Direct estimation of effective horizontal and vertical stress from acoustic wave velocities of shales using machine learning

Accurate estimation of in-situ effective stresses is essential for geomechanical modelling, wellbore stability, hydraulic fracturing design, and carbon storage integrity. Conventional empirical and numerical models frequently encounter nonlinear and anisotropic stress behaviour in shale formations, resulting in significant estimation errors. This study analyzed and compared Adaptive Boosting (AdaBoost) machine learning (ML) model against four baseline models; Gradient Boosting(GB), Artificial Neural Network (ANN), Random Forest (RF), Support Vector machine(SVM) and 3 classical empirical methods (Linear, polynomial and dynamic -modulus poroelastic models) in estimating directly effective minimum horizontal stress (σh′) and effective vertical stress (σv′) from compressional (V p ) and shear (V s ) wave velocities, thereby eliminating the need for empirical transformation equations. Model evaluation was accomplished using evaluated R, R², RMSE, MAE, MAPE, and MAD. A curated dataset of 2,636 high-quality shale core test samples from the Northern Lights CCS well (31/5-7 EOS) was used for model training (70%) and testing (30%). Sample-size sensitivity was further examined across four additional train/test partitions (75/25, 80/20, 85/15, and 90/10). All models except SVM achieved R² > 0.99 for both stress targets. AdaBoost outperformed all models for σ h ′ prediction, achieving a test R² of 0.9984, RMSE of 0.071 MPa, and MAPE of 0.35%. For σ v ′, GB and ANN jointly led with R² = 0.9998, whilst AdaBoost returned a competitive R² of 0.9995. . Sample-size learning curves confirmed GB as the most stable and data-efficient model, whilst AdaBoost showed non-monotonic sensitivity to training data composition. All empirical benchmarks substantially underperformed relative to the ML models, with the best empirical result (polynomial regression, R² = 0.989 for σ v ′) failing to match any ensemble model. These findings establish that AdaBoost is a viable and competitive model for direct in-situ stress estimation from acoustic velocities in shale formations.