Unraveling Depositional Conditions Using Machine Learning Tools: Application to Tethyan Phosphorite Deposits.

Abstract

This study applies machine learning techniques to predict geochemical anomalies indicative of paleoenvironmental conditions, particularly redox conditions. Cerium (Ce) is utilized as a key geochemical proxy for redox. This work focuses on phosphate-rich deposits within the Tethyan phosphorite belt, which spans the Middle East, North Africa, and Turkey. Major oxide geochemical data were used to develop predictive models for Cerium distribution. Two machine learning models including Extreme Gradient Boosting (XGBoost) and Random Forest (RF) were employed to predict both Ce concentrations and Ce anomalies (Ce/Ce*). The datasets were geochemically classified into oxic and sub-oxic redox patterns, which were then evaluated individually and in combination using both models. Feature importance analysis for Ce concentration consistently identified MnO as the most influential predictor across models. For Ce anomaly prediction, different predictors dominated depending on the redox context: in the oxic dataset, K₂O, SiO₂, and Fe₂O₃ were the most significant; in the sub-oxic dataset, MnO and Fe₂O₃ were dominant, indicating possible terrigenous input in addition to the redox state of environment. In the combined dataset, Al₂O₃ and Fe₂O₃ emerged as key predictors. Among the models tested, XGBoost yielded the lowest prediction errors for both Ce concentration and Ce anomaly across all datasets, confirming its superior performance. The findings underscore the utility of Ce concentration and Ce/Ce* anomaly as robust indicators for paleoenvironmental reconstructions. These geochemical proxies offer valuable insights into ancient redox conditions and the environmental conditions of phosphorite deposition in the Tethyan region.