A multiscale graph neural network for predicting the properties of high-density cycloalkane-based diesel and jet range biofuels

Abstract

Predicting the fuel properties using computer techniques can speed up the search for alternatives to replace fossil-based diesel and jet fuels and lower research costs. However, previously reported graph neural network (GNN) models are not suitable for the fuel property prediction of biofuels with ring structures, such as cycloalkane-based high-density biofuels, because GNNs with a limited number of layers are inadequate for capturing the global structure of compounds. In this work, we proposed a multiscale graph neural network (MGNN) model to estimate the fuel properties of cycloalkane-based diesel and jet-range biofuels. The MGNN model increased the receptive field of each node, allowing nodes to perceive topological and feature information from a larger neighborhood, which enhanced the complexity and capacity of the model, thereby improving its fitting ability. Traditional over-smoothing issues in the MGNN were overcome by introducing dense connections, which maintained the distinctiveness of vertex embedding and preserved substructure details. The coefficients of determination of the linear regressions (R²) were all in the range of >0.98 with smaller mean relative errors (MREs) and a narrower range of error distribution compared to conventional GNN models. A detailed analysis of the relationship between these properties and various topological descriptors was discussed. The results show a promising and accurate method for estimating the fuel properties of cycloalkane-based diesel and jet-range biofuels.