The Yield Sooting Index (YSI) is a bench-scale measurement of sooting tendency that is useful in assessing the potential soot emissions from biofuel blendstocks. While fast to measure, acquiring YSI measurements for each component in a blended fuel is a difficult task. Predictive models for YSI are therefore critical to enabling more robust prediction of a fuel's engine-out emissions. In this talk, we improve upon previous state-of-the-art methods in YSI prediction by leveraging a recently developed model for bond dissociation energy prediction. Previous work has emphasized the role of bond strengths, and the radicals which form when the weakest bonds in a molecule are broken, as a critical feature in influencing soot formation. Hydrogen abstraction or other bond breaking reactions results in the formation of two radicals. We construct a graph neural network model that predicts a molecule's overall YSI as a function of the bond strengths and resulting radicals by breaking each of the molecule's weakest bonds. It does this by learning a suitable contribution from each carbon atom in the product radicals and summing over all the carbon atoms in each radical to form a single contribution from that bond breaking reaction. The breaking of different bonds is then combined to a single molecule prediction by learning a weighted average contribution from each bond, based on the relative bond dissociation free energies from the weakest predicted bond in the molecule. The model is validated through a 10-fold cross-validation and compared to a previous linear model approach that encodes molecules based on their carbon fragments. The GNN achieved an MAE &gt;25% lower than the linear model. Mechanistically-inspired architectures for property prediction can not only improve accuracies, but also offer interesting scientific insight. Without explicit instruction, the model learns that weak bonds that retain oxygen functional groups lead to lower sooting index predictions, while those that lead to the radical recombination of PAH rings are predicted to have a higher YSI.

Chemistry-informed neural networks for the prediction of yield sooting index

ACS Spring 2022

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