Performance comparison of newly developed hydro-mechanical (hybrid) models for real-time induced seismicity forecasting

The Adaptive Traffic Light System (ATLS) is a seismic risk mitigation tool that can be used during geoenergy exploitation projects such as in the creation of Enhanced Geothermal Systems (EGS). In real-time applications, ATLS needs to include a data assimilation scheme in order to produce adaptive and time-dependent probabilistic seismicity forecasts by using the maximum available information at the moment of the assessment during such industrial operation. Numerical models capable to robustly forecast in real-time the temporal evolution of induced seismicity while properly accounting for uncertainties are the core elements of a functioning ATLS. In this respect, hydro-mechanical (HM) hybrid models are suitable models since they combine both statistical and physics-based assumptions to forecast induced seismicity. They include an accurate modeling of the physical processes involved in the generation of seismicity caused by human activity by keeping the associated computational cost low. In this work, we have developed two classes of simplified hybrid models that are based on 1D radial pore-fluid diffusion and fluid injection. The first class (HM0) accounts for linear pore-fluid diffusion from the fluid source, while the second class (HM1) accounts for the non-linear response of the medium due to pressure-dependent permeability variations upon fluid injection. Each class is sequentially coupled to two stochastic seismicity models that trigger seismicity if the respective space-time pressure solution reaches a critical value according to the Mohr-Coulomb criterion. However, both seismicity models differ in the way they simulate seismicity. One model uses an analytical approach based on probability density functions of model parameters to simulate seismicity (CAPS), while the other model simulates seismicity via stochastic seed approach (SEED). In a hindcast experiment, we test the models’ real-time forecasting capabilities within a data assimilation scheme using datasets from in-situ injection experiments that encompass different spatial scales. An essential step towards building a reliable ATLS is to properly weight each candidate model according to its respective contribution to the seismic hazard calculation. Therefore, an important part in our calculations is to evaluate and compare the forecasting performance of each model for each of the presented cases. Our findings show that models that take into account reliable distribution of uncertainties on a-priori selected model parameters as well as models accounting for a physically more accurate pressure solution in space and time generate better forecasting performances.