ChemCNet: An Explainable Integrated Model for Intelligent Analyzing Chemistry Synthesis Reactions

Abstract

Palladium (Pd)-catalyzed cross coupling reactions are of great significance in organic synthesis. However, the reaction route is more complex, time-consuming and costly. For addressing the above problems, a model-related feature selection strategy is introduced, focusing on iterative optimization of feature description and prediction to guide and strengthen each other. Then, we combine the lightweight convolution neural network (CNN) driven by attention mechanism with CatBoost to build an intelligent chemical synthesis reaction analysis model-ChemCNet. Moreover, we conduct the interpretability analysis based on ChemCNet model. The results show that ChemCNet model has achieved relatively high prediction accuracy and generalization, and it is helpful to provide reliable decision-making information for the experimenter or institution.