Using the molecular transformer model for predicting hydrotreating reactions

Data-driven deep learning (DL) models in organic chemistry and computational chemistry have gained attention in recent years to facilitate the work of designing and predicting chemical reactions. For hydrotreating (HT) reactions that occur during conversion of biomass-derived liquids into hydrocarbons, we propose that DL models will enable prediction of chemical transformations without requiring full-scale expensive hydrotreating experimentation. We applied an attention-based transformer model, Molecular Transformer (MT), to test HT reaction prediction that incudes compound product prediction for a set of reactants and process conditions such as temperature and pressure. The model uses datasets based on string representation of molecules. Publicly available chemical reaction datasets derived from the US Patent Office were initially used to pretrain the model, which was then augmented by reactions extracted from relevant HT literature. It is observed that underlying hydrotreating chemistries can potentially be predicted by learning the chemical transformation from the reaction data. Also, the model is likely to learn the non-intuitive trends and correlations between existing reaction data and predicts the outcome accordingly.