The seismic horizon is a critical input for the structure and stratigraphy modeling of reservoirs. It is extremely hard to automatically obtain an accurate horizon interpretation for seismic data in which the lateral continuity of reflections is interrupted by faults and unconformities. The process of seismic horizon interpretation can be viewed as segmenting the seismic traces into different parts and each part is a unique object. Thus, we have considered the horizon interpretation as an object detection problem. We use the encoder-decoder convolutional neural network (CNN) to detect the “objects” contained in the seismic traces. The boundary of the objects is regarded as the horizons. The training data are the seismic traces located on a user-defined coarse grid. We give a unique training label to the time window of seismic traces bounded by two manually picked horizons. To efficiently learn the waveform pattern that is bounded by two adjacent horizons, we use variable sizes for the convolution filters, which is different than current CNN-based image segmentation methods. Two field data examples demonstrate that our method is capable of producing accurate horizons across the fault surface and near the unconformity which is beyond the current capability of horizon picking method.

Faults in the subsurface can be an avenue of, or a barrier to, hydrocarbon flow and pressure communication. Manual interpretation of discontinuities on 3D seismic amplitude volume is the most common way to define faults within a reservoir. Unfortunately, 3D seismic fault interpretation can be a time-consuming and tedious task. Seismic attributes such as coherence help define faults, but suffer from “staircase” artifacts and nonfault-related stratigraphic discontinuities. We assume that each sample of the seismic data is located at a potential fault plane. The hypothesized fault divides the seismic data centered at the analysis sample into two subwindows. We then compute the coherence for the two subwindows and full analysis window. We repeat the process by rotating the hypothesized fault plane along a set of user-defined discrete fault dip and azimuth. We obtain almost the same coherence values for the subwindows and the full window if the analysis point is not located at a fault plane. The “best” fault plane results in maximum coherence for the subwindows and minimum coherence for the full window if the analysis point is located at a fault plane. To improve the continuity of the fault attributes, we finally smooth the fault probability attribute along the estimated fault plane. We illustrate the effectiveness of our workflow by applying it to a synthetic and two real seismic data. The results indicate that our workflow successfully produces a continuous fault attribute without staircase artifacts and stratigraphic discontinuities.

Predicting elastic parameters based on digital rock images is an interesting application of a convolutional neural network (CNN), which can improve the efficiency of prediction. Predicting elastic parameters by a conventional CNN, which is used for image classification such as LeNet and AlexNet, lacks geophysical constraints, and its accuracy in predicting elastic parameters is poor, with limited training data available. The combination of a U-Net and a convolutional neural network (CUCNN) is proposed to predict the elastic parameters from digital rock images with limited training data. In CUCNN, the rock matrix and pore types segmented from gray-scale images are treated as constraints that induce the convolutional kernels to extract the global as well as the local-scale rock features. The loss function, designed in a composite form to accelerate the convergence speed, contains the segmentation error and elastic parameters predicted from the gray-scale images. By adding geophysical constraints to the CNN, an implicit representation from the gray-scale image to the elastic parameters can be gained, which can improve the accuracy and efficiency of parameter prediction. Our method was tested using training and verification data derived from 1800 2D image slices of Berea sandstone samples, and the results were compared against the CNN model. The VP and VS were calculated by the finite-element method as the control to test the performance of both models. Our results show that CUCNN’s R2 score is 0.84, which increased by as much as 0.21 compared to the conventional CNN.