Coupling data-driven geochemical analysis and ensemble machine learning for automatic identification of oceanic anoxic events

Abstract

Oceanic Anoxic Events (OAEs) have been recorded across the Phanerozoic and linked with catastrophic events in geological records, including the massive release of CO2 into the atmosphere and mass extinctions of marine animals, particularly during the Cretaceous period (e.g., OAE-2). Overall, the occurrence of OAEs were typically identified based on the deposition of organic-rich black shales, carbon isotopic excursions, and enrichment of redox-sensitive elements. While various OAE intervals have been extensively studied across the Tethys using multiproxy geochemical records, recognizing the expression, and understanding the duration of these events are rather challenging and a subject of active debate. This is further compounded by the time-consuming and expert-demanding analysis to interpret complex geochemical records associated with OAEs. To address these issues, we propose a novel approach by coupling data-driven geochemical analysis and ensemble machine learning to recognize and predict the occurrence of OAE-2 in the Upper Cretaceous based on key geochemical records (δ13Corg, TOC, Mo, V, U) collected from different areas geographically. Considering variation in data availability and completeness, we performed machine learning-based data imputation to fill the gaps in geochemical records without perturbing the overall trends and patterns. With this, our prediction of OAE in various locations using ensemble machine learning, achieving an accuracy of up to 90% in the validation and 78% in the blind test predictions. The model could also match the interpreted OAE-2 intervals from different locations with higher resolution prediction based on the δ13Corg and the TOC as the most important parameters followed by the sensitive redox elements. This suggests that the model utilized similar parameters used by geologists in identifying OAEs, increasing the model interpretability. Application of machine learning and data-driven geochemical analysis could help in providing a robust and time-efficient identification of OAE and find new unexplored OAEs along the stratigraphic records.