Supplementary Materials for

Earth’s deepest earthquake swarms track fluid ascent beneath nascent arc volcanoes

Lloyd T. White, Nicholas Rawlinson, Gordon S. Lister, Felix Waldhauser, Babak Hejrani, David A.

Thompson, Dominique Tanner, Colin G. Macpherson, Hrvoje Tkalčić, Jason P. Morgan,

Correspondence to: lloydw@uow.edu.au

This file includes:

Supplementary Text

Supplementary Figures 1 to 11

Supplementary Tables 1 to 3

Captions for Supplementary Video 1 to 3

Captions for Supplementary Datasets 1 to 5

Captions for Interactive Map Data

Captions for Interactive 3D Model

1

1

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

Supplementary Text

Visualisation of hypocenter data using eQuakes

The location (longitude, latitude and depth) of hypocenter data with magnitudes 1.0 < x <10.0,

and depths between 1 to 1000 km were visualised using eQuakes. The software was further used to

produce cross-sections of the different hypocenter catalogues perpendicular to the strike of oceanic

trenches as well as to visualize the hypocenter data at discrete time intervals (e.g. yearly, monthly,

daily). Multiple cross sections were generated along strike of the trench to interpret the 3D

geometry of the Wadati-Benioff Zone. The interpretation of the 3D slab geometry was cross-

checked with anomalous P-wave velocity structures in seismic tomographic data (Li et al., 2008).

Earthquake location using a fully non-linear grid search algorithm

For the Mariana sub-vertical pipe swarm, a total of 47,322 picks from 637 earthquakes are

used (Supp. Fig. 1a), whereas for the Izu-Bonin pipe swarm, 45,452 picks from 631 earthquakes are

used (Supp Fig 1b). The geographic distribution of picks is generally good (Supp. Fig. 2), but there is

a distinct lack of recording stations in the Pacific region to the east, which will have a detrimental

effect on our ability to accurately locate events. The vast majority of picks correspond to direct P-

and S-arrivals, but where available, other phases such as pP, sP, PcP and PcS are retained to

improve depth location accuracy.

The non-linear inverse problem that we solve is to find the location of each earthquake such

that the differences between the observed and predicted arrival times (through a prescribed Earth

model) are minimized. This is formalized by specifying an objective function S(m), where m is a

vector that defines the latitude, longitude and depth of the earthquake, such that:

2

2

23

24

25

26

27

28

29

30

31

32

33

34

35

36

37

38

39

40

41

42

43

44

45

46

S (m )=(dobs−dm (m) )TC0−1 (dobs−dm (m ) ) (1)

where dobsis a set of N observed arrival times (from a given hypocenter) with the mean subtracted,

C0−1 is a data covariance or weighting matrix which accounts for data picking uncertainty, and

dm (m ) is a set of predicted traveltimes with the mean subtracted. By removing the mean, we can

eliminate the origin time as an unknown in the inversion, since source location can be constrained

by the relative on-set times of a phase at a set of stations.

We locate the minimum of S(m) using a nested grid-search, which is a fully non-linear

technique capable of locating the global minimum of S(m) and providing valuable information on

location uncertainty. The method works by computing S(m) at a regular grid of points in latitude,

longitude and depth centered on the ISC location. The grid with the minimum value of S(m) is then

chosen as the updated location, and the grid search is repeated using a finer grid spacing. The final

solution provided by this two-step procedure is then used as the new hypocenter. For both the

Mariana and Izu-Bonin sub-vertical pipe swarms, we use an initial grid spacing of 10 km in all three

dimensions, and a grid which extends 150 km in latitude and longitude and 200 km in the depth

direction from the reference ISC (2013) location. For the second stage, the refined grid has a

spacing of 2 km in all three dimensions and a grid which extends 30 km in latitude and longitude

and 40 km in the depth direction from the initial source location. We found that the final locations

were quite robust with respect to variations in the grid spacing and grid extent. To account for data

outliers, we remove arrival time picks which differ from the corresponding prediction by more than

two standard deviations.

Although a nested grid search is likely to be less computationally efficient than other direct

search methods such as simulated annealing, genetic algorithms and the neighborhood algorithm

(e.g., Billings et al., 1994a; 1994b; Sambridge and Kennett 2001), it does heavily sample model

3

3

47

48

49

50

51

52

53

54

55

56

57

58

59

60

61

62

63

64

65

66

67

68

69

70

space, so information on hypocenter uncertainty can be achieved with little additional

computational effort. Here, we adopt the approach described by Sambridge and Kennett (1986) for

defining the 95% confidence region around the solution m¿

obtained from the nested grid search.

This is achieved by identifying the region of parameter space that satisfies:

T (m ,m¿ )=S (m )−S (m

¿ )S (m¿ )

≤χ32 (0.95 )

(N−3 )(2)

where N is the number of data picks and χ32 (0.95 )=7.815 is the Chi-square statistic for a p-value of

0.05 and three degrees of freedom. This approach is valid since the objective function of Equation 1,

with [C0−1 ]ij=δ ijσ ij where δ ij=1 when i=j and δ ij=0 otherwise, can be written as:

S (m )=∑i=1

N [dobsi −dmi (m ) ]2

σ i2

(3)

which has a chi-square distribution with N-M degrees of freedom, where M=3 is the number of

model parameters. In equation 3, σ i represents the picking uncertainty associated with diobs.

