Seismicity precursors to large earthquakes unified in a stress accumulation framework



[1] Various seismicity patterns before large earthquakes have been reported in the literature. They include foreshocks (medium-term acceleration and short-term activation), quiescence, doughnut patterns and event migration. The existence of these precursory patterns is however debated. Here, we develop an approach based on the concept of stress accumulation to unify and categorize all claimed seismic precursors in a same physical framework. We first extend the Non-Critical Precursory Accelerating Seismicity Theory (N-C PAST), which already explains most precursors, to additionally include short-term activation. Theoretical results are then compared to the time series observed prior to the 2009Mw = 6.3 L'Aquila, Italy, earthquake. We finally show that different precursory paths are possible before large earthquakes, with possible coupling of different patterns or non-occurrence of any. This is described by a logic tree defined from the combined probabilities of occurrence of the mainshock at a given stress state and of precursory silent slip on the fault. In the case of the L'Aquila earthquake, the observed precursory path is coupling of quiescence and accelerating seismic release, followed by activation. These results provide guidelines for future research on earthquake predictability.

1. Introduction

[2] Although numerous seismicity precursors have been proposed in the past decades (see, e.g., reviews by Kanamori [1981], Wyss and Habermann [1988], and Mignan [2011]), our ability to predict large earthquakes remains highly doubted [Jordan et al., 2011]. A number of studies have shown that claimed precursory anomalies are not observed systematically, which questions their existence as valid precursors [e.g., Zechar and Zhuang, 2010]. Others have shown that observed patterns are undistinguishable from the normal behavior of seismicity [e.g., Hardebeck et al., 2008; Marzocchi and Zhuang, 2011], which is well described by the Epidemic-Type Aftershock Sequence (ETAS) model where multiple aftershock sequences are superposed on a constant background rateμ [Ogata, 1988]. Although such models based on aftershock statistics [e.g., Gerstenberger et al., 2005] are promoted for Operational Earthquake Forecasting (OEF) [Jordan and Jones, 2010], they provide probabilities too low for operational forecasting of large mainshocks [van Stiphout et al., 2010]. Further investigation of potential precursors is therefore crucial.

[3] The aim of this paper is to extend the Non-Critical Precursory Accelerating Seismicity Theory (N-C PAST) [Mignan et al., 2007; Mignan, 2008, 2011] to explain, in a unified framework, all basic seismicity precursors proposed in the literature and to provide new guidelines for earthquake predictability research. Considered precursory patterns are: quiescence [e.g., Wyss and Habermann, 1988], Accelerating Seismic Release (ASR) [e.g., Bufe and Varnes; 1993; Mignan, 2011], short-term activation [e.g.,Jones and Molnar, 1979], Mogi doughnut and event inward migration [e.g., Kanamori, 1981], outward migration [e.g., Jones et al., 1982], decelerating seismicity [e.g., Papadimitriou, 2008], and b-value decrease (i.e., increase in event size [e.g., Rundle et al., 2000]). We will show in this study that the N-C PAST makes specific predictions for the space-time-rate characteristics and couplings of these various patterns.

2. The Non-Critical Precursory Accelerating Seismicity Theory (N-C PAST)

[4] The N-C PAST [Mignan et al., 2007; Mignan, 2008, 2011] is a theoretical framework based on the concepts of elastic rebound [Reid, 1910] and of static stress transfer [e.g., King, 2007] which predicts that various precursory patterns may be observed days to months before a large earthquake due to loading on the main fault [King and Bowman, 2003]. A large earthquake is defined as any event due more likely to tectonic loading than to nearby triggering [Mignan et al., 2006; King, 2007], which corresponds to faults with a length/width ratio greater than one, i.e., about Mw ≥ 6 [Mignan, 2008]. In the N-C PAST, the background seismicity rate is non-stationary, the variations of which are function of the stress state surrounding the fault [e.g.,Ogata, 2005]. This hypothesis is supported by independent geodetic measurements [Ogata, 2007]. Aftershocks can be superposed on the background rate μ(t) following the ETAS model. This process is however considered as noise that masks precursory anomalies, in agreement with Ogata [1988, 2005, 2007].

[5] The regional stress state (Figure 1 and see 360° rotation animation in the auxiliary material) depends on the following static stress field

display math

function of time t and distance r along the stress field gradient with boundary conditions σ(r → ∞, t) = σ*0 and σ(r = 0, t) = σ0 +  inline imaget [Mignan et al., 2007]. σ0 is the stress drop associated to a hypothetical silent slip occurring at time t0 on the fault [Wyss and Habermann, 1988; Ogata, 2005, 2007], h the depth of the fault segment base, inline image the secular stress loading rate on the fault and σ*0 the crustal background stress. Only horizontal variations in σ(r,t) are considered with changes with depth d assumed negligible (dr). In the original N-C PAST, two categories of background events were defined with densityδbm and δb0 for quiescence regime and normal background regime, respectively [Mignan et al., 2007; Mignan, 2008]. Here we introduce a third category, with density δbpfor activation regime, to include short-term foreshocks in the proposed framework, withδbmδb0δbp. Each event category is associated with a given stress range [e.g., Ogata, 2005; King, 2007]. Two stress thresholds are defined: σ*0 − Δσ* the limit between quiescence and normal activity, and σ*0 + Δσ* the limit between normal activity and activation. The boundary layer σ*0 ± Δσ* corresponds to the net stress change in the crust with σ*0 = 0 [e.g., King, 2007]. Δσ* is expected to be small since changes of the order of millibars can trigger relative quiescence and activation [e.g., Ogata, 2005].

