Indonesia is arguably one of the tectonically most complex regions on Earth today due to its location at the junction of several major tectonic plates and its long history of collision and accretion. It is thus an ideal location to study the interaction between subducting plates and mantle convection. Seismic anisotropy can serve as a diagnostic tool for identifying various subsurface deformational processes, such as mantle flow, for example. Here, we present novel shear wave splitting results across the Indonesian region. Using three different shear phases (local S, SKS, and downgoing S) to improve spatial resolution of anisotropic fabrics allows us to distinguish several deformational features. For example, the block rotation history of Borneo is reflected in coast-parallel fast directions, which we attribute to fossil anisotropy. Furthermore, we are able to unravel the mantle flow pattern in the Sulawesi and Banda region: We detect toroidal flow around the Celebes Sea slab, oblique corner flow in the Banda wedge, and sub-slab mantle flow around the arcuate Banda slab. We present evidence for deep, sub-520 km anisotropy at the Java subduction zone. In the Sumatran backarc, we measure trench-perpendicular fast orientations, which we assume to be due to mantle flow beneath the overriding Eurasian plate. These observations will allow to test ideas of, for example, slab–mantle coupling in subduction regions.
 Located at the quadruple junction of four major tectonic plates, namely the Indo-Australian, Eurasian, Philippine Sea, and Pacific plates, the Indonesian region is shaped by subduction, collision and accretion. It can hence be viewed as a tectonic setting analogous to primitive accretionary orogens [e.g.,Silver and Smith, 1983; Snyder and Barber, 1997; Hall, 2009]. Several studies have been undertaken in the past few decades to shed light on this Rubik's cube of terranes and microplates; these include studies of surface geology [e.g., Bader and Pubellier, 2000], paleomagnetics [e.g., Fuller et al., 1999], active source seismic reflection and refraction experiments [e.g., Kopp et al., 1999], and GPS measurements [e.g., Walpersdorf et al., 1998]. Although this provides us with a general understanding of the tectonic evolution of this complex region, a coherent picture of the subsurface processes, especially mantle flow through the maze of subducting slabs in eastern Indonesia, is still lacking.
1.1. Seismic Anisotropy and Shear Wave Splitting
 One approach to understanding deformational processes associated with a tectonic area is studying seismic anisotropy, the directional dependence of seismic wave speed. Seismic anisotropy is currently our best tool for determining in situmantle flow direction. In the upper mantle, olivine crystals will align preferentially with finite strain (referred to as lattice-preferred orientation; LPO). The alignment is such that under dry mantle conditions, the olivinea-axis is parallel to the direction of mantle flow and theb-axis is perpendicular to the shear plane (A-type LPO [e.g.,Ribe, 1989; Babuška and Cara, 1991; Mainprice, 2007]). The crust and the lithosphere, however, can also harbor numerous sources of anisotropy, both as LPO and SPO (shape-preferred orientation); these could be aligned cracks, potentially fluid- or melt-filled, oriented melt-pockets, or thin layers of materials with varying elastic properties [e.g.,Crampin and Booth, 1985; Kaneshima et al., 1988; Kaneshima and Ando, 1989; Kendall et al., 2005; Holtzman and Kendall, 2010; Bastow et al., 2010].
 A direct diagnostic of seismic anisotropy is shear wave splitting: upon encountering an anisotropic medium, a shear wave splits into two orthogonally polarized shear waves traveling at different velocities. Shear wave splitting is described by two parameters: the polarization of the faster shear wave (ϕ) and the delay time between the two waves (δt), which can both be measured, as outlined below. These two parameters can provide information about deformational processes as well as compositional properties. For example, in the case of A-type LPO of olivine in the upper mantle, the fast direction for a vertically propagating phase is parallel to the direction of mantle flow. Or, for SPO of cracks in the crust, the fast polarization parallels the alignment of fractures (provided their spacing is smaller than the seismic wavelength).
 The magnitude of δtis a function of both the strength of the anisotropy and the thickness of the anisotropic layer. The trade-off between these two parameters complicates placing spatial constraints on the anisotropy. Furthermore, shear wave splitting is a path-integrated, aggregate measurement, i.e., the exact location and extent of the anisotropy along the raypath cannot be determined from the measurement itself. The challenge, therefore, is to constrain the anisotropic fabrics both causally and spatially in light of the tectonic processes present. We approach this problem by using three different shear phases (localS, SKS, and downgoing S), as outlined in Di Leo et al. . Using different phases can improve vertical resolution beneath a receiver. In addition to that, source-side splitting can detect anisotropic fabrics where it is not possible to place seismic stations [e.g.,Nowacki et al., 2012].
 This is the first study of seismic anisotropy in the Banda and Flores Sea region and Borneo. We combine these novel results with observations of previous seismic anisotropy studies to gain a more coherent picture of the tectonic processes in the Indonesian region.
