letters to nature

been as high as 7,000 m3 s-1 early in the eruption, or ,550 m3 s-1 if 18. Thordarson, T. & Self, S. The Laki (Skaftar Fires) and Grmsvotn eruptions in 1783-1785. Bull.
Volcanol. 55, 233263 (1993).
averaged over the 72 minutes of ow ination. These numbers are at 19. Mogi, K. Relations between the eruptions of various volcanoes and the deformation of the ground
the upper end of discharge rates reported for basaltic volcanoes on surfaces around them. Bull. Earthquake Res. Inst. Univ. Tokyo 36, 99134 (1958).
land1418. 20. West, M. E., Menke, W. & Tolstoy, M. Massive magma reservoir beneath Axial volcano, Juan de Fuca
After the ination/drain-back signal in the VSM2 pressure ridge. Nature (submitted).

record, the instrument also recorded a further gradual subsidence

of the sea oor by 1.4 m over the next 5 days (Fig. 3). This part of the Acknowledgements
signal is due to the deation of the summit of Axial volcano as We thank H. Milburn, C. Meinig and P. McLain for helping to build, maintain and rescue
the VSM2 instrument; the ROPOS group and the crews of the NOAA ship Ron Brown and
magma moved from the reservoir beneath the caldera and into the
RV Thompson for their help with successful recovery of VSM2; and S. Carbotte and J. Fink
south rift zone. This deation signal was also observed by two other for comments on the manuscript. This work was supported by the NOAA Vents Program
sea-oor instruments during the same time period: VSM1, and the West Coast and Polar Regions Undersea Research Center..
deployed near the centre of the caldera, recorded a subsidence of
Correspondence and requests for materials should be addressed to W.W.C.
3.2 m (ref. 5), and an array of extensometers on Axial's north rift (e-mail: [email protected]).
zone measured a 4-cm decrease of a 405-m-long horizontal
baseline6. This distance decrease can be explained if the baseline
endpoints moved 68 cm radially inward toward the centre of the
caldera during deation6. These are, to our knowledge, the rst in
situ ground deformation measurements made during a submarine
volcanic eruption. The VSM1 and extensometer data were used
.................................................................
before the data were recovered from VSM2to calculate6 the depth Spanwise ow and the attachment
of the magma reservoir (3.8 km) and the volume of magma removed
from the reservoir beneath the caldera (207 106 m3), on the basis of of the leading-edge vortex
a point-source elastic deformation model19. We note that the third
data point from VSM2 ts exactly on the same subsidence curve on insect wings
predicted by this deformation model (Fig. 3c). This agreement gives
increased condence in the model results, which are also consistent James M. Birch & Michael H. Dickinson
with estimates of the depth of Axial's magma reservoir based on
seismic tomography20. Department of Integrative Biology, University of California, Berkeley,
Little is known about deep-sea eruptions, because we have only California 94720, USA
been able to detect them for about a decade and none have ever been ..............................................................................................................................................
witnessed. The data we report here, recorded by the VSM2 instru- The ow structure that is largely responsible for the good
ment caught in the 1998 lava ow at Axial volcano, were obtained by performance of insect wings has recently been identied as a
fortuitous circumstance. The instrument was simply in the right leading-edge vortex1,2. But because such vortices become detached
place at the right time, with the right sensors, and happened to have from a wing in two-dimensional ow1, an unknown mechanism
the right physical design to survive the eruption. M must keep them attached to (three-dimensional) apping wings.
The current explanation, analogous to a mechanism operating on
Received 17 April; accepted 28 June 2001.
delta-wing aircraft, is that spanwise ow through a spiral vortex
1. Fox, C. G., Dziak, R. P., Matsumoto, H. & Schreiner, A. E. Potential for monitoring low-level drains energy from the vortex core3. We have tested this hypoth-
seismicity on the Juan de Fuca Ridge using military hydrophone arrays. Mar. Technol. Soc. J. 27, 2230
esis by systematically mapping the ow generated by a dynami-
(1994).
