The extensive progress in computing power coupled with advances in computer algorithms allow the application of deep learning (DL) into various fields. In the conversion of biomass-derived liquids into hydrocarbons, DL models can be used to provide non-intuitive insight in considering alternatives to a disrupted supply chain or when feedstock diversification is needed without requiring full-scale expensive hydrotreating experimentation. An ensemble of DL models may need to be developed to address specific portions of the conversion process. Given that chemical reactions can be expressed in a linguistic form using SMILES, we applied the Molecular Transformer algorithm to model and predict the outcomes of hydrotreating (HT)-related reactions. Publicly available datasets derived from the US Patent Office were initially used to pretrain a model. The training dataset was then augmented by adding reactions extracted from HT-relevant literature. Preliminary results show that predictive accuracy increased from 17% to 71% by augmenting the original generic training set with a much smaller number of targeted reactions. We developed experiments to better understand how the algorithm “learned” the additional information and to increase model interpretability. Future work will entail inclusion of computed and additional project-derived experimental data into the training and validation sets.

Modeling hydrotreating reactions using transformer networks

ACS Fall 2021

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.