Machine Learning for Physicochemical Property Prediction of Complex Hydrocarbon Mixtures

Machine learning has proven effective for predicting properties of pure compounds from molecular structures, but mixtures, in particular oil fractions, are rarely dealt with. At best, the bulk estimated based on compound properties, linear mixing rules, and a reconstructed composition feedstock. As detailed such mixtures is well determined often approximated by lumps, accuracy can be improved. In this work, we demonstrate naphtha case study our property estimation method. First, PIONA delumped into molecule-level composition, machine learning-based approach used to predict those molecules, which further combined another deep neural network prediction properties. The latter models trained mixture using vectors that represent mixture. first vector combination representation geometries make up second applies rules boiling temperatures, critical liquid densities, vapor pressures predicted with learning. last consists learned distillation curve. We show an integrated starts structures offers significant improvements over existing approaches applied industry academia.