2. Tectonic Setting
 Indonesia consists of the Sundaland core and a patchwork of amalgamated terranes (Figure 1). Sundaland, the continental core of SE Asia [e.g., Hamilton, 1979] is of Gondwanian origin and marks the present-day southeastern edge of the Eurasian plate. It encompasses the Thai–Malay Peninsula, Indochina, SW Borneo, Sumatra, Java, as well as the Sunda Shelf. The shelf was exposed during the Pleistocene and is still quite shallow today with water depths of around 200 m or less [e.g.,Hall, 2009]. Due to the paucity of volcanic and seismic activity in the Sundaland core, several authors have described it as a craton or shield [e.g. Ben-Avraham and Emery, 1973; Tjia, 1996; Barber et al., 2005]. However, unlike better established cratons or shields, such as the African craton or the Canadian shield, the Sundaland core is not underlain by cold thick lithosphere that has been consolidated since the Precambrian. Instead, the Sundaland core has undergone considerable internal deformation throughout the Meso- and Cenozoic, including the formation of several sedimentary basins and extensive faulting [Hall and Morley, 2004]. Furthermore, elevated surface heat flow of >80 mW/m2 [Artemieva and Mooney, 2001] and low lithospheric and upper mantle velocities [e.g., Widiyantoro and van der Hilst, 1997; Ritsema and van Heijst, 2000] are indicative of a weak lithosphere, which, again, is uncharacteristic for a craton [Hall and Morley, 2004].
 Although we know that Indonesia is at the junction of several major tectonic plates, drawing exact plate boundaries remains challenging. Gordon labeled these ‘diffuse plate’ boundaries, noting they have, in fact, a width of between 200 and 1000 km, the exception being the Sunda trench, where the Indo-Australian plate is subducting beneath Eurasia. In eastern Indonesia, however, where there is a plethora of active and relict subduction zones, plate boundaries are more ambiguous.
 Since the mid-Mesozoic, several fragments that rifted from the Australian plate have successively attached themselves to the Sundaland margin. As a result, the SE Asian margin resembles an accretionary orogen in its early stages [Hall, 2009]. During the early Cretaceous, SW Borneo was added, followed by east Java, east Borneo, and western Sulawesi in the late Cretaceous. Further fragments, such as east Sulawesi, were added to the eastern margin in the Cenozoic [Hall, 2009].
 The oldest and most prominent subduction zone in the region is that of the Indo-Australian plate subducting beneath Eurasia at the Sunda trench with the Sunda arc having been active since the early Cenozoic [e.g.,Hamilton, 1988; Whittaker et al., 2007]. The exact timing of subduction initiation remains unknown; however, evidence suggests that it was not before 45 Ma [Hall, 2009]. Although this subduction zone appears to be continuous along Sumatra and Java, there is a sharp change in several characteristics at the Sunda strait, which separates Java from Sumatra. The age of the downgoing slab sharply increases from ∼60 Ma to ∼100 Ma between Sumatra and Java, respectively [Sdrolias and Müller, 2006]. Furthermore, the Sunda Strait marks a change in bulk sound speed within the slab [Gorbatov and Kennett, 2003], a decrease of deep seismicity below 500 km in Sumatra [e.g., Engdahl et al., 1998], as well as a change in the distance between the volcanic front and the top of the slab from ∼90 km in Sumatra to ∼150 km in Java [Syracuse and Abers, 2006]. All these changes occur without being evident in the shape of the slab [Gudmundsson and Sambridge, 1998; Syracuse and Abers, 2006].
 The extremely curved Banda arc extends from Buru (south of Halmahera) over Seram and Timor to Flores (south of Sulawesi). It features an inner active volcanic arc and an outer non-volcanic arc. There has been a debate about whether the strong curvature is due to subduction of two slabs dipping toward each other [e.g.,Cardwell and Isacks, 1978; McCaffrey, 1989] or bending of a single slab [e.g., Hamilton, 1979; Hall, 2002]. Spakman and Hall favor the view that the Banda arc formed during the mid-Miocene (∼15 Ma ago) by subduction of a Jurassic ocean embayment [e.g.,Hamilton, 1979; Charlton, 2000; Hall, 2002] and that the shape is a result from the pre-existing bend in the impinging Australian plate margin as well as rollback of the Jurassic, hence old and cold, slab back to the curved shape of the embayment.
 The Sulawesi region northwest of the Banda arc exhibits the highest density of subduction zones in Indonesia. A unique feature of the area is double-sided subduction of the Molucca Sea plate in northern Sulawesi. This microplate is subducting westward at the Sangihe trench and eastward at the Halmahera trench. The two respective volcanic arcs are in the processes of colliding. This is the only arc–arc collision we can observe today [Jaffe et al., 2004]. The northern part of the Molucca sea slab is flanked by two additional subduction zones dipping toward each other: the Philippine Sea plate is subducting westward beneath Mindanao peninsula [e.g., Aurelio, 1992], and the Celebes Sea plate, itself part of the Eurasian plate [Hall and Nichols, 1990], is subducting eastward beneath Mindanao [e.g., Sajona et al., 1997]. Furthermore, the southern end of the Celebes Sea plate is subducting southwards beneath northern Sulawesi. Evidence suggests that the Celebes Sea slab and the Molucca Sea slab are colliding at a depth of ∼200 km [Walpersdorf et al., 1998; Kopp et al., 1999].
