[M-IS04] Interdisciplinary studies on pre-earthquake processes

[MIS04-07] Nowcasting Earthquakes: Machine Learning and the Information Content of Earthquake Catalogs

★Invited Papers

*John B Rundle¹ (1.University of California, Davis, CA USA)

キーワード：Earthquake Nowcasting, Machine Learning, Shannon Information, Earthquake Catalogs

In a famous paper, Gardner and Knopoff (1974) found that progressively removing aftershocks, (correlated seismicity) from earthquake catalogs resulted in Poisson interval statistics which are known to describe random processes having no memory of previous events. This result, and other more recent papers, have documented that de-clustered catalogs generally obey Poisson interval statistics. Recently, we have proposed a series of methods under the name "earthquake nowcasting" to evaluate the current risk of major earthquake activity. Using machine learning, these methods develop proxy timeseries for California that appear similar to the long- hypothesized cycle of tectonic stress accumulation and release in seismically active regions. In the present paper, we analyze the information content of the California earthquake catalog using Receiver Operating Characteristic (ROC) and Shannon Information Entropy (SIE) methods. By using the Fltered amplitude of the monthly rate of small earthquakes, rather than interval statistics, we show that there is signiFcant information content in the catalog. We further construct a simple simulation similar to actual data that illustrates that a timeseries having Poisson recurrence statistics can nevertheless contain signiFcant information and nowcast skill. For that reason, we conclude that catalogs of small earthquakes contain information value regarding current risk of major earthquake occurrence. In addition, nowcasting methods using machine learning present a new approach to the construction of testable earthquake scenarios for use in peak ground velocity and acceleration calculations.