Data-driven machine learning models for predicting engineering properties in deep-sea sediments

Abstract

Predicting the properties of deep-sea sediments offers critical insights into past oceanic conditions, including sediment composition, stratigraphy, and geochemical signals. However, accurate prediction is hindered by the high spatial variability of these sediments. This study presents a data-driven machine learning framework to predict five key sediment properties. Five prediction scenarios were developed with tailored preprocessing and hyperparameter tuning, and Shapley additive explanations were employed to assess feature importance and the relationships between depth and sediment properties. Among the five tested algorithms, the extreme gradient boosting (XGBoost) model achieved the highest predictive performance. Depth and compressional wave velocity emerged as the most and second most influential features for estimating porosity, grain density, calcite content, and thermal conductivity. The depth-dependent predictions with quantified uncertainties generated by the XGBoost model demonstrate that the proposed framework provides a robust approach for predicting deep-sea sediment properties.