Reconstruction of Missing Sonic Logs Using a Hybrid Deep-Learning Model with Bayesian Optimization

Transport in Porous Media · Springer Nature · 2026-09-23 · DOI: 10.1007/s11004-026-10344-3

Аннотация

Well logs play a fundamental role in reservoir identification and lithological interpretation. As a critical bridge linking well-log measurements to seismic responses, sonic logs are essential for log-constrained seismic inversions. However, sonic log data frequently suffer from incompleteness or missing intervals owing to instrumental limitations and complex geological conditions, which limit reliable seismic inversion and reservoir characterization. The re-acquisition of logging data is cost-prohibitive and operationally challenging, which significantly hinders subsequent seismic interpretation. To address this issue, a convolutional neural network (CNN)-feature-wise linear modulation (FiLM)-bidirectional long short-term memory (BiLSTM) framework incorporating Bayesian optimization (BO) is proposed to reconstruct sonic logs from natural gamma-ray (GR) and resistivity (RES) logs. In this framework, a CNN extracts local depthwise features, and a FiLM mechanism integrates lithological information via conditional feature modulation. A BiLSTM network is further incorporated to capture long-range sequential dependencies, enabling the reconstruction of missing sonic log intervals. In addition, BO is used to optimize the weights assigned to the individual terms in the composite loss function. Experiments were conducted using data from 10 wells in the Tongliao region, with six wells used for training and validation, and four for cross-well prediction. The results demonstrate that the proposed method consistently outperforms multiple benchmark models, including random forest (RF), extreme gradient boosting (XGBoost), temporal convolutional network (TCN), time-series transformer (TST), BiLSTM, and CNN-BiLSTM. On the normalized scale, the model achieved a root mean squared error (RMSE) of 0.0536, a mean absolute error (MAE) of 0.0390, and a coefficient of determination ($$R^2$$) of 0.8785. Furthermore, applying the reconstructed sonic logs to the log-constrained seismic inversion improved the continuity and fine-scale structural details of the inversion sections, thereby providing a valuable reference for seismic inversion under incomplete well-log conditions.

Читать у издателя →

Похожие статьи