Updates on the campus response to coronavirus (COVID-19)

Ph.D. Dissertation Defense - Motaz Alfarraj

Event Details

Thursday, November 7, 2019

9:30am - 11:30am

Room 5234, Centergy

For More Information


Event Details

TitleLearning from Seismic Data to Characterize Subsurface Volumes


Dr. Ghassan AlRegib, ECE, Chair , Advisor

Dr. James McClellan, ECE

Dr. Zhigang Peng, EAS

Dr. David Anderson, ECE

Dr. Ying Zhang, ECE


The exponential growth of collected data from seismic surveys makes it impossible for interpreters to manually inspect, analyze and annotate all collected data. Deep learning has proved to be a potential mechanism to overcome big data problems in various computer vision tasks such as image classification, image captioning, and semantic segmentation. However, the applications of deep learning are limited in the field of subsurface volume characterization due to the limited availability of consistently-annotated seismic datasets. Obtaining annotations of seismic data is a labor-intensive process that requires field knowledge. Moreover, since seismic data are indirect low-resolution measurements of the subsurface, seismic interpreters rely on the few direct high-resolution measurements of the subsurface from well-logs and core data to confirm their interpretations. Different interpreters might arrive at different valid interpretations of the subsurface, all of which are in agreement with well-logs and core data. Therefore, to successfully utilize deep learning for subsurface characterization, one must address and circumvent the lack or shortage of consistent annotated data. In this dissertation, we introduce a learning-based physics-guided subsurface a volume characterization framework that can learn from limited inconsistently-annotated data. The introduced framework integrates seismic data and the limited well-log data to characterize the subsurface at a higher-than-seismic resolution. The introduced framework takes into account the physics that governs seismic data to overcome noise and artifacts that are often present in the data. Integrating a physical model in deep-learning frameworks improves their generalization ability beyond the training data. Furthermore, the physical model enables deep networks to learn from unlabeled data, in addition to a few annotated examples, in a semi-supervised learning scheme. Applications of the introduced framework are not limited to subsurface volume characterization, it can be extended to other domains in which data represent a physical phenomenon and annotated data is limited.

Last revised October 29, 2019