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

Area of Science:

• Marine geology and geophysics
• Data-driven predictive modeling
• Oceanographic sedimentology

Background:

• Deep-sea sediment properties provide crucial insights into paleoceanographic conditions.
• High spatial variability in deep-sea sediments complicates accurate property prediction.
• Understanding sediment composition, stratigraphy, and geochemistry is essential for climate reconstruction.

Purpose of the Study:

• To develop and validate a data-driven machine learning framework for predicting key deep-sea sediment properties.
• To identify the most influential features impacting sediment property prediction.
• To quantify prediction uncertainties and assess model robustness.

Main Methods:

• Development of a machine learning framework utilizing five prediction scenarios with tailored preprocessing and hyperparameter tuning.
• Application of the extreme gradient boosting (XGBoost) algorithm as the primary predictive model.
• Utilizing Shapley additive explanations (SHAP) for feature importance analysis and understanding relationships between depth and sediment properties.

Main Results:

• The extreme gradient boosting (XGBoost) model demonstrated superior predictive performance compared to four other algorithms.
• Depth and compressional wave velocity were identified as the most significant predictors for porosity, grain density, calcite content, and thermal conductivity.
• The XGBoost model provided depth-dependent predictions with quantified uncertainties, highlighting its robustness.

Conclusions:

• The proposed machine learning framework offers a robust and accurate method for predicting deep-sea sediment properties.
• Feature importance analysis reveals key drivers for sediment property estimation, particularly depth and seismic velocity.
• The framework's ability to provide quantified uncertainties enhances its utility in paleoceanographic research.