On the Cause of the1930s Dust Bowl

Siegfried D. Schubert,1* Max J. Suarez,1 Philip J. Pegion,1,2

Randal D. Koster,1 Julio T. Bacmeister1,3

During the 1930s, the United States experienced one of the most devas-tating droughts of the past century. The drought affected almost two-thirdsof the country and parts of Mexico and Canada and was infamous for thenumerous dust storms that occurred in the southern Great Plains. In thisstudy, we present model results that indicate that the drought was causedby anomalous tropical sea surface temperatures during that decade and thatinteractions between the atmosphere and the land surface increased itsseverity. We also contrast the 1930s drought with other North Americandroughts of the 20th century.

In the United States, the 1930s werecharacterized by a decade of rainfall defi-cits and high temperatures that desiccatedmuch of the land surface of the GreatPlains. The drought and its associated duststorms created one of the most severe en-vironmental catastrophes in U.S. historyand led to the popular characterization ofmuch of the southern Great Plains as the“Dust Bowl” (1, 2).

While progress has been made in under-standing some of the important processescontributing to drought conditions (3–7 ),the mechanisms by which a drought can bemaintained over many years are not wellestablished. A number of studies have usedthe historical record of meteorological andoceanographic observations to identify sta-tistical relations between slowly varyingPacific Ocean sea surface temperatures(SSTs) and precipitation over the GreatPlains (8, 9). The record of observations,however, is too short to provide definitiveresults for long-term drought. Understand-ing the causes of the 1930s drought isparticularly challenging in view of the scar-city of upper-air meteorological observa-tions prior to about 1950.

Several recent studies using state-of-the-art atmospheric general circulationmodels (AGCMs) have shown how SSTanomalies can produce prolonged droughtconditions. Tropical SST anomalies, in par-ticular, were found to contribute to recentprolonged drought conditions over much ofthe northern middle latitudes (10), todrought in the Great Plains (11), and to

drought conditions in the African Sahelregion during the 1970s and 1980s (12).

The importance of the Pacific SSTs (thepan-Pacific pattern) in forcing long-term pre-cipitation variations in the Great Plains led us toexpect that this pattern would be an importantfactor during the Dust Bowl when drought wasmost severe. SST anomalies, however, weresurprisingly weak throughout the tropical Pacif-ic during the 1930s. This prompted a muchcloser look at the relationship between SSTanomalies and the generation of the Dust Bowl.

Our study is based on a number ofcentury-long simulations carried out withthe NASA Seasonal-to-Interannual Predic-tion Project (NSIPP) atmospheric generalcirculation model (13), the same modelused in (11) and (12), although run here ata somewhat coarser horizontal resolution(14 ). The basic model simulations are anensemble of fourteen 100-year (1902 to2001) runs forced by observed monthlySSTs (15). These simulations will be re-ferred to as the C20C runs, because theywere carried out as part of the Climate ofthe 20th Century project (16 ). The runsdiffer only in their initial atmospheric con-ditions. As such, the degree of similarity inthe runs (the “signal”) provides us with anassessment of how much the SSTs controlGreat Plains climate variations, while the dis-agreement among the runs (the “noise”) pro-vides us with an estimate of the unpredictablecomponent of the climate variability.

Figure 1 shows time series of the pre-cipitation averaged over the Great Plainsand filtered to retain time scales longer thanabout 6 years, using data from observations(17 ) and from the 14 simulations. The timeseries of the observed and simulated anom-alies show considerable variability, withextended periods of both above- and below-normal conditions throughout the century.

1Earth Sciences Directorate, National Aeronautics andSpace Administration/Goddard Space Flight Center,Greenbelt, Maryland 20771, USA. 2Science Applica-tions International Corporation, Beltsville, Maryland20705, USA. 3Goddard Earth Sciences and TechnologyCenter (GEST ), University of Maryland, Baltimore,MD 21250, USA.