Equation 3 accounts for the effect of errors in the velocity model (which are usually unclear) by

rescaling S so that S (m¿ )=N−3 which is simply the expectation value of a chi-square distribution

with N-3 degrees of freedom. This is obtained by multiplication with the factor (N−3 ) /S (m¿ ). Both

the objective function (Equation 1) and method for computing confidence intervals assume that the

data errors are Gaussian, which will influence the final hypocenter locations and uncertainty

estimates. The validity of this assumption is tested below, when a comparison is made between

locations obtained using an L2 and L1 measure of misfit.

4

4

71

72

73

74

75

76

77

78

79

80

81

82

83

84

85

86

87

88

89

90

The method described above is applied to the ISC arrival time dataset in the presence of the

ak135 global reference model (Kennett et al., 1995), which is the default standard model for global

earthquake location (Engdahl et al., 1998). The standard deviation of the picking error for P-phases

and S-phases are set to a uniform value of σ P=0.5 s and σ S=1.0 s which is towards the higher end

of what is considered the normal range for manual picks of teleseismic phases for a trained analyst

(e.g., Douglas et al., 1997; Leonard 2000).

Supp. Fig. 1 compares the relocated events for both the Mariana and the Izu-Bonin sub-vertical

pipe swarms with their respective ISC locations; in this case, the relocated events are colored

according to the limits of the 95% confidence region in depth, which is here defined as the average

distance between source ms=(rs , θs , ϕs ) (radius, latitude and longitude) and the two intersection

points of (r , θs, ϕs ) with the boundary of the 95% confidence region defined by T (m,ms ). In the

horizontal direction, the limit of the 95% confidence region is calculated as the mean of (1) the

average distance between ms=(rs , θs , ϕs ) and the two intersection points of (r s ,θ , ϕs ) with the

boundary of the 95% confidence region and (2) the average distance between ms=(rs , θs , ϕs ) and

the two intersection points of (r s ,θs , ϕ) with the boundary of the 95% confidence region. In the

following tests, we use these measures as proxies for the vertical and horizontal uncertainty in

earthquake location.

5

5

91

92

93

94

95

96

97

98

99

100

101

102

103

104

105

106

107

108

109

110

Depth distribution of well constrained events

In general, the more picks available for an earthquake, the more accurate the final location,

although data noise, phase type and geographic distribution of recording stations also plays a role.

Supp. Fig. 3 compares two subsets of located events for the Mariana and Izu-Bonin earthquake

pipes; the first (Supp. Fig. 3, top) only displays earthquake locations which were constrained by a

minimum of 25 picks, while the second (Supp. Fig. 3, bottom) only displays earthquake locations

with a depth uncertainty no larger than 10 km (Mariana) or 15 km (Izu-Bonin). The advantage of

using uncertainty estimates is that they account for the sensitivity of the location parameters to the

arrival time picks, and where redundancy occurs, will take into account inconsistencies in the data.

However, when both the statistical distribution of the noise and its standard deviation are largely

assumed, uncertainty estimates become less reliable. As such, it is useful to consider both sets of

results. For the Mariana and Izu-Bonin sub-vertical pipes, both subsets of events quite clearly

delineate the pipe, although it could be argued, in the case of the Mariana, that gaps are present

between ~120-140 km and ~160-220 km depth. If we gradually relax the depth uncertainly

criterion, these gaps become filled with events that are less accurately located. The common feature

of these events is that they are relatively low magnitude, which makes it unsurprising that they

have fewer high-quality picks and hence are more poorly constrained in depth. Nevertheless, with a

maximum depth uncertainty set to ~20 km, these “gaps” are largely filled, so whether they actually

exist or are zones in which only smaller earthquakes occur (which in itself would be worthy of

further investigation in the future) is unclear. It is worth noting, however, that the double difference

relocation method does not identify any gaps in the Mariana earthquake pipe (see Figure 2 of main

manuscript). These tests support the existence of the two pipe-like swarms of seismicity, since

removing the more poorly constrained events does not result in a loss of the sub-vertical pipe-like

geometry.

6

6

111

112

113

114

115

116

117

118

119

120

121

122

123

124

125

126

127

128

129

130

131

132

133

134

135

Influence of pP and sP phases on focal depth

The arrival times of direct P and S phases can be used to obtain accurate epicenters, but

control on depth is generally poor (largely due to a trade-off between event depth and origin time).

