Controls of seismogenic zone width and subduction 1

velocity on interplate seismicity: insights from 2

analog and numerical models 3

FABIO CORBI1,*, ROBERT HERRENDÖRFER2, FRANCESCA FUNICIELLO3, YLONA 4

VAN DINTHER2 5

6

1 Géosciences Montpellier Laboratory, University of Montpellier, Montpellier, France. 7

2 Institute of Geophysics, ETH Zurich, Sonneggstrasse 5, CH-8092, Zurich, Switzerland. 8

3 Dipartimento di Scienze, University “Roma Tre”, L. S.L. Murialdo, 1, 00146, Rome, Italy. 9

*Correspondence to: fabio.corbi3@gmail.com 10

Keywords: subduction megathrust earthquakes, analog modelling, numerical modelling, subduction 11

velocity, seismogenic zone width 12

13

Abstract 14

Correlations between geodynamic parameters and interplate seismicity characteristics in subduction 15

zones are generally weak due to the short instrumental record and multi-parameter influences. To 16

investigate the role of subduction velocity Vs and the width of the seismogenic zone W on 17

maximum magnitude Mmax, seismic rate τ, charateristic recurrence rate τc, and moment release rate 18

MRR we add synthetic data sets from simplified analog- and numerical models to a global database 19

of natural subduction zones seismicity. Our models suggest that Mmax increases with W and is 20

unaffected by Vs, τ increases with Vs, τc increases with Vs/W, MRR increases with Vs*W. In 21

nature, only the positive correlation between Vs and τ is significant. Random sampling of our time 22

series suggests that a minimum span of at least the characteristic recurrence time would be required 23

to witness Mmax and to observe similarly strong correlations in nature. 24

AUTHOR’SPOST-PRIN

T

25

1. Introduction 26

The world largest earthquakes occur along the subduction megathrusts: the large faults between the 27

subducting and overriding plates (Figure 1a). Along megathrusts, stress is locally built up as a 28

consequence of friction that acts against plate convergence. When the fault strength is overcome, 29

stress is released through a variety of slip modes that range from slow slip events to regular 30

earthquakes [e.g., Ide et al., 2007; Peng and Gomberg 2010]. During regular earthquakes, slip can 31

quickly reach tens of meters and involve hundreds of kilometers of the fault, resulting in moment 32

magnitude Mw 8 or larger earthquakes. These earthquakes, in combination with tsunamis that they 33

may trigger, can cause extensive human losses and severe damages in densely populated areas, as 34

for the 2011 Mw 9.0 Tohoku-Oki event (Japan). 35

The seismic signature differs between subduction zones under various aspects, including the 36

maximum magnitude and earthquake productivity [e.g., Ide 2013]. Such variability has been 37

initially attributed to the combination of age of subducting plate and plate-rate convergence [Uyeda 38

and Kanamori 1979; Ruff and Kanamori 1980]. However, this former idea failed in explaining the 39

occurrence of events like the 2004 Sumatra and the 2011 Tohoku-Oki, pushing the scientific 40

community to find other possible links between subduction interplate seismicity and long-term 41

geodynamic parameters [e.g., Heuret et al., 2011; Schellart and Rawlinson 2013]. 42

Two parameters, namely the width of the seismogenic zone W and subduction velocity Vs, have 43

been proposed to exert a key role on interplate seismicity for their first order control on deformation 44

rate and coupling area between plates, respectively [e.g., Kanamori and Brodsky 2001; Scholz and 45

Campos, 2012]. However, previous studies focusing on the possible relationships between 46

seismicity and W or Vs lead to contradicting conclusions. While Kelleher et al. [1974] concluded 47

that the maximum magnitude increases with the width of the contact zone between the converging 48

plates, Pacheco et al. [1993] and Heuret et al. [2011] found no significant correlation between these 49

two parameters. Similarly for Vs, Ruff and Kanamori [1980] and Jarrard [1986] noted that the 50

earthquake magnitude potential of subduction zones is positively correlated with relative plate 51

motions, but then it is not clear why the fastest subduction zones (i.e., Tonga and the New 52

Hebrides) have not experienced a Mw > 8.0 earthquake since 1900 [e.g., Heuret et al., 2011]. 53

These observational studies provide a snapshot of an intertwined truth, as they are confronted with a 54

too short instrumental time span within which multiple, intercorrelated geodynamic parameters 55

affect the seismicity at the same time. Our instrumental record dates back to 1900, while thousands 56

of years of data are required to cover several complete seismic cycles and reveal each subduction 57

zone seismic characteristics [e.g., McCaffrey, 2008]. Unfortunately paleoseismological 58 AUTHOR’SPOST-PRIN

