Cross-well machine learning prediction of sonic logs in Newfoundland and Labrador

Key Points

• Evaluate the ability to predict compressional slowness from non-sonic logs using machine learning.
• Used a leakage-free, features-only strategy for blind cross-well prediction.
• Built causal lag windows from past non-sonic logs without including sonic-derived channels.
• Implemented depth conditioning and rank-aggregated feature selection with time-aware validation.
• Compared three model types: Random Forest, Extreme Gradient Boosting, and BiLSTM.
• XGBoost achieved an R² of 0.895 and MAE of 11.38 µs/m in cross-well validation.
• Random Forest was competitive in several configurations.
• BiLSTM underperformed compared to the other models.
• Highlighting the importance of rigorous leakage control and feature engineering.