Event details

Christopher Johnson Los Alamos National Laboratory will discuss how seismic noise is the signal: learning the earthquake activity on the central San Andreas Fault". Earthquake forecasting is still an elusive goal; however, advances in machine learning based data processing have proven successful in predicting the timing of slip in controlled laboratory experiments. Scaling laboratory observations to Earth experiments is needed to establish what governs the timing of earthquakes. Using machine learning to extract and identify hidden signals in seismic noise provides new insight to the fault physics that control the accumulation and release of stress in the crust. I will present results from experiments performed on the San Andreas fault that test the ability of machine learning to provide predictive information about stress accumulation and release in the crust. For brittle failure event, time-to-failure models hint at weak indicators prior to earthquakes, but rigorous testing does not convey confidence the model is extracting a systematic signature from the waveforms. I look deeper in the fault zone at low frequency earthquakes (LFEs), which originate in the more ductile portion of the lower crust and have shown some evidence of activity changes before large seismogenic events that produce strong ground shaking. A catalog containing >1 million detections over 15 years provides a training data set to build a model that predicts LFE activity. Applying the trained model to 7 years of seismic waveforms reproduces the burst-like LFE behavior, with the greatest misfit occurring during the largest burst sequences within the 1-hour time intervals. During periods of low LFE activity, the model indicates more activity than originally detected using template matching. The most informative features are in the higher frequency bands, suggesting the model is utilizing new information that is typically not associated with LFEs. The results underscore the power of machine learning in seismic signal analysis and provide insight to the occurrence of ongoing slow-slip activity, which incentivizes new efforts to continue characterizing potential signals occurring prior to larger seismogenic event.