Predictive Skill: Using the Bayesian Inference to Study RSQSim and UCERF3

Jordan Cortez, Guillermo Beas, Shengyu Wang, Cynthia Tong, Sebastien D. Rossouw, & Katia Ascencio

Submitted August 1, 2018, SCEC Contribution #8248, 2018 SCEC Annual Meeting Poster #313

The undergraduate studies in earthquake information technology (USEIT) probabilistic forecasting team at the Southern California Earthquake Center (SCEC) worked closely with the High Performance Computing (HPC) team to calibrate parameters within the Rate State earthQuake Simulator (RSQSim), a physics-based multi-cycle earthquake simulator. The parameters varied included normal stress, earthquake slip rate, and rate- and state-friction coefficients. Java code was written to analyze a battery of catalogs generated from distinct variations of these parameters to calculate rates of earthquakes of specified magnitudes and regions. These rates were then compared to those of the statistics-based Uniform California Earthquake Rupture Forecast version 3 (UCERF3) in order to find the catalog (and set of parameters) most closely resembling the best understanding of seismic activity in California. The team then established a probabilistic framework for analyzing larger catalogs being generated with the chosen parameters. For a set of 9 fault sections on major faults in Southern California, we computed transition probabilities (defined here as the probability of any combination of the 9 fault sections rupturing within a fixed time window) directly from RSQSim catalogs by counting. While RSQSim catalogs often show strong coupling of neighboring faults, catalogs with finite lengths cannot fully sample this probability space. Even with long (1 million year) synthetic catalogs, event counts for many permutations of the 9 considered faults were low or zero. To address this and provide more robust probabilistic estimates, we also computed marginal probabilities assuming independence of each fault section. We then took a Bayesian approach using Laplace smoothing with marginal probabilities as our prior, and transition counts from RSQSim catalogs as data. The Laplace smoothing alpha parameter (α0) controls the relative weight assigned to transition and marginal probabilities. We found optimal values for α0 through log-likelihood testing using probabilities computed from one catalog to predict another. Finally, we compared the skill of UCERF3 with the forecasts from RSQSim data. From our statistical analysis, we provide a framework for analyzing the strengths of each forecasting system and evaluating physics-based earthquake models. This process improves our understanding of earthquake behavior, thus contributing to earthquake hazard and risk awareness.

Cortez, J., Beas, G., Wang, S., Tong, C., Rossouw, S. D., & Ascencio, K. (2018, 08). Predictive Skill: Using the Bayesian Inference to Study RSQSim and UCERF3. Poster Presentation at 2018 SCEC Annual Meeting.

Related Projects & Working Groups
Communication, Education, and Outreach (CEO)