Under the Macroscope: Constraining Spatiotemporal Characteristics of Triggered Seismicity at the Reservoir Scale Using Data-driven Models

Figure: Model of CO2 injection into a structurally confined dome in
(St Johns, in Arizona) with calculation of non-local bending stresses induced
in the caprock above the injection sites.

 

One challenge facing triggered seismicity hazard assessment is
an apparent lack of readily-available, strong predictors: it is difficult or
expensive to obtain the necessary data to determine whether injection at a
particular site is likely to induce earthquakes. Susceptibility to induced
seismic activity is influenced by a host of site- and operation-specific
parameters, such as ambient stress magnitudes, presence of faulting, and the
volume of fluid injection. Modern machine learning techniques can help us make
sense of these multi-dimensional spaces, first, by dividing the parameter space
into discrete regions of seismic susceptibility and, second, by probing the
functional (predictive) relationships between predictors and outcome.

This investigation will help direct efforts to improve
numerical tools for modeling induced seismicity at the reservoir scale. Modern
reservoir flow simulators are equipped for modeling the injection operation,
reservoir pressure build-up and associated geomechanical deformation, and fault
failure criteria; however, they typically neglect the earthquakes themselves
or, if capable, focus on individual triggered events. We are implementing new
methods for modeling multiple earthquake events – sequences – into existing
reservoir simulators. These new tools will be applied to several recent induced
earthquake sequences with a focus on understanding their bulk behavior, for
instance, the correlation between injection and seismicity rate, and the
variable delay between the beginning of injection at a site and the subsequent onset
of seismic activity (if any.)