Remote Connection via ReadyTalk
Access Code: 4866455
About the Speaker: Ben Holtzman, Geophysicist, Lamont Doherty Earth Observatory
Ben Holtzman is a geophysicist at the Lamont Doherty Earth Observatory in New York. His research involves experimental, theoretical and observational studies of upper mantle rheology, with an emphasis on partially molten rocks and quantitative interpretation of seismic measurements, with recent migrating into geothermal energy problems. (For more information: Seismic Sound Lab page and Columbia University page)
Machine learning methods applied to about 46,000 small earthquakes (0.5<M<1.5) over three years in the Geysers Geothermal Field, CA, reveal annual cyclic changes in the source properties that closely track the fluid injection history of the reservoir. Our methods treat seismic signals as spectrograms, inspired by approaches in “Machine Listening”, with the intent of isolating spectral variations associated with changes in faulting process. The approach is empirical: unsupervised learning methods require that no information is provided to the algorithm to guide feature detection or to seed the clustering of earthquakes based on their spectral-temporal content. Other authors have found changes in moment tensor associated with fluid injection rate (or total volume) in the NW Geysers. Our methods identify similar temporal patterns in events with slightly different spectral properties, but an order of magnitude smaller than in previous studies and extending over the entire Geysers Field. Based on previous studies, we interpret these patterns to reflect changes in faulting process, due to interplay between thermal stresses and poroelastic stresses varying with time. This general approach opens new paths to near real time identification of spatio-temporal patterns in microseismicity as indicators of changing thermo-mechanical conditions, for both reservoir monitoring and earthquake forecasting.
Host: Kurt Nihei