Shale anisotropy model building based on deep neural networks

Shale anisotropy is essential for processing seismic data with long offsets, wide azimuths and multiple components. However, direct measurements of shale anisotropy are not available in routine downhole measurements. We propose to train a deep neural network to estimate the nonlinear relationship between well logs and anisotropy parameters. Since it is impossible to collect massive labelled field data, we generate paired synthetic data of features (e.g., velocities) and labels (mainly elastic constants) based on the Hudson-Cheng’s model assuming crack-induced anisotropy. By tuning hyperparameters such as the number of hidden layers, the number of nodes in each layer, etc., we obtain a fully connected neural network with 5 hidden layers that fits well with the synthetics. The neural network is applied to published laboratory measurements and the field data of a China well and a US well. We show that the estimated anisotropy using the neural network agrees reasonably well with the published laboratory measurements, and is consistent with the mineralogy log for the China well. Moreover, the predictions of the neural network coincide well with the traditional inversion results for both field data with no modifications to the neural network. Therefore, the neural network is reliable and generally applicable to shales. Lastly, an unsupervised learning algorithm – Gaussian mixture model is applied to the selected outputs of the deep neural network, and is demonstrated to be a potential method for lithological classification.