Seminar of the Department of Seismology
Title: From Stability to Intermittent Criticality: Forecasting Anthropogenic Seismicity in Complex Fracture Networks
Presenter: dr hab. Grzegorz Kwiatek
Link: https://teams.microsoft.com/meet/363959519319362?p=Tqi6dydJNdBh4BxzIL
Abstract:
As part of the St1 Deep Heat project (Helsinki, Finland), two hydraulic campaigns were carried out in 2018 and 2020 in adjacent 6-km-deep wells just 500 m apart. Both produced stable, pressure-controlled seismicity with no signs of runaway behavior, which could be kept in check simply by adjusting the injection. Because the two campaigns were so close together and so well recorded, they seemed an ideal case for calibrating (adaptive) traffic-light systems.
Yet when we tried to hindcast the seismicity rates and the largest expected magnitude using a new physics-based method, the two campaigns behaved very differently. The models we fit to each one disagreed on their key parameters, including the seismogenic index, the b-value, and the magnitude-frequency distribution, and their magnitude forecasts were inconsistent, ranging from large overestimates to large underestimates. As a result, we could not simply carry the settings from one stimulation over to its neighbor. The difference came down not to the injection rates alone but to fine structural variations within a heavily fractured reservoir that contained no major faults: small-scale complexity that controlled whether the seismicity stayed stable or began to drift.
To improve the forecasts, we applied machine learning, adding seismo-mechanical parameters that capture how earthquakes cluster and interact and how seismic energy release balances the hydraulic input. We compared the St1 case with two other studies that do contain distinct faults: the Cooper Basin (Australia) and laboratory fluid-injection experiments. Where distinct faults were present, seismo-mechanical parameters improved the forecasts; at St1, however, the distributed fracture network produced seismicity close to random in time and space, leaving little structure for the ML models to learn from. Forecasting deviations from stable behavior is therefore inherently harder in distributed fracture networks than on localized faults.
Laboratory stick-slip experiments on complex faults reveal that as stress builds, the fault enters a state of intermittent criticality. Small-scale asperities gradually break down and interact, collectively preparing the fault surface for a system-size slip by progressively smoothing the short- (mm-to-cm) scale stress field. The runaway earthquake on a complex fault is a statistical event that cannot be predicted deterministically. Nevertheless, the seismo-mechanical parameters and ML techniques may enable detection of the critical state of the system, when a runaway event becomes possible.