 West of Sulawesi lies the isle of Borneo, the perhaps most tectonically stable region in the study area today, as there is currently no active subduction around the island and, hence, little seismicity compared to the rest of Indonesia. The mid-Eocene (∼45 Ma ago), however, saw the onset of subduction of the Proto-South China Sea in the Sarawak Orogeny [Hutchinson, 2006]. Subduction lasted until the Early Miocene [e.g., Hall et al., 2008]. Although Borneo has been a static part of the Sundaland core throughout the Cenozoic, palaeomagnetic data suggest that it has undergone a series of counter-clockwise rotations in the past as a consequence of convergence between the Eurasian and Indo-Australian plates: by ∼40° in the Early to mid-Cretaceous and again by ∼45–50° between 25 and 10 Ma ago [Fuller et al., 1999].
 The Makassar Straits separating Borneo and Sulawesi formed during the Eocene through rifting [e.g., Hamilton, 1979; Cloke et al., 1999]. The Wallace Line, i.e., the biogeographical boundary between the ecozones of Asia and Australasia, follows the Makassar Straits.
 This is, of course, only a brief outline of the tectonics of one of the geologically most complex regions on Earth today. For a more in-depth review, we refer the reader toLee and Lawrence , Hall  or Hall  and references therein, for example. Nonetheless, it becomes clear that the region in its tectonic complexity and diversity cannot and should not be treated as one homogeneous landmass, especially when conducting a comprehensive study of seismic anisotropy, as we do here.
3. Previous Studies of Seismic Anisotropy in Indonesia
 The anisotropic structure of the Java–Sumatra subduction system was investigated by Hammond et al.  using measurements of local S and SKS splitting (Region 2 in Figure 1). They observed anisotropy which they ascribed to trench-parallel aligned vertical cracks in the deforming overriding Eurasian plate as well as fossil anisotropy in the subducting slab. They noted several differences in slab characteristics beneath Sumatra and Java, such as location of the volcanic front [Syracuse and Abers, 2006], bulk sound speed [Gorbatov and Kennett, 2003], and slab age [Sdrolias and Müller, 2006], which they inferred to be the cause of differing splitting patterns for Java and Sumatra.
 In a recent study [Di Leo et al., 2012], we investigated the anisotropic characteristics of the Sangihe subduction zone, which extends from the southern Philippines to northern Sulawesi (Region 3 in Figure 1). As in this study, we used three different shear phases in the splitting analysis, and were thus able to distinguish three regions of anisotropy at this subduction zone: one in the overriding Eurasian plate due to fossil anisotropy as well as SPO of melt-filled cracks, one atop the subducting slab due to shear layers with very strong olivine LPO in the deep mantle wedge, and one beneath the slab due to trench-parallel sub-slab mantle flow beneath the Molucca Sea microplate.
 In this study, we present new shear wave splitting measurements for parts of Indonesia - such as the Banda Sea region, southern Sulawesi, and Borneo (Regions 1 and 4 inFigure 1) - previously unstudied with seismic anisotropy. However, we also revisit aspects of the studies mentioned above and add new shear wave splitting results in the backarc regions of Java and Sumatra.
4. Data and Methodology
 We analyze three different shear phases, local S, SKS, and teleseismically recorded S waves, to attain an improved spatial resolution of anisotropic fabrics in the Indonesian subduction system and Borneo.
 The majority of the data used for local S and SKS splitting analysis comes from the temporary JISNET (Japan Indonesia Seismic Network) array [Ohtaki et al., 2000], which recorded from 1998 to 2007, with some stations having been reactivated recently. Recordings from some stations of this network were examined in the previous studies of Hammond et al.  and Di Leo et al. . In this study, we use 5 years of data (1998–2003) recorded by JISNET stations TOLI, PCI, and KDI on Sulawesi, TARA, BKB, PKBI, and PTK in Kalimantan (the Indonesian part of Borneo), TPI on Belitung Island, TPN on the Riau Islands. In addition to that, we use recordings from GSN (Global Seismographic Network) stations BTDF in Singapore (1998–2006) and KAPI on Sulawesi (1999–2010), as well as GEOFON station LUWI on Sulawesi (2008–2010) (Figures 2, 5, and 8 and Table S1 in the auxiliary material).
 The data used in the source-side splitting analysis were recorded by five GEOSCOPE stations located in Russia (SEY), China (WUS), India (HYB), Australia (CAN), and Antarctica (DRV) and one K-NET station in Kyrgyzstan (AAK) (Figure 3 and Table S1 in the auxiliary material).
 Data were filtered using Butterworth bandpass filters with corners between 0.1 and 1 Hz in the case of source-sideS and local S data, and 0.03 and 0.4 Hz for SKS data. Overlapping filter bands for the different shear phases should decrease any influence of frequency dependent anisotropy. Nonetheless, we applied low frequency filters to local S data to test for frequency dependence, but could not detect any evidence for it.
