Prestack time migration velocity analysis using recurrent neural networks

Abstract

We present a new efficient method to perform prestack time migration velocity analysis (MVA) of seismic data based on recurrent neural nets (RNNs). We assume that there exits a mapping from each local data subvolume, which we term an analysis volume, to each single velocity point value in space. Under this hypothesis, as RNNs are capable of learning structural information both in time and in space, we propose designing a net that can learn this mapping, and exploit it to produce the entire root-mean-square (rms) velocity field. The performance of the method is evaluated via real data experimental results in comparison to existing technique for MVA. We present different aspects of the system's behavior with various training and testing scenarios and discuss potential advantages and disadvantages. We also recast the same RNN to learn and automate migration velocity picking from constant velocity migration (CVM) panels and compare real data results for this alternative. Our approach is extremely efficient in terms of computational complexity and running time, and therefore can be potentially applied to large volumes of three-dimensional (3D) seismic data, and significantly reduce work load. An extension to four-dimensional (4D) data is also possible.