Missing Well-Log Data Prediction Using a Hybrid U-Net and LSTM Network Model

Well logs play a crucial role in geological exploration and the development of oil and gas resources. However, in practice, well logs are often missing or incomplete. Since re-logging is often impractical given how well logs are obtained, they are usually estimated from existing well logs. The complexity and variability of underground structures, along with the highly nonlinear relationships among different log types, make it challenging for empirical regression and traditional machine-learning methods to deliver accurate predictions. Existing methods have difficulty in extracting and integrating the spatial, sequential, and multiscale dependencies among known well logs to predict unavailable well logs. In this study, we developed a hybrid U-Net and long short-term memory (LSTM) model that combines U-Net’s ability to extract multiscale spatial features with LSTM’s strength in modeling sequential depth-wise trends. The architecture uses convolutional blocks and skip connections to extract and map the multiscale spatial features of logging curves, while the LSTM component models the trend changes of well-log measurements with depth. We evaluated this model alongside the state-of-the-art hybrid convolutional neural network (CNN) and LSTM model (CNN-LSTM) on well-log data from the Force 2020 machine-learning competition to predict missing compressional slowness (DTC) logs. Results from training and testing experiments in eight blind wells show that the U-Net-LSTM outperforms the benchmark CNN-LSTM, achieving lower mean absolute error (MAE), root mean square error (RMSE), and higher correlation coefficients R. This demonstrates the potential of the proposed model as an effective tool for reliable and geologically consistent missing well-log reconstruction.