An amalgamation of clustering and random forest algorithms for P-wave velocity log prediction: A Volve field case study
Vikram Kumar, Sayantan Ghosh · Interpretation · 2025
Abstract The P-wave velocity (VP) plays a crucial role in petroleum exploration and production. We focus on accurately predicting VP to identify porous hydrocarbon-bearing intervals for enhanced reservoir characterization and rock mechanical property calculations. The VP logs are available for a limited number of wells because acquisition is expensive. In this study, we predict VP by combining unsupervised K-means clustering with a supervised random forest (RF) machine-learning algorithm. This improved method of VP prediction provides results that are on par with other existing methods. Our approach demonstrates consistent performance in cross validation with strong predictive accuracy. Using conventional well logs and additional features obtained from feature engineering techniques offers an effective and reliable solution for VP prediction. We obtain four clusters from K-means using well-log data, wherein each cluster refers to a different rock type. Therefore, we choose four clusters. The log data within each cluster (rock type) are used for training RF models, with and without outliers. We perform exploratory data analysis for visualization and apply a one-class support vector machine to remove the outliers. We evaluate the model performance using R2, the mean-squared error, and the mean absolute error. Our cross-validation result indicates that R2 scores of each cluster are 0.87, 0.60, 0.88, and 0.87 for the training data set and 0.88, 0.55, 0.87, and 0.88 for the test data set, demonstrating the model’s robustness. The model trained without outliers achieved an R2 of 0.95 on the 30% test data set, compared to 0.94 with the outliers. Cross-validation results confirm the model’s stability, as it maintains consistent R2 performance across each fold, indicating its reliability across all clusters. Furthermore, our trained model, both with and without outliers, demonstrates robustness in predicting the blind well across all depths.