Spatial Verification of Earthquake Simulators Using Self-Consistent Metrics for Off-Fault Seismicity

Abstract:

We address the problem of verifying the self-consistency of earthquake simulators with the data from which their parameters are drawn. Earthquake simulators are a class of computational simulations which attempt to mirror the topological complexity of the earthquake fault system on which the earthquakes occur. In addition, the physics of friction and elastic interactions between fault elements can be included in these simulations as well. In general, the parameters are adjusted so that natural earthquake sequences are matched in their scaling properties in an optimal way. Generally, these parameters choices are based on paleoseismic data extending over many hundreds and thousands of years. However, one of the problems encountered is the verification of the simulations applied to current earthquake seismicity. It is this problem, for which no currently accepted solution has been proposed, that is the objective of the present paper. Physically-based earthquake simulators allow the generation of many thousands of years of simulated seismicity, allowing for robust capture of statistical properties of large, damaging earthquakes that have long recurrence time scales for observation. Following past simulator and forecast model verification efforts, we approach the challenges in spatial forecast verification fo simulators; namely, that simulator output events are confined to the modeled faults, while observed earthquakes often occur off of known faults. We present two methods for overcoming this discrepancy: a simplistic approach whereby observed earthquakes are shifted to the nearest fault element and a variation of the Epidemic-type aftershock (ETAS) model, which smears the simulator catalog seismicity over the entire test region. To test these methods, a Receiver Operating Characteristic (ROC) plot was produced by comparing the rate maps to observed m>6.0 earthquakes since 1980. We found that the nearest-neighbor mapping produced poor forecasts, while the modified ETAS method produced rate maps that agreed with observations. These ETAS results were further analyzed by comparing them against catalogs of earthquakes randomly generated from the ETAS rate map. The observed ROC scores for all simulators were near or exceeded the 95th percentile of the distribution of random scores.