Complex reaction processes in Kerogen pyrolysis unraveled by deep learning-based molecular dynamics simulation

Abstract

The performance of the instruments and operational handling inevitably introduce experimental errors, leading to an incomplete understanding of the kerogen pyrolysis mechanism. Data-driven deep learning offers the potential to address these challenges. Based on conventional experiments (¹³C-NMR, XPS, and FT-IR), this study constructed a molecular model of kerogen from Huadian oil shale. Subsequently, reactive molecular dynamics simulations were employed to simulate the pyrolysis process of the kerogen macromolecular model, yielding 40,483 stable and complex non-equilibrium structures during the pyrolysis process. For these structures, we utilized deep learning combined with quantum chemistry calculations to establish, for the first time, a high-precision pyrolysis potential energy model specific to kerogen molecules. This model reveals the pyrolysis mechanism at the atomic scale with significantly enhanced accuracy. By employing a data-driven approach, we reduced errors in the study of pyrolysis mechanisms caused by experimental instrumentation and manual operations. Furthermore, this method overcomes computational limitations inherent to traditional quantum chemistry and molecular dynamics, achieving a balance between computational accuracy and speed. This study not only provides a new perspective for using deep learning to tackle the challenges of non-equilibrium complex structures in computational chemistry but also unveils the kerogen pyrolysis mechanism at the atomic scale. It further advances the theoretical computational framework for regulating oil shale reaction processes.