Performance Testing, Comparison, and Selection of Induced Seismicity Rate Models for the Groningen Gas Field

Reliable seismic hazard and risk assessment for the Groningen gas field requires seismicity rate models that are consistent with observations and demonstrate predictive performance. We present a framework to evaluate, compare, and combine such models, applying it to four variants of the extreme threshold failure model. These variants combine either an elastic or elastic–viscoplastic stress model with one of two fault-weighting schemes, conditional on throw-thickness ratios: hard-cutoff or Bernstein-polynomial weighting. We extend the Collaboratory for the Study of Earthquake Predictability consistency testing framework with multiresolution and adaptive-resolution spatial tests to diagnose scale-dependent performance, and we employ leave-one-out cross-validation (LOO-CV) with stacking to derive ensemble weights. Our analysis shows that while all models are temporally consistent with the observed seismicity, they exhibit systematic spatial discrepancies at intermediate scales (4–8 km) in retrospective tests. The introduction of Bernstein weighting reduces these spatial inconsistencies and improves predictive performance. Although differences in LOO-CV scores are modest, stacking analysis assigns all ensemble weight to the Bernstein-weighted models, with a slight preference for the elastic–viscoplastic formulation. This study establishes a workflow for the ongoing testing and combination of induced seismicity models.