Depth-sensitive phases such as pP and sP can therefore play an important role in estimating

earthquake depth. As was demonstrated earlier, the use of even a single differential arrival time

such as pP-P can provide a well-constrained estimate of depth, and the consistency between such

estimates at different stations can help improve confidence in the result. However, there are issues

surrounding the correct identification of pP, including the potential to misidentify sP and pwP (P-

wave reflection from the ocean surface instead of the ocean bottom) as pP. Waveform modelling

performed using the CMT solutions derived in this paper indicate that pwP occurs as a low

amplitude reverberation following the main pP phase, and are unlikely to be mis-identified. These

same tests also show that the amplitude of pP varies with azimuth, with maximum amplitudes

generally occurring to the north and south, and minimum amplitudes to the east and west, where

sP becomes dominant. Therefore, it is possible that pP phases identified in these regions are

actually sP, which would have the effect of increasing the estimate of focal depth. Our pP-P results

don't indicate that this is a major problem, due to the consistency in depth estimates from stations

at different azimuths. In addition, a comparison between hypocenter depths determined using our

single-station pP-P scheme and the non-linear location scheme with all available phases indicates

that there is little evidence of inconsistency between the two sets of results (Supp. Fig. 4). Another

independent test is to perform the non-linear location without pP and sP phases, and compare the

results with those illustrated in Supp. Fig. 1. As shown in Supp. Fig. 5A-B, there is little difference

between the two sets of results, although it is clear that a majority of locations do not have sP or pP

phase picks. However, if we compare the depths of only those events that have at least one sP or pP

pick, the mean difference in depth is only 20 km for the Mariana pipe and 18 km for the Izu-Bonin

pipe, which given the depth extent of the pipes is not significant.

7

7

136

137

138

139

140

141

142

143

144

145

146

147

148

149

150

151

152

153

154

155

156

157

158

159

160

Exclusion of S-wave phases

If P and S-wave arrival times have similar uncertainty, then the S-phase will likely be more

sensitive to location due to S-wavespeeds being substantially lower than P-wavespeeds. In practice,

the higher levels of noise in S-wave arrival time datasets tends to nullify this advantage (see

Gomberg et al., 1990 for more discussion), but the increased data coverage they provide means that

their inclusion in the non-linear location scheme may provide improved results. To test the

influence of S-arrivals, we perform a test in which events are located using only the arrival times of

P-waves. A comparison between these locations and those of Supp. Fig. 1, in which both P and S

arrival times are exploited, is shown in Supp. Fig. 5C-D. For both sub-vertical pipes, the differences

in event distribution are negligible (see Supp. Table 1), with the exception of a small cluster of

events at approximately 400 km depth in the Mariana pipe, which appear in the inversion of P-

arrivals only, but not in the inversion of P and S. These are likely to be artifacts resulting from

insufficient constraints in the P-wave arrival time dataset.

Use of L2 misfit function

The use of an L2 norm to define the objective function (Equation 1 in main text), assumes that

the data errors have a Gaussian distribution, but in general, this may not be the case. In particular,

incorrect phase identification may result in large outliers, which can significantly influence

earthquake location if an L2 measure of misfit is used. The use of a fully non-linear location scheme

makes it relatively simple to test whether the L2 norm is suitable, via comparison with an L1

definition of the objective function which we know is robust with respect to outliers. For both pipes,

a comparison between the use of an L2 and L1 norm shows that the effect on the final earthquake

locations is minimal (see Supp. Fig. 5E-F and Supp. Table 1).

8

8

161

162

163

164

165

166

167

168

169

170

171

172

173

174

175

176

177

178

179

180

181

182

183

184

185

Influence of 3D velocity structure

Apart from data picking uncertainties, another major source of location error comes from our

assumption that the Earth's velocity structure conforms to the ak135 global reference model

(Kennett et al., 1995). The principal variations in wavespeed in the Earth occur with depth, but it is

well known that lateral variations in velocity structure can affect the arrival times of earthquakes

significantly enough that they need to be taken into account during location (e.g. Rawlinson et al.,

2010).

In a study involving a global earthquake dataset of 450,000 events, Amaru (2007) compared

the results of locations made using the ak135 global reference model (Kennett et al., 1995) and a 3D

global velocity model obtained from body wave tomography. The results showed that while some

events were shifted in location in the 3D model by as much as 50 km with respect to location in

ak135 (Kennett et al., 1995), the vast majority experienced a location shift of less than 20 km, with

an average shift of just 12.8 km. The average location shift in depth was only 3.1 km, with the

largest shifts occurring at shallow depths, where crustal anomalies are relatively large compared to

those in the underlying mantle. If these shifts are roughly indicative of the scale of uncertainty

caused by using ak135 (Kennett et al., 1995) in the location of the Mariana and Izu-Bonin events,

then, like the location uncertainty caused by picking error, they are unlikely to change the seismic

delineation of the two pipes. In fact, any such shift in this case is likely to contain a dominant

component that is common to every event, given that they are clustered in a very similar location. A

DC shift of this kind is unlikely to impact the interpretation of the earthquake clusters at all, unless

it is extremely large. However, both earthquake pipes do lie west of a subduction zone, and the

faster velocities associated with a descending slab may have had the effect of applying an easterly

DC location shift to both clusters of events. A well-known example of location bias due to

9

9

186

187

188

189

190

191

192

193

194

195

196

197

198

199

200

201

202

203

204

205

206

207

208

209

210

subduction was the Longshot explosion, which took place in the Aleutian islands in 1965. The

sampling of the Aleutian slab by nearly all recorded phases caused a 26 km location error (Herrin

and Taggart, 1968). However, even if the DC shift was this large (unlikely, since there is a better

azimuthal distribution of stations in our case), it is hard to see the interpretation changing, given

the scale of the features that are considered.

