A Missing Well-Logs Generation Method Based on Generative Adversarial Networks Models

Well logs are important for reservoir evaluation and geological analysis in the petroleum industry. During the logging process, equipment malfunctions, harsh drilling conditions, operational limitations, and other factors may lead to missing of logging data. Due to the inherent complexity of geological conditions and the limitations of logging data, it is often very difficult to complete logging curves using traditional technology. The paper focuses on developing a method for generating missing well logs, which may enhance the completeness and accuracy of logging data. Generative Adversarial Networks (GANs) have shown remarkable success as a framework for training models to produce realistic-looking data. The Gated Recurrent Unit (GRU) network processes sequence data through cycles of time steps. Unlike traditional feedforward neural networks, GRU uses recurrent layers to capture long-term dependencies in sequences. Therefore, GRU is more suitable for processing time series data. Well logging curves can be considered as continuous time series data. A method for generating missing well logs is proposed, based on the combination of the GRU and GANs. This method can learn the potential patterns of logging curves from incomplete data based on the deep learning models. Through adversarial training, the model's generator and discriminator mutually enhance each other. Well logs will be generated once the generator has progressively learned the complex distribution characteristics. The predictive accuracy of the model will increase through feedback iteration because the discriminator assesses the authenticity of the generated data. The results of the study show that well logs generated by the GAN have high accuracy compared to traditional machine learning methods, even in complex geological environments. The generated logging data matches well with the real data. For the GR curve, the correlation between the generated data and the actual logging curve reached 0.90. Qualitatively and quantitatively, we find that the generative method has significant advantages in generating well logs. This research highlights the potential of deep generative models in petroleum engineering applications.