Operational aftershock forecasting during the Ridgecrest sequence

The 2019 Ridgecrest earthquake sequence was the first significant ‘real-world’ application of the Third Uniform California Earthquake Rupture Forecast with Epidemic Type Aftershock Sequences (UCERF3-ETAS) model in the California forecasting region since it was released in 2017. For the one-year anniversary of the 2019 Ridgecrest sequence, we share our experiences and responsibilities as SCEC researchers as this sequence unfolded, and what we’ve learned since that has resulted in numerous operational improvements (Milner et al., 2020) and the first “out of sample” evaluation of the UCERF3-ETAS model (Savran et al., 2020).

After any significant earthquake, scientists are usually asked, “Was that the big one? Will there be more?” We know from observations that the first event of a sequence may not be the largest in terms of magnitude, and although the sequence will eventually die off, its maximum magnitude and duration cannot be known in advance with certainty. Earthquakes cannot be predicted, as this implies knowing exactly the time, location, and magnitude of a future earthquake. Instead, scientists focus their efforts on forecasting earthquakes, which provide the probability that earthquake(s) of some size(s) may occur in a future timespan. Therefore, earthquake forecasts could answer the question “What is the probability of a bigger one occurring?”

A significant effort is being directed to improving aftershock forecasting models in order to answer this important question. The simplest model combines Omori’s Law for the decay of the rate of aftershocks with time and Gutenberg-Richter’s relation between the number of earthquakes and their magnitudes to forecast the expected numbers of earthquakes with various magnitudes following a mainshock. This forecasting model, known simply as “Reasenberg-Jones”, is used to provide authoritative forecasts made by the U.S. Geological Survey for earthquakes with M≥5.5, including the M7.1 Ridgecrest earthquake. You may have heard a scientist on TV state after a big earthquake has occurred, “the chance of a bigger one is about 5% and decays over time if nothing else happens.” Such statements are based on models like Reasenberg-Jones.

The more advanced epidemic type aftershock sequence (ETAS) models hypothesize that each earthquake generates its own aftershocks, and that the number of aftershocks depends on the magnitude of the earthquake. Although ETAS-type models are more complex, they better reflect our understanding of aftershock sequences. The most advanced ETAS-type model is an iteration of the Third Uniform California Earthquake Rupture Forecast known as UCERF3-ETAS. Unlike standard ETAS models that assume forecasted earthquakes can be described by their epicenters (or hypocenters), UCERF3-ETAS allows for large earthquakes to occur on a set of large known faults (such as the San Andreas or Garlock faults). This means questions such as “What is the probability that the M6.4 Ridgecrest earthquake would trigger an aftershock on the Garlock fault?” or “What is the probability that a M4.5 near the San Andreas would turn into something larger?” can be answered using UCERF3-ETAS. The former question was asked by officials and the concerned public immediately following the M6.4 Searles Valley earthquake.

Within 33 minutes of this earthquake, SCEC computer scientist Kevin Milner started running aftershock simulations on a high-performance computing cluster at the University of Southern California, and preliminary results were posted to SCEC’s post-earthquake science coordination website within an hour of the M6.4 event. The model forecasted a 2.8% chance of an M ≥ 6.4 aftershock in the next week. A larger M7.1 event occurred the next day, 5 July 2019 at 08:19 p.m. (local time), when the primary UCERF3-ETAS developer was unavailable. Another member of the SCEC research team, Bill Savran stepped in and initiated simulations of the larger earthquake (then reported as M6.9) less than an hour later and ran follow-up simulations with the updated magnitude (M7.1) and depth at 04:25 a.m. the following morning. Those revised simulations, which used a simple (point-source) representation of the M7.1, were completed and posted for review just before California Earthquake Prediction Evaluation Council (CEPEC), a committee that advises the governor of California about earthquake predictions, forecasts, and hazards, convened at 10:30 am on July 6.

Research has shown that large earthquakes like the M6.4 and the M7.1 events are better described using finite-faults instead of point sources, because their rupture lengths span tens of kilometers. Prior to Ridgecrest, only ruptures on modeled UCERF3 faults could be specified as inputs to the model. However, neither the M6.4 nor the M7.1 ruptures occurred on the pre-existing UCERF3 faults. Within a day of the M7.1, Kevin Milner implemented a new feature in the UCERF3-ETAS model to allow for input events to be defined on fault surfaces not previously defined. Using this new capability, he computed new simulations using a fault surface drawn based on the observed aftershocks. Within a week, Kevin implemented an additional capability to fetch the authoritative finite-fault surfaces used by ShakeMap provided by the U.S. Geological Survey Comprehensive Catalog (ComCat). Subsequently, he also implemented the ability to automatically fetch input data from ComCat for configuring new simulations and post-processing that can be used to visually assess the performance of the forecasts. Using these new operational capabilities, weekly simulations have been run that update the forecasts with newly observed seismicity during the Ridgecrest earthquake sequence.

