Evaluation of Earthquake Early Warnings as External Forecasts

Evaluation of Earthquake Early Warnings as External Forecasts and Predictions (EFPs) Philip Maechling Information Technology Architect Southern California Earthquake Center (SCEC) 7 May 2013 SCEC: An NSF + USGS Research Center

Premise: EEW In California Is Imminent

Earthquake Parameters Forecast by EEW Systems may forecast (1) event-specific parameters (mag, location) and site-specific earthquake parameters (peak ground motion). – Event Parameters • Magnitude • Location – Site-specific Parameters: • Site specific ground motion intensity based on earthquake location, earthquake magnitude, and attenuation relation to forecast peak ground motions at sites of interest. 3

Warning time depends on your location’s distance from where the earthquake begins. The slanted red line shows how warning time increases with distance from the epicenter. In this case, warning time increases beyond the 21 mileradius blind zone with, for instance, approximately 10 seconds warning at 40 miles distance.

CISN Shake Alert Single sensor Sensor network τc-Pd On-site Algorithm Virtual Seismologist (VS) Sensor network Elarm. S Decision Module (Bayesian) CISN EEW Testing Center USER Module - Single site warning - Map view Test users • predicted and observed ground motions • available warning time • probability of false alarm • …

Problems in Assessing Forecasts Shake. Alert Forecast Evaluation Problems: – Scientific publications provide insufficient information for independent evaluation – Data to evaluate forecast experiments (e. g. catalog used to evaluate forecasts) are often incompletely defined – Active researchers are constantly tweaking their codes and procedures, which become moving targets – Difficult to find resources to conduct and evaluate long term forecasts – Standards are lacking for testing forecasts against reference observations 8

EEW Algorithm Differences Different algorithms produce different forecast parameters • Some (e. g. On-site) produce site-specific information (PGA), event magnitude, but no origin time or distance to event • Some (e. g. Vs) produces full event parametric information. • Some (e. g. Elarms) produce site specific ground motion estimates on a regular grid. • Some produce single values (On-site) • Some produce time-series with updates (Vs, Elarms) 9

Applying the CSEP Testing Approach to Earthquake Early Warning Forecasts CISN and SCEC received funding from USGS to develop and evaluate prototype network-based EEW: SCEC has implemented the CISN Testing Center (CTC) to evaluate the system and seismological performance of the CISN and USGS Shake. Alert earthquake early warning prototype system.

Scale of SCEC CTC Activity CISN Testing Center Design Goals and Constraints: – – Establish scientific framework for Shake. Alert Testing Simple and inexpensive to develop and operate Provide value to USGS and Shake. Alert developers Communicate value of EEW testing to SCEC community and CISN 15

CISN Testing Document

Design of an Experiment Many CSEP testing principles are applicable to CISN EEW Testing. The following definitions need to be made to evaluate forecasts: – – – Exact definition of testing area Exact definition of a forecast Exact definition of input data used in forecasts Exact definition of reference observation data Measures of success forecasts 17

Selection of a Testing Region Decide if the 3 CSEP regions valid for EEW • Region Under Test • Catalog Event Region • Buffer to avoid catalog issues 18

Experiment Design Summary 1. 1: Magnitude XY Diagram Measure of Goodness: Data points fall on diagonal line Relevant: T 2, T 3, T 4 Drawbacks: Timeliness element not represented Which in series of magnitude estimates should be used in plot. 19

Experiment Design Summary 2. 2: Magnitude and Location error by time after origin Measure of Goodness: Data points fall on horizontal zero line Relevant: T 3, T 4 Drawbacks: Eventspecific not cumulative 20

CISN Testing Document

CISN EEW Performance Summary Processing Retrieve Data ANSS Earthquake Catalog Filter Catalog Earthquake Catalog Filtered Earthquake Catalog Observed ANSS Data Produce Web Summaries UCB/Elarm. SNI EEW Data Source EEW Trigger Reports Load Reports EEW Trigger Reports CIT/On. Site EEW Data Source CISN EEW Trigger Data CISN EEW Testing Center and Web Site

Why Use CSEP Software for EEW Testing

CISN Testing Document

Shake. Alert Performance Testing System Overview at USC CISN Testing Center (top right) retrieves a daily earthquake catalog from ANSS Data Center (bottom right) and Shake. Alert performance logs from U. C. Berkeley (left). It then matches new ANSS events to Algorithm Alerts and Decision Module Alerts and plots (1) Event Performance summaries, and (2) Cumulative Performance Summaries

Types of Shake. Alert Performance Summaries Currently Available Summaries posted online at: http: //scec. usc. edu/scecpedia/CTC_Results

Cumulative Shake. Alert Performance Results for all ANSS catalog events M 3. 5+ with Network Codes CI and NC between 4 March 2012 and 31 March 2013.