Depth phase analysis using USArray data

In order to independently identify depth phases for the Mariana pipe events, broadband

seismic data from USArray, which was deployed near the west coast of USA in 2006-2007, were

analyzed for the duration of the Mariana pipe seismic activity. This provided between 185 and 366

geographically similar stations covering a narrow epicentral distance range (75°-93°). Vertical

component data, acquired from the IRIS Data Management Centre, were initially cut 300s before

and after the predicted P-wave arrival time for all of the Mariana pipe events. Traces were then

tapered and filtered using a 2nd order two-way Butterworth bandpass filter with corner

frequencies of 0.5 Hz and 3 Hz. The data underwent a visual inspection to remove noisy traces and

only data with a clear P-wave arrival were retained for further processing. Initial alignment of

traces was achieved by making a moveout correction using the ak135 global model (Kennett et al.,

1995). Utilizing the similarity of the direct P-waveform across the array, trace alignment was then

refined using the adaptive stacking technique (Rawlinson and Kennett, 2004). Differential moveout

between the P and pP phases varies by less than 1s across the epicentral distance range of the

USArray data (Crotwell et al., 1999), meaning alignment along the direct P-wave and stacking

should also allow identification of the associated depth phases at the frequencies considered in this

study. However, we also align traces along time windows associated with pP using the same

adaptive stacking approach to increase coherency of this phase.

10

10

211

212

213

214

215

216

217

218

219

220

221

222

223

224

225

226

227

228

229

230

231

232

233

234

235

Several of the pipe events exhibit pP-P times (7s to >30s) consistent with earthquake focal

depths within the mantle wedge (40 km to greater than 200 km) as evidenced by clear arrivals both

in the linear and quadratic stacks. These results are in agreement with depths reported by ISC as

well as the non-linear and double-difference methods utilized in this study (Supp. Table 2). We note

that the GCMT catalogue states that some of the events listed below occur at depths of 12–18 km.

However, other depth solutions were obtained following the CSCMT-3D solution (discussed above)

and these are more faithful to the various depths obtained for the equivalent hypocenters (Supp.

Data File 1; Supp. Table 2).

Calculation of potential fluid/melt ascent rate within the sub-arc mantle

We inferred that the rate of ascent of fluid/melt within the sub-arc mantle could be calculated

from each “burst” of seismicity that was identified in Supp. Figs. 9 and 10. A velocity value was

calculated for every pair of hypocenters where one was younger and shallower than another

hypocenter for each “burst” that was identified in Figure 6 of the main paper as well as in Supp.

Figs. 9 and 10. The ascent rate calculations generated numerous velocity values within each “burst”

and we report the median value for each “burst”. The velocity values and the corresponding

calculations are provided in Supplementary File 5.

Our intention of using this cumbersome calculation was to ensure we did not subconsciously

infer shallowing of hypocenters within each “burst”. However, we also considered whether the

same ascent rates could be calculated using depth and time values generated from random number

generators. First of all we created a test to determine if “bursts” of activity could be identified

between randomly generated depths of 0 and 250 km (Haarhr, 2019), as well as randomly

generated dates and times over a two year period (http://random-date-generator.com/). This test

was created to simulate the Mariana sub-arc earthquake pipe (i.e., similar depth and period of

activity). We repeated the test five times by regenerating the depth and time values (using the same

11

11

236

237

238

239

240

241

242

243

244

245

246

247

248

249

250

251

252

253

254

255

256

257

258

259

260

limits). In doing so, we found no evidence of individual “bursts” (Supp. Fig. 12). We are therefore

confident that the bursts that we identified do not reflect a random process.

We conducted additional tests to determine if the ascent rate velocities that we obtained from

real-world hypocenter data were valid and how much these were influenced by different depth

ranges, time ranges and the number of nodes (i.e., hypocenters) used in the calculations. We found

that ascent rates could be calculated using number obtained from random number generators if we

restricted the depth range and time range when requesting randomly generated numbers. We also

found that the calculation is sensitive to the depth and time range used in the calculation (Supp. Fig.

13).

Finally, we calculated 20 pseudo-ascent rates using randomly generated depths between 0 km

and 200 km, a period of 10 days and using 20 nodes – these values were selected to simulate the

approximate length, duration and number of events within the Mariana ‘bursts’). We then

calculated the median ascent rate for each iteration and took an average of 19 out of 20 the results

and the standard deviation to obtain a statistically significant estimate of an ascent rate generated

from randomly generated data (see Supplementary File 5 for the data and calculations). We used

the same test, but restricted the depth range to values between 0 and 60 km to simulate the Izu-

Bonin pipe. The median velocity generated for the Izu-Bonin and Mariana pipes is compared with

the values obtained for each ‘burst’ in Supp. Fig. 14. This work demonstrates that the median

velocity obtained from the randomized data is lower than all velocities obtained for the Izu-Bonin

example, but is within error of Burst 2 and 3. For the Mariana example, similar velocities were

obtained from the randomized data and Bursts 1, 2 and 3. This analysis implies that there is

stronger statistical support for the faster ascent rate velocities (e.g., Izu-Bonin Burst 1 and 4, as well

as Mariana Burst 4).

