A novel genetic inversion workflow based on spectral decomposition and convolutional neural networks for sand prediction in Xihu Sag of East China Sea

Marine seismic data often suffer from low resolution due to the challenges posed by high burial depth and wave interference during data acquisition. Seismic inversion serves as a crucial method to improve the resolution of seismic data and predict the sand bodies distribution. A complicated underground reservoir makes the conventional inversion method difficult to achieve. Some advanced inversion methods require substantial computational resources. This paper applied spectral decomposition and convolutional neural networks (CNN) in a genetic algorithm (GA) inversion. By utilizing convolutional neural networks (CNN), we can effectively learn and capture the spatial structure present in the data and establish a nonlinear relationship between these seismic attributes and the distribution of sand bodies. Integrating the CNN with a genetic algorithm (GA) allows us to achieve a high-resolution interpretation of sand bodies that aligns with the geological patterns at a fast computational speed. The application result shows that the predicted sand thickness has a high correlation with actual sand thickness at wells. A new horizontal well conforms to the prediction result at 94.1% accuracy (7534/ 8000 samples are predicted correctly).

Automated summary

This paper applied spectral decomposition and convolutional neural networks (CNN) in a genetic algorithm (GA) inversion to improve the resolution of marine seismic data and predict the distribution of sand bodies. By capturing the spatial structure present in the data, a high-resolution interpretation of sand bodies aligned with geological patterns was achieved.