M 4. 5 at 2. 4 km (1. 5 mi) NE of The Geysers, CA 38. 8123, -122. 786, 2 km Mar 14 2013 09: 23 UTC 71954065

4. 0 at 21. 3 km (13. 3 mi) SE of Markleeville, CA 38. 5598, 119. 616, 7 km Jan 24 2013 23: 25: 51 UTC 71928201

EEW Scientific Testing Lessons Learned as External Forecasts and Predictions (EFPs) – Difficult to determine (real-time) information used in forecast, and especially when Bayesian approach is fully implemented – Forecasts are not evaluated before earthquake occurs and final results are known. Tests based on forecast logs are treated as prospective testing. – Performance on an event by event basis are higher priority, catalog delays have significant impact. – Performance of systems changes as ANSS catalog changes. Single measurement less useful than ongoing evaluations over time. 35

End

Forecast Testing Should Increase Along with Forecast Impact SCEC Computational Forecast Users Scientific and Engineering Requirements for Forecast Modeling Systems Public and Governmental Forecasts Automated prospective testing of forecast models over time within collaborative forecast testing center. Engineering and Interdisciplinary Research Automated retrospective testing of forecast models using community defined validation problems. Collaborative Research Project Computational codes, structural models, and simulation results versioned with associated tests. Individual Research Project Development of new computational, data, and physical models.

Testing Center System Requirements The goals of both an EEW and Earthquake Forecast Testing Center Goals (as outlined by Schorlemmer and Gerstenberger (2007)) describe what is needed to build trust in results: Controlled Environment Transparency Comparability Reproducibility

Proposed Performance Measures Summary Reports for each M ≥ M-min: Key documents is 3 March 2008 document which specifies six types of tests. – – – Summary 1: Magnitude Summary 2: Location Summary 3: Ground Motion Summary 4: System Performance Summary 5: False Triggers Summary 6: Missed Triggers 39

Testing Center Conclusions 1. Broad impact of seismological technologies like EEW are great enough to warrant significant effort for evaluation. 2. Independent evaluation for EEW provides valuable service to agencies including CISN, USGS, CPEC, NEPC, and others. 3. Prospective must be done to before techniques will be accepted. 4. Similarities between problems lead to similar scientific techniques. 5. Similarities between problems lead to similar technology approach and potentially common infrastructure. 6. “Neutral” third party testing has significant benefits to the science groups involved in forecasting. 7. CSEP infrastructure can be adapted for use in CISN EEW Testing Centers.

Useful Information to Help Interpretation of CTC Performance Summary Results • CTC performance summaries are focused on Shake. Alert speed of operation and earthquake event parameter forecasts. No site specific ground motion performance summaries have been developed. • CTC Shake. Alert performance summaries compare ANSS event parameters to Shake. Alert alert event parameter forecasts. No searching for “false alerts” is currently done. • Because ANSS Catalog is considered the correct observation, the CTC waits 48 hours after events to let the ANSS catalog settle down before generating CTC performance summaries. • Cumulative summaries depend on the catalog filter criteria (date range, magnitude, and region) used to select ANSS catalog events. Currently, ANSS events are selected for Mag >= 3. 5, Network ID NC or CI. And, we currently run for two time frames Sept 2011 through present and March 3, 2012 through present (considered the Elarms 2 era). • One pager event summaries are produced for new all ANSS events with Network ID NC or CI and Mag >= 3. 0

EEW As An Earthquake Forecast Earthquake Warning systems predict final earthquake magnitude before it is known, possibly before earthquake rupture is completed. - Most rigorous type of forecast testing in prospective testing. - We consider prospective EEW testing if the EEW algorithms log their forecasts in real-time, before final forecasts parameters (e. g. final magnitude) is known.