2. Pert, M. R. & Chadwick, W. W. Jr in Faulting and Magmatism at Mid-Ocean Ridges (eds Buck, W. R., cally scaled model insect while simultaneously measuring the
Delaney, P. T., Karson, J. A. & Lagabrielle, Y. 59116 (AGU Geophysical Monograph 106, American resulting aerodynamic forces. Here we report that, at the Reynolds
Geophysical Union, Washington DC, 1998). numbers matching the ows relevant for most insects, apping
3. Dziak, R. P. & Fox, C. G. The January 1998 earthquake swarm at Axial Volcano, Juan de Fuca Ridge:
Hydroacoustic evidence of seaoor volcanic activity. Geophys. Res. Lett. 26, 34293432 (1999).
wings do not generate a spiral vortex akin to that produced by
4. Embley, R. W., Chadwick, W. W. Jr, Clague, D. & Stakes, D. The 1998 eruption of Axial Volcano: delta-wing aircraft. We also nd that limiting spanwise ow with
Multibeam anomalies and seaoor observations. Geophys. Res. Lett. 26, 34252428 (1999). fences and edge bafes does not cause detachment of the leading-
5. Fox, C. G. In situ ground deformation measurements from the summit of Axial Volcano during the edge vortex. The data support an alternative hypothesisthat
1998 volcanic episode. Geophys. Res. Lett. 26, 34373440 (1999).
6. Chadwick, W. W. Jr, Embley, R. W., Milburn, H. B., Meinig, C. & Stapp, M. Evidence for deformation
downward ow induced by tip vortices limits the growth of the
associated with the 1998 eruption of Axial Volcano, Juan de Fuca Ridge, from acoustic extensometer leading-edge vortex.
measurements. Geophys. Res. Lett. 26, 34413444 (1999). The failure of conventional theory to explain the hovering ight
7. Baker, E. T., Fox, C. G. & Cowen, J. P. In situ observations of the onset of hydrothermal discharge
of insects prompted the search for unsteady aerodynamic mechan-
during the 1998 submarine eruption of Axial Volcano, Juan de Fuca Ridge. Geophys. Res. Lett. 26,
34453448 (1999). isms that could account for the good performance of apping
8. Grifths, R. W. & Fink, J. H. Solidication and morphology of submarine lavas: A dependence on wings. One such mechanism, dynamic stall, is manifest by the
extrusion rate. J. Geophys. Res. 97, 1972919737 (1992). formation of a leading-edge vortex (LEV) on the top surface of the
9. Gregg, T. K. P. & Fink, J. H. Quantication of submarine lava-ow morphology through analog
experiments. Geology 23, 7376 (1995).
wing at high angles of attack15. However, in both experimental and
10. Gregg, T. K. P. & Chadwick, W. W. Jr. Submarine lava-ow ination: A model for the formation of lava computational studies of two-dimensional (2D) wings started from
pillars. Geology 24, 981984 (1996). rest, the LEV builds in strength until it detaches from the wing and is
11. Chadwick, W. W. Jr, Gregg, T. K. P. & Embley, R. W. Submarine lineated sheet ows: A unique lava shed into the wake, being replaced by a trailing-edge vortex of the
morphology formed on subsiding lava ponds. Bull. Volcanol. 61, 194206 (1999).
12. Fox, C. G. Evidence of active ground deformation on the mid-ocean ridge: Axial Seamount, Juan de
opposite sign6,7. This pattern of vortex growth and shedding repeats,
Fuca Ridge, April-June, 1988. J. Geophys. Res. 95, 1281312822 (1990). giving rise to a trail of counter-rotating vortices known as a Karman
13. Gregg, T. K. P. & Fornari, D. J. Long submarine lava ows: Observations and results from numerical street. In three-dimensional (3D) models of both hawkmoths
modeling. J. Geophys. Res. 103, 2751727532 (1998).
(Manduca sexta) and fruities (Drosophila melanogaster), the LEV
14. Wadge, G. The variation of magma discharge during basaltic eruptions. J. Volcanol. Geotherm. Res. 11,
139168 (1981). is not quickly shed, but rather remains attached to the wing
15. Lockwood, J. P. & Lipman, P. W. Holocene eruptive history of Mauna Loa Volcano. Prof. Pap. US Geol. throughout the stroke3,8,9. There are currently two hypotheses to
Surv. 1350, 509536 (1987). explain this prolonged attachment of the LEV in 3D ows. One
16. Wolfe, E. W., Neal, C. A., Banks, N. G. & Duggan, T. J. Geologic observations and chronology of
eruptive events. Prof. Pap. US Geol. Surv. 1463, 198 (1988).
possibility, analogous to a mechanism thought to stabilize an
17. Rowland, S. K. & Walker, G. P. L. Pahoehoe and a'a in Hawaii: Volumetric ow rate controls the lava attached vortex on delta-wing aircraft such as Concorde10, is that
structure. Bull. Volcanol. 52, 615628 (1990). base-to-tip spanwise ow in the form of a spiral vortex limits the