T

investigations [e.g., Cisternas et al., 2005; Sieh et al., 2008] do not provide enough spatio-temporal 59

coverage to extend our global analysis that far back in time. Additionally, contemporary variations 60

of several geodynamic parameters within tens of subduction zones makes it difficult to identify one-61

to-one correlations and cause-effect relationships [e.g., Heuret et al., 2011]. 62

To overcome these limitations, we use complementary analog- and numerical models to investigate 63

how W and Vs control the seismic behavior of subduction interplate events. Both analog and 64

numerical models have recently become a robust tool of investigation [e.g., Corbi et al., 2013; van 65

Dinther et al., 2013; van Dinther et al., 2014a,b; Funiciello et al., 2013]. Their main advantage is the 66

capability to simulate tens to hundreds of seismic cycles per model. Properly scaled analog models 67

are physically self-consistent (i.e., stresses and strain evolve spontaneously in response to applied 68

boundary conditions), while numerical models have also the advantage of being more flexible and 69

efficient for a parametric study. A previous benchmark ensures the similarity between the two 70

modelling techniques [van Dinther et al., 2013]. To describe the seismic behavior of our models we 71

use the following parameters: maximum magnitude Mmax, seismic rate τ, characteristic seismic 72

rate τc and moment release rate MRR (Supporting information Text S1). These parameters allow us 73

to answer the following questions: a) how large is the strongest event? b) how many events occur in 74

a given time period? c) what is the recurrence time of the largest events; d) how fast is the release of 75

seismic energy for a given (geometric and kinematic) subduction configuration in a given time 76

window? We finally compare the modelling results with a database that includes geometrical, 77

mechanical and seismological properties of worldwide subduction zones [Heuret et al., 2011]. The 78

aim of this comparison is to identify potential cause-effect relationships among the investigated 79

parameters that may be flawed by the short observation interval and multi-parameter influence. 80

81

2. Methods 82

We use three complementary sources of information to analyze the role of W and Vs. Besides a 83

global database of natural subduction zones [Heuret et al., 2011], new insights are provided by a 84

systematic parameter study executed with an analog and numerical models that have been described 85

in detail by Corbi et al. [2013] and van Dinther et al. [2013], respectively. Here we recall their 86

basics, while a summary of model performance and scaling is provided in Supporting information 87

Text S2. The setup of our models (Figure 1b,c) mimics the subduction environment (Figure 1a). A 88

10° dipping, flat and undeformable plate, analog of the downgoing slab, underthrusts a visco-elastic 89

wedge (i.e., the overriding plate). The frictional interaction between converging plates leads to a 90

stress build up, which is episodically released by frictional instabilities propagating along the base 91

of the model (i.e., the analog earthquakes). A velocity-weakening zone (analog of the seismogenic 92 AUTHOR’SPOST-PRIN

T

zone) is confined between two velocity-strengthening zones at its updip and downdip limits [e.g., 93

Scholz 1998; Marone and Saffer, 2007]. Vs is applied as a boundary condition in the numerical 94

model and via a stepping motor in the analog models. W is varied together with the depth Dz of the 95

downdip limit according to the worldwide geometry of subduction zones (Supporting information 96

Figure S1 and Table S1). The numerical models systematically investigate the role of Dz at constant 97

Vs. Since Dz plays only a secondary role on the selected seismological parameters, related results 98

are shown in Supporting information Text S3. The numerical models are 2D and the analog models 99

are quasi-2D as the geometric, kinematic and frictional properties of the setup are constant along the 100

width. Both analog and numerical models are monitored for a time scaled to nature of > 105 yr, 101

allowing to recognize tens of seismic cycles per model. 102

Correlations between the investigated parameters (i.e., W and Vs) and seismological ones (i.e., 103

Mmax, τ, and MRR) are provided by means of Spearman correlation coefficient R and p-value 104

[Press et al., 1996] both for our models and for natural subduction zones. The Spearman correlation 105

coefficient is a non-parametric measurement that indicates how well two variables follow a 106

monotonic function. With respect to the more common Pearson correlation coefficient, the 107

Spearman correlation coefficient has the advantage of dealing with skewed data or outliers. This is 108

particularly important for the natural dataset where correlations may be flawed by few fast 109

subduction zones (i.e., North Tonga and New Hebrides; Heuret et al., 2011). 110

2.1 Analog models 111

The wedge is made of 2.5%wt Pig Skin gelatin (see Di Giuseppe et al. [2009] for details on 112

rheological properties and preparation and Corbi et al. [2011] for frictional behavior of gel against 113

sandpaper). The seismogenic zone is modeled with sandpaper while the updip and downdip 114

velocity-strengthening zones are modeled with acetate plastic sheets. The setup is designed 115

maintaining Dz constant during the experimental run. Monitoring is performed via particle image 116

velocimetry PIV method [MatPIV; Sveen, 2004]. The velocity field is used to calculate model 117

deformation time series and earthquake source parameters [Corbi et al., 2013]. 118

