Estimating shear wave velocity of geologically complex reservoirs using a hybrid machine learning approach, integrating core measurements and geophysical logs

Shear wave velocity (Vs) is a fundamental rock property for reservoir characterization, geomechanical analysis, and seismic interpretation in hydrocarbon exploration. However, direct Vs measurements are often impeded by high costs, technical challenges, and limitations of empirical rock-physics models in heterogeneous tight gas reservoirs. Therefore, this research presents an innovative hybrid Group Method of Data Handling (GMDH) with differential evolution (DE) (GMDH-DE) to address these challenges. A total of 6253 well-log datasets from six wells, each with seven input parameters, were used to accurately predict Vs in the S6 block of the Sulige gas field, China. The GMDH-DE algorithm was selected for its ability to resolve complex nonlinear relationships between variables, optimize parameters via evolutionary strategies, and decrease computational time. The results show that the GMDH-DE model demonstrated superior performance to the GMDH, Convolutional Neural Network (CNN), and Decision Tree (DT) models, achieving high precision with a coefficient of determination R2 of 0.9995, low root mean square error RMSE: 0.02259, and mean absolute error MAE: 0.0115 during training. During testing, it achieved an R2 of 0.9969, RMSE: 0.0301, and MAE: 0.0211. The GMDH-DE reduced computational processing time by 59.56% for GMDH and DT.Additionally, Shapley additive explanations (SHAP) analysis highlighted acoustic log (AC), deep resistivity (RLLD), and porosity (POR) as the most influential input parameters for Vs prediction. The GMDH-DE was further tested on well 7 from the same basin, with no target data, and successfully predicted Vs with high accuracy. The GMDH-DE framework enables accurate Vs estimation, enhancing reservoir characterization, mitigating geomechanical risk assessment, and optimizing hydrocarbon recovery in geologically complex reservoirs worldwide.