NATURE | VOL 412 | 16 AUGUST 2001 | www.nature.com 2001 Macmillan Magazines Ltd 729
letters to nature
growth of leading-edge vorticity by removing energy from the strength of vorticity along the length of the wing. In contrast to the
vortex core3,11. This model, based on ow visualizations of a low velocity of spanwise ow within the core, DPIV showed a broad
dynamically scaled mechanical hawkmoth, has been supported by sheet of base-to-tip spanwise ow on the dorsal surface of the rear
a computational uid-mechanics simulation12. An alternative two-thirds of the wing, with peak velocities approaching 40% of
hypothesis is that induced downward ow from tip vortices and wing-tip velocity (Fig. 1c). The source of this ow is apparent when
wake vorticity reduces the effective angle of attack, attenuating the the wing was viewed from the rear (Fig. 1d). The laterally directed
growth of the LEV. The wing then rotates, reverses direction, and spanwise ow along the wing surface is induced by chord-wise
sheds the accumulated vorticity before the LEV can enlarge to an vorticity, most of which rolls up into a coherent tip vortex that is
unstable condition and detach13,14. contiguous with the LEV after it detaches from the wing and bends
We tested both these hypotheses using a dynamically scaled backward. In sum, the visualizations indicate that the ow structure
robotic model of a Drosophila wing apping in mineral oil8 (see generated by a model fruity wing at a Reynolds number of 160 is
Methods). One wing of the model was equipped with a sensor to substantially different from that described for a model hawkmoth
measure aerodynamic forces. In addition we characterized the ow wing. We found no structure analogous to the spiral vortex of a
around the wing using digital particle velocimetry (DPIV; see Fig. 1a delta-wing aircraft, and the core of the LEV, as dened by the region
and Methods). Flow visualization indicated that uid separates at of rotation ow, contained axial ow of very low velocity.
the leading edge of the wing, forming a shear layer of vorticity that Although the magnitude of axial ow within the core was small
concentrates into a large attached LEV (Fig. 1b). Velocity transects for the model fruity wing, it might still be critical for limiting the
through the region of concentrated vorticity indicate a central core growth of the LEV and maintaining attachment throughout the
of rotational ow as expected of an idealized vortex15 (Fig. 1b, inset). stroke. To test this hypothesis, we constructed a wing with teardrop-
Plots of spanwise velocity (uz) reconstructed from rear views shaped fences at 0.40R and 0.60R (where R is the distance from base
indicated that the axial ow velocity within the core of the LEV to tip), which were designed to extend within the core region of the
was quite small, only 25% of the average wing-tip velocity. Some LEV and limit spanwise ow. When viewed from the rear, DPIV
axial ow within the core of the LEV is expected, driven by a showed that the fences lowered the axial ow of uid within the core
spanwise pressure gradient that results from the variation in the below detectable levels, but had very little effect on the structure of

b,c
a b
A 0.15
u
y C
0.0
0.01
d uy ux 0.12
A B
u
uz x 0.0
z 0.1
x B C D

5 cm D

Light sheets
d 0.15 c
uz 0 0
.0
A 0.15
3
0
.03
A B

0.07

0.
16

0
.2
0 6
0. 0 .30
07 .3
5

0.
17
B
0.02

0.02

0.15 0.05 0 0.05 0.15 0.15 0.05 0 0.05 0.15

Vorticity (s 1) Vorticity (s 1)

