In this work, we solve the seismic inversion problem of obtaining an elastic model of the subsurface from recorded seismic data using a convolutional neural network (CNN). For simplicity we consider a 1D layered earth model and normal incidence seismic data. We systematically test the robustness of the network in predicting P-impedance (Ip) of new, previously unobserved, earth models when the input to the network consisted of seismograms generated with (1) different source wavelets; (2) earth models that had different geostatistical spatial correlations; and (3) earth models that had different underlying rock physics relation than that in the training data. Results show that the CNN successfully predicts impedances generated with both variograms ranges on which it was trained and variogram ranges on which it was not trained. The CNN was able to predict with medium success samples generated with rock physics model parameters and source wavelet phase outside of the training range. The CNN was not able to predict either the training set or any of the testing sets in the presence of various source wavelet frequencies, showing the importance of knowing a-priori the value of the wavelet frequency when generating the synthetic seismic data. Overall, the CNN has shown great promise in predicting a high frequency impedance model from a low frequency seismic signal, given appropriate training data.
Presentation Date: Wednesday, October 17, 2018
Start Time: 1:50:00 PM
Location: 204B (Anaheim Convention Center)
Presentation Type: Oral