2.2 Numerical models 119

The continuum-mechanics based numerical model solves for the conservation equations of mass 120

and momentum for an incompressible medium with a visco-elasto-plastic rheology [Gerya and 121

Yuen, 2007]. The governing equations are discretized on a fully-staggered finite-difference grid in 122

combination with a marker-in-cell technique, in which advected markers carry material properties. 123

The interface between the wedge and the subducting plate is modelled as a frictional boundary 124

layer, in which a non-associative Drucker-Prager plastic flow law is applied with a pressure-125 AUTHOR’SPOST-PRIN

T

dependent yield strength. To simulate analog earthquakes, van Dinther et al. [2013] introduced a 126

velocity dependent friction and the inertial term to the momentum equations. 127

2.3 Natural database 128

We compare modelling results with the natural global database compiled by Heuret et al. [2011]. 129

This database includes Mw≥5.5 interplate events occurred worldwide in the 1976-2007 interval 130

sampled from the Harvard CMT catalog [Dziewonski and Woodhouse, 1981] and Mw>7 in the 131

1900-1975 interval sampled from the Centennial catalog [Engdahl and Villasenor, 2002]. Besides 132

seismological information, the database includes W and Vs of worldwide subduction zones. As our 133

models have a subduction angle of 10°, we highlight the subsample of the original database with 134

shallowly dipping seismogenic zones (dip angle <15°). 135

136

3. Results and Analysis 137

3.1 Control on Mmax 138

In our models, Mmax overall increases with W, while it is mostly insensitive to Vs (Fig. 2a,b, Fig. 139

3a,b). The correlations between W and Mmax are high for both the numerical and analog models 140

(i.e., R= 0.93 and R=0.73 for numerical and analog models, respectively), while the correlations 141

between Vs and Mmax are low (R=0.15 and R=0.22 for numerical and analog models, 142

respectively). The largest events rupture at least the entire seismogenic zone width (i.e., a rupture 143

width equal or larger than W, dashed line in Fig. 2a,b) and also propagate into the up- and downdip 144

velocity-strengthening regions. This penetration distance depends on plates interface frictional 145

properties and downdip location of the seismogenic zone with respect to the trench and backstop. 146

For W>180 km, Mmax saturates at Mw~ 9.5. 147

Statistically insignificant correlations between Mmax and W (R=0.15, p>0.05) and between Mmax 148

and Vs (R=0.09, p>0.05) are observed in nature. 149

3.2 Control on τ and τc 150

The seismic rate τ increases mainly as a function of Vs (Fig. 2c,d; Fig 3c), which leads to a high 151

correlation between Vs and τ (R= 0.82 and R=0.95 for numerical and analog models, respectively). 152

Note that τ is lower in the analog models than in the numerical models (Fig. 3c). This difference is 153

due to the lower interseismic locking of the analog models (~40%) with respect to the numerical 154

ones (~100%), which reduces the effective strain rate. Almost identical τ is obtained taking into 155

account this reduced interseismic locking of the analog models (Fig 3c). W has no consistent effect 156 AUTHOR’SPOST-PRIN

T

on τ in our models (R=-0.4 and R= 0.09 for the numerical and analog models, respectively; Fig. 157

2c,d). 158

τc, in contrast to τ, clearly decreases as a function of both W and Vs (R=-0.75 and R=-0.14 for 159

numerical and analog models, respectively; Fig. 2g,h). Higher correlation is observed when 160

comparing τc to the Vs/W ratio (R=0.95 and R=087 for numerical and analog models, respectively; 161

Fig. 3e). 162

In nature, the correlation between τ and Vs is the highest observed and significant (R=0.57; p<0.05 163

Fig. 3h). 164

3.3 Control on MRR 165

MRR increases both with Vs and W. Similarly to Mmax, MRR increases with W for W<180km 166

due to the strong influence of the increasing size of the largest earthquakes in wider seismogenic 167

zones. For W>180km MRR approaches our setup’s limit of Mmax (Fig. 2g,h). The correlations 168

between MRR and Vs are higher (R=0.73 and R=0.72 for numerical and analog models, 169

respectively) than correlations between MRR and W (R=0.45 and R=0.54 for numerical and analog 170

models, respectively). Correlations are higher when comparing MRR with the product W*Vs 171

(R=0.93 and R=0.90 for numerical and analog models, respectively; Fig. 3d), while in nature the 172

MRR versus W*Vs correlation appears to be insignificant (R=0.22; p>0.05 Fig. 3i). 173