These tests were a useful exercise, particularly for assessing the sensitivity of the ascent rate

velocity calculations with respect to different inputs. However, these results also need to be taken

12

12

261

262

263

264

265

266

267

268

269

270

271

272

273

274

275

276

277

278

279

280

281

282

283

284

285

with some caution as the randomly generated numbers were generated with particular boundary

conditions (depth range, time range, number of nodes) and Supp. Figure 12 (and Supp. Data 5)

shows that the ‘bursts’ are unlikely to be generated using randomized data.

We expect that this investigation will prove to be of some use for future workers. Our overall

aim here was to develop an unbiased method to infer shallowing as well as to calculate a velocity of

transport. Ultimately, the best test we have at present is to compare our calculated ascent rates of

the HypoDD datasets with ascent rates obtained from independent data (as are discussed in the

main paper).

13

13

286

287

288

289

290

291

292

293

294

Supp. Fig. 1.

Non-linear location results for (a) Mariana pipe; (b) Izu-Bonin pipe. Colored dots represent the non-

linear location results, which in this case are colored according to uncertainty in depth (a measure

14

14

295

296

297

298

based on the boundary of the 95% confidence region). The size of the coloured dots is proportional

to the magnitude of the event (see scale range in top two panels). Small black dots represent the

earthquake locations determined by the ISC (2013). The Mariana example is defined by a total of

637 events, and the Izu-Bonin example is defined by a total of 631 events.

Supp. Fig. 2.

Locations of recording stations used in this study for (a) Mariana sub-vertical earthquake pipe; (b)

Izu-Bonin sub-vertical earthquake pipe. Stations are colored according to the number of arrival

time picks available.

15

15

299

300

301

302

303

304

305

306

307

308

309

310

311

312

Supp. Fig. 3.

Results of two tests which only plots events when there are more than 25 picks available (top two

panels), or when the uncertainty in depth is less than 10 km (bottom left panel) or 15 km (bottom

right panel). The size of the coloured dots is proportional to the magnitude of the event (see scale

range in top two panels). Small black dots represent the earthquake locations determined by the

ISC (2013). The number of events in each plot is as follows (clockwise from top left): 201, 258, 197,

249.

16

16

313

314

315

316

317

318

319

320

321

Supp. Fig. 4.

Comparison of depths between the non-linear location results (solid black line) and pP-P inversions

(red dots). The error bars are equal to one standard deviation of the spread of depth values when

more than one station has P and pP picks. When an event has P and pP picked at only one station,

no error bar is shown. In the case of the Mariana sub-vertical pipe (A) 51 events are compared,

whereas for the Izu-Bonin sub-vertical pipe (B), 88 events are compared. The grey area denotes the

95% confidence region associated with the non-linear location result (solid black line).

17

17

322

323

324

325

326

327

328

329

Supp. Fig. 5.

Comparison of non-linear event locations that make use of all available data (red dots), versus (as

shown by small black dots) (a), (b) those that exclude pP and sP arrivals; (c), (d) those that exclude

S arrivals; and (e), (f) those that use an L1 measure of misfit. The size of the red dots is proportional

to the magnitude of the event (see scale range in top two panels). The Mariana sub-vertical pipe is

defined by a total of 637 events, and the Izu-Bonin sub-vertical pipe is defined by a total of 631

events in each of the three tests.

18

18

330

331

332

333

334

335

336

337

Supp. Fig. 6.

Histogram of initial (ISC, grey) delay time residuals for Mariana (A) and Izu-Bonin (B) pipe events,

and residuals for the final DD locations (blue). (C-D) Histograms of lateral and vertical relative

location errors, computed from the major axes of the horizontal and vertical projection of the 90%

confidence ellipsoids obtained from a bootstrap analysis of the final double-difference vector based

on 200 samples with replacement. Phase travel time curves for the Mariana (E) and the Izu-Bonin

19

19

338

339

340

341

342

343

344

(F) pipe events showing all available phase arrival time picks in the ISC bulletin (2013) (grey dots),

and phase arrival time picks observed pair wise at common stations and used to form delay time

data (blue dots) that went into the double-difference analysis.

Supp. Fig. 7.

Bathymetric maps (Ryan et al., 2009) that show published geochemical data (71) that have been

collected from the seafloor above and within the vicinity of the Izu-Bonin (top panels) and Mariana

(lower panels). The samples that directly overlie the two sub-vertical earthquake pipes are

primarily basalts and basaltic andesites according to their bulk rock compositions (middle panels).

The Izu-Bonin example also has several andesites and dacites above the pipe (middle panels). The

trace element data for these samples demonstrates typical volcanic arc signatures with respect to

N-MORB compositions (Sun and McDonough, 1989).

20

20

345

346

347

348

349

350

351

352

353

354

355

356

Supp. Fig. 8.

The location of the Izu-Bonin and Mariana sub-arc earthquake pipe swarms compared with the

location of arc volcanoes and their recorded eruptive histories (http://volcano.si.edu). The

bathymetric maps are taken from Global Multi-resolution Topographic database (Ryan et al., 2009).

21

21

357

358

359

360

361

Supp. Fig. 9.

