In this paper, we introduce a deep neural network method to solve the seismic inversion problem. After standard processing steps on acquired seismic data, the resulting low frequency seismic traces can be considered as filtered version of the earth’s high frequency reflectivity by convolving reflectivity with a wavelet. This wavelet is not a physical source. Instead, it is a filter that accounts for the difference between reflectivity and seismic trace. Seismic inversion involves removing the imprints of the wavelet in the seismic data and then converting the results into acoustic or elastic impedance. For this reason, seismic inversion often starts with estimating wavelet from wells where reflectivity is known. The quality of wavelet estimation directly impacts the quality of inversion products. In the latest machine learning approaches for seismic inversion, wavelet estimation is also required to generate pairs of synthetic reflectivity and seismic trace to train a neural network. Not surprisingly, predictions made by the network are sensitive to the wavelet used to generate training data. In field applications, wavelets often vary vertically and spatially, due to attenuation, change of geology and rock properties. This non-stationarity nature imposes additional challenge in estimating wavelet from seismic data and wells. To overcome all these obstacles related to wavelet, we propose a new training strategy to train deep neural networks (DNN) to solve the seismic inversion problem. The biggest advantage of our method is that it does not need wavelet estimation. As a result, it represents a true end-to-end machine learning approach to seismic inversion.
Presentation Date: Wednesday, September 18, 2019
Session Start Time: 1:50 PM
Presentation Start Time: 2:15 PM
Presentation Type: Oral