Jonathan Daniel Smith, Elias Rafn Heimisson , Stephen Bourne, Jean-Philippe Avouac 

In this study we use the Groningen gas field to test a new method to assess stress changes due to gas extraction and forecast induced seismicity. We take advantage of the detailed knowledge of the reservoir geometry and production history, and of the availability of surface subsidence measurements and high quality seismicity data. The subsurface is represented as a homogeneous isotropic linear poroelastic half-space subject to stress changes in three-dimensional space due to reservoir compaction and pore pressure variations. The reservoir is represented with cuboidal strain volumes. Stress changes within and outside the reservoir are calculated using a convolution with semi-analytical Green functions. The uniaxial compressibility of the reservoir is spatially variable and constrained with surface subsidence data. We calculate stress changes since the onset of gas production. Coulomb stress changes are maximum near the top and bottom of the reservoir where the reservoir is offset by faults. To assess earthquake probability, we use the standard Mohr-Coulomb failure criterion assuming instantaneous nucleation and a non-critical initial stress. The distribution of initial strength excess, the difference between the initial Coulomb stress and the critical Coulomb stress at failure, is treated as a stochastic variable and estimated from the observations and the modelled stress changes. The exponential rise of seismicity nearly 30 years after the onset of production, provides constraints on the distribution of initial strength. The lag and exponential onset of seismicity are well reproduced assuming either a generalized Pareto distribution, which can represent the tail of any distribution, or a Gaussian distribution, to describe both the tail and body of the distribution. The Gaussian distribution allows to test if the induced seismicity at Groningen has transitioned to the steady-state where seismicity rate is proportional to the stressing rate. We find no evidence that the system has reached such a steady-state regime. The modeling framework is computationally efficient making it possible to test the sensitivity to modeling assumptions regarding the estimation of stress changes. The forecast is found robust to uncertainties about the ability of the model to represent accurately the physical processes. It does not require in particular a priori knowledge of the location and orientation of the faults that can be activated. The method presented here is in principle applicable to induced seismicity in any setting provided deformation and seismicity data are available to calibrate the model.