3.4 Observational time window length for capturing Mmax 174

In our models as well as in nature, Mmax evolves with observation length due to earthquake 175

magnitude variability. Short observation intervals may cause an underestimation of Mmax, while 176

long enough time series contain with high probability at least one event with magnitude close or 177

equal to the “real” Mmax. 178

Aiming to quantify the observation interval for capturing Mmax, we recorded the maximum rupture 179

widths mRw in 104 time windows of certain durations with random initial starts. We proceed in two 180

steps: 181

First we check for the evolution of the temporary maximum rupture width. We normalize the 182

temporary mRw by the final mRw, and the time window length Tw, by the characteristic recurrence 183

time Tc (where Tc=1/τc). In most models, a plateau is reached for rupture widths Rw> ~0.8 * mRw 184

(Figure 4a) justifying the use of this threshold (i.e., 0.8*mRw) as the criterion for random time 185

window search. 186

Then, we determined the percentage of time windows that contain at least one event with rupture 187

width Rw>0.8*mRw (Fig. 4a). Figure 4b shows the percentage of successful time windows as a 188 AUTHOR’SPOST-PRIN

T

function of Tw/Tc. In the case of perfect periodicity, we would expect a 100% probability of 189

finding the largest earthquake when Tw/Tc=1. For the majority of our analog and numerical 190

models, which have a stronger variability in seismic behavior, the percentage of successful time 191

windows is around 50% for Tw/Tc=1 and approaches a plateau at 90% for 1≤Tw/Tc≤5. For the 192

analog models with a clear outlier in the time series, a time window is unlikely (i.e. <50%) to be 193

successful with a length of Tw=5*Tc. In this cases, the time series is not long enough to experience 194

events of similar size. 195

196

4. Discussion 197

The combination of our models with natural observations, each with their own non-overlapping 198

uncertainties and limitations (Supporting information Text S4), leads to the following discussion 199

points. 200

Our models show that Mmax is mainly controlled by W as the rupture potential increases with W. 201

This is in agreement with studies suggesting that a large downdip extent of the seismogenic zone is 202

required for the occurrence of mega-events [Kelleher et al., 1974; Uyeda and Kanamori, 1979; Lay 203

et al., 1982]. In our models, the largest ruptures saturate the entire seismogenic zone and propagate 204

into the velocity-strengthening regions, resulting in Mmax larger than those predicted by simple 205

Mw-Rw scaling for Rw=W [Blaser et al., 2010]. The location of several subduction zones (e.g., 206

East Alaska, Japan; Figure 3f) above this Mw-W scaling suggests that the occurrence of larger than 207

expected Mmax may also occur in nature. Regarding the rupture potential to propagate outside the 208

seismogenic zone, an alternative mechanism to the inefficient arresting effect of velocity 209

strengthening regions observed in our models is dynamic fault weakening, as suggested for the 210

2011 Mw9.0 Tohoku earthquake [Noda and Lapusta, 2013]. 211

Random sampling of our time series suggests that an observation history > 5 Tc would provide high 212

chances (i.e., > 90%) to observe at least one event that ruptured the 80% of the downdip of the 213

megathrust, and in turn displaying nearly the full seismic potential. A similar conclusion (i.e., Tw > 214

5 Tc) for determining recurrence time of Mw>9 events was constrained statistically for natural 215

megathrust earthquakes [McCaffrey, 2008]. This limited observation span with respect to the 216

seismic cycle duration might explain: a) why Mmax for the majority of natural subduction zones is 217

smaller than Mmax predicted on the base of Mw-W scaling (Fig. 3e); and b) why the observed 218

Mmax of the majority of subduction zones are more similar to the median magnitude of our models 219

(Fig 3a). A striking example is the 2011 M9.0 Tohoku earthquake with a rupture width of 200 km 220

[Romano et al., 2014]. Before this event the instrumentally recorded Mmax on the Japan segment 221

was 8.3, which is lower than expected from its W=161km [Heuret et al., 2011; Fig. 3e]. It should be 222 AUTHOR’SPOST-PRIN

T

noted, however, that by using the scaling law [Blaser et al., 2010], we assume that Mmax scales 223

with Rw. Although a large downdip Rw is linked to a wide along-strike rupture lengths [e.g., Wells 224

and Coppersmith, 1994], it is debatable whether this scaling law holds at very high Mw since 225

unrealistically high widths would be expected for Mw>9.2. Such great megathrust earthquakes have 226

a large along-strike component (e.g., the 2004 Sumatra-Andaman earthquake; Shearer and 227

Burgmann, 2010), which is not taken into account in our 2D models. Therefore, Mmax might be 228

controlled by other factors that have been suggested to control the along-strike rupture propagation, 229

such as the trench-parallel extent of a subduction zone segment, the upper plate strain [Heuret et al., 230

