Molecular graph-based deep learning method for predicting multiple physical properties of alternative fuel components

Abstract

The rapid development of novel fuels has spawned the demand for fast and precise prediction of fuel physical properties. In this work, we coupled the molecular features and atom features extracted from Graph Isomorphism Network (GIN) to propose a Molecular Graph-based Deep Learning (MGDL) method with richer information features. The prediction results of flash point (FP) were compared through three Graph Neural Network (GNN) models, among which the GIN had the best prediction performance, with R2 and mean absolute error (MAE) of 0.991 and 3.952 K, respectively. In addition, the maximum absolute error was 28.99 K and only 5.49% of the compounds had an absolute error greater than 10 K. Other physical properties e.g., normal boiling point (NBP), density, were also accurately predicted demonstrating the wide applicability of the method.