*To whom correspondence should be addressed: E-mail: schubert@gmao.gsfc.nasa.gov

Fig. 1. Time series of precipitation anomalies averaged over the U.S. Great Plains region (30°N to50°N, 95°W to 105°W; see box in insets). A filter (28) is applied to remove time scales shorter thanabout 6 years. The thin black curves are the results from the 14 ensemble members from the C20Cruns. The green solid curve is the ensemble mean. The red curve shows the observations. The mapsshow the simulated (left) and observed (right) precipitation anomalies averaged over the Dust Bowlperiod (1932 to 1938). Units, mm/day.

R E P O R T S

www.sciencemag.org SCIENCE VOL 303 19 MARCH 2004 1855

The correlation between the observed andensemble mean anomalies is 0.57. The cor-relation between the ensemble mean andthe individual simulations, a measure of thehighest correlation we can expect with theobservations, ranges from 0.53 to 0.79.While there is considerable scatter among

the ensemble members, there are periodsduring which all the curves tend to followone another. In particular, during the 1930salmost all of the runs show a tendency fordry conditions, consistent with the obser-vations. The dry conditions of the 1930s arefollowed, in the early 1940s, by a rapid

transition by all ensemble members to wet-ter conditions, again consistent with theobservations. In general, the simulationsagree with the observations to the extentthat the observed anomalies fall within thescatter of the ensemble members.

Figure 1 includes maps of the ensemblemean and observed precipitation anomalies av-eraged over the Dust Bowl period (1932 to1938). The observations show deficits exceed-ing 0.1 mm/day covering much of the centralUnited States, with peak deficits exceeding 0.3mm/day centered on Kansas. The simulatedanomalies are similar to the observed, withpeak deficits of similar magnitude and againcentered over Kansas. The main discrepancy isthe large deficit simulated over Mexico thatdoes not occur in the observations (18). Thesimulation also fails to capture the full spatialextent of the drought, particularly the dry con-ditions that were observed over the northernGreat Plains and parts of Canada. An inspectionof the individual ensemble members shows thatthis discrepancy occurs as a result of the ensem-ble averaging, which acts to filter out the un-predictable (noise) component of the simulatedanomalies. In fact, there is wide variability inthe spatial pattern of the dry conditions amongthe ensemble members, with some showingnegative precipitation anomalies (�–0.1) ex-tending well into Canada and covering an areathat exceeds that of the observed anomalies.

Figure 1 also reveals some rather pecu-liar model behavior during the mid-1970s.While the observations and 2 of the en-semble members show a tendency forslightly wet or neutral conditions, 12 of theensemble members predict a droughteven more severe than that during the1930s. This highlights the probabilistic na-ture of the drought prediction problemand suggests, if we believe the model re-sults, that the central United States waslucky to have had near normal conditionsduring the 1970s, because the probability ofhaving a major drought was rather high(12 chances out of 14). In fact, the historicaltendency for droughts to occur in the GreatPlains roughly every 20 years (1910s, 1930s,and 1950s), together with the very dry condi-tions that existed over parts of the central Unit-ed States by the mid-1970s, collectively led tospeculation at the time that we were about toenter an extended dry period (19, 20).

Figure 2 shows our best estimate of theglobal SST anomalies (15, 21) that occurredduring the Dust Bowl period. It must beemphasized that this time-averaged field isbased on extrapolations of a rather limitednumber of ship observations using empiricalorthogonal functions. Figure 2 shows that theanomalies are negative in most places, in-cluding the tropical Pacific and North Pacific,as well as much of the Southern Ocean. Pos-itive anomalies occur in the tropical and

Fig. 2. The global SST anomalies averaged for the Dust Bowl period (1932 to 1938). The boxesdelineate the various subregions (tropical oceans, Indian Ocean, Pacific Ocean, Atlantic Ocean)used in designing the idealized SST forcing experiments. The anomalies are the differences from the1902 to 1999 SST climatology (15). Units, °C.

Fig. 3. (Top left) Ensemble mean precipitation anomalies averaged for the period 1932 to 1938from the C20C runs. (Top right) The 100-year mean precipitation anomalies from the Global SSTrun (Fig. 2 and Table 1). (Bottom left) Same as the top right panel except for the Tropical SST run.(Bottom right) The difference between the precipitation anomalies from the Global SST and theTropical SST runs. Units, mm/day. In all but the C20C panel, values are shaded only if they aresignificant at the 10% level based on a t test. The contour intervals are the same as the shadingintervals (see color bar), with dashed contours indicating negative values.

