Large scale seismic data interpolation with matrix completion

Seismic surveys amass large and incomplete data sets, and designing algorithms to interpolate the missing data at very large scales poses a daunting and critical challenge. We study how to apply scalable matrix completion methods to such interpolation problems. Recent studies in matrix completion have shown that a matrix that has low rank can be exactly completed when only a small number of observations are available. However, there are two challenges to applying matrix completion to seismic data. Matrix completion is typically applied to two dimensional or dyadic data whereas seismic data is often tensorial. Also successful matrix completion requires a low-rank matrix structure. We address these problems by organizing the seismic data on a matrix grid which exhibits a low-rank structure. This encoding allows us to apply the Jellyfish algorithm, developed at the University of Wisconsin, which achieves state-of-the-art performance for large-scale matrix completion. The proposed framework makes it possible to complete high-SNR interpolations of gigabytes of 4-D seismic data in minutes on standard multicore workstations. Our preliminary experimental results suggest that matrix completion provides a promising new approach to the seismic data interpolation problem.