The distribution of reservoir facies is a crucial factor for geoscientists to consider when determining drilling locations. However, due to the complexity of the earth’s subsurface, gathering the facies information is often a very challenging task.

We propose to develop a Convolutional Neural Networks (CNN) model that can accurately map the facies distribution using only the 3D Post-Stack seismic data. We modified CNN with U-Net architecture for mapping the facies on all seismic points. We have improved the model by implementing transfer learning that allows to use many learned parameters from previous model.

The display of a) input from seismic data, b) the true facies distribution, c) the seismic overlay with predicted facies distribution, and d) predicted facies distribution. The model give a highly-confident prediction or detection of the facies distribution compare with the original one. However, we note that there is an absence of facies distribution on the small and noisy part, indicate with blue circle

Skills: Python, Numpy, Pandas, scikit-learn, Keras, Scipy