A weighted machine learning ensemble for predicting shear slowness in the Tensleep Formation, SE Wyoming (USA)

Accurate determination of shear slowness (DTS) is essential for well placement optimisation, mechanical rock property estimation, and reservoir characterisation. However, direct DTS measurement is expensive, limited in availability, and lacking repeatability. Traditional empirical methods require extensive calibration and are only valid for specific rock types. This study presents the voting regressor (VR), an ensemble machine learning (ML) technique combining Extra Trees, Random Forest, Gradient Boosting, LightGBM, support vector regressor, and multi-layer perceptron to enhance DTS prediction. The model uses well logs including gamma ray (GR), bulk density (RHOB), porosity index (PHIX), and compressional slowness (DT), applied to the Tensleep Formation in the Teapot Dome field, Wyoming, USA. Each model was evaluated using the coefficient of determination (R2), mean absolute error (MAE), mean-squared error (MSE), and root mean-squared error (RMSE). A weighted averaging approach based on R2 performance was used to build the VR model, achieving an R2 of 0.96, RMSE of 0.19, MAE of 0.12, and MSE of 0.037. DT, PHIX, and RHOB were the most important features, while GR showed minimal impact. Validation on two unseen wells confirmed a strong generalisation of the VR (R2 = 0.91-0.92). This work highlights the potential of ML for accurate DTS prediction and improved subsurface characterisation in data-limited settings.