2012] and interplate roughness [e.g., Wang and Bilek 2014]. 231

In our models as well as in natural subduction zones, no significant correlation between Mmax and 232

Vs is obtained (Fig. 2a,b; Fig 3b,g). Based on natural observations, it is unclear whether Vs 233

individually controls Mmax. On one hand, Heuret et al. [2011] identified a feedback between slow 234

subduction zones, shallow dipping slab and wide seismogenic zones, which is proposed to lead to 235

the generation of the largest Mmax. On the other hand, Uyeda [1983] associated the largest 236

magniudes with fast subduction zones due to their high strength of the mechanical coupling of the 237

subduction megathrust. In our models high Vs is associated to high τ and τc increasing the 238

probability to observe the largest events. Therefore, previously suggested Mmax-Vs correlations for 239

nature might have been affected by this potential observational bias. This is in agreement with 240

statistical simulations suggesting that the occurrence of Mw>9 earthquakes cannot be excluded at 241

any subduction zone, independently of Vs [McCaffrey, 2008]. However, it should be noted that our 242

models do not take into account the thermal evolution of a subduction zone, which is suggested to 243

control the vertical extent of the seismogenic zone [e.g., Heuret et al., 2011; Dal Zilio et al. 2016]. 244

More complete physical models self-consistently resolving both subduction dynamics and 245

seismogenesis are thus required to provide further insights. 246

Vs is found to exert a primary control on τ and τc both in analog and in numerical models, which 247

supports the direct correlation between Vs and τ that is found when considering subduction 248

megathrust events only [e.g., Heuret et al., 2011] and the whole convergent margin seismicity [Ide 249

2013]. When taking the different interseismic locking in the analog and numerical models into 250

account, the τ versus Vs and MRR versus Vs relationships of both models tend to overlap (Figure 251

3c-d). The reduced interseismic locking in the analog models may be explained by creeping that 252

occurs at the base of the gelatin wedge, which in turn reduces the seismic coupling. Natural 253

subduction zones display a wide range of seismic coupling [e.g., Scholz and Campos, 2012], whose 254

calculation is, however, affected by the short observation time. Moreover, interseismic locking may 255

also change through subsequent seismic cycles as suggested for the Mentawai segment of the Sunda 256

megathrust [Philibosian et al., 2014]. Aseismic slip transients belonging to the slow slip phenomena 257 AUTHOR’SPOST-PRIN

T

observed in subduction zones [e.g., Peng and Gomberg, 2010], release periodically a fraction of the 258

accumulated elastic energy of convergent margins, and, thereby reduce the long-term locking [e.g., 259

Radiguet et al, 2016]. Therefore, different sources affecting the amount of locking in subduction 260

zones may explain the scatter of τ and Vs in the natural database (Fig. 3h). 261

In our models, MRR is controlled by both Vs and W (Fig 2g,h). In nature, MRR is mainly 262

influenced by the contribution of the largest event. For example, the great 1960 Mw9.5 Chile 263

earthquake alone accounts for about 30% of the total seismic energy released during the last century 264

[Heuret et al., 2011]. Our models furthermore demonstrates that the frequency of the largest events 265

is a crucial factor to estimate the long-term MRR. This frequency, τc, is shown to be controlled by 266

the ratio of Vs/W, which means that larger W reduces the recurrence rate of the largest events. This 267

weakens the positive impact of W on MRR (Fig 3e). Consequently, the fastest subduction zones 268

with medium to large W are expected to have the largest MRR. A comparison of the correlations 269

with respect to MRR and τc would require multiple cycles of the largest events, which is clearly 270

beyond the available data. Paleoseismological studies [e.g. Cisternas et al., 2005; Sieh et al., 2008] 271

might provide such estimates in some subduction zones to allow for a comparison in future studies. 272

273

5. Conclusions 274

The following conclusions are drawn linking our models to natural observations: 275

- Mmax is strongly influenced by W. Subduction zones where the largest observed rupture 276

widths are smaller than W, may have shown only a fraction of their seismic potential. 277

However, the role of lateral megathrust extent (not taken into account in our models) might 278

become dominant for ruptures with Mw>9. 279

- Mmax is likely independent from Vs. The previously observed Mmax versus Vs correlation 280

in nature is likely due to the short observation interval and/or feedbacks between Vs and W. 281

