Decoding Molecular Controls on Carbon Sequestration in Lake Sediments Using Machine Learning

Lake sediments are critical archives of past environmental changes, yet interpreting the complex biogeochemical processes governing carbon sequestration remains a significant challenge. Our previous work in Lake Liangzi (Hou et al., 2026, Water Research) highlighted a paradox: eutrophication-driven ecological regime shifts increased organic matter inputs but ultimately weakened long-term carbon burial. This motivates the need for integrative approaches that can resolve the molecular-level controls governing the fate of sedimentary organic matter.
Here, we propose a machine-learning-assisted framework that integrates ultra-high-resolution mass spectrometry (FT-ICR MS) with data-driven analysis to explore these controls. We analyze a sediment core spanning nearly two centuries to construct a comprehensive matrix of thousands of unique organic molecules and their intrinsic chemical properties (e.g., O/C, H/C, AI_mod). Based on operational persistence criteria, molecules are classified into distinct fate categories reflecting their stability.
Using these fate classifications, our primary goal is to train and interpret an XGBoost model to test the hypothesis that molecular fate can be inferred from chemical properties alone, and to identify candidate molecular characteristics that may govern carbon persistence. In parallel, molecular transformations between adjacent sediment layers are examined to reveal potential shifts in dominant biogeochemical reaction pathways associated with historical ecosystem changes.
Finally, we explore the use of model performance itself as a diagnostic proxy for ecosystem stability. This approach is designed to assess the hypothesis that periods of ecological transition are associated with reduced predictability of biogeochemical processes. Overall, this study presents a transferable analytical strategy aimed at extracting process-oriented insights from lake sediment archives, highlighting the potential of machine-learning-guided approaches to advance limnogeology beyond descriptive reconstruction.