 We measure shear wave splitting from three-component seismic data following the method ofSilver and Chan . It grid-searches for the two splitting parameters, direction of the fast shear wave (ϕ) and delay time (δt), that minimize the second eigenvalue of the covariance matrix of the particle motion, thereby correcting for the splitting. The cluster analysis method of Teanby et al.  allows removal of subjectivity regarding the choice of window lengths around the shear phases [Wüstefeld et al., 2010]. Examples of shear wave splitting results are shown in Figure 4.
 For source-side splitting, we use events that originate in the study area and are recorded at teleseismic distances between 50° and 85° [Wookey et al., 2005; Di Leo et al., 2012]. The method relies on the assumption that the lower mantle above D″ is insignificantly anisotropic. A number of studies have suggested this to be the case [e.g., Meade et al., 1995; Montagner and Kennett, 1996; Panning and Romanowicz, 2006; Kustowski et al., 2008]. Therefore, we assume that the measured anisotropic fabrics are located in the upper 660 km beneath the receiver and the source. Provided the receiver-side anisotropy is well constrained (i.e., sufficient measurements with results showing little to no backazimuthal variation, thus indicating simple one-layer anisotropy [Silver and Savage, 1994]), it can be removed in the splitting analysis. The remaining splitting must therefore stem from the source-side. Splitting parameters for these upper mantle corrections were taken fromSKS splitting studies by Barruol and Hoffmann  and Wolfe and Vernon .
 In order to make the source-side splitting results comparable with those ofSKS and local Smeasurements, one has to rotate the measured fast direction (after the receiver-side correction) from the geographical reference frame of the receiver to that of the source. This is done withϕ″ = azimuth + backazimuth − ϕ, where ϕ″ is the source-side fast direction andϕis corrected receiver-side fast direction. Only events that have been recorded by two or more stations and yield consistent results are used (Figure 3 and Table S4 in the auxiliary material).
 In the following, we discuss results separately for the different regions. Results are shown in Figures 5–6 and listed in Tables S2–S4 in the auxiliary material.
5.1. Region 1: Southern Sulawesi and the Banda Sea
 We analyzed recordings from two broadband stations in southern Sulawesi: KAPI in the southwest and KDI in the southeast. For KAPI, we obtained 8 good quality SKS splitting measurements with a stacked average of δt = 1.13 s (calculated using the stacking technique of Wolfe and Silver ) (Figure 5). Fast directions are consistently E–W trending, parallel to the Sunda–Banda trench. Unfortunately, all 8 rays have similar backazimuths, thereby making it impossible to rule out any backazimuthal variation in splitting to indicate whether there is one horizontal or dipping layer of anisotropy or several layers [Silver and Savage, 1994]. The two SKS measurements at KDI have δt of 1.32 and 1.38 s, and ϕ is NE–SW, approximately parallel to the curvature of the Banda Arc. These two rays have very similar backazimuths as well.
 We obtained 7 local Ssplitting measurements for KAPI and 11 for KDI. All 18 events are deep with depths of ∼460–690 km. Time-lags measured at KAPI are 0.38–0.95 s and 0.34–1.08 s at KDI. Fast orientations are predominantly NW–SE trending, oblique to the subducting slab with a trend toward perpendicular.
 Source-side splitting analysis yielded 18 measurements for 9 events in the Banda Sea region. Five events are fairly shallow with depths between ∼33 km and 54 km, two earthquakes are of intermediate depth (∼97 and 143 km), and two originate within the mantle transition zone at depths 533 km and 596 km. Delay times vary between 0.50 s and 1.70 s. Apart from the two deep events, fast polarizations roughly follow the curvature of the Banda Arc.
5.1.1. Anisotropy in the Banda Sea Slab?
 In order to determine if there are measurable anisotropic fabrics in the subducting Banda Sea slab, we test whether there is a correlation between measured delay time and path length of each respective ray within the slab. We use the same approach as Di Leo et al. . We built a 3-D model of the slab based on the seismicity-derived Regional Upper Mantle (RUM) model byGudmundsson and Sambridge and the iasp91 1-D velocity model [Kennett and Engdahl, 1991]. Although there are more recent slab models by Syracuse and Abers  and Hayes et al. , they do not image the westernmost extent of the slab (i.e., the Banda arc), which is of great importance for the interpretation of our splitting measurements in that area. Our slab model has a thickness of 100 km, which is generally considered to be a typical value for most slabs [e.g., Bercovici and Karato, 2003]. Furthermore, this makes our results comparable to those of Hammond et al. , who also used 100 km slab thickness in their modeling. Using the TauP Toolkit of Crotwell et al.  and Winchester and Crotwell , we raytrace through the model and compare the estimated path length of each ray within the slab to the measured delay time (Figure 9). We assume the 1-D raypaths to be approximate to the actual path. If the slab harbors significant seismic anisotropy,δt should increase with increasing slab path length. Local S rays traversing the slab show no increase in δt with slab path. For those downgoing Sphases traveling through the slab, however, there appears to be an increase in time-lag. This indicates that the deep slab, i.e., that part of the slab not sampled by upgoing localS waves, may well be anisotropic.
