Prediction of Sonic Well Logs Using Deep Neural Network: Application to Petroleum Reservoir Characterization in Mexico

The sonic log is a key tool for assessing the mechanical properties of rocks, identifying structural features, calibrating seismic data, and monitoring well integrity. However, sonic data are often incomplete due to time and cost constraints, tool failures, or unreliable measurements. Traditional approaches to generate synthetic sonic logs usually rely on empirical relationships or statistical methods. In this study, we applied an artificial intelligence approach in which a deep neural network was trained with real data from an oilfield in Mexico to reconstruct sonic logs based on their relationships with other geophysical well logs. Three models, each using different input logs, were trained to predict the sonic response. The models were validated on wells excluded from training, and performance was evaluated using the root mean square error (RMSE) and mean absolute percentage error (MAPE), showing satisfactory accuracy. The models achieved RMSE values between 1.4 and 1.7 [μs/ft] and MAPE values between 2.1 and 2.6% on independent test wells, confirming robust predictive performance. We also generated synthetic sonic logs for wells where no sonic data were originally acquired, demonstrating the practical value of the proposed method. This work integrates convolutional (CNN) and recurrent (GRU) layers in a single deep-learning architecture, trained under strict well-level validation. The workflow is demonstrated on wells from the Tabasco Basin, representing a field-scale deployment not previously reported in similar studies.