A spatio-temporal analysis of seismic activity within the Izu-Bonin earthquake pipe demonstrates

that there are periods of increased seismic activity over one to several day periods that also occur

over a range of depths. Top panel: Time of occurrence plotted against depth of events. The

subfigures represent zoomed in views of the individual burst events as row 1: Depth vs. Time, row

2: Map View, row 3: Depth vs. Longitude (i.e. a west to east section) and row 4: Depth vs. Latitude

(i.e. a north-south section). The hypocenters are sized according to their magnitude and each burst

22

22

362

363

364

365

366

367

368

369

is colored with respect to time. The grey hypocenters represent the total double-difference re-

located hypocenters.

Supp. Fig. 10.

A spatio-temporal analysis of seismic activity within the Mariana earthquake pipe demonstrates

that there are periods of increased seismic activity over one to several day periods that also occur

over a range of depths. Top panel: Time of occurrence plotted against depth of events. The

23

23

370

371

372

373

374

375

376

subfigures represent zoomed in views of the individual burst events as row 1: Depth vs. Time, row

2: Map View, row 3: Depth vs. Longitude (i.e. a west to east section) and row 4: Depth vs. Latitude

(i.e. a north-south section). The hypocenters are sized according to their magnitude and each burst

is colored with respect to time. The grey hypocenters represent the total double-difference re-

located hypocenters.

24

24

377

378

379

380

381

25

25

382

Supp. Fig. 11. A summary of the depth (and 90% confidence intervals) compared with timing of the

hypocenters (relocated using HypoDD) that define ‘bursts’ within the broader (a-d) Mariana and (e-

h) Izu-Bonin sub-vertical earthquake pipe swarms. Note that the apparent difference in error bar

lengths between the Mariana and Izu-Bonin examples is simply due to different scale of the y-axis.

Supp. Fig. 12. An example of randomly generated time and depth values covering a two-year period

(the approximate duration of the Izu Bonin and Mariana earthquake pipes) and over a depth range

of 250 km (the approximate length of the Mariana pipe).

26

26

383

384

385

386

387

388

389

390

391

392

Supp. Fig. 13. A series of box-and-whisker plots showing the variation in ascent rate velocities

obtained using different depth ranges and time ranges, as well as different number of ‘nodes’ (i.e.

hypocenters) used within the calculations. These calculations were performed in an attempt to

understand whether the ascent rates calculated using real-world hypocenter data (i.e., Mariana and

Izu Bonin ‘Bursts’ 1 to 4) actually reflected random values. The variation in median velocity values

demonstrates that the outcomes of the calculations are influenced by the depth range of the values

within an individual burst and the duration of activity within an individual burst.

27

27

393

394

395

396

397

398

399

400

401

Supp. Fig. 14. A series of box-and-whisker plots showing the range, median and standard deviation

of velocities calculated for each ‘burst’ (see Supp. Data 5 for data and calculations). These are

compared with the average of 19 median velocities (black line) and their standard deviation (grey

box) obtained for randomly generated values used to simulate activity within (a) the Izu-Bonin pipe

(i.e., depth range of 60 km, duration of 10 days and with 20 nodes/hypocenters), and (b) the

Mariana pipe (i.e., depth range of 200 km, duration of 10 days and with 20 nodes/hypocenters).

Note that the grey box would shift if different boundary conditions were used to obtain the

randomly generated numbers (as per Supp. Fig. 12 and 13).

28

28

402

403

404

405

406

407

408

409

410

411

412

Supp. Table 1.

Differences in hypocenter determination obtained by comparing the results of the non-linear

location scheme using all available phases with several other scenarios.

Mariana sub-vertical pipe Izu-Bonin sub-vertical pipe

Depth

(km)

Latitude

(km)

Longitude

(km)

Depth

(km)

Latitude

(km)

Longitude

(km)

Mean difference between ISC

and non-linear locations

20 6 12 42 14 17

Mean difference between non-

linear locations with and

without sP and pP phases

20 5 5 18 5 5

Mean difference between non-

linear locations with and

without S phases

10 3 6 8 2 3

Mean difference between non-

linear locations using L2 and

L1 measures of misfit

10 3 6 15 7 9

Supp. Table 2.

Comparison of the depth of the 20 earthquakes temporally and spatially associated with the

Mariana mantle-pipe earthquake swarm – this includes hypocenter depths as reported in the ISC

29

29

413

414

415

416

417

418

419

420

bulletin (2013) and centroid moment tensors as reported in the GCMT (Ekström et al., 2012)

database, as well as the depths determined from various relocation techniques employed in this

study*. These are listed in order from deeper to shallower events (according to the HypoDD

solution) and color coded according to the level of confidence in the assigned depth (high: ≥4

solutions within 15 km depth; moderate: 3 solutions within 15 km depth; low: ≤2 solutions within

15 km depth).