5.2. Region 2: Sumatra and Java
 Recordings from stations in the backarc region of Sumatra (BTDF, TPN, TPI) yielded 5 SKS splitting measurements: 3 for BTDF and one each for TPN and TPI (Figure 6). Delay times are quite variable, ranging from 0.68 to 2.73 s. At BTDF alone, lag-times are between 0.90 and 2.73 s with a stacked average of ∼1.10 s. At all three stations, fast directions are consistently NE–SW trending.
 Three deep events, with depths between ∼535 and 636 km, in the backarc region of Java fulfilled our criteria and rendered source-side splitting results. Delay times are between 0.33 and 0.79 s. Fast polarizations are NW–SE trending for the two westernmost events and NE–SW trending for the event at the southern end of the Makassar Straits.
 No local S measurements were obtained for these three stations.
5.3. Region 3: Northern Sulawesi and the Molucca Sea
 We analyzed data from three stations in northern Sulawesi: TOLI, PCI, and LUWI. We obtained 10 good-qualitySKS splitting measurements for these stations: 4 for TOLI, 4 for LUWI, and 2 for PCI (Figure 7). Lag-times show stacked averages 1.25 s, 1.30 s, and 0.98 s, respectively. Fast directions are roughly, but consistently, E–W trending.SKS phases recorded at PCI have similar backazimuths. At TOLI and LUWI, however, the rays have different backazimuths, but similar ϕ, which is indicative of a single, horizontal layer of anisotropy.
 We obtained 6 local S splitting results for the region: 4 for PCI and 2 for LUWI; none for TOLI, however. Fast directions measured at both stations are predominantly NNW–SSE trending. Delay times for the PCI results are between 0.36 and 0.54 s; for LUWI, 0.30 and 0.60 s.
5.4. Region 4: Borneo
 Data recorded at 4 stations in Kalimantan on Borneo, TARA, BKB, PKBI, and PTK, yielded 7 SKS splitting results (Figure 8). Due to the lack of seismicity in the region, no local Sor source-side splitting was measured. Delay times vary between 0.55 s and 1.35 s. Fast directions tend to be parallel to the coast line. Backazimuths of the events used are all too similar to determine whether there is more than one horizontal layer of anisotropy present.
 Our new measurements of shear wave splitting in Indonesia, in conjunction with previous studies, contribute to unraveling various tectonic processes in this complex area, highlighting the fact that the landmass conveniently labeled Indonesia ought not to be treated as one homogenous region, e.g., when conducting studies of global mantle flow.
6.1. Region 1: Southern Sulawesi and the Banda Sea
 We believe the anisotropy detected by both the SKS and local Ssplitting to be caused by mantle flow in the Banda wedge. Due to the strong curvature and obliquity of the Banda subduction zone, corner flow in the wedge is not as two-dimensional as in regular subduction zones (Figure 10). The interior of the wedge generally does not produce large amounts of anisotropy [e.g., Buttles and Olson, 1998; Di Leo et al., 2012]. Thus, the two different phases may be sensitive to two different aspects of this same corner flow: The deep local events may be detecting shear layers at the bottom of the wedge, immediately above the slab. In that region, strain and shear stress are extremely high and progressive simple shear can thus produce very strong olivine LPO [Ribe, 1989]. Several studies have predicted these supra-slab shear layers numerically [McKenzie, 1979; Ribe, 1989; Kendall and Thomson, 1993], and they have been observed in analog models [Buttles and Olson, 1998] and seismically in the Sangihe subduction zone [Di Leo et al., 2012]. In the latter study, delay times measured for these shear layers are similar to those measured at KAPI and KDI, making it possible that they are being caused by the same mechanism. This region of strong LPO at the slab–wedge interface is only fully developed below ∼200 km depth and has a thickness of ∼100 km [Buttles and Olson, 1998] or an angular width of ∼5–10° [Ribe, 1989]. SKSrays recorded at KDI and KAPI do not traverse this part of the subduction zone. Corresponding time-lags are of similar magnitude as those arising from mantle flow at the Sangihe subduction zone in northern Sulawesi [Di Leo et al., 2012]. Fast directions are approximately parallel to the curvature of the Banda arc. We therefore suggest that the anisotropy measured at the two stations in southern Sulawesi is due to mantle flow around the continental keel of the island, parallel to the arcuate Banda slab. The resulting picture is that of oblique corner flow - i.e., neither trench-parallel nor perpendicular - in the wedge of the Banda subduction zone.
