Integration of Machine Learning Techniques for Structural-Level Modelling of Fuel Molecules

Abstract

Traditional refining processes are transitioning towards more efficient molecule utilisation, due to environmental regulations. Concurrently, advancements in computer technology and analytical chemistry are enhancing refining process simulation techniques. The inclusion of renewable energy sources, such as biomass-derived oils, poses challenges to traditional hydrocarbon-focused simulation methods. This study aims to develop a data-driven approach to refining process simulation by integrating cheminformatics and machine learning techniques, leveraging abundant molecular data from advanced analytical methods. Current refining models vary in molecular structure representation, creating integration challenges. To address this, we introduce a novel connectivity matrix extraction method for molecular substructure representation and the Mol2vec model for substructure similarity calculation. We enhance pure component property prediction accuracy using machine learning regression techniques to associate molecular structure with properties. Moreover, to address the complexities of handling heteroatom-containing compound systems, we propose a method for reconstructing molecular compositions based on bulk properties, leveraging the machine learning property prediction model developed in this study. This approach not only offers higher accuracy compared to previous techniques but also facilitates seamless integration with existing composition models, thereby expanding the model's applicability. Furthermore, compared to template-based chemical reaction rules, machine learning-based reaction pathway prediction model offers broader applicability and flexibility. The molecular substructure description method developed in this study accurately captures the molecular substructure environment compared to methods directly based on SMILES. We integrated the molecular substructure description method with machine learning techniques to develop versatile reaction prediction models.

Date of Award	19 Jun 2024
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Nan Zhang (Main Supervisor) & Xiaolei Fan (Co Supervisor)