Explainable machine learning assisted design of tailor-made fuels using conjoint fingerprints

Abstract

This work presents an advanced computer-aided molecular design framework by considering the molecular structure of fuels as the fundamental design degree of freedom. Initially, machine learning algorithms were trained with conjoint fingerprints along with extended connectivity fingerprints, molecular access system keys and functional groups fingerprint separately. Conjoint fingerprints demonstrated the highest predictive accuracy, with R² values generally exceeding 0.9 for physicochemical properties. Subsequently, a novel structure-constrained molecular generator was introduced to systematically generate chemical structures by exploring all rule-based possible isomers of a given target molecule, aiming to produce high-performance fuels. Computational property prediction was employed to virtually screen the generated structures against desired physicochemical fuel property constraints. Finally, explainable artificial intelligence techniques were then applied to achieve atomic-level visualization and quantitative analysis by overlaying the target molecular structure with the atomic contribution values gained for specific properties, providing detailed insights that improve the understanding of how fuel molecular structures affect properties and aid in designing new molecules compared to traditional qualitative analysis. Two case studies were dedicated to illustrating the framework for (i) molecular generation tailored for compression-ignition engines and (ii) analyzing cetane number attributions based on atomic contributions.