Figure 1 Maximum axial ow occurs behind the leading-edge vortex. a, Close-up of wing red star represents location of maximum vorticity. White lines indicate ux (m s-1) transects
and the position of one light sheet during the separate 22 side-view and 34 rear-view graphed in inset. c, The spanwise velocity plot at 0.48R superimposed as a contour map
digital particle image velocimetry (DPIV) experiments. u, velocity; v, vorticity (subscripts on the vorticity plot. Values are normalized to instantaneous wing-tip velocity. Negative
x, y and z indicate vector components in Cartesian space). DPIV performed every 1 cm in values represent ow out of the page. d, Rear view of vorticity at the trailing edge (black)
both views. Letters indicate position of the following panels. b, Vorticity and ow pattern at with superimposed velocity vectors. Wing tip is to the left, and the white line represents
0.48R in chord-wise view at midstroke; leading edge to the left. Velocity vectors are the uz (m s-1) velocity transect (inset).
superimposed on a 15 15 cm area around the wing. The black bar represents the wing,

730 2001 Macmillan Magazines Ltd NATURE | VOL 412 | 16 AUGUST 2001 | www.nature.com
 letters to nature
the LEV (Fig. 2a). This result, together within the absence of any ow over the wing surface, we constructed a clear acrylic wall that
structure resembling a spiral vortex, suggest that the maintenance of exactly matched the sweep of the wing to block spanwise ow at
vortex attachment on fruity wings does not occur through a the tip (Fig. 2c). As with the fences, this manipulation had no
mechanism analogous to that operating on delta-wing aircraft. effect on the dynamics of vortex attachment. However, the pres-
However, one possibility is that the axial ow behind the leading ence of the cylindrical wall increased the total strength of the LEV
edge plays the same role in removing energy from the growing by 14%, resulting in an 8% increase in force during the stroke.
vortex as the ow within a spiral vortex on a Concorde or These results show that the LEV can remain attached throughout
hawkmoth. To test this hypothesis, we disrupted the broad posterior the stroke, with no signs of initiating Karman shedding, even when
sheet of spanwise ow by constructing teardrop fences that pro- its strength is articially increased by a solid boundary at the tip.
jected rearward towards the trailing edge (Fig. 2b). But rather than Thus, the net aerodynamic performance is limited by the magni-
increasing the strength of the LEV as expected if the spanwise ow tude of the LEV and not by its stable attachment throughout the
limits vortex growth, this manipulation caused instead a small stroke. Further, the ow structure generated by the fruity wing
decrease in vortex strength as manifest by a 25% drop in net appears more robust than that described for a hawkmoth at higher
force. More importantly, as with the anteriorly directed fences, Reynolds numbers4.
the rearward fences did not alter the time course of force generation If spanwise ow does not remove energy from the LEV in 3D,
relative to the fenceless wing, indicating that the experimental what then explains its constant attachment relative to the 2D case?
manipulations had no effect on the dynamics of vortex attachment One possibility, suggested by the rise in vortex strength in the
throughout the stroke. In a third experiment aimed at manipulating presence of a wall, is that downward ow induced by both the tip
vortex and previously shed wake vorticity substantially lowers the
effective angle of attack thereby limiting the growth of the LEV
compared to the 2D case. On conventional xed-wing aircraft,
a downwash from tip vortices reduces the `effective' or `aerodynamic'
