Improving Physics-Based Earthquake Forecasts for the 2016-2017 Central Italy Earthquake Sequence

Date: 4/24/2019

Time: 02:15 PM

Room: Pike

After a devastating earthquake, the ensuing cascade of aftershocks can be even more destructive than the mainshock. Operational earthquake forecasting seeks to provide reliable real-time information about the time-dependence of seismic hazard. Two forecasting approaches are commonly used: statistical short-term clustering models, such as ETAS, that employ empirical laws to predict aftershock patterns and physics-based models (CRS) that combine the stress transfer hypothesis with the Dieterich’s rate-and-state friction law.

Here we assess the effect of model choices at real-time conditions and quantify the influence of input data quality on the predictive skills of CRS forecasts during the 2016-17 Central Italy sequence with 7 M ≥ 5.4 events that occurred within less than 5 months. We develop 7 physics-based models with progressively increasing level of refinement as part of a pseudo-prospective experiment with 1-year time horizon. The preliminary CRS forecasts include data available few minutes after each M ≥ 5.4 event and feature synthetic source models with empirically determined fault length and fault constitutive parameters from literature. The sophisticated CRS models gradually incorporate optimized rate-state parameters, spatially heterogeneous receiver faults, best available slip models, and M3+ secondary triggering effects.

We evaluate model performance using CSEP metrics over different time horizons and compare against a benchmark ETAS model. We find that revised catalog data and a realistic representation of crustal heterogeneity boost CRS models’ performance. The more complex and data-wealthy CRS forecasts reach probability gains per earthquake higher by 3 orders of magnitude when compared against preliminary models. The results confirm that CRS models are as informative as ETAS only when secondary triggering effects are combined with realistic slip models, spatially heterogeneous receiver planes and optimized fault constitutive parameters. Our results support and extend other recent retrospective experiments on the predictive power of CRS forecasts.

Presenting Author: Simone Mancini

Authors