Physics-based and probabilistic modelling of induced seismicity: where we stand and where to go next

Induced seismicity, that is, seismicity associated with subsurface operations, has been reported over the last 50 years in different sites world-wide. The knowledge gained from those empirical case studies have helped to pinpoint the main co-factors leading to fault instability, being related to the tectonic stress state of the reservoir, its local geology, induced pore pressure and thermo-chemical changes, local stress redistribution from event-event interactions and feedback from (a)seimic slip on faults. Despite the progress made, forecasting of the induced seismic risk during each stage of a reservoir project remains a challenge. Studies relying on classical earthquake catalogue parameters, which are fed into advanced traffic light system (ATLs) have been partly successful in linking the seismic risk to operational parameters. However, these datasets are generally „biased“ toward more energetic sequences (i.e. few high magnitude events) and are limited by the highly variable quality of operational data available for each particular site. A growing number of field studies (e.g. the Mw5.5 Pohang earthquake, the Mw3.9 Strasbourg earthquake) have been challenging the general validity of usually adopted log-linear frequency-magnitude correlation statistics, thereby calling for a critical revision of our current understanding of the dynamics leading to induced earthquakes.

In this contribution, we will discuss how integrating information derived from physics-based quantitative models into existing stochastic frameworks can help to overcome the shortcomings of purely statistical approaches to induced seismicity.

We will dedicate a first part of this contribution to showcase via diverse field and laboratory examples how physics-based models can be effective to improve our understanding of structure-property relationships under transient loading and across-scale, and to discriminate intrinsic spatio-temporal footprints of the driving dynamics leading to induced earthquakes including expected magnitudes, their spatial and temporal distribution with respect to site-specific conditions and operational parameters. We will then open a second chapter where we will discuss how front ending these software solutions to existing state-of-the-art HPC platforms can enable near real-time model calibration and to additionally explore the sensitivity in terms of either reservoir performance and/or induced seismicity to a wide spectrum of reservoir parameters therefore outperforming classical forecasting models based on field data alone. We will then conclude with an open discussion on how these hybrid approaches can benefit from novel methods from the field of artificial intelligence and machine learning and where existing knowledge gaps remain.