Figure 1.

Space-time volume representing the quiescence envelope (in blue) and the activation envelope (in red) predicted by the N-C PAST before the occurrence of a mainshock at timetf and location (x0, y0). The spatial extent of the loading stress field is assumed circular for illustration purposes, with inline image. The blue surface corresponds to the stress threshold σ*0 − Δσ* and the red surface to σ*0 + Δσ*, following equation (1). Background events are represented by spheres with density δbm below σ*0 − Δσ*, δbp above σ*0 + Δσ* and δb0 in between, with δbmδb0δbp. (See 360° rotation animation in the auxiliary material.)

[6] The spatiotemporal extent of the quiescence envelope (i.e., stress shadow, blue surface in Figure 1) is described by

display math

by applying the conditions σ(rQ, t) = σ(0, t*m) = σ*0 − Δσ* to equation (1) [Mignan et al., 2007]. It is the decrease of rQ with time t due to stress accumulation on the fault, which explains the phenomenon of ASR, as well as inward migration of events and the Mogi Doughnut. The abrupt appearance of quiescence at t0 is in agreement with observations [Wyss and Habermann, 1988]. The spatiotemporal coupling of quiescence and ASR predicted by the N-C PAST has been observed prior to several large earthquakes [Mignan and Di Giovambattista, 2008; Mignan, 2011]. It should be emphasized that observation of these different patterns is not systematic but that the N-C PAST can also explain the possible non-occurrence of specific patterns (seeSection 4).

[7] The activation envelope (red surface in Figure 1) is similarly described by

display math

where σ(rA, t) = σ(0, t*p) = σ*0 + Δσ*. Δσ* ∼ 0 yields t*m ≈ t*p ≈ t*, the time limit between quiescence/ASR and short-term activation. While the original N-C PAST predicts the occurrence of the mainshock when the stress on the fault reaches backσ*0 [Mignan et al., 2007], a delay of the rupture would lead to more stress accumulation, potentially up to σ > σ*0 + Δσ* (see Section 4). If the fault is in a state of overloading, the delay is expected to be short. The N-C PAST also predicts that these short-term foreshocks concentrate in the epicentral area, as typically observed [e.g.,Jones and Molnar, 1979]. The observation of repetitive seismic bursts in short-term foreshock sequences [Bouchon et al., 2011] is also compatible with N-C PAST activation due to overloading. Event inward migration and activation in the epicentral area can also explain the potential increase in event size (i.e.,b-value decrease) prior to a large earthquake if theb-value is lower near the main fault than in outer regions [Page et al., 2011]. Noteworthy, the N-C PAST also predicts an outward migration of events during activation, in agreement with observations in relocated foreshock sequences [e.g.,Jones et al., 1982].

[8] The non-stationary background rate (Figure 2) then takes the form

display math

for a fixed volume of maximum extent rmax = max(rQ), with k a geometrical parameter, d the spatial dimension and rQ and rA described respectively by equations (2) and (3) (Mignan et al. [2007] only described the case t ≤ t*m). Figure 2a represents a stochastic iteration of equation (4) obtained by thinning [Lewis and Shedler, 1979] as well as the theoretical curve. ASR follows a power-law as originally proposed byBufe and Varnes [1993]. Activation however differs from the simple t−1 type behavior originally proposed by Jones and Molnar [1979]. The inverse Omori law is however empirical and the N-C PAST leads to a behavior visually close to thet−1 type behavior. Noteworthy, the phenomenon of decelerating seismicity [e.g., Papadimitriou, 2008] is consistent with the appearance of quiescence, as can be seen in Figure 2.

Figure 2.

Precursory time series. (a) Stochastic iteration of equation (4), in black and N-C PAST theoretical curve (cumulative and non-cumulative forms), in orange. (b) Time series observed prior to the 2009 L'Aquila earthquake, in black and N-C PAST fit, in orange. The yellow envelope represents the 0.01 and 0.99 percentiles obtained from 1,000 simulations. See explanation of the top bar color scheme inFigure 3.