 Splitting measurements of the two deep events (>530 km; off the southeastern tip of Sulawesi) as well as the splitting results for the two earthquakes around 7.5°S and 129°E, on the other hand, are sensitive to deeper anisotropy, presumably deep within the slab (Figure 9). Since old, cold, Jurassic continental crust is being subducted at the Banda trench, it is possible that the slab - unlike younger, thinner slabs - has retained measurable fossil anisotropy. Alternatively, the two deep events may be detecting anisotropy in the mantle transition zone, as discussed below (seesection 6.2.2). The fast directions of these two deep events are nearly perpendicular to those measured with local S phases in the same area, indicating that the two phases are detecting two different regions of anisotropy: above the slab and deep within or below. Irrespective of the source of this deep anisotropy, this locality is an excellent example of how using different shear phases can significantly improve spatial resolution of seismic anisotropy.
6.2. Region 2: Sumatra and Java
6.2.1. The Sunda–Banda Transition: Evidence From Fossil Anisotropy
 The transition from the Sunda to the Banda subduction zone, i.e. from subduction of oceanic to continental lithosphere, is not evident in seismicity-derived slab models [Gudmundsson and Sambridge, 1998; Syracuse and Abers, 2006; Hayes et al., 2012]. Hammond et al.  measured highly varied fast directions from local S splitting at JISNET station BMI (Figures 5 and 6), which is located atop this transition zone between the two subduction zones. Therefore, these contorted anisotropic fabrics may perhaps stem from fossil anisotropy frozen into this boundary zone between the two slab fragments. The joining of the two plate fragments is sure to have been accompanied by high stresses and strains, and the fossil anisotropy is the remaining evidence thereof.
6.2.2. Anisotropy in the Mantle Transition Zone
 Three deep-focus events (between 535 and 636 km depth) yield source-sideS splitting results with delay times between ∼0.33 s and 0.79 s, giving evidence to deep anisotropic fabrics within the transition zone (TZ). Three mineral phases dominate volumetrically in the TZ: wadsleyite, ringwoodite, and garnet. Of these, only wadsleyite has significant intrinsic elastic anisotropy [Mainprice et al., 1990] and the potential to induce S wave anisotropy of ∼1% [Tommasi et al., 2004; Demouchy et al., 2011]. Ringwoodite and garnet do not appear to produce significant anisotropy [e.g., Carrez et al., 2006; Tommasi and Mainprice, 2008]. However, wadsleyite is only abundant in the upper TZ, i.e., above 520 km. With the source-side events being below ∼535 km, another mechanism or mineral phase need to be invoked.
 A likely candidate is a dense hydrous magnesium silicate (DHMS): the hydrous D phase [e.g., Frost, 2006; Kawamoto, 2006]. Iwamori  and Ohtani et al. have shown that, along a cold slab geotherm, the hydrous D phase can replace ∼50% of subducted peridotites, making it volumetrically abundant enough to potentially produce significant anisotropy. At TZ pressure–temperature conditions, a D phase single-crystal has aVS anisotropy of ∼17% [Mainprice et al., 2007]. However, there are as yet no studies on deformation of the hydrous D phase and consequently no information about its LPO development. Thus, we cannot confirm that the D phase in the subducted slab is the source of anisotropy; it is, however, a possible option.
Wookey et al.  detected similar deep anisotropy beneath the Tonga–Kermadec and New Hebrides subduction zones, with δtranging from 0.6 s to 7.1 s. They proposed three different models to account for the measured splitting: pooling of slab material in the topmost lower mantle, strain-induced anisotropy around the leading edge of the slab, or an anisotropic boundary layer near the 660 km discontinuity. In an effort to explain these findings,Nippress et al.  modeled subduction body force stresses in the uppermost lower mantle due to a viscosity increase at the 660 km discontinuity and predicted resultant shear wave splitting. The viscosity contrast impedes a slab as it descends [Kusznir, 2000], and this could potentially induce LPO, and thus seismic anisotropy, in the surrounding minerals, i.e., perovskite. Nippress et al.  favor this scenario as an explanation for the deep seismic anisotropy observed by Wookey et al. . As the slab has reached a depth of ∼700 km at the Sunda trench [Gudmundsson and Sambridge, 1998], we view this as a viable mechanism to account for the splitting that we measure here.
 Although we detect clear evidence for sub-520 km anisotropy, we cannot with certainty identify its cause, as this is beyond the scope of this study. A broader, global study of shear wave splitting in the mantle transition zone could potentially shed more light on this phenomenon.
6.2.3. Seismic Anisotropy in the Sumatran Backarc
 At the three stations in the Sumatran backarc (BTDF, TPN, TPI), we measure SKSsplitting with time-lags between 0.68 s and 2.73 km. Althoughδtis quite varied, fast directions are consistently NE–SW trending, i.e., trench-perpendicular. Unfortunately, no other shear phases rendered splitting results at these stations for improved spatial constraints. Fast orientations are nearly perpendicular to those measured byHammond et al. , indicating that we are detecting a different region of anisotropy. There are no geological surface expressions of prominent tectonic features present which would suggest that the splitting is caused by fossil anisotropy of ancient lithospheric-scale faults [e.g.,Bastow et al., 2007, 2011]. Although we cannot rule out a small contribution from the crust, delay times caused by crack-induced crustal anisotropy are typically not more than ∼0.2–0.3 s [Gledhill, 1991; Gledhill and Gubbins, 1995; Morley et al., 2006]. We suggest that the bulk of the splitting is due to olivine LPO in the upper mantle, either due to basal asthenospheric drag or corner flow in the mantle below the overriding Eurasian plate.
