Fully Reversible Neural Networks for Large-Scale 3D Seismic Horizon Tracking

Summary Neural networks are a successful tool for horizon tracking for the interpretation of seismic images. So far, most research works with 2D inputs. In 3D, most work is restricted to relatively small 3D inputs because of memory limitations. Training a neural network typically requires the storage of the network states for every layer. This becomes a problem for deep networks and large data inputs. We avoid this problem by employing recently introduced fully reversible convolutional networks that require storage of the network states for a few layers only. Therefore, we can use much larger 3D input data than in the case of non-reversible networks. A field data example illustrates that fully reversible neural networks are suitable for horizon tracking and allow the input size to increase by a least an order of magnitude, such that we can also learn from structures with a larger length scale.