*********************************
There is now a CONTENT FREEZE for Mercury while we switch to a new platform. It began on Friday, March 10 at 6pm and will end on Wednesday, March 15 at noon. No new content can be created during this time, but all material in the system as of the beginning of the freeze will be migrated to the new platform, including users and groups. Functionally the new site is identical to the old one. webteam@gatech.edu
*********************************
Title: Learning from Seismic Data to Characterize Subsurface Volumes
Committee:
Dr. AlRegib, Advisor
Dr. McClellan, Chair
Dr. Peng
Abstract:
The objective of the proposed research is to develop a framework to characterize subsurface volumes. Despite the success of deep learning in various vision tasks on natural images, its applications in the seismic domain are limited due to the small amount of available labeled data. Furthermore, seismic data lack color and edge information, which are essential to characterize natural images. Therefore, we investigate alternative approaches to characterize seismic data such as texture analysis and sequence modeling depending on the type of labels. For instance, with image-level annotations of subsurface structures, we utilize texture-based similarity measures to capture the subsurface structures. Such similarity measures enables the development of a large labeled dataset of subsurface structure images a few labeled examples. Such dataset can then be used to train a learning-based structure characterization framework. However, image-level labels are not suitable for fine-grained characterization of subsurface volumes due to their limited resolution. Therefore, we investigate the use of the well-log data as labels for seismic data. Well-logs are integrated into a sequence modeling framework that captures the temporal dynamics of seismic traces. The sequence modeling framework models seismic traces as time series, and utilizes recent advances in recurrent neural networks to find a suitable non-linear mapping from seismic to rock property that can generalize beyond well locations. In addition, we propose a novel framework that uses geophysical and learned forward models to enable the neural networks to learn from unlabeled pre-stack and post-stack data in semi-supervised learning scheme.
