We present a framework for efficiently performing 3D seismic inversion of reservoir facies under geologically realistic geostatistical models of prior uncertainty. Our approach is based on directly learning the inverse mapping between 3D seismic data and reservoir models using 3D convolutional neural networks. We generate training dataset for the learning problem by simulating facies from object-based geostatistical model and forward modeling seismic data. We present a method for performing falsification of prior geologic uncertainty with seismic data, which is critical to ensure reliability during prediction with real data. We demonstrate the efficacy of our approach by successfully inverting a large-scale model of a real deltaic reservoir from 3D post and partial stack seismic data.
Presentation Date: Tuesday, October 13, 2020
Session Start Time: 1:50 PM
Presentation Time: 1:50 PM
Location: 351F
Presentation Type: Oral