angle of attack, resulting in a phenomenon termed induced drag15.
As it approaches the undersurface of the apping wing, uid from
the free stream is gradually deected downward until it is oriented
parallel to the surface of the wing near the boundary layer (Fig. 3a).
We estimated the aerodynamic angle of attack by measuring the
mean orientation of ow vectors in the region between the free
stream and the undersurface of the wing (box B, Fig. 3a). The
aerodynamic angle of attack changes as a function of wing span
from 438 at the tip to as low as 218 at 0.40R (Fig. 3b). Even in 2D, the
angle of attack measured within this transition region would be
lower than estimated from the free stream due to downwash
induced by spanwise vorticity attached to the wing. However, the
b effect in 3D will be substantially worse due to the ow induced by
chord-wise tip vorticity and the free vorticity within the wake shed
during previous strokes.
In order to examine the inuence of this additional downwash on
force production, we examined the ows and forces generated by a
sequence of successive downstrokes starting from rest (Fig. 3ce). In
the rst stroke, when there is a tip vortex present but the wake is not
yet established, both the lift and aerodynamic angle of attack are
high (Fig. 3d, e). By the third stroke, the aerodynamic angle of attack
and the mean lift reach an asymptotic level as the full wake structure
becomes established. In the second stroke, the measured values for
both the angle of attack and mean lift were markedly lower than in
c the rst stroke or all subsequent strokes. One possible explanation
for this transient attenuation, which we consistently observed, is
that the deleterious inuence of the wake is particularly strong in the
Acrylic
second stroke because of the increased vorticity generated and shed
sheet during the rst stroke. However, no matter what is responsible for
especially large downwash in the second stroke, there is a strict
Model correlation between the aerodynamic angles of attack and lift
wing (r2 = 0.998), which suggests that downwash has a potent inhibitory
effect on the strength of the LEV and the magnitude of force
production. In total, the ow visualizations and force measure-
ments suggest that the downward ow induced by the tip vortex and
wake vorticity limit the growth of the LEV and contribute to its
0.15 0.10 0.05 0 0.05 0.10 0.15
prolonged attachment.
Vorticity (s 1)
Because a spiral vortex is clearly present on a robotic model
hawkmoth, our results suggest that the precise ow structure of a
Figure 2 The leading edge vortex (LEV) remains attached despite experimental LEV may depend critically on Reynolds number. Fruities ap their
manipulation. Left column shows wing morphology, dotted line is position (0.48R) wings within a range of Reynolds number of 100250, whereas the
of vorticity plots (right column). a, Forward-pointing fences. b, Rearward-pointing much larger hawkmoth operates at Reynolds numbers in excess of
fences. c, Wall. Compared to the two fence manipulations, the presence of a wall 2,000. One possible explanation for the difference in ow structure
causes the LEV to enlarge and extend farther along the wing before detaching into a is that the pressure gradients within the vortex core are too small to
tip vortex. drive a substantial axial ow on the smaller wing of Drosophila.

NATURE | VOL 412 | 16 AUGUST 2001 | www.nature.com 2001 Macmillan Magazines Ltd 731
letters to nature
a
c Stroke 1 Stroke 2 Stroke 3

A B C

0.1 0.0 0.1

1
u y velocity (m s )
0.05 0.1 0.15 0.2 0.25
Velocity (m s1)

b d e
50
0.6
Angle of attack ()

40

Angle of attack ()
40 0.51
0.4
Lift (N)

Mean lift (N)

30
36 0.49
20 0.2
32 0.47
10 0.0
28 0.45
0
1.0 0.8 0.6 0.4 0.2 Start Flip 1 2 3 4
Position along wing (% R) Downstroke Stroke

