(253bz) Prediction of the Sooting Tendencies of Candidate Biofuels from Molecular Structure Via an Artificial Neural Network

The tendency of candidate bio-based fuel blendstocks to produce soot in traditional engine designs is a significant constraint in screening compounds for next generation biofuels. Experimental methods have thus been developed for rapidly screening the sooting tendencies of small quantities of compounds. While such methods go a long way in easing the experimental burden associated with testing soot formation, it is difficult to obtain an estimate of a molecules tendency to form soot for expensive or otherwise difficult to obtain molecules. We have therefore analyzed the Yield Sooting Index (YSI) measurements from 295 fuel-like compounds recently compiled by McEnally & Pfefferle (2010) to determine if sooting index can be predicted accurately from molecular structure. Molecular descriptors were calculated for each compound and used to train a feed-forward artificial neural network (FF-ANN). Prior to training, 20% of the dataset was reserved as a test set to validate the predictive capacity of the model. The trained model showed median absolute errors of a similar range to the expected experimental error range, indicating that YSI can indeed be predicted by molecular structure. Furthermore, validations of the network architecture indicate that the model generalizes well to new classes of molecules, and subsequent experimental testing has indicated that the model indeed correctly predicts general sooting trends. This work should provide an easy framework for screening the viability of large lists of potential bio-blendstocks, as high-sooting species can be identified directly from molecular structure.