Fig. 1: Overview of the approach.

The learning process is divided into three phases that alternate continuously: (1) the wake phase, (2) the abstraction phase, and (3) the dreaming phase. a The wake phase searches for synthesis routes from a set of purchasable molecules to synthesize the target product. The search process is guided by neural models that select the most promising molecules and rank candidate templates based on the given intermediate molecule. b The abstraction phase grows the library of reaction templates by extracting frequently used reaction chains from molecule synthesis routes. In this work, we mainly focus on two types of reaction chains: (1) cascade reactions, which we defined as “cascade chains” and (2) complementary reactions, which we defined as “complementary chains'', and abstracting out these reaction chains into a new reaction template. c The dreaming phase augments the training dataset by simulating retrosynthetic experiences through bottom-up and top-down approaches, termed “fantasies''. Then, the neural models are refined on replayed experiences from the wake phase and fantasies for future use.