Figure 3 Induced downwash lowers both the aerodynamic angle of attack and the lift leading edge. Mean Uy velocity within box B from (a): stroke 1, 0.0026 m s-1; stroke 2,
generated by the LEV. a, Wing viewed from the side (white bar) at 0.60R. Shown are ow 0.0056 m s-1; stroke 3, 0.0054 m s-1. d, Lift forces during four downstrokes. Note the
velocity magnitudes (pseudocolour) with velocity vectors superimposed during a fourth attenuating effect of the wake on forces during translation (middle third of stroke). First
stroke. We calculated effective angle of attack in three areas along the wing (boxes A, B stroke, blue; second stroke, red; third stroke, green; fourth stroke, black. e, Mean
and C). b, Although the wing was set to an angle of attack of 458 throughout the stroke, effective angle of attack measured in box B (blue) and mean lift force (red) for the central
the effective angle of attack changes along the wing from tip to base; closed circles third of the wing and central third of the stroke (dashed lines in d), respectively, for each of
represent mean values from box A; open circles, box B; triangles, box C. c, Pseudocolour four strokes.
plots represent vertical uid velocity when viewing wing from the rear, 2 cm behind the

However, because the body length of a typical adult insect is in the Digital particle image velocimetry
range of 45 mm (refs 16, 17), fruities (body length ,2.54 mm) To visualize wake structure, the tank of mineral oil was lled with bubbles by pumping air
might represent a better general model for the uid mechanics of through a ceramic water-purier lter. After larger bubbles had risen to the top of the tank,
smaller bubbles were effectively stationary relative to wing motion and these created a
hovering ight than the much larger hawkmoths (body length highly concentrated and stable seed for DPIV. We used a dual Nd:YAG laser system
,42 mm; ref. 18). In addition, the observation that a apping (Insight v. 3.0, TSI) to create two identically positioned light sheets ,2.5 mm thick
wing at lower Reynolds number can generate comparable perfor- separated by 2,500 ms. A two-frame cross-correlation of pixel intensity peaks with 50%
mance in the absence of a spiral vortex raises the possibility that overlap of 64 64 pixel interrogation areas resulted in 900 velocity vectors per sheet
position. Whether near the clear wing or in the uid, vector validation resulted in removal
while such a structure is present at higher Reynolds numbers, it is of an average of 1.2 vectors per slice (less than 0.2% of vectors in each image); these missing
not necessary for the continued attachment of the LEV. We suggest values were lled by interpolation from a 3 3 nearest neighbour matrix. For each wing
instead that the prolonged attachment of the LEV on insect wings (regular, front-fence, rear-fence, wall) we acquired 24 views, each 20 20 cm, perpen-
may be due to the attenuating effect of the downwash induced by dicular to the long axis of the wing (that is, from the side) at mid-downstroke, moving out
wake vorticity. M the wing from 0.24R to 1.12R (where R is wing length) in 1-cm increments. We then
moved the laser and camera to acquire 34 views parallel to the long axis of the wing (that is,
from the rear), starting 6 cm in front of the leading edge and moving in 1-cm increments to
Methods 14 cm behind the trailing edge of the wing. From these 34 rear views, we constructed new
Robotic model arrays from the wing base to tip that held yz velocity and x vorticity. Although the
morphological angle of attack (AOA) was 458, we calculated the effective AOA (a9) at each
This consists of two robotic arms immersed in a 1 1 2 m Plexiglas tank lled with point in the uid as a9 = 45 - atan(uy/ux). All calculations were made using custom
mineral oil (density, 0.88 103 kg m-3; kinematic viscosity, 115 cSt)8. In this experiment, software programs written in Matlab (v.5.3, MathWorks).
we removed one arm and describe the pattern from a single apping wing. We constructed
a model Drosophila wing from 2.25-mm-thick clear acrylic, and apped it at 168 mHz in a
symmetric pattern (stroke amplitude, 1808; angle of attack at midstroke, 458 up and Received 11 October 2000; accepted 17 May 2001.
downstroke), which resulted in a mean Reynolds number of 160. A force sensor at the wing 1. Dickinson, M. H. & Gotz, K. G. Unsteady aerodynamic performance on model wings at low Reynolds
base measured forces orthogonal and parallel to the wing surface, from which we numbers. J. Exp. Biol. 174, 4564 (1993).
calculated lift and drag coefcients. To limit axial ow, we built two wings with teardrop 2. Ellington, C. P. The aerodynamics of hovering insect ight. IV. Aerodynamic mechanisms. Phil. Trans.
fences at 0.40R and 0.60R. One wing had fences extending toward the leading edge, the R. Soc. Lond. B 305, 79113 (1984).
other wing had fences toward the trailing edge. We also built an acrylic wall with radius 3. Ellington, C. P., Van den Berg, C., Willmott, A. P. & Thomas, A. L. R. Leading-edge vortices in insect
equal to wing length (25 cm). When in position, this wall (extending 21 cm above and ight. Nature 384, 626630 (1996).
35 cm below the wing) covered 2108 of arc, and was within 1 mm of the wingtip 4. Van den Berg, C. & Ellington, C. P. The three-dimensional leading-edge vortex of a `hovering' model
throughout the stroke. hawkmoth. Phil. Trans. R. Soc. Lond. B 352, 329340 (1997).

