Background: Human-induced earthquakes have become an important topic of political and scientifi c discussion, owing to the concern that these events may be responsible for widespread damage and an overall increase in seismicity. It has long been known that impoundment of reservoirs, surface and underground mining, withdrawal of fl uids and gas from the subsurface, and injection of fl uids into underground formations are capable of inducing earthquakes. In particular, earthquakes caused by injection have become a focal point, as new drilling and well-completion technologies enable the extraction of oil and gas from previously unproductive formations.

We present a technique that greatly improves the precision in measuring temporal variations of crustal velocities using an earthquake doublet, or pair of microearthquakes that have nearly identical waveforms and the same hypocenter and magnitude but occur on different dates. We compute differences in arrival times between seismograms recorded at the same station in the freqency domain by cross correlation of short windows of signal. A moving-window analysis of the entire seismograms, including the coda, gives g(t), the difference in arrival times versus running time along the seismogram. The time resolution of the method is an order of magnitude better than the digitization interval. The g(t) technique is illustrated with a pair of microearthquakes, M = 1.7 and 2.0, that occurred before and after the Coyote Lake, California, earthquake (M = 5.9) of August 6, 1979, and on the same segment of the Calaveras fault that ruptured during the earthquake. The coda wave arrivals for some stations are progressively delayed for the second earthquake in the doublet, so that its seismogram appears as a stretched version of the earlier event. We interpret this systematic variation in 6(t) along the coda as a change in the average S velocity in the upper crust in the time interval between the two doublets. S wave velocities appear to have decreased by 0.2% in an oblong region 5-10 km in radius at the south end of the aftershock zone.

Near-source observations show that earthquakes initiate with a distinctive seismic nucleation phase that is characterized by a low rate of moment release relative to the rest of the event. This phase was observed for the 30 earthquakes having moment magnitudes 2.6 to 8.1, and the size and duration of this phase scale with the eventual size of the earthquake. During the nucleation phase, moment release was irregular and appears to have been confined to a limited region of the fault. It was characteristically followed by quadratic growth in the moment rate as rupture began to propagate away from the nucleation zone. These observations suggest that the nucleation process exerts a strong influence on the size of the eventual earthquake.

Earthquake signal detection and seismic phase picking are challenging tasks in the processing of noisy data and the monitoring of microearthquakes. Here we present a global deep-learning model for simultaneous earthquake detection and phase picking. Performing these two related tasks in tandem improves model performance in each individual task by combining information in phases and in the full waveform of earthquake signals by using a hierarchical attention mechanism. We show that our model outperforms previous deep-learning and traditional phase-picking and detection algorithms. Applying our model to 5 weeks of continuous data recorded during 2000 Tottori earthquakes in Japan, we were able to detect and locate two times more earthquakes using only a portion (less than 1/3) of seismic stations. Our model picks P and S phases with precision close to manual picks by human analysts; however, its high efficiency and higher sensitivity can result in detecting and characterizing more and smaller events.

We construct a probability model for rupture times on a recurrent earthquake source. Adding Brownian perturbations to steady tectonic loading produces a stochastic load-state process. Rupture is assumed to occur when this process reaches a critical-failure threshold. An earthquake relaxes the load state to a characteristic ground level and begins a new failure cycle. The load-state process is a Brownian relaxation oscillator. Intervals between events have a Brownian passage-time distribution that may serve as a temporal model for time-dependent, long-term seismic forecasting. This distribution has the following noteworthy properties: (1) the probability of immediate rerupture is zero; (2) the hazard rate increases steadily from zero at t ‫ס‬ 0 to a finite maximum near the mean recurrence time and then decreases asymptotically to a quasi-stationary level, in which the conditional probability of an event becomes time independent; and (3) the quasi-stationary failure rate is greater than, equal to, or less than the mean failure rate because the coefficient of variation is less than, equal to, or greater than In addition, the model provides 1/ 2 Ϸ 0.707. Ί expressions for the hazard rate and probability of rupture on faults for which only a bound can be placed on the time of the last rupture. The Brownian relaxation oscillator provides a connection between observable event times and a formal state variable that reflects the macromechanics of stress and strain accumulation. Analysis of this process reveals that the quasi-stationary distance to failure has a gamma distribution, and residual life has a related exponential distribution. It also enables calculation of "interaction" effects due to external perturbations to the state, such as stress-transfer effects from earthquakes outside the target source. The influence of interaction effects on recurrence times is transient and strongly dependent on when in the loading cycle step perturbations occur. Transient effects may be much stronger than would be predicted by the "clock change" method and characteristically decay inversely with elapsed time after the perturbation.