6.3. Region 3: Northern Sulawesi and the Molucca Sea
 With stacked averages of 1.25 s and 1.30 s, delay times measured at TOLI and LUWI, respectively, are similar in magnitude to those measured at MNI with δt of ∼1.53 s [Di Leo et al., 2012]. We ascribed this to trench-parallel sub-slab mantle flow due to double-sided subduction of the Molucca Sea plate. With the entire microplate sinking, the mantle is being squeezed outwards on both sides, parallel to the two trenches. Consequently, it seems likely thatSKS phases recorded at LUWI are detecting this same mantle flow as it escapes the sinking plate. In the case of TOLI, fast directions are perpendicular to the strike of the Molucca Sea slab and oblique to the North Sulawesi trench. Placing depth constraints on this anisotropy is difficult, as no local S measurements were obtained at TOLI. The SKS rays are within the influence sphere of both subduction zones, so it is possible that both are contributing to the measured splitting. We relate the splitting measured at PCI to toroidal flow around the SE dipping Celebes Sea slab. Such toroidal flow around lateral slab edges has been observed both in analog [Buttles and Olson, 1998; Schellart, 2004] and numerical models [Lowman et al., 2007]. Delay times are smaller than those recorded at TOLI, LUWI, and MNI with a stacked average of ∼0.98 s. Unlike the Molucca Sea slab, however, the Celebes Sea slab has only reached a depth of ∼200–250 km, and subduction is impeded by collision with the Molucca Sea slab [Walpersdorf et al., 1998; Kopp et al., 1999]. Therefore, it is not surprising that the anisotropic fabrics due to this toroidal flow are less extensive and less pronounced.
 Local S splitting recorded at PCI and LUWI complements measurements taken at MNI, located on the volcanic arc of the Sangihe subduction zone [Di Leo et al., 2012]. In our recent study, we were able to distinguish two different regions of anisotropy in the Sangihe subduction zone due to the excellent depth distribution of local earthquakes. Events originating below ∼380 km detected shear-layers atop the subducting slab, resulting in trench-normal fast directions. LocalSphases of events shallower than ∼380 km were sensitive to trench-parallel aligned cracks, perhaps melt-filled, in the overriding plate, thereby giving trench-parallel fast orientations. We suggest that localS splitting measured at LUWI and PCI is caused by these same phenomena in the Sangihe subduction zone.
6.4. Region 4: Borneo
 Delay times measured at the four Borneo stations are quite variable, ranging from 0.55 to 1.35 s. However, fast directions are fairly consistently parallel to the island's coast line. It is probable that this is due to fossil anisotropy, i.e., anisotropic fabrics frozen into the lithosphere, stemming from extensive counter-clockwise rotation of Borneo in the past. Paleomagnetic data suggest this rotation to have amounted to ∼40° in the Late Cretaceous and ∼45–50° between 25 and 10 Ma ago in the Miocene [Fuller et al., 1999]. Fossil anisotropy bears evidence of the last significant episode of deformation and can remain intact over a long time in the absence volcanism or subsequent deformation [e.g., Silver and Chan, 1991; Helffrich, 1995; Bastow et al., 2007]. Given the lack of subduction, volcanism or other significant deformation in Kalimantan since the Miocene, it is likely that the SKS phases are detecting fossil anisotropic fabrics created during rotation of Borneo in the past.
6.5. The Relation Between Subduction and Mantle Flow
 Despite progress having been made in the past few decades in comprehending subduction zone dynamics, several aspects are still not fully understood. This includes the role of subducted lithosphere in mantle convection or the coupling between plates and the underlying mantle. This set of observations in a tectonically complex region with several active plate margins provides us with a unique opportunity of studying the behavior of the mantle in subduction settings. We infer several different flow features in the upper mantle beneath Indonesia: toroidal flow around the Celebes Sea slab, oblique corner flow in the Banda wedge, trench-parallel flow beneath the arcuate Banda slab, trench-perpendicular flow in the Sumatran backarc, and potentially flow around the leading edge of the Indo-Australian slab. This implies that the discussion about whether mantle flow in subduction zones is either trench-perpendicular or -parallel, which has arisen from studies of seismic anisotropy, is perhaps oversimplified and that flow at convergent plate boundaries is more complex than previously believed. In most regions of oceanic lithosphere, observations of seismic anisotropy suggest mantle flow direction to be well correlated with the surface motion of oceanic plates [e.g.,Becker et al., 2003; Conrad et al., 2007; Becker, 2008; Kreemer, 2009]. However, that may not be true in regions of density driven mantle flow, such as hot spots or subduction zones [e.g., Gaboret et al., 2003; Hammond et al., 2005]. In the subduction zones of Indonesia, fast directions are neither uniformly perpendicular nor parallel to apparent plate motion. This has been observed in several other subduction zones as well, e.g., Kamchatka [Peyton et al., 2001], South America [Russo and Silver, 1994], and the Marianas [Pozgay et al., 2007]. Numerical studies of three-dimensional mantle convection in subduction systems have suggested as much [e.g.,Lowman et al., 2007; Jadamec and Billen, 2010, 2012]. They show that various factors - including slab geometry, morphology, and density; mantle viscosity; temperature gradients along the slab; plate configuration - can all lead to complex mantle flow patterns, including oblique or toroidal flow alongside trench-parallel and -perpendicular flow.Jadamec and Billen [2010, 2012]have argued that when assuming a composite viscosity of the mantle (i.e., Newtonian and non-Newtonian viscosity) and a power law rheology (which allows for deformation by dislocation creep), a partial decoupling between mantle and plates is possible. This is due to the fact that near subducting slabs, dislocation creep deformation rates are higher for a given stress (due to the power law relationship between strain rate and stress), which reduces the effective viscosity [e.g.,Karato et al., 2008]. A consequence of this is that at subduction zones, mantle flow velocities can be up to around eight times faster than surface plate motions, and mantle flow directions do not necessarily parallel plate trajectory [e.g., Jadamec and Billen, 2010, 2012].
