Neural Network Simplex Encoders to Monitor Fracture Processes using Transportable Acoustic Sources
David D. Nolte
Purdue University, USA

Biosketch
David D. Nolte is the Edward M Purcell Distinguished Professor of Physics and Astronomy at Purdue University performing research in the fields of optical interferometry and holography. He received his baccalaureate from Cornell University in 1981, his PhD from the University of California at Berkeley in 1988 and was a post-doctoral member of AT&T Bell Labs before joining the physics faculty at Purdue. He has been elected Fellow of the Optical Society of America, Fellow of the American Physical Society and Fellow of the AAAS. In 2005 he received the Herbert Newby McCoy Award of Purdue University. He has founded two biotech startup companies in the area of medical diagnostics. David is the author of four books, and he blogs regularly on topics of nonlinear physics at https://galileo-unbound.blog/
Introduction of the Lecture
Transportable acoustic sources, known as chattering dust, are a novel geophysical tool to study laboratory-based fracture properties and processes. These sources mimic natural acoustic emission with the added advantage that they can transport through fractures and fracture networks. A challenge posed by the chattering dust is the uncontrolled character of the acoustic signals that masks arrival-time and amplitude information, which are usually the two most valuable information channels for seismic characterization of fractures. Fortunately, late-arriving coda waves caused by multiple scattering in fracture networks provide subtle information channels that can be extracted using machine learning to identify fracture processes.
This talk will describe the design and performance of a general class of neural networks called simplex encoders. These encoders reduce the dimensionality of the dataset and accurately classify fracture saturation properties deduced from the chattering dust data. One type of simplex encoder, the so-called Siamese neural network, has gained popularity for its ability to discern subtle differences among different classes of behavior. The classification accuracy of the Siamese encoder applied to the fracture data will be related to more general simplex encoders. Simplex encoders are a potentially powerful tool for the study of coupled processes in the earth sciences and engineering.