Reservoir characterization reimagined: a hybrid neural network approach for direct three-dimensional petrophysical property characterization

Abstract

Reservoir characterization, crucial for oilfield development, aims to unravel intricate non-linear relationships within real-world data. Conventional methods, rooted in simplistic theories, often lead to uncertainties and inaccuracies in workflows. Leveraging the power of deep learning, this study introduces a pioneering approach: a hybrid neural network model merging convolutional and Long Short-Term Memory (LSTM) RNN layers. Focused on effective porosity modeling for the Ghar Member of the Asmari Formation in western Iran, the study utilizes post-stack seismic data and well-log information. By effectively deciphering spatio-temporal information within the data, our methodology allows for spatially aware predictions of effective porosity values, a capability not addressed by previous studies. The hybrid neural network model predicts effective porosity values for the entire reservoir, creating a 3D grid of porosity. It leverages CNN and RNN layers to decipher spatio-temporal information within the data, thereby enabling the model to make spatially aware predictions. The model achieved a mean squared error (MSE) of 0.005, generating clear 3D porosity models with greater detail compared to traditional machine learning and geostatistical methods. This innovative methodology represents a step forward in reservoir characterization, offering improved precision and efficiency. It holds promise for advancing oilfield development practices in the future.