R E P O R T S

19 MARCH 2004 VOL 303 SCIENCE www.sciencemag.org1856

North Atlantic Oceans and in some regions ofthe South Pacific. Surprisingly, the tropicalanomalies tend to be small, generally lessthan 0.3°C. The largest anomalies occur inthe North Atlantic and just off the coast ofAsia, where they exceed 0.5°C.

To understand the importance of each ofthese features to the 1930s drought, wecarried out a number of idealized experi-ments in which SST anomalies, averagedover the 1932 to 1938 Dust Bowl period,were applied only to particular subregions(see outlines in Fig. 2). The remainder ofthe ocean was assigned climatologicalSSTs. Table 1 summarizes the experiments.Our aim was to separate the contributionsto the drought from each of the three trop-ical basins (Indian, Pacific and Atlantic)and the extratropics (22).

We first examined whether forcing themodel with the 1932 to 1938 time–meanSST anomalies produces the same timemean response in the Great Plains precipi-tation as that obtained in the originalC20C ensemble. The top two panels ofFig. 3 compare the ensemble mean precip-itation anomalies from the C20C runsaveraged from 1932 to 1938 with thosefrom the Global run. The results are quitesimilar. In addition to the dry anomalies,the idealized forcing also reproduces someof the wet anomalies in the Pacific North-west and along the southeast coast. Thereare some discrepancies between the runs.Most notably, the rainfall deficits that ex-tend into Alabama in the Global run areabsent in the C20C runs. Despite such minordifferences, which must be artifacts of theGlobal run, it appears that the basic droughtconditions simulated by the model in the GreatPlains during the 1930s can be explained as aresponse to the time–mean global SST anoma-lies and that the year-to-year variations of theSSTs in that decade played at most a secondaryrole in shaping the drought.

We next tried to distinguish between trop-ical and extratropical effects. The results ofthe tropical run (lower left panel of Fig. 3)show that the main features of the drought arereproduced with the tropical SST forcingalone. The Global-minus-Tropical differencemap (lower right panel) shows that the impactof the extratropical SSTs is much smaller.Extratropical anomalies tend to broaden theregion of drought conditions, especially tothe east and south of the central Great Plains.They also increase the region of wet anoma-lies in the Pacific Northwest.

The precipitation anomalies from thedifferent model runs, including the contri-butions from the different tropical basinsand the extratropics, averaged over the coreDust Bowl region (the rectangles in Fig. 4)are shown in fig. S1. The results show thatthe contributions from the tropical Pacific

and tropical Atlantic are significant, whilethose from the tropical Indian Ocean andthe extratropics are not (22). Figure S1also shows the results of repeating theGlobal run, but in this case disabling theinteractions between the atmosphere andland surface [the Fixed Beta run(22)].Preventing this feedback reduces theprecipitation deficit by 50%. Thus, land-atmosphere interaction appears to be re-sponsible for much of the drought sever-ity. The much tighter confidence intervalsassociated with the Fixed Beta results (fig.S1) show that without soil moisture feed-backs, precipitation variability is greatlyreduced, consistent with previous studiesemploying the same model (7, 11). Thespatial distribution of the precipitation

anomalies from the Fixed Beta run showsthat the reduced deficits span the GreatPlains (Fig. 4). These results are consistentwith previous work (7 ) showing that theGreat Plains region is particularly sensitiveto soil moisture changes.

The results presented so far have beenfor annual-mean conditions. It is wellknown, however, that most of the rain inthe Great Plains tends to fall during thespring and summer seasons. We show inFig. 5 the seasonal distribution of the area-averaged precipitation anomalies from ob-servations, the Global run, and the FixedBeta run. The results from the Global runare quite similar to the observed (23),though there is a general tendency to un-derestimate the deficits. The largest deficits

Fig. 4. (Top) Meanprecipitation anoma-lies from the GlobalSST run. (Bottom)Mean precipitationanomalies from theFixed Beta run. Shad-ing indicates signifi-cance at the 10%level based on a ttest. The contour in-tervals are the sameas the shading inter-vals (see color bar),with dashed con-tours indicating neg-ative values. Theboxes indicate thecore Dust Bowl re-gion (36°N to 39°N,99.5°W to 105°W ).Units, mm/day.

