks (DCNN) have been made great achievements in the engineering fields of computer vision and natural language processing, but its application in chemical reactions remains limited. Herein, an effective DCNN model has been proposed and exemplarily applied to yield prediction of Buchwald-Hartwig amination on the basis of the characteristics of chemical datasets. By designing a pooling-free network framework and dealing with previously reported 3690 chemical reactions and 120 descriptors, a satisfied yield prediction with R2 = 0.96 and RMSE = 4.95% (random 70% of the all data for training) has been achieved. Even if the size of training set decreased to 50%, the yield prediction with acceptable accuracy (R2 = 0.94 and RMSE = 6.15%) has been also realized by the newly designed DCNN model. Besides, the result of principal component analysis highlights that only 64 descriptors are needed to enable a good yield prediction and that catalyst ligands (48 of 64 descriptors) could have a larger impact on yield than additives and aryl halide substrate.