Machine Learning for Fuel Property Predictions: A Multi-Task and Transfer Learning Approach

Abstract

<div class="section abstract"><div class="htmlview paragraph">Despite the increasing number of electrified vehicles the transportation system still largely depends on the use of fossil fuels. One way to more rapidly reduce the dependency on fossil fuels in transport is to replace them with biofuels. Evaluating the potential of different biofuels in different applications requires knowledge of their physicochemical properties. In chemistry, message passing neural networks (MPNNs) correlating the atoms and bonds of a molecule to properties have shown promising results in predicting the properties of individual chemical components. In this article a machine learning approach, developed from the message passing neural network called Chemprop, is evaluated for the prediction of multiple properties of organic molecules (containing carbon, nitrogen, oxygen and hydrogen). A novel approach using transfer learning based on estimated property values from theoretical estimation methods is applied. Moreover, the effect of multi-task learning (MTL) on the predictions of fuel properties is evaluated. The result show that both transfer learning and multi-task learning are good strategies to improve the accuracy of the predicted values, and that accurate predictions for multiple fuel properties can be obtained using this approach.</div></div>