The Prediction Problems of Earthquake System Science

Editor’s note: The following is the text of the SSA Presidential Address presented at the Annual Luncheon of the Seismological Society of America (SSA) Annual Meeting on 30 April 2014. The Seismological Society of America (SSA) has always been dedicated to understanding and reducing the earthquake threat. The Society was founded in 1906 “for the acquisition and diffusion of knowledge concerning earthquakes and allied phenomena.” According to our new strategic plan, approved by the Board in 2012, the core purpose of SSA is to “advance seismology and the understanding of earthquakes for the benefit of society.” This plan lays out the vision for SSA to be “the primary forum for the assembly, exchange, and dissemination of scientific knowledge essential for an earthquake-aware and safer world.” In the past twenty years or so, the study of earthquakes has become a true system science, offering new pathways for the advancement of seismology. Today I would like to explore what the rise of earthquake system science might imply for the future of our field and for SSA’s mission in earthquake research. System science seeks to explain phenomena that emerge from nature at the system scale, such as global climate change or earthquake activity in California or Alaska. The “system” is not a physical reality, but a hypothetical representation of nature, typically a numerical model that replicates an emergent behavior and predicts its future course. The choice of target behavior determines the system model, as can be illustrated by two representations of earthquake activity in California. One is UCERF3, the latest uniform California earthquake rupture forecast of the Working Group on California Earthquake Probabilities, which represents future earthquake activity in terms of time-dependent fault-rupture probabilities. Another is the Southern California Earthquake Center (SCEC)’s CyberShake ground-motion model, which uses simulations to represent the probability of future earthquake shaking at geographic sites, conditional on the fault rupture. These two system-level models can be combined to generate site-specific hazard curves, the main forecasting tool of probabilistic seismic-hazard analysis (PSHA). The first point to emphasize is that earthquake system science is all about forecasting and prediction. For many years now, earthquake prediction has remained an awkward topic in polite seismological company, primarily because it has been defined in the public mind by something we cannot do, which is to predict with high probability the regional occurrence of large earthquakes over the short term. Yet the “P-word” is too central to our science to be banned from our working vocabulary. From a practical perspective, we must be able to predict earthquake hazards in order to lower seismic risk. From the basic-research perspective of system science, testing a model’s predictions against new data is the principle means by which we can gain confidence in the hypotheses and theories on which the model is built. For example, many interesting problems of contingent predictability can be posed as physics questions in a system-specific context. What will be the shaking intensity in the Los Angeles basin from a magnitude 7.8 earthquake on the southern San Andreas fault? By how much will the strong shaking be amplified by the coupling of source directivity to basin effects? Will deep injection of waste fluids cause felt earthquakes near a newly drilled well in Oklahoma? How intense will the shaking be during the next minute of an ongoing earthquake in Seattle? SSA should stake its claim as the central forum for the physics-based study of earthquake predictability, and its publications should be the place where progress in understanding predictability is most rigorously documented. My second point is that forecasting and prediction are all about probabilities. The deep uncertainties intrinsic to earthquake forecasting are most coherently expressed in terms of two distinct types of probability: the aleatory variability that describes the randomness of the system, and the epistemic uncertainty that characterizes our lack of knowledge about the system. In UCERF3, the former is cast as the time-dependent probabilities of fault ruptures, of which there are over 250,000, whereas the latter is expressed as a logic tree with 5760 alternative branches. Similarly, CyberShake represents the aleatory variability in wave excitation through conditional hypocenter distributions and conditional slip distributions, and it characterizes the epistemic uncertainty in the wavefield calculations in terms of alternative 3D seismic-velocity models. The full-3D treatment of seismic-wave propagation has the potential to improve our PSHA models considerably. A variance-decomposition analysis of the recent CyberShake results indicates that more accurate earthquake simulations could reduce the aleatory variance of the strong-motion predictions System science offers a brick-by-brick approach to building up our understanding of earthquake predictability.