Powerful and Reliable Prediction using Latent Variables of Experimentally Unobservable Reactions in Organic Synthesis

In this study, a novel machine learning algorithm was designed to assist in the development of organic reactions. This algorithm addresses the complexities inherent in batch- type organic reactions, including the necessity for numerous experiments and the effects of intricate characteristics of reaction pathways. By integrating molecular relationships and actual yields from observable reactions, the algorithm is used to estimate untested yields via extrapolation. An approach based on Bayesian optimization and dual annealing optimization is employed to compute expected values and evaluate plausibility. The algorithm’s dual-loop 2 structure, incorporating latent variables and experimental values, maximizes the coefficient of determination. Physicochemical aspects of the algorithm are validated using natural bond orbital charges, and its utility in synthesizing perfluoroiodinated naphthalenes is demonstrated. The algorithm exhibits potential for application in predicting experimentally unobservable reactions, thereby advancing the field of synthetic organic chemistry.