Machine Learning Applications to Earthquake Source Characterization and Hazard Analysis
Daniel Trugman
Los Alamos National Lab
- Date & Time
- Location
- Building 3, Rambo Auditorium
- Host
- Jessica Murray
- Summary
Observational seismology is an increasingly data-rich field in which machine learning techniques show significant potential. In this seminar I focus on two specific examples where simple algorithms and concepts from machine learning prove useful in characterizing earthquake source properties and hazard. I will begin by describing a new framework for empirical Ground Motion Prediction Equations (GMPEs) based on a supervised learning algorithm known as a Random Forest. To demonstrate its utility in earthquake hazard analysis, I use the Random Forest GMPE to measure the influence of dynamic stress drop on the observed peak ground accelerations of moderate magnitude earthquakes in the San Francisco Bay Area. In the second part of this seminar, I will discuss ongoing collaborative efforts to systematically analyze the P-waveform features of large magnitude earthquakes for potential signatures of nucleation and rupture onset characteristics that correlate with event size. Many real-time earthquake early warning algorithms implicitly assume a form of rupture determinism in which the earthquake magnitude can be rapidly estimated from a small snapshot of waveform data, yet it remains unclear how viable these approaches are for large-magnitude events. Through analyses of a large dataset of Japanese Kik-Net borehole waveforms, we hope to gain insight into the physics of earthquake rupture and to provide unbiased observational constraints for real-time hazard applications.