Table 1. The idealized SST experiments. The anomalies are the time-averaged (1932 to 1938) deviationsfrom the 1902 to 1999 mean (22). All runs are 100 years in length. The regions are defined in Fig. 2.

Run SST

Control Climatological (average of 1902 to 1999)Global Global anomaliesTropical Anomalies confined to tropics, climatological elsewherePacific Anomalies confined to tropical PacificPacAtl Anomalies confined to tropical Pacific and tropical AtlanticPacInd Anomalies confined to tropical Pacific and tropical Indian OceanFixed Beta Global anomalies and atmosphere–land surface interaction disabled

R E P O R T S

www.sciencemag.org SCIENCE VOL 303 19 MARCH 2004 1857

occur during the warm season, with abouthalf the deficit occurring during the sum-mer months. Somewhat surprisingly, thefall season has larger deficits than thespring season. The winter season precip-itation anomalies are by comparison rathersmall; in fact, the observed winter anomalyis slightly positive. The main impact ofdisabling the interactions with the land sur-face is to reduce dramatically the deficitduring the summer. Figure 5 includes theensemble mean and 90% confidence inter-vals of the precipitation anomalies fromthe C20C runs. Comparing those resultswith those from the Global run providesfurther evidence that the Global run repro-duces the basic features of the C20C run—in this case, the annual cycle of the anom-alies. The fact that the confidence intervalsfor the C20C anomalies encompass zerofor winter and spring confirms that thesignificant precipitation anomalies for thecore region are confined to the summerand fall season.

These results suggest that we must look tothe summer and fall seasons to understand themechanisms linking tropical SST anomaliesto the Dust Bowl region (24). An analysis ofthe summer circulation changes (not shown)suggests that the role of the cold Pacific SSTanomalies was to generate a global-scale re-sponse in the upper troposphere (negativeheight anomalies in the tropics and a tenden-cy for positive height anomalies in the middlelatitudes) that suppressed rainfall over theGreat Plains. The warm Atlantic SST anom-alies produced two upper-level anticycloniccirculation anomalies on either side of theequator, with the northern anomaly extendingacross the Gulf of Mexico and the southernUnited States. In the lower troposphere, theresponse to the warm Atlantic SST anomalieswas a cyclonic circulation anomaly posi-tioned to suppress the supply of moistureentering the continent from the Gulf of Mex-

ico. It is noteworthy that the Atlantic re-sponse is confined almost exclusively to thesummer and fall season.

While the severity, extent, and durationof the 1930s drought was unusual for the20th century, proxy climate records indi-cate that major droughts have occurred inthe Great Plains approximately once ortwice a century over the past 400 years(25). There is evidence for multidecadaldroughts during the late 13th and 16th cen-turies that were of much greater severityand duration than those of the 20th century(25). For example, tree-ring analyses inNebraska suggest that the drought that be-gan in the late 13th century lasted 38 years(26 ). An analysis of the other major centralU.S. droughts of the 20th century (11) sug-gests that a cool tropical Pacific is commonto all. Only the Dust Bowl drought, how-ever, combined cool Pacific SSTs with awarm Atlantic Ocean. Figure 1 shows thatsince the early 1980s (with the exception of1987 to 1989), the Great Plains generallyexperienced above-normal precipitation.On the other hand, much of the western(especially the southwestern) UnitedStates, including some parts of the GreatPlains, experienced below-normal precipi-tation during the past 5 years, leading tomoderate or extreme drought conditions(27 ). The cause of this most recent droughtis unclear, although a preliminary look atthe relation between SSTs and long-termprecipitation variations over the southwest-ern United States from our C20C runs sug-gests a strong link to the pan-Pacific patterndiscussed earlier. One difference comparedwith the Great Plains is that the southwest-ern United States appears to have a strongerlink to the Indian Ocean SSTs.