ISC EVENT

Number

Depth (km) of earthquake hypocentre or moment tensor:

Confidence of

HypoDD*

depth

ISC

database

pP-P

arrival

time*

HypoDD* GCMT

catalogue

CSCMT-3D*

solution 1

(“shallow”)

CSCMT-

3D*

solution 2

(“deep”)

Indicative

depth from

adaptive

stacking of

depth

sensitive

phases

(USArray)

8956511 109.5 107.5 105.6 12 - 58–127 ~110 High

8956586 90.3 102.3 91.4 12 15–19 41–123 ~90 High

12662964 83.8 69.8 77.9 12 15–22 36–109 - High

30

30

421

422

423

424

425

426

427

11705384 35.0 165 71.6 17.9 15–17 45–120 ~79 Moderate

11705445 65.2 49.2 70.4 12.6 - 44–121 - Moderate

12664872 65.2 67.2 61.6 12 15–20 33–114 ~65 High

12652669 51.8 51.8 49.6 12.2 - 48–130 - High

8956585 50.4 68.4 47.5 12 15–19 55–120 - High

12660440 42.0 48 44 12 15–19 57–106 - High

8823380 41.4 47.4 43.4 12 15–19 48–136 ~41 High

8767893 43.8 59.8 37.8 12 - 30–108 - Moderate

11705431 45.9 29.9 36 12 - 46–125 - High

12670232 28.9 - 32.1 12 15–19 42–115 - High –

although

uncertainty

with which

GSCMT-3D

solution fits

best

8956569 34.2 34.2 31.1 12 15–19 41–123 - High –

although

uncertainty

with which

GSCMT-3D

31

31

solution fits

best

12661269 42.3 32.3 29.8 12 15–18 37–106 - High –

although

uncertainty

with which

GSCMT-3D

solution fits

best

12664883 37.3 - 29.8 12 - 31–111 - Moderate

9137769 26.9 - 24.3 12 - 43–127 - Low

10396345 19.7 21.7 20.9 12 15–17 42–125 25–30 High

11505273 20 18 16.1 12 - 36–106 - High

12652398 4.7 18 15.1 12 15–19 46–115 - High

32

32

428

429

430

Supp. Table 3.

Summary of the results of the calculated ascent velocities for each “burst” that was identified in the

Izu-Bonin and Mariana sub-vertical earthquake pipe swarms.

Izu-Bonin sub-vertical

pipe

Mariana sub-vertical pipe

Median Velocity (km/hr) Median Velocity (km/hr)

Burst 1 33.98 1.36

Burst 2 0.47 0.69

Burst 3 0.78 1.17

Burst 4 3.11 7.27

Combined 1.95 1.27

33

33

431

432

433

434

435

436

437

Supp. Video 1. (separate file)

Animation showing the temporal nature of the seismic events within the most clearly defined

earthquake pipe located beneath the Mariana arc/back-arc. This animation also shows that the

majority of hypocenters that define the sub-vertical, pipe-like structure occur within 2006 and

2007, primarily within several distinct month-long pulses of activity. The hypocenter data were

sourced from ISC (2013).

Supp. Video 2. (separate file)

Animation showing a series of cross-sections generated perpendicular to the plate boundary to

show that that the earthquake pipes are not part of the same structure as the down-going Pacific

Plate. This animation also shows regions of aseismicity along the length of the slab that may

indicate the location of slab tears. The hypocenter data were sourced from ISC (2013).

Supp. Video 3. (separate file)

Animation showing both the spatio-temporal nature of the Izu-Bonin and Mariana earthquake

pipes. These animations show the cumulative location of hypocenters across the region between

1970 and 2009. As time passes, the subducted slab geometry becomes more apparent in both cross-

sections, but in 1985-1986 and in 2006-2007 there are pulses of earthquake ruptures that are

clearly distinguished from the subducted Pacific slab. The hypocenter data were sourced from ISC

(2013).

34

34

438

439

440

441

442

443

444

445

446

447

448

449

450

451

452

453

454

455

456

457

458

459

Supp. Data File 1. (separate file)

Earthquake location, arrival time and relocated data from the Mariana pipe that were used and

were generated as part of this project.

Supp. Data File 2. (separate file)

Earthquake location, arrival time and relocated data from the Izu-Bonin pipe that were used and

were generated as part of this project.

Supp. Data File 3. (separate file)

Selection of stacked events from the Mariana sub-vertical earthquake pipe as recorded by USArray.

In each plot, the top trace shows the quadratic stack, while the trace immediately below shows the

linear stack. All remaining traces are the traces from each station which went into the stack. Global

phase arrival time predictions based on the ISC bulletin (2013) are shown in red. Possible pP and

sP arrivals in the stack are highlighted in yellow and green respectively, although when there is no

second arrival marked, the highlighted phase could be pP or sP. In most cases, the observations are

broadly consistent with the predictions, which provides independent evidence that the earthquakes

are distributed between the subducted slab and the surface, and cannot all be clustered near the

surface. The title of each plot contains the following information: stack filename, latitude, longitude

and depth of event, origin date and origin time (based on ISC Bulletin (2013)).

35

35

460

461

462

463

464

465

466

467

468

469

470

471

472

473

474

475

476

477

478

Supp. Data File 4. (separate file)

Summary figures showing the CSCMT-3D solutions for 20 earthquakes within the mantle wedge

pipe. These show contour plots of correlation between the observed and synthetic waveforms in

the depth versus time, the double-couple percentage for each moment tensor, the broadband

stations that were used for the calculation and the fit between the observed waveforms (thick gray

line), CSCMT-3D solution synthetics (black line) and the GCMT solution synthetics (red line). The

numbers above and below the waveforms show the correlation between the observed and CSCMT-

3D synthetics (black text) and the correlation between the observed and GCMT synthetic

waveforms (red text). A collection of depth vs. correlation plots are also included to show where

one or two depth solutions were possible (green lines indicate values within 5% of the maximum

correlation value).

