Deep Learning for Multichannel Log Prediction Tool using Drilling Data and Gamma-Ray Log

This abstract presents a deep learning workflow for multichannel synthetic log prediction using surface drilling data and gamma-ray logs in a complex carbonate reservoir. A two-stage architecture was developed: an initial ensemble of fully connected neural networks (FCNN) generated baseline predictions of compressional and shear sonic logs (DTCO and DTSM). A second-stage Bidirectional Long Short-Term Memory (BiLSTM) network with a Conv1D encoder and sequence-to-sequence (seq2seq) architecture was applied to enhance temporal consistency and fill measurement gaps. Additionally, an ahead-of-the-bit LSTM model was trained to forecast log trends for real-time geosteering support.

The approach was validated on 10 wells from two adjacent development zones using leave-one-out cross-validation. Results showed significant performance improvement over simpler ensemble models, particularly in vertical build-up sections where LWD data was sparse or missing. In blind test wells, the model achieved RMSEs of 6.88% for DTCO and 8.24% for DTSM with corresponding R² scores of 0.63 and 0.73. The model successfully captured both macro-trends and fine-scale variations. This workflow enables real-time subsurface characterization and supports operational decision-making in unconventional environments.