Comparative analysis of ensemble and linear models for geophysical Delta-T compressional (DTC) well log imputation and reconstruction
Open Access
- 6 November 2026
- journal article
- research article
- Published by Frontiers Media SA in Frontiers in Earth Science
- Vol. 14, 1915184
Abstract
Delta-T Compressional (DTC) well logs provide important sonic information for formation evaluation and subsurface characterization, but incomplete or unavailable sonic-log measurements can limit subsequent interpretation. This study evaluates six supervised machine-learning models Multiple Linear Regression, SGD Regressor, K-Nearest Neighbors (KNN), Multi-Layer Perceptron (MLP), Random Forest, and XGBoost for DTC log reconstruction using well-log data from the Taranaki Basin, New Zealand. The dataset was cleaned, depth-aligned, and divided into an 80% training set and a 20% internal test set. Three independent wells were reserved for blind testing. Hyperparameter optimization was performed using randomized search with 2-fold cross-validation for the selected nonlinear models. On the internal test dataset, Random Forest achieved the highest R2 of 0.96, followed by KNN and XGBoost, both with an R2 of 0.93. Independent blind testing showed that model performance varied between wells. SGD achieved an R2 of 0.85 for Trapper-A1, XGBoost achieved 0.71 for Ahuroa South B-1, and MLP achieved 0.81 for Kahili-1C. MLP and XGBoost showed relatively better performance across the independent wells. The predicted DTC profiles reproduced the overall trends of the observed logs, although larger deviations occurred in some depth intervals. An uncertainty analysis based on MLP P5–P95 prediction intervals showed empirical coverage of 33.30%, 25.31%, and 29.32% for the three blind-test wells, indicating that the estimated uncertainty intervals were narrower than the observed prediction variability. The proposed workflow can support DTC reconstruction in wells with incomplete or unavailable sonic-log measurements and may provide useful input for formation evaluation, geomechanical analysis, and hole-cleaning studies. Further validation using additional wells and improved uncertainty-calibration methods is required before broader application.Keywords
This publication has 48 references indexed in Scilit:
- ML-Based Geophysical Well Log Prediction for Subsurface Energy Applications: A Stratified Analysis of Feature Sets and Model PerformancePure and Applied Geophysics, 2025
- Automatic Lithology Classification of Whole Core Images Using Multi-Input Convolutional Neural Networks: Integrating Visible Light and Ultraviolet PhotographyNatural Resources Research, 2025
- Rock Slope Stability Prediction: A Review of Machine Learning TechniquesGeotechnical and Geological Engineering, 2025
- Deep learning-aided simultaneous missing well log prediction in multiple stratigraphic units: a case study from the Bhogpara oil field, Upper Assam, Northeast IndiaEarth Science Informatics, 2024
- Missing log prediction using machine learning perspectives: A case study from upper Assam basinEarth Science Informatics, 2024
- Evaluation and Development of a Predictive Model for Geophysical Well Log Data Analysis and Reservoir Characterization: Machine Learning Applications to Lithology PredictionNatural Resources Research, 2022
- A review of tracer testing techniques in porous media specially attributed to the oil and gas industryJournal of Petroleum Exploration and Production Technology, 2022
- Contrasting machine learning regression algorithms used for the estimation of permeability from well log dataArabian Journal of Geosciences, 2021
- Prediction of sonic log and correlation of lithology by comparing geophysical well log data using machine learning principlesGeoJournal, 2021
- The Elements of Statistical Learning2009