- τ is mainly influenced by Vs. Differences in interseismic locking may explain the scattered 282

relationship between Vs and τ in nature. 283

- W does not play a relevant role in controlling τ in our models and in nature, but our models 284

suggest that larger W is associated with lower τc. 285

- MRR is both strongly influenced by Vs and W due to their control on τc and Mmax. Time 286

series larger than τc may be needed to observe the full seismic potential of a megathrust and 287

posibly improve correlations in nature. 288 AUTHOR’SPOST-PRIN

T

Figures and captions 289

290

Figure 1: Schematic representation of a subduction zone in nature (a) compared to our analog (b) 291

and numerical models (c). The red lines highlights the seismogenic zone. The dashed purple 292

polygon in (b) highlights the cross sectional area analyzed by PIV. The numerical setup, including 293

gravity vector, is rotated by the subduction angle of 10˚ with respect to the analog setup. Boundary 294

conditions are given on the respective sides. 295

AUTHOR’SPOST-PRIN

T

296

Figure 2: Modelling results: Maximum magnitude Mmax (a,b), seismic rate τ (c,d), characteristic 297

rate τC (e,f), and moment release rate MRR (g,h) as a function of W and Vs (colorbar) in numerical 298

models (top row) and analog models (bottom row). The black dashed line in panels a and b 299

represents the Mw-Rw scaling [Blaser et al., 2010] for Rw=W. 300

301

AUTHOR’SPOST-PRIN

T

302

Figure 3: Summarized role of W and Vs on the maximum (squares) and medium magnitude 303

(triangles; a,b). Role of Vs on τ (c). Control of Vs*W on MRR (d) and control of Vs/W on τc (e). 304

The roles of W and Vsn in the dataset of natural subduction zones [Heuret et al., 2011] is shown for 305

comparison in terms of Mmax (f,g), τ (h), and MRR (i). Each panel report the Spearman correlation 306

coefficient R and the p-value. Subduction zones with a dip angle≤15˚: Sumatra (SUM), Cocos 307

(CO), Andaman (AN), Java (JV), E-Alaska (EA), Western Aegean (WA), Antilles (AT), Timor 308

(TI), Nankai (NA), Southern Chile (SC). Green stars in (f) represents Mmax for the Japan 309

subduction zone segment before (i.e., Mw=8.3) and after the 2011 Mw=9.0 Tohoku earthquake, 310

respectively. 311

312

AUTHOR’SPOST-PRIN

T

313

Figure 4: Random sampling of the model time series. Evolution of the temporary maximum rupture 314

width Rw normalized by the final maximum rupture width mRw as a function of normalized time 315

window length Tw/Tc (a). Percentage of successful time windows sampling an event with 316

Rw>0.8*mRw as a function of Tw/Tc (b). 317

318

Acknowledgements 319

FC and RH ran and analyzed the analog and numerical models, respectively, while together taking the lead in 320paper writing. All authors contributed into concept development and further writing of the paper. Riccardo 321Lanari is acknowledged for his support in the laboratory. Serge Lallemand is acknowledged for fruitful 322discussions on a preliminary version of the manuscript. FC received funding from the European Union’s 323Horizon 2020 research and innovation programme under the Marie Sklodowska-Curie grant agreement No 324658034 (AspSync). RH was supported by SNSF grant 200021-153524. Analog and numerical modelling data 325are available contacting the corresponding author and RH (robert.herrendoerfer@erdw.ethz.ch), respectively. 326

AUTHOR’SPOST-PRIN

T

References 327

Blaser, L., F. Krüger, M. Ohrnberger, and F. Scherbaum (2010), Scaling relations of earthquake source 328

parameter estimates with special focus on subduction environment, Bull. Seismol. Soc. Am., 100(6), 329

2914–2926, doi:10.1785/0120100111. 330

Cisternas, M. et al. (2005), Predecessors of the giant 1960 Chile earthquake., Nature, 437(7057), 404–7, 331

doi:10.1038/nature03943. 332

Corbi, F., F. Funiciello, C. Faccenna, G. Ranalli, and a. Heuret (2011), Seismic variability of subduction 333

thrust faults: Insights from laboratory models, J. Geophys. Res., 116(B6), 1–14, 334

doi:10.1029/2010JB007993. 335

Corbi, F., F. Funiciello, M. Moroni, Y. van Dinther, P. M. Mai, L. a Dalguer, and C. Faccenna (2013), The 336

seismic cycle at subduction thrusts: 1. Insights from laboratory models, J. Geophys. Res., 118, 1–19, 337

doi:10.1029/2012JB009481. 338

Dal Zilio, L. Y. van Dinther, T. Geria (2016). Plate convergence rate controls earthquake-size distribution of 339

mountain belts. AGU abstract T22B-02. 340

Di Giuseppe, E., F. Funiciello, F. Corbi, G. Ranalli, and G. Mojoli (2009), Gelatins as rock analogs: A 341

systematic study of their rheological and physical properties, Tectonophysics, 473(3–4), 391–403, 342

doi:10.1016/j.tecto.2009.03.012. 343

Dziewonski, A. M., and J. H. Woodhouse (1981), Determination of earthquake source parameters from 344

waveform data for studies of global and regional seismicity, J. Geophys. Res.,86, 2825–2852, 345

doi:10.1029/JB086iB04p02825. 346

Engdahl, R., and A. Villaseñor (2002), Global seismicity: 1900–1999, in International Handbook of 347

