Impact of molecular composition on viscosity of heavy oil: Machine learning based on semi-quantitative analysis results from high-resolution mass spectrometry

Abstract

The primary impediment to the recovery of heavy oil lies in its high viscosity, which necessitates a deeper understanding of the molecular mechanisms governing its dynamic behavior for enhanced oil recovery. However, there remains a dearth of understanding regarding the complex molecular composition inherent to heavy oil. In this study, we employed high-resolution mass spectrometry in conjunction with various chemical derivatization and ionization methods to obtain semi-quantitative results of molecular group compositions of 35 heavy oils. The gradient boosting (GB) model has been further used to acquire the feature importance rank (FIR). A feature is an independently observable property of the observed object. Feature importance can measure the contribution of each input feature to the model prediction result, indicate the degree of correlation between the feature and the target, unveil which features are indicative of certain predictions (Zien et al., 2009). We have developed a framework for utilizing physical insights into the impact of molecular group compositions on viscosity. The results of machine learning (ML) conducted by GB show that the viscosity of heavy oils is primarily influenced by light components, specifically small molecular hydrocarbons with low condensation degrees, as well as petroleum acids composed of acidic oxygen groups and neutral nitrogen groups. Additionally, large molecular aromatic hydrocarbons and sulfoxides also play significant roles in determine the viscosity.