Machine learning for FT-ICR MS analysis: predicting origin, thermal evolution, and biodegradation of crude oils

Abstract

This study provides a detailed assessment of the integration of machine learning with Fourier transform ion cyclotron resonance mass spectrometry (FT-ICR MS) for the characterization of crude oil properties. Utilizing data from both positive-ion atmospheric pressure photoionization (APPI) and negative-ion electrospray ionization (ESI) techniques, this study systematically modeled geochemical characteristics (origin, thermal evolution, and biodegradation) using partial least-squares for discriminant analysis (PLS-DA) and ordered predictor selection for discriminant analysis (OPSDA). The models were initially trained using subsets in which the variable characteristic under investigation was isolated. Under these conditions, very high classification performance was obtained for several heteroatom classes, with some models showing complete class separation within the evaluated data set. This performance illustrates the ability of the models to discern subtle geochemical signatures within the oil samples. The subsequent application of these models to the entire data set confirmed their ability to generalize and maintain high predictive accuracy across a broader spectrum of samples. In the final stages of analysis, interpretative insights were drawn from the most significant variables identified by the models. This discussion delves into the molecular constituents that play pivotal roles in the geochemical differentiation of oils, shedding light on potential markers for oil classification. These findings underscore the effectiveness of APPI (+) and ESI (−) in identifying molecular features and demonstrate the synergy between advanced mass spectrometry techniques and machine learning. Such integrative approaches promise to enhance oil classification processes, offer valuable insights for the petroleum industry, and aid environmental management by providing a robust framework for identifying molecular markers that distinguish different types of crude oils. Overall, this study confirms the robust capabilities of machine learning for high-resolution petroleum molecular characterization and paves the way for future innovations in analytical and data-driven approaches for complex crude oil analysis.