732 2001 Macmillan Magazines Ltd NATURE | VOL 412 | 16 AUGUST 2001 | www.nature.com
 letters to nature
5. Sunada, S., Sakaguchi, A. & Kawachi, K. Airfoil section characteristics at low Reynolds number. tion and transfer the concept of modulation direction to novel
J. Fluids Eng. 119, 129135 (1997).
6. Kyia, M. & Arie, M. A contribution to an inviscid vortex-shedding model for an inclined at plate in
stimuli6. Transfer to novel stimuli is considered to be the most
uniform ow. J. Fluid Mech. 82, 223240 (1977). decisive behavioural index for category learning12,13. The training
7. Sarpkaya, T. An inviscid model of two-dimensional vortex shedding for transient and asymptotically consists of six sequential `training blocks' (see Methods), in each of
steady separated ow over an inclined plane. J. Fluid Mech. 68, 109128 (1975).
which a different pair of rising and falling frequency-modulated
8. Dickinson, M. H., Lehmann, F. O. & Sane, S. P. Wing rotation and the aerodynamic basis of insect
ight. Science 284, 19541960 (1999). tones is presented (Fig. 1a, numbers). In the early blocks, discrimi-
9. Willmott, A. P., Ellington, C. P. & Thomas, A. L. R. Flow visualization and unsteady aerodynamics in nation performance gradually improves in each block as the animals
the ight of the hawkmoth Manduca sexta. Phil. Trans. R. Soc. Lond. B 352, 303316 (1997). learn to discriminate between the rising and falling tones over a
10. Kucheman, D. The Aerodynamic Design of Aircraft (Pergamon, Oxford, 1978).
11. Maxworthy, T. Experiments on the Weis-Fogh mechanism of lift generation by insects in hovering
series of daily training sessions (discrimination learning; Fig. 1b). A
ight Part 1. Dynamics of the `ing'. J. Fluid Mech. 93, 4763 (1979). sudden transition in behaviour then occurs, after which novel
12. Liu, H., Ellington, C. P., Kawachi, K., Van den Berg, C. & Willmott, A. A computational uid dynamic stimuli are immediately identied as belonging to their correct
study of hawkmoth hovering. J. Exp. Biol. 201, 461477 (1998). category: discrimination performance is already high in the rst
13. Dickinson, M. H. The effects of wing rotation on unsteady aerodynamic performance at low Reynolds
numbers. J. Exp. Biol. 192, 179206 (1994).
session of a training block, even though the stimulus pair to be
14. Wang, J. Two dimensional mechanism for insect hovering. Phys. Rev. Lett. 85, 22162219 (2000). discriminated is novel (categorization; Figs 1c, 2). The point during
15. Milne-Thomson, L. M. Theoretical Aerodynamics (Macmillan, New York, 1966). the training when this transition occurs differs between animals
16. May, R. M. in Diversity of Insect Faunas (eds Mound, L. A. & Waloff, N.) 188204 (Blackwell Scientic,
(Fig. 2), although in the 10 animals studied it was always after the
London, 1978).
17. Dudley, R. The Biomechanics of Insect Flight. Form, Function, Evolution (Princeton Univ. Press, second training block.
Princeton, 2000). In addition to the sudden emergence of concept transfer we nd a
18. Willmott, A. P. & Ellington, C. P. The mechanics of ight in the hawkmoth Manduca sexta. II. sudden alteration in the perceptual scaling characteristics of the
Aerodynamic consequences of kinematic and morphological variation. J. Exp. Biol. 200, 27232745
(1997).
animals: measuring psychometric functions for frequency-modula-
tion rate after each training block shows a generalization gradient
for modulation rate during the discrimination-learning phase; that
Acknowledgements
is, the greater the difference in modulation rates (rates of change of
We thank C. Ellington, M. Gahrib, G. Lauder, S. Sane and J. Wang for comments and
suggestions on this manuscript. This work was supported by the National Science
frequency) between the training stimuli used in the training block
Foundation, Ofce of Naval Research, and DARPA. and the test stimuli, the smaller the conditioned response rate
Correspondence and requests for materials should be addressed to M.H.D.
(e-mail: [email protected]).
a 12
rising falling
10 2
Frequency (kHz)

