Optimizing downdip interface prediction by using seismic features to train a neural network

Advancements in seismic imaging using neural networks have significantly improved geological characterization. Fully convolutional networks (FCNs), particularly UNet-based architectures, have demonstrated high accuracy in predicting acoustic velocity models from raw seismic data. However, traditional approaches often overlook critical spectral features, such as wavelet positioning and phase rotation, which are essential for refining seismic predictions. A spectral recomposition approach for estimating the temporal position of wavelets in a seismogram is employed so that this information can be incorporated to train a FCN to predict velocity models of geological structures. In addition, critical angle information to enhance the training dataset is used with the objective of predicting downdip layers more accurately. Both critical angle information and wavelet positions in time are incorporated in the training process for the neural network as additional features. As the critical angle is barely reached in seismic acquisition on deeper structures, reflections in downdip layers present a wider reflection angle, which might be identified as a sharper critical angle by some techniques. This information, treated as a feature for increasing information to train a FCN, has been shown to be capable of predicting acoustic velocity models more accurately, especially for the shape of downdip structures.