3. Real Data Analysis

[9] We fitted equation (4) to the time series observed prior to the April 2009 Mw = 6.3 L'Aquila earthquake (Figure 2b), for which the existence of seismic precursors has been ardently debated [Hall, 2011]. Earthquake data was retrieved from the Bollettino Sismico ( for magnitudes m ≥ 1.8 (i.e., conservative completeness magnitude Mc estimate) in a cylinder of radius rfix = 80 km centered on the mainshock epicenter (13.380°E, 42.342°N). We assume the role of aftershocks to be negligible, as 95% of events are expected not to produce any aftershock (using the parameters of Marzocchi and Zhuang [2011]). The largest event of the time series is expected to produce 18 aftershocks in normal conditions, but its occurrence during the activation stage makes it difficult to access the role of triggering in a state of overloading. To fit the data, we used k = π and d = 2 assuming a stress field of circular extent, h = 12.9 km [Anzidei et al., 2009], inline image = 2.5/300 = 0.0083 MPa/yr [Maercklin et al., 2011; Walters et al., 2009], the bin width Δt = 0.02 yr (∼1 week) and Δσ* = 4 10−5 MPa such that t*m ≈ t*p (i.e., inline image). Parameters derived from the dataset are t0 = 2008.1, t*mt*p ≈ t* = 2009.234, tf = 2009.264, δbm = 0.0075, δb0 = 0.0160 and δbp = 7.4403 with δ in events/yr/km2. Based on the computation of the temporal joint log-likelihood, we found that the theoretical time series provides a good match to the observations (Figure 2b) with a corrected Akaike Information Criterion AICc(N-C PAST, 6 parameters) = 709 versus AICc(Poisson with rateμ = 6.29 events/Δt, 1 parameter) = 856. Only for a minimum magnitude cutoff Mco ≥ 3.4 is the time series better fitted by a Poisson process. It shows that a better evaluation of microseismicity is crucial [e.g., Mignan et al., 2011; Mignan, 2012]. Studies that claim that precursory patterns are unreliable commonly use Mco = 4.0 [Hardebeck et al., 2008; Marzocchi and Zhuang, 2011].

4. Precursory Paths

[10] The occurrence of the different precursory patterns in the N-C PAST depends on the probability of occurrence of silent slip on the fault 1-Pr(σ0 = σ*0 = 0) with Δσ* ∼ 0, and on the probability of occurrence of the mainshock at a given stress state Pr(σf) = Beta(α, β) bounded on the stress interval [σ0, σmax] with σf the failure stress on the fault (Figure 3a). The Beta distribution, chosen for its flexibility, represents the combination of aleatoric and epistemic uncertainties on the process of rupture. It follows that the probability of occurrence of quiescence, ASR and activation can be roughly described by Pr(Q) = 1-Pr(σ0 = 0), Pr(ASR) = Pr(Q)Pr(σf ≥ 0) and Pr(A) = Pr(σf > 0), respectively. Based on this set of combinations, we defined a logic tree (referred to as N-C PAST precursory tree) that describes all possible precursory paths before a large earthquake due to stress accumulation on the fault in the range [σ0, σf] (Figure 3b). It is an analogue to the established probabilistic volcanic hazard assessment based on event trees [Newhall and Hoblitt, 2002], both earthquake and volcanic eruption processes being based on the concept of stress accumulation. The logic tree includes five different precursory paths including one where no precursor is observed. The L'Aquila earthquake precursory time series corresponds to path #1 (coupling of quiescence and ASR, followed by activation).

Figure 3.

(a) Cumulative probability of failure versus failure stress, described by the cumulative Beta distribution bounded on the interval [σ0, σmax]. The fault (at r = 0) is more likely to rupture at the background stress level σ*0 = 0 but may also occur at lower or higher stresses due to the combination of aleatoric and epistemic uncertainties. Only precursors expected in the range [σ0, σf] occur, with σfthe failure stress. Precursors considered are quiescence (Q), ASR and activation (A). (b) N-C PAST precursory tree. 5 different precursory paths are predicted, including the scenario where no precursor is observed (in green). Each precursor is represented by a different color, yellow (Q), orange (ASR) or red (A), in an attempt to represent the degree of apparent stress accumulation in an OEF perspective.

5. Conclusions

[11] The N-C PAST unifies and categorizes all basic seismicity precursors claimed in the literature in a same physical framework. Precursory quiescence is characterized by event densityδbmδb0 and depends on the occurrence of silent slip at time t0 < tf. ASR depends on the occurrence of quiescence and occurs in the same space-time window with event density per unit of time accelerating fromδbm to δb0. Short-term activation is characterized byδbpδb0, is independent of the occurrence of quiescence/ASR and is expected to occur only if the fault is in a state of overloading. Event inward migration and the Mogi doughnut are associated with quiescence/ASR, all forming medium-term far-field patterns (∼10 fault lengths). Event outward migration andb-value decrease are associated with activation, all forming short-term near-field patterns (∼1 fault length). The N-C PAST thus provides new guidelines to reinvestigate seismicity precursors in a systematic way.


[12] The author thanks Alan Kafka and John Rundle for their constructive comments. The work leading to this publication has received funding from the European Union's Seventh Framework Programme (FP7) under contracts no. 262330 (Network of European Research Infrastructures for Earthquake Risk Assessment and Mitigation, NERA) and no. 282862 (Strategies and tools for Real-Time Earthquake Risk Reduction, REAKT).

[13] The Editor thanks Alan Kafka and John Rundle for their assistance in evaluating this paper.