................................................................. 8

Change in pattern of ongoing cortical

6
5
4
activity with auditory category 2
1
3

learning
4
0 6
0 250 0 250
F. W. Ohl*, H. Scheich* & W. J. Freeman Time (ms) Time (ms)
b 100 c
* Leibniz-Institut fur Neurobiologie, Brenneckstrae 6, D-39118 Magdeburg,
Response rate (%)

Germany 80
Department of Molecular and Cell Biology, University of California,
60
129 LSA, Berkeley 94720-3200, USA
.............................................................................................................................................. 40
Hit rate
Humans are able to classify novel items correctly by category1,2; False-alarm rate
20
some other animals have also been shown to do this37. During
category learning, humans group perceptual stimuli by abstract- 0
ing qualities from similarity relationships of their physical 1 6 12 1 6
Session number
12
Session number
properties1,2,8. Forming categories is fundamental to cognition9 d e
and can be independent of a `memory store' of information
Conditioned response

100
about the items or a prototype10. The neurophysiological mechan- Generalization Categorization
80
isms underlying the formation of categories are unknown. Using gradient boundary
rate (%)

60
an animal model of category learning6, in which frequency-
40
modulated tones are distinguished into the categories of `rising'
and `falling' modulation, we demonstrate here that the sorting of 20

stimuli into these categories emerges as a sudden change in an 0

32 16 8 4 2 0 2 4 8 16 32 3216 8 4 2 0 2 4 8 16 32
animal's learning strategy. Electro-corticographical recording Modulation rate (kHz s1) Modulation rate (kHz s1)
from the auditory cortex11 shows that the transition is accompa-
nied by a change in the dynamics of cortical stimulus representa- Figure 1 Stimuli and behavioural measures of category learning. a, Rising (red) and falling
tion. We suggest that this dynamic change represents a (blue) frequency-modulated tones used in the six sequential training blocks (numbers).
mechanism underlying the recognition of the abstract quality b, Sample learning curve of gerbil 3 before transition to categorization. c, Sample learning
(or qualities) that denes the categories. curve of same animal after transition to categorization. d, Psychometric function for
Mongolian gerbils (Meriones unguiculatus) can be trained with modulation rate obtained after training block shown in b. Peak modulation rate of
pairs of linearly rising and falling frequency-modulated tones 8 kHz s-1 (arrow) corresponds to modulation from 2 kHz to 4 kHz in 250 ms used in this
(Fig. 1a) to establish the categories of `rising' and `falling' modula- block. e, Sigmoid psychometric function obtained after training block shown in c.

NATURE | VOL 412 | 16 AUGUST 2001 | www.nature.com 2001 Macmillan Magazines Ltd 733