30 day Garlock M≥7 probabilities, expressed as a function of time since the M7.1 mainshock. Probabilities from the short-term UCERF3-ETAS model are depicted with a thick, black line; they decrease from 3.2% immediately following the M7.1 to 0.044% just before the 3 June, 2020, M5.5, at which point the probability increases to 0.25%. Sampling uncertainties (95% bounds) for UCERF3-ETAS simulations are depicted with a light gray shaded region. Model update times, at which the model was updated to include recent seismicity (weekly for the first two months, and then sporadically after), are marked with vertical dashed lines. Probabilities from the long-term time-dependent UCERF3 model are nearly constant in this time period at 0.015% (blue, thick line).


The chance of triggering a large Garlock rupture was of particular concern to the scientific community and the public at large in the days following the Ridgecrest events. UCERF3-ETAS is uniquely able to quantify the probability of this scenario. The figure above shows a 3.2% probability of triggering a M≥7 on the Garlock fault within 30 days of the M7.1. That probability generally decreases with time, but can also increase after large aftershocks. This is apparent after the 3 June, 2020 M5.5 event, which increased the 30 day Garlock probability from 0.044% to 0.25%. Even a year after the M7.1 mainshock, the probability of a M≥7 Garlock event is greater than three times the long-term probability of such an event according to the time-dependent UCERF3 model (which incorporates elastic rebound but not recent seismicity). It is important to note that these probabilities are sensitive to the fault surface representation used. Initial point-source calculations put the 30 day probability of a M≥7 Garlock at 0.65%, and finite-fault surface probabilities, which are higher because those surfaces extend toward the Garlock fault from the epicenter, ranged from 1.5% to 5.3%.

Over the last year since the Ridgecrest sequence began, Bill Savran has been working with the Collaboratory for the Study of Earthquake Predictability (CSEP) to develop techniques to statistically evaluate simulation-based models like UCERF3-ETAS. The near real-time UCERF3-ETAS forecasts provide the first opportunity to evaluate UCERF3-ETAS using new data. To first order, UCERF3-ETAS captures the spatiotemporal evolution of seismicity during the Ridgecrest sequence. This adds to the growing body of evidence that ETAS models can be informative forecasting tools, and supports the use of UCERF3-ETAS for operational aftershock forecasting following significant earthquakes. However, UCERF3-ETAS with the parameters employed mildly overpredicts the seismicity rate, on average, aggregated over the forecasts made during the sequence. Further testing indicates the forecasts do not include enough variability in the forecasted magnitude-number distributions to match observations. Also, more frequent updating of the model, say after all events with M≥3.5, could help improve the spatial seismicity forecast. Therefore, accounting for uncertainty in the model parameters could improve test results, and thus forecasts, for future aftershock sequences. Details on the evaluation experiment including the evaluation metrics used are explained in detail by Savran et al. (2020).

We have all heard the common adage “prior proper planning prevents poor performance”, which could not be more relevant to the experience issuing UCERF3-ETAS forecasts during the Ridgecrest sequence. Proper training and extensive documentation of the process for initiating and running the UCERF3-ETAS models meant that we could produce forecasts in a timely manner that allowed scientists and others to assess the potential of aftershocks. Dealing with real-time issues, such as data outages and the unpredictability of nature, led to numerous improvements in the operational capabilities of the model. The Ridgecrest sequence enabled an earthquake forecasting experiment that provided us with confidence in the forecasting process and introduced suggestions on possible improvements to the model. Every opportunity to study an earthquake sequence is an opportunity to advance research that can improve the operational capabilities of forecasting models—and to help answer important questions following a significant earthquake.

About the Authors

William Savran is a SCEC software engineer at the University of Southern California. He is the lead developer of the Collaboratory for the Study of Earthquake Predictability, and works with researchers around the world to develop and implement methods for unbiased evaluations of earthquake forecasting models in California and beyond.
Kevin Milner is a SCEC computer scientist at the University of Southern California. He develops earthquake forecasting models for California spanning time scales from minutes to centuries, with the aim of reducing uncertainties and turning cutting-edge research into societally relevant and useful products.


This research was supported by the Southern California Earthquake Center. SCEC is funded by NSF Cooperative Agreement EAR-1600087 and USGS Cooperative Agreement G17AC00047. Additional support was provided by W.M. Keck Foundation Grant 005590-00001 and the California Earthquake Authority. We thank our colleagues from the Working Group on California Earthquake Probabilities (WGCEP), the Collaboratory for Study of Earthquake Predictability (CSEP), and the Collaboratory for Interseismic Modeling (CISM) for collaborating in this research.