 The architecture of the tectonic setting in Indonesia is largely prescribed by the geometry of pre-existing plate boundaries; the region, especially eastern Indonesia, is an amalgamation of various terranes and microplates. As they converged and coalesced, subduction was an inevitable consequence. This implies that the tectonic setting is a precondition for the resulting mantle flow. However, although the mantle is partially decoupled from the slabs, its movement is not simply a passive response to plate movement. The degree of decoupling between plates and mantle is still poorly constrained. Observations of mantle flow directions as inferred from seismic anisotropy can therefore provide predictions that can be modeled to explore the parameters controlling slab–mantle coupling.
 In this study, we present new shear wave splitting results for previously unstudied (with respect to seismic anisotropy) areas of Indonesia: the Sumatran backarc, Borneo (Kalimantan), southern Sulawesi, and the Banda Arc. We put these results into the broader context of the complex tectonic setting of Indonesia and previous studies by Hammond et al.  and Di Leo et al. . Like the latter study, we use three different shear phases (local upgoing S, SKS, and downgoing Sfor source-side splitting) to obtain an improved spatial resolution of anisotropic fabrics. We demonstrate how this method can complement other techniques (e.g., studies of palaeomagnetics or GPS measurements) in unraveling the tectonic history of a region. For example,SKSmeasurements of Borneo show coast-parallel fast directions. This fossil anisotropy, frozen into the lithosphere below Borneo, reflects the island's history of counter-clockwise block rotation as first detected by paleomagnetism [e.g.,Fuller et al., 1999]. Moreover, seismic anisotropy may identify tectonic boundaries that cannot be detected when, for example, examining seismicity-derived slab surface models [Gudmundsson and Sambridge, 1998; Syracuse and Abers, 2006; Hayes et al., 2012]. Hammond et al.  have shown this to be the case for the Sunda Strait separating Sumatra from Java. Similarly, the highly varied fast directions recorded at station BMI southwest of Sulawesi reflect the boundary between the Sunda and the Banda Arc and thus the transition from subduction of oceanic to continental lithosphere. Furthermore, using upgoing and downgoing Sphases, we are able to detect three different domains of anisotropy in the Banda subduction zone: trench-perpendicular to sub-perpendicular shear-layers atop the slab, as detected by localSphases, as well as deep-slab anisotropy and trench-parallel sub-slab anisotropy paralleling the curvature of the Banda arc, as detected by source-side splitting. In northern Sulawesi, localS and SKS splitting results complement findings by Di Leo et al. for the Sangihe subduction zone. We attribute the trench-parallel anisotropy measured at LUWI to sub-slab mantle flow associated with double-sided subduction of the Molucca Sea microplate. We ascribe E–W trending fast directions measured at PCI to toroidal flow around the edge of the Celebes Sea slab at the North Sulawesi subduction zone. Finally, we measure trench-perpendicular fast directions in the Sumatran backarc that we ascribe to mantle flow below the overriding Eurasian plate.
 By using careful, multiphase studies of seismic anisotropy in conjunction with other geophysical and geological evidence, we can pick apart the signal associated with even a very complex tectonic history. We observe mantle flow which is different than - but apparently related to - these complex tectonics. This has implications for the nature of the coupling between subducting plates and the underlying mantle.
 The research leading to these results has received funding from Crystal2Plate, an FP7-funded Marie Curie Action under grant agreement number PITN-GA-2008-215353, and from the European Research Council under the European Union's Seventh Framework Program (FP7/2007-2013)/ERC Grant agreement 240473 ‘CoMITAC’. Data used in this study was provided by GEOFON and the Japan Indonesian Seismic Network (JISNET), courtesy of the Earthquake Research Institute, University of Tokyo. Several figures were created using the GMT software [Wessel and Smith, 1991]. We thank T. Becker, B. Foley and an anonymous reviewer for their comments.