Fig. 2

The architecture of the reaction context recommendation model. A The initial component is the candidate generation model, comprising a feedforward neural network. This model encodes reaction fingerprints and predicts the probabilities of the solvents and reagents that might be relevant to the reaction as a multi-label classification problem. The predicted relevant solvents and reagents are then enumerated combinatorically to generate a list of possible reaction contexts for the reaction. B Subsequently, the ranking model predicts the temperature and relevance score for each generated reaction context from the first model