AI System for Supporting Generation of Optimal Synthetic Pathways Based on Chemical Reaction Big Data

The selection and design of the appropriate reaction paths has a significant impact on the economics and productivity of the chemical process, enhanced by milder operating conditions, use of cheaper reactants and fewer reaction steps. However, exploration of reaction information is difficult even with reaction databases available, causing path explosion problem due to huge search space. In this study, we propose an AI system (ASICS), which supports synthetic path design at the basic stages of research and process design, based on the hybrid generative exploration and exploitation of reaction knowledge graphs encoding big data of patented reactions and machine learning-based retrosynthetic prediction. ASICS generates an optimal synthetic path that satisfies the given constraints (regulated compounds, etc.), based on A* search using synthetic accessibility and retrosynthetic prediction scores. The preference in searching between confirmed reaction spaces and unexplored reaction spaces through prediction can be selected by the user. The fusion of reaction knowledge base and retrosynthetic prediction model enables to generate optimal synthetic paths beyond the accumulated reaction information.