Stacking velocity estimation using recurrent neural network
Reetam Biswas, Anthony Vassiliou, Rodney Stromberg, Mrinal Kanti Sen · 2018
We describe a new method based on the Machine Learning (ML) technique for normal moveout correction (NMO) and estimation of stacking velocity. A Recurrent Neural Network (RNN) is used to calculate stacking velocity directly from the seismic data. Finally, this velocity is used for NMO correction of the data. We used the Adam optimization algorithm to train the network of neurons to estimate stacking velocity for a batch of seismic gathers. This velocity is then compared with the correct stacking velocity to update the weight. The training method minimizes a cost function defined as the mean squared error between the estimated and the correct velocities. The trained network is then used to estimate stacking velocity for rest of the gathers. Here we illustrate out method on a noisy real data set from Poland. We first trained the network using only 18 percent of gathers and then used the network to calculate stacking velocity for the remaining gathers. We used these stacking velocity to perform Normal moveout correction and finally we stacked to get the post-stack seismic section. We also show comparison between the stacks generated from the two velocities Presentation Date: Wednesday, October 17, 2018 Start Time: 9:20:00 AM Location: Poster Station 1 Presentation Type: Poster