Interpolation through machine learning

Hongliang Zhang, Amr Ibrahim, Daniel O. Trad, Kristopher A. Innanen

Inspired by the image super-resolution problem, a CNN-based residual dense network (RdNet) is utilized to interpolate missing seismic traces within 2D synthetic seismic data. For the sake of comparison, interpolations are also implemented with a previously proposed residual network (ResNet) and a minimum weighted norm inversion (MWNI). As demonstrated by a series of synthetic experiments in this study, the contiguous memory mechanism, residual learning and feature fusion in both local and global levels enable RdNet to interpolate regularly missing traces with relatively high recovered S/N and accommodate spatial aliasing. In cases of randomly missing traces, RdNet produces comparable though slightly degraded results relative to the conventional minimum weighted norm inversion. Reliable results are obtained with less missing data, e.g., recovered S/N of ~40 dB and 30 dB for 10% and 30% randomly missing traces, respectively. As the missing-trace percentage increases, errors accrue in regions of the data with big gaps (typically larger than five consecutive traces). We expect that this will be improved by including more training data, which is currently being examined.