Earthquake and Engineering Seismology, Part A, edited by W. H. K. Lee et al., chap. 41, pp. 665–690, 348

Academic, Amsterdam. 349

Funiciello, F., F. Corbi, Y. van Dinther, and A. Heuret (2013), Unraveling Megathrust Seismicity, Eos, 350

Trans. Am. Geophys. Union, 94(51), doi:10.1029/2010JB007993.Corbi. 351

Gerya, T., and D. Yuen (2007), Robust characteristics method for modelling multiphase visco-elasto-plastic 352

thermo-mechanical problems, Phys. Earth Planet In., 163(1-4), 83–105. 353

Herrendoerfer, R., Y. van Dinther, T. Gerya, and L. A. Dalguer (2015), Earthquake supercycle in subduction 354

zones controlled by the width of the seismogenic zone, Nat. Geosci, 8(6), 471–474, 355

doi:10.1038/ngeo2427. 356

AUTHOR’SPOST-PRIN

T

Heuret, A., S. Lallemand, F. Funiciello, C. Piromallo, and C. Faccenna (2011), Physical characteristics of 357

subduction interface type seismogenic zones revisited, Geochemistry Geophys. Geosystems, 12(1), 1–358

26, doi:10.1029/2010GC003230. 359

Heuret, a., C. P. Conrad, F. Funiciello, S. Lallemand, and L. Sandri (2012), Relation between subduction 360

megathrust earthquakes, trench sediment thickness and upper plate strain, Geophys. Res. Lett., 39(5), 361

1–6, doi:10.1029/2011GL050712. 362

Ide, S., G. C. Beroza, D. R. Shelly, and T. Uchide (2007), A scaling law for slow earthquakes., Nature, 363

447(7140), 76–9, doi:10.1038/nature05780. 364

Ide, S. (2013), The proportionality between relative plate velocity and seismicity in subduction zones, Nat. 365

Geosci., 6(8), 1–5, doi:10.1038/ngeo1901. 366

Jarrard, R. D. (1986), Relations among subduction parameters, Rev. Geophys., 24, 217–284, 367

doi:10.1029/RG024i002p00217. 368

Kanamori, H., and E. E. Brodsky (2001), The Physics of Earthquakes, Phys. Today, 54(6), 34, 369

doi:10.1063/1.1387590. 370

Kelleher, J., J. Savino, H. Rowlett, and W. McCann (1974), Why and where great thrust earthquakes occur 371

along island arcs, J. Geophys. Res., 79, 4889–4899, doi:10.1029/ JB079i032p04889. 372

Lay, T., H. Kanamori, and L. Ruff (1982), The asperity model and the nature of large subduction zone 373

earthquakes, Earthquake Predict. Res., 1, 3–71. 374

Marone, C., and D. M. Saffer (2007), Fault friction and the upper transition from seismic to aseismic 375

faulting, in The Seismogenic Zone of Subduction Thrust Fault, edited by T. H. Dixon and C. Moore, 376

pp. 346–369, Columbia Univ. Press, New York. 377

McCaffrey, R. (2008), Global frequency of magnitude 9 earthquakes, Geology, 36(3), 263, 378

doi:10.1130/G24402A.1. 379

Noda, H., and N. Lapusta (2013), Stable creeping fault segments can become destructive as a result of 380

dynamic weakening., Nature, 493(7433), 518–21, doi:10.1038/nature11703. 381

Pacheco, J. F., L. R. Sykes, and C. H. Scholz (1993), Nature of Seismic Coupling Along Simple Plate 382

Boundaries of the Subduction Type, J. Geophys. Res., 98(B8), 14133–14159, doi:10.1029/93JB00349. 383

Peng, Z., and J. Gomberg (2010), An integrated perspective of the continuum between earthquakes and slow-384

slip phenomena, Nat. Geosci., 3(9), 599–607, doi:10.1038/ngeo940. 385

AUTHOR’SPOST-PRIN

T

Philibosian, B., K. Sieh, J. Avouac, D. H. Natawidjaja, H. Chiang, C. Wu, H. Perfettini, C. Shen, M. R. 386

Daryono, and B. W. Suwargadi (2014), Rupture and variable coupling behavior of the Mentawai 387

segment of the Sunda megathrust during the supercycle culmination of 1797 to 1833. Journal of 388

