Inverse distance weighted Random Forest for S-wave velocity prediction

Abstract

S-wave velocity (Vs) serves as a crucial link between rock physics and seismic exploration. It plays an essential role in unveiling the complexity of subsurface structures and their dynamic behavior. It plays an essential role in unveiling the complexity of subsurface structures and their dynamic behavior. It is essential for revealing the complexity of underground structures and their dynamic characteristics. However, existing artificial intelligence methods typically build Vs prediction models using common sensitive parameters from multiple wells, which overlook the differences of sensitive parameters among wells and lack spatial constraints among wells. These limitations reduce the prediction performance of artificial intelligence models. Additionally, the necessity to select common logging parameters as input may result in the loss of sensitive logging parameters specific to certain wells. To address these issues, we propose an inverse distance weighted Random Forest method for Vs prediction. In the proposed method, decision trees are first employed to rank the logging parameters of each training well based on their sensitivity to Vs, acknowledging that sensitive parameters may vary between wells. Next, individual Vs prediction models are trained using the top-ranked sensitive parameters for each well. Finally, by considering the spatial distance and geological similarity between test and training wells, comprehensive estimates for the test wells are obtained through inverse distance weighted integration of predictions from multiple single-well models. We use three wells as training data and select one blind well in each working area for testing. Datasets from two different working areas demonstrate that the proposed method is effective for Vs prediction. The performance is quantified by comparing the predicted Vs with the true Vs using mean absolute percentage error and coefficient of determination. Compared to support vector regression, extreme gradient boosting, and Random Forest, the testing results show that the proposed method has the highest accuracy and stronger robustness in predicting Vs in blind wells.