References and Notes1. D. Worster, Dust Bowl: The Southern Great Plains in

the 1930s (Oxford University Press, New York, 1979).2. The “Dust Bowl” was loosely defined to be the region

encompassing the western third of Kansas, south-eastern Colorado, the Oklahoma Panhandle, thenorthern two-thirds of the Texas Panhandle, andnortheastern New Mexico.

3. J. Namias, Mon. Weather Rev. 83, 199 (1955).4. K. E. Trenberth, G. W. Branstator, P. A. Arkin, Science

242, 1640 (1988).5. R. Atlas, N. Wolfson, J. Terry, J. Clim. 6, 2034

(1993).6. K. C. Mo, J. Nogues-Paegle, R. W. Higgins, J. Clim. 10,

3028 (1997).7. R. D. Koster, M. J. Suarez, M. Heiser, J. Hydrometeor 1,

26 (2000).8. M. Ting, H. Wang, J. Clim. 10, 1853 (1997).9. M. Barlow, S. Nigam, E. H. Berbery, J. Clim. 14, 2105

(2001).10. M. P. Hoerling, A. Kumar, Science 299, 691 (2003).11. S. D. Schubert, M. J. Suarez, P. J. Pegion, R. D. Koster,

J. T. Bacmeister, J. Clim. 17, 485 (2004).12. A. Giannini, R. Saravanan, P. Chang, Science 302,

1027 (2003).13. J. Bacmeister, P. J. Pegion, S. D. Schubert, M. J. Suarez,

“An atlas of seasonal means simulated by the NSIPP1 atmospheric GCM” (NASA Tech. Memo. No.104606, volume 17, Goddard Space Flight Center,Greenbelt, MD, 2000).

14. The model was run with a 3° latitude by 3.75°longitude horizontal grid as compared with the 2°latitude by 2.5° longitude grid used in the earlierstudy (11). The resolution change was made forcomputational efficiency and does not appear tohave much impact on the model’s basic climatologyor its variability.

15. N. A. Rayner, et al., J. Geophys. Res. 108, 4407(2003).

16. C. Folland, J. Shukla, J. Hinter, M. Rodwell, in CLIVARExchanges, vol. 7, no. 2 (June 2002), pp. 37–39.

17. R. Vose et al., “The global historical climatologynetwork: long-term monthly temperature, precip-itation, sea level pressure, and station pressuredata” (ORNL/CDIAC-53, NDP-041. Carbon DioxideInformation Analysis Center, Oak Ridge NationalLaboratory, Oak Ridge, Tennessee, 1992).

18. Mexico experienced somewhat above-normal precip-itation during the 1930s. This may have been due, inpart, to a number of hurricanes that made landfallover Mexico, especially during 1933 and 1936. Thecoarse resolution of the model runs precludes thesimulation of hurricanes; this may account for someof the unrealistic dry conditions simulated by themodel over Mexico.

19. J. R. Borchert, Ann. Assoc. Am. Geogr. 61, 1 (1971).20. N. J. Rosenberg, in North American Droughts, N. J.

Rosenberg, Ed. (AAAS Selected Symposia Series,Westview Press, Boulder Colorado, 1978), pp. 1–7.

21. The SST anomalies used to force the model werecomputed separately for each calendar month as thesum of a mean annual cycle (computed for the period1902 to 1999) and the SST anomaly field for thatcalendar month. In Fig. 2, we show only the time-averaged anomalies.

22. Materials and methods are available as supportingmaterial on Science Online.

23. It should be noted that current AGCMs (includingthis model) tend to have difficulty simulatingwarm-season continental rainfall. In the GreatPlains, this is primarily due to a poor representa-tion of nighttime precipitation and appears to bestrongly tied to deficiencies in how the modelsrepresent convective rain. The good comparisonwith observations shown in Fig. 5 indicates thatdespite any such deficiencies in the warm seasonprecipitation climatology, the NSIPP model pro-duces a realistic annual cycle in the response to theSST forcing.

