Geology-Constrained Time Series Generative Adversarial Network for Well Log Curve Reconstruction

Complex geological and downhole engineering conditions—such as borehole enlargement, fracture development, and mud invasion—often induce anomalous logging responses, leading to missing key curves and compromising reservoir evaluation accuracy. Traditional interpolation and statistical methods struggle to capture the non-stationarity and strong nonlinearity of log curves; conventional models often neglect sequential dependencies along depth, while deep sequence models are limited to point-by-point regression, restricting their ability to maintain overall geological consistency. To address these challenges, this study proposes a Geology-Constrained Time Series Conditional Generative Adversarial Network (GC-TSGAN). Lithological information is embedded as prior conditions into both the generator and discriminator. The model leverages LSTM to capture sequential dependencies along depth, while an LSGAN-based adversarial loss enforces distributional consistency and local morphological fidelity. Random search and Bayesian optimization are applied for efficient hyperparameter tuning. Experiments on logging data from 41 wells in the B Basin, Chad, show that GC-TSGAN outperforms baseline models including RF, XGBoost, LSTM, and ANN across RMSE, MAE, and R². Results confirm that the proposed model achieves high-precision log curve reconstruction under complex geological conditions, providing a reliable data foundation for geological modeling and reservoir evaluation.