Supp. Data File 5. (separate file)

Ascent rate calculations derived from analyses of the double-difference relocated hypocenters

generated as part of this study.

Caption for Interactive Map Data (separate file)

KMZ file showing the location of earthquake hypocenters (ISC) that define the Mariana and Izu-

Bonin mantle earthquake pipes. Note that these are projected to the surface, but depth information

and other metadata is captured for individual points.

36

36

479

480

481

482

483

484

485

486

487

488

489

490

491

492

493

494

495

496

497

498

499

500

501

Caption for Interactive 3D Model (separate file)

An interactive 3D PDF was built using the ISC hypocenter locations that define the Mariana sub-arc

mantle earthquake pipe. This file shows the location of the earthquake hypocenters that define the

pipe-like structure with respect to the subducted Pacific Plate and the overriding Philippine Sea

Plate.

37

37

502

503

504

505

506

507

508

References

Amaru, M.L. 2007. Global travel time tomography with 3-D reference models”, PhD thesis, Utrecht

University, Utrecht, Netherlands

Billings, S.D. 1994. Simulated annealing for earthquake location, Geophysical Journal International

118, 680-692

Billings, S.D., Kennett, B.L.N., Sambridge, M.S. 1994. Hypocentre location: genetic algorithms

incorporating problem-specific information, Geophysical Journal International 118, 693-706

Crotwell, H.P., Owens, T.J., Ritsema, J. 1999. The TauP toolkit: flexible seismic travel-time and ray-

path utilities, Seismological Research Letters 70, 154-160

Douglas, A., Bowers, D., Young, J.B. 1997. On the onset of P seismograms, Geophysical Journal

International 129, 681-690

Ekström, G., Nettles, M., Dziewonski, A.M. 2012. The global CMT project 2004-2010: Centroid-

moment tensors for 13,017 earthquakes, Physics of the Earth and Planetary Interiors 200-201,

1-9, doi:10.1016/j.pepi.2012.04.002

Gomberg, J.S., Shedlock, K.M., Roecker, S.W. 1990. The effect of S-wave arrival times on the accuracy

of hypocenter estimation, Bulletin of the seismological Society of America 80, 1605-1628

Haahr, M. 2019. RANDOM.ORG: True Random Number Service. Available at:

https://www.random.org [last accessed 23 May 2019]

Herrin, E., Taggart, J. 1968. Source bias in epicenter determinations. Bulletin of the Seismological

Society of America 88, 722-743

International Seismological Center (ISC). 2013. On-line Bulletin, International Seismological Center,

Thatcham, United Kingdom, http://www.isc.ac.uk

Kennett, B.L.N., Engdahl, E.R., Buland, R. 1995. Constraints on seismic velocities in the Earth from

traveltimes. Geophysical Journal International 122, 108-124

38

38

509

510

511

512

513

514

515

516

517

518

519

520

521

522

523

524

525

526

527

528

529

530

531

532

Leonard, M. 2000. Comparison of manual and automatic onset time picking, Bulletin of the

Seismological Society of America 90, 1384-1390

Li, C., van der Hilst, R. D., Engdahl, E.R., Burdick, S. 2008. A new global model for P wave speed

variations in Earth’s mantle. Geochemistry Geophysics Geosystems 9, Q05018, doi:

10.1029/2007GC001806

Rawlinson, N., Kennett, B.L. 2004. Rapid estimation of relative and absolute delay times across a

network by adaptive stacking. Geophysical Journal International 157, 332-340

Rawlinson, N., Pozgay, S., Fishwick, S. 2010. Seismic tomography: a window into deep Earth, Physics

of the Earth and Planetary Interiors 178, 101-135

Ryan, W.B.F., Carbotte, S.M., Coplan, J.O., O’Hara, S.O., Melkonian, A., Arko, R., Weissel, R.A., Ferrini,

V., Goodwillie, A., Nitsche, F., Bonczkowski, J., Zemsky, R. 2009. Global multi-resolution

topography synthesis. Geochemistry, Geophysics, Geosystems 10 doi:10.1029/2008GC002332

Sambridge, M.S. Kennett, B.L.N. 1986. A novel method of hypocenter location, Geophys. J. R. Astr. Soc.

87, 679-697

Sambridge, M.S., Kennett, B.L.N. 2001. Seismic event location: nonlinear inversion using a

neighbourhood algorithm, Pure and Applied Geophysics 158, 241-257

Sun, S.-s, McDonough, W.F. 1989. Chemical and isotopic systematics of oceanic basalts: implications

for mantle composition and processes in Magmatism in the Ocean Basins, A. D. Saunders, M. J.

Norry, Eds., (Special Publication 42, Geological Society London), 313-345

39

39

533

534

535

536

537

538

539

540

541

542

543

544

545

546

547

548

549

550

551

552

553