24. It can be argued that precipitation deficits duringsummer are the result of anomalies that are “re-membered” from the previous spring and winter. Inthese calculations, any such memory of the previ-ous season’s condition would come from the landsurface. The Fixed Beta run suggests, that at most,such an impact would account for 50% of thesummer precipitation deficit. Furthermore, the lackof substantial ensemble mean precipitation deficits

Fig. 5. The seasonal variationof the mean precipitationanomalies averaged over theU.S. Great Plains (see box ininsets in Fig. 1). The black barsshow the observed values aver-aged over the Dust Bowl period(1932 to 1938). The light graybars are the 100-year averagedresults from the Global SSTrun. The dark gray bars are the100-year averaged results fromthe Fixed Beta run. The smallboxes and associated thin barsare the ensemble mean and90% confidence intervals, re-spectively, of the 1932 to 1938precipitation anomalies from the C20C runs. The labels on the abscissa indicate the seasons,where DJF is December, January, February; MAM is March, April, May; JJA is June, July, August;SON is September, October, November. Note that for DJF, the Fixed Beta run producesanomalies that are too small to show up in the figure.

R E P O R T S

19 MARCH 2004 VOL 303 SCIENCE www.sciencemag.org1858

over the continental United States during thespring and winter seasons (not shown) suggeststhat the impact of the land is primarily a contem-poraneous feedback and not the result of memoryof the previous season’s anomalies. We note thatthe observations do show substantial precipitationdeficits to the north and east of the core regionduring spring, but similar anomalies are also obtainedfor some of the individual model runs, suggestingthat the observed precipitation anomalies in thespring may not have been forced by SST anomalies.

25. C. A. Woodhouse, J. T. Overpeck, Bull. Am. Meteorol.Soc. 79, 2693 (1998).

26. D. L. Bark, in North American Droughts, N. J. Rosen-berg, Ed. (AAAS Selected Symposia Series, WestviewPress, Boulder Colorado, 1978), pp. 9–23.

27. U.S. Drought Monitoring Page (www.drought.unl.edu/dm/monitor.html).

28. Y. Zhang, J. M. Wallace, D. Battisti, J. Clim. 10, 1004(1997).

29. This work was supported by the NASA Earth Science

Enterprise’s Global Modeling and Analysis Program andthe NASA Seasonal-to-Interannual Prediction Project.

Supporting Online Materialwww.sciencemag.org/cgi/content/full/303/5665/1855/DC1Materials and MethodsFig. S1References and Notes

23 December 2003; accepted 20 February 2004

Laboratory Earthquakes: TheSub-Rayleigh–to–Supershear

Rupture TransitionKaiwen Xia,1,2 Ares J. Rosakis,2 Hiroo Kanamori1*

We report on the experimental observation of spontaneously nucleatedsupershear rupture and on the visualization of sub-Rayleigh–to–supershearrupture transitions in frictionally held interfaces. The laboratory experi-ments mimic natural earthquakes. The results suggest that under certainconditions supershear rupture propagation can be facilitated during largeearthquake events.

The surface-wave magnitude (Ms) 8.1 (Mw

7.8) central Kunlunshan earthquake thatoccurred on 14 November 2001 was an ex-traordinary event from the point of view ofdynamic rupture mechanics. The rupture oc-curred over a long, near-vertical, strike-slipfault segment of the active Kunlunshan faultand featured an exceptionally long (400 km)surface rupture zone and large surface slipdisplacements (1). Modeling of the rupturespeed history (2) suggests rupture speeds thatare slower than the Rayleigh wave speed, cR,for the first 100 km, transitioning to super-shear (speed higher than the shear wavespeed, cS) for the remaining 300 km of rup-ture growth. Other events, such as the 1979Imperial Valley earthquake (3, 4), the 1992Landers earthquake (5), the 1999 Izmit earth-quake (6), and the 2002 Denali earthquake(7), may also have featured supershearspeeds. Supershear was also predicted theo-retically (8, 9) and numerically (10, 11). Evenwith these estimates and predictions at hand,the question of whether natural earthquakeruptures can propagate at supershear speedsis still a subject of active debate. In addition,the exact mechanism for transition of aspontaneously nucleated rupture from sub-Rayleigh to supershear rupture speed is notclear. One answer to this question was pro-vided by the two-dimensional Burridge-

Andrews mechanism (BAM) (10). Recentnumerical investigations of frictional rup-ture have suggested alternative, asperity-based, three-dimensional mechanisms (12–14). Whether and how supershear ruptureoccurs during earthquakes has importantimplications for seismic hazards, becausethe rupture speed influences the characterof near-field ground motions.

