Density Logging Curve Reconstruction Method Based on Bayesian-Optimized CNN-LSTM

Abstract

Density logging curves are critical evaluation indicators in oil and gas exploration and development, with their accuracy directly affecting the reliability of reservoir lithology identification, porosity calculation, and fluid property analysis. During logging operations, instrument malfunctions or borehole enlargement often lead to missing or distorted density logging data, compromising reservoir evaluation accuracy. Traditional reconstruction methods, constrained by insufficient model representation capabilities, fail to effectively capture relationships between logging curves under complex geological conditions, resulting in reconstructed curve accuracy that struggles to meet reservoir evaluation requirements. To address this issue, this study proposes a density logging curve reconstruction method based on Bayesian-optimized convolutional neural networks (CNN) combined with long short-term memory networks (LSTM). This approach integrates CNN's advantages in local spatial feature extraction with LSTM's capability in temporal dependency modeling, while introducing Bayesian optimization to automatically search for optimal hyperparameters. Through this integration, the model can fully leverage the strengths of both components when processing logging data under complex geological conditions, thereby enhancing overall performance. Applied to density logging curve reconstruction in three exploration wells in the Sichuan Basin, experimental results demonstrate that compared to standalone CNN or LSTM models, the Bayesian-optimized CNN-LSTM model exhibits superior performance in both accuracy and stability, significantly mitigating the impact of logging curve distortion or data loss on reservoir evaluation.