A Hybrid Deep-Learning Method for Accurate Well Log Reconstruction in Complex Formations

Abstract

Reconstructing missing or distorted well logs in geophysical well logging is crucial for accurate reservoir evaluation, particularly in tight sandstone formations like those in the Ordos Basin. Traditional methods often falter due to nonlinear couplings and noise, resulting in unreliable predictions. This study presents a hybrid deep-learning framework that integrates particle swarm optimization (PSO) for hyperparameter tuning, Inception modules for multiscale feature extraction, bidirectional long short-term memory (BiLSTM) networks for capturing bidirectional spatiotemporal dependencies, and efficient channel attention (ECA) for adaptive feature weighting. The model was trained and validated using well logging data from 10 exploration wells in the Ordos Basin, with inputs including true formation resistivity (RT), natural gamma ray (GR), and acoustic time (AC). The proposed model enables accurate density log reconstruction, showing significant improvements and low errors in derived reservoir parameters when validated against core data. This approach offers a cost-effective solution for reconstructing incomplete logs.