To answer the above-stated questions, weconducted experiments that mimic the earth-quake rupture processes. Our goal was toexamine the physical plausibility and condi-tions under which supershear ruptures can begenerated in a controlled laboratory environ-ment. We studied spontaneously nucleateddynamic rupture events in incoherent, fric-tional interfaces held together by far-fieldtectonic loads. Thus, we departed from ex-perimental work that addresses the super-shear ruptures of coherent interfaces loadedby impact-induced stress waves (15, 16).

An exploding wire-triggering mechanism(Fig. 1C), which simulates a localized pres-sure release, was used to trigger the rupture(17). This triggering mechanism is inspiredby recent numerical work on rupture alongfrictional interfaces (18, 19). Experimentally,it is a convenient way of triggering the sys-tem’s full-field, high-speed diagnostics (Fig.1A) that would otherwise be unable to cap-ture an event with a total duration of �50 �s.

More than 50 experiments, featuring arange of angles � and far-field pressures P,were performed, and the symmetric bilateralrupture process histories were visualized inintervals of 2 �s. Depending on P and �,rupture speeds that are either sub-Rayleigh or

supershear were observed. The maximumshear stress field for an experiment with � �25° and P � 7 MPa (Fig. 2A) shows that thespeed of the rupture tip is very close to cR andfollows closely behind the circular shearwave front that is emitted at the time ofrupture nucleation. The same was found to betrue for smaller angles and lower pressures.For an experiment with � � 25° and P � 15MPa (Fig. 2B), the circular trace of the shearwave is also visible and is at the same loca-tion as in Fig. 2A. However, in front of thiscircle a supershear disturbance, featuring aMach cone (pair of shear shock waves), isclearly visible. For this case, the sequences ofimages before 28 �s have a similar form tothe image displayed in Fig. 2B and reveal adisturbance that was nucleated as supershear.Its speed history, v(t), is determined indepen-dently by either the rupture length record orby measuring the angle, �, of the shearshocks with respect to the fault plane andwith the use of the relation v � cS/sin �. Itsspeed was 1970 m/s, which is close to thelongitudinal wave speed cP. In previousexperiments involving strong, coherent in-terfaces and stress wave loading, stablerupture speeds near �2cS were observed(15). This apparent discrepancy can be ex-plained by referring to the rupture velocitydependence on the available energy per unitcrack advance within the supershear regime(16). This energy attains a maximum valueat speeds closer to �2cS for strong inter-faces. For weaker interfaces, this maximummoves toward cP.

To visualize a transition within our fieldof view (100 mm), we kept � � 25° butreduced P to 9 MPa (Fig. 3, A to C). Thecircular traces of P and S waves are visible,followed by a rupture propagating initially atcR (Fig. 3A). A small secondary rupture ap-pears in front of the main rupture and prop-agates slightly ahead of the S wave front (Fig.3B). The two ruptures coalesce, and the lead-ing edge of the resulting rupture grows at aspeed close to cP. The transition length L hereis �20 mm (Fig. 3D).

The above transition phenomenon is com-parable with BAM, which is described in(20). Andrews investigated this transition in aparameter space spanned by a normalizedsupershear transition length, L/Lc, and a non-dimensional driving stress parameter, s [s �

1Seismological Laboratory, MC 252-21, California In-stitute of Technology (Caltech), Pasadena, CA 91125,USA. 2Graduate Aeronautical Laboratories, MC 105-50, Caltech, Pasadena, CA 91125, USA.

*To whom correspondence should be addressed. E-mail: hiroo@gps.caltech.edu

R E P O R T S

www.sciencemag.org SCIENCE VOL 303 19 MARCH 2004 1859