Reconstructing Missing Acoustic Log with Multilevel Wavelet Decomposition and Gated Recurrent Unit Networks

Summary Acoustic (AC) logging data is essential for reservoir characterization. However, data gaps due to equipment failure or borehole instability hinder reliable interpretation. To overcome these challenges, we propose a novel multilevel wavelet decomposition network (mWDN) integrated with a bidirectional gated recurrent unit (BiGRU) (mWDN-BiGRU) to reconstruct missing AC well logs. By leveraging wavelet-based multiresolution analysis, the model effectively isolates high- and low-frequency components, and, when combined with the gated recurrent unit’s (GRU’s) capacity to capture temporal dependencies, it delivers superior reconstruction fidelity. Comparative evaluations on logging data from 100 wells in the Shengli Oil Field against baseline models—extreme gradient boosting (XGBoost), convolutional long short-term memory (ConvLSTM), a weight-dropped long short-term memory (AWD-LSTM), BiGRU, and multi-head Atte attention with BiGRU (MHA-BiGRU)—demonstrate that mWDN-BiGRU consistently achieves lower mean absolute errors (MAEs) and exhibits robust generalization across training, validation, and test data sets. In well-specific case studies, including challenging conditions in water-producing, reservoir, and thick mudstone sections, mWDN-BiGRU not only registers the highest Pearson correlations but also provides sharp, stable predictions with only a modest increase in training time relative to BiGRU. Furthermore, we introduce a sequence-to-sequence uncertainty quantification method along with feature importance analysis, which underscores the pivotal roles of logs, such as compensated neutron log (CNL) and gamma ray (GR), under varying geological conditions. Overall, the proposed method offers a powerful, efficient tool for reconstructing missing well logs, thereby enhancing subsurface characterization and supporting more informed decision-making in advanced reservoir management.