Elastic Pre-stack Seismic Inversion in Stratified Media Using Machine Learning

The similarity between seismic inversion and machine learning lies in the fact that both are essentially optimization problems. This study describes a workflow in which the pre-stack seismic inversion is framed as a deep learning problem. The earth medium of interest is elastic with weak lateral variations. A convolutional neural network (CNN) is trained and tested on a set of synthetic earth models and seismic records. The trained network is then used to make predictions of subsurface elastic parameters directly from field data from a land survey. The predicted elastic parameters, even in the absence of low frequency content in the input seismic data, show a good match with the well logs. The P-wave velocity model appears to have better resolution and accuracy than the result derived from travel-time tomography. Adequate predictions from the CNN is achieved by the careful construction of training samples and the conditioning of seismic data so that the synthetic data used for training approximately represent the statistics of the geology and field data. The computational efficiency of using deep learning in such a way for waveform inversion may be desirable for elastic model building.