Deep Learning Applied to Seismic Data Interpolation

Summary Physical and economical constraints cause seismic data to suffer from the incompleteness on regular acquisition grids. The objective of this work is to explore the capabilities of Artificial Neural Networks to recover 3D seismic data where values are missing. The results indicate that it is possible to recover seismic data having only 20% of the measurements. The best reconstruction is achieved with a deep auto-encoder. The artificial neural network was trained and tested on 3D data patches extracted from the initial data cube. Significant improvement in the quality was gained while combining predictions for a single missing value from multiple overlapping patches. Computational speed is increased while exploiting high performance software i.e. OpenMP on multi-core CPU. To the best of our knowledge this is the first successful attempt to recover seismic signal with artificial neural networks.