Property prediction of fuel mixtures using pooled graph neural networks

Renewable fuels offer a sustainable option for engine applications where electrification is more challenging, not possible. To evaluate the potential of novel fuels it is crucial to first determine their combustion and spray related properties. This can be done experimentally, but during screening of multiple fuel candidates this can be cost and time expensive. Machine learning can be used for rapid, inexpensive, and accurate predictions of fuel mixture properties. To this end a novel function for pooling molecular representations called MolPool has been developed, which was combined with graph neural networks. The new approach processes the input permutation invariant, allowing for application to a varying number of components in the mixture. In this article, three different compression ignition engine related properties were investigated: derived cetane number (DCN), flashpoint, and viscosity. The results show that this novel neural network approach is able to increase the prediction accuracy and the generalizibility compared to traditional blending laws for all investigated properties. MolPool improves the prediction if oxygenated species are present in the mixture resulting non-linear mixture behavior, which is common for renewable fuels. Thus, MolPool shows great potential for prediction of various properties and fuel mixtures.