Geophysical Research : Solid Earth, , 7258–7287, doi:10.1002/2014JB011200. 389

Press W. H., Teukolsky S. A., Vetterling W. T., Flannery B. P.(1996): Numerical Recipes in C, Cambridge 390

University Press. 391

Radiguet, M., Perfettini, H., Cotte, N., Gualandi, A., Valette, B., Kostoglodov, V., Lhomme, T., 392

Walpersdorf, A., Cabral Cano, E., and Campillo, M. (2016), Triggering of the 2014 Mw 7.3 Papanoa 393

earthquake by a slowslip event in Guerrero, Mexico. Nat. Geosci., advanced online 394

Romano, F., E. Trasatti, S. Lorito, C. Piromallo, a Piatanesi, Y. Ito, D. Zhao, K. Hirata, P. Lanucara, and M. 395

Cocco (2014), Structural control on the Tohoku earthquake rupture process investigated by 3D FEM, 396

tsunami and geodetic data., Sci. Rep., 4, 5631, doi:10.1038/srep05631. 397

Ruff, L., and H. Kanamori (1980), Seismicity and the subduction process., Phys. Earth Planet. Inter., 23(3), 398

240–252, doi:10.1016/0031-9201(80)90117-X. 399

Schellart, W. P., and N. Rawlinson (2013), Global correlations between maximum magnitudes of subduction 400

zone interface thrust earthquakes and physical parameters of subduction zones, Phys. Earth Planet. 401

Inter., 225, 41–67, doi:10.1016/j.pepi.2013.10.001. 402

Scholz, C. H. (1998), Earthquakes and friction laws, Nature, 391(6662), 37–42, doi:10.1038/34097. 403

Scholz, C. H., and J. Campos (2012), The seismic coupling of subduction zones revisited, J. Geophys. Res., 404

117(B5), 1–22, doi:10.1029/2011JB009003. 405

Shearer, P., and R. Bürgmann (2010), Lessons Learned from the 2004 Sumatra-Andaman Megathrust 406

Rupture. Annu. Rev. Earth Planet. Sci. 2010. 38:103–31. doi: 10.1146/annurev-earth-040809-152537. 407

Sieh, K., D. H. Natawidjaja, A. J. Meltzner, C.-C. Shen, H. Cheng, K.-S. Li, B. W. Suwargadi, J. Galetzka, 408

B. Philibosian, and R. L. Edwards (2008), Earthquake supercycles inferred from sea-level changes 409

recorded in the corals of West Sumatra., Science, 322(5908), 1674–8, doi:10.1126/science.1163589. 410

Sveen, J. K. (2004), An introduction to matpiv v.1.6.1 Eprint no. 2, ISSN 0809-4403, Department of 411

Mathematics, University of Oslo, http:// www.math.uio.no/jks/matpiv. 412

Uyeda, S. (1983), Comparative Subductology, EPISODES o.93 No 2. 413

Uyeda, S., and H. Kanamori (1979), Back-Arc Opening and the Mode of Subduction, J. Geophys. Res., 414

84(B3), 1049–1061, doi:10.1029/JB084iB03p01049. 415

AUTHOR’SPOST-PRIN

T

van Dinther, Y., Gerya, T., Dalguer, L., Corbi, F., Funiciello, F., and Mai, P. (2013). The seismic cycle at 416

subduction thrusts: 2. Dynamic implications of geodynamic simulations validated with laboratory 417

models, J. Geophys. Res. Solid Earth, 118(4), 1502–1525, doi:10.1029/2012JB009479. 418

van Dinther, Y., Mai, P.M., Dalguer, L.A., and Gerya, T.V. (2014). Modeling the seismic cycle in 419

subduction zones: the role of off-megathrust earthquakes, Geoph. Res. Lett., 41 (4), 1194-1201, 420

doi:10.1002/2013GL058886. 421

van Dinther, Y., Gerya, T.V., Dalguer, L.A., Mai, P.M., Morra, G., and Giardini, D. (2013). The seismic 422

cycle at subduction thrusts: insights from seismo-thermo-mechanical models, J. Geophys. Res., 118, 423

183-6202, doi:10.1002/2013JB010380. 424

Wang, K., and S. L. Bilek (2014), Fault creep caused by subduction of rough seafloor relief, Tectonophysics 425

610 (2014) 1–24. doi: 10.1016/j.tecto.2013.11.024. 426

Wells, D. L., and K. J. Coppersmith (1994), New Empirical Relationships among Magnitude, Rupture 427

Length, Rupture Width, Rupture Area, and Surface Displacement, Bull. Seismol. Soc. Am., 84(4), 974–428

1002. 429

AUTHOR’SPOST-PRIN

T