Supercooling of Peltier cooler using a current

pulse
Cite as: Journal of Applied Physics 92, 1564 (2002); https://doi.org/10.1063/1.1489713
Submitted: 15 February 2002 • Accepted: 07 May 2002 • Published Online: 18 July 2002

G. Jeffrey Snyder, Jean-Pierre Fleurial, Thierry Caillat, et al.

ARTICLES YOU MAY BE INTERESTED IN

The transient behavior of Peltier junctions pulsed with supercooling

Journal of Applied Physics 112, 014514 (2012); https://doi.org/10.1063/1.4735469

Measurement of Thermal Conductivity by Utilization of the Peltier Effect

Journal of Applied Physics 30, 1351 (1959); https://doi.org/10.1063/1.1735334

Improved supercooling in transient thermoelectrics

Applied Physics Letters 85, 3247 (2004); https://doi.org/10.1063/1.1806276

Journal of Applied Physics 92, 1564 (2002); https://doi.org/10.1063/1.1489713 92, 1564

© 2002 American Institute of Physics.

JOURNAL OF APPLIED PHYSICS VOLUME 92, NUMBER 3 1 AUGUST 2002

Supercooling of Peltier cooler using a current pulse

G. Jeffrey Snyder,a) Jean-Pierre Fleurial, and Thierry Caillat
Jet Propulsion Laboratory, California Institute of Technology, Pasadena, California 91109
Ronggui Yang and Gang Chen
Department of Mechanical Engineering, Massachusetts Institute of Technology, Cambridge,
Massachusetts 02139
共Received 15 February 2002; accepted for publication 7 May 2002兲
The operation of a Peltier cooler can be temporarily enhanced by utilizing the transient response of
a current pulse. The performance of such a device, using (Bi,Sb) 2 Te3 -based thermoelectric
elements, was examined from ⫺70 to 55 °C. We establish both theoretically and experimentally the
essential parameters that describe the pulse cooling effect, such as the minimum temperature
achieved, maximum temperature overshoot, time to reach minimum temperature, time while cooled,
and time between pulses. Using simple theoretical and semiempirical relationships the dependence
of these parameters on the current pulse amplitude, temperature, thermoelectric element length,
thermoelectric figure of merit and thermal diffusivity is established. At large pulse amplitudes the
amount of pulse supercooling is proportional to the maximum steady-state difference in
temperature. This proportionality factor is about half that expected theoretically. This suggests that
the thermoelectric figure of merit is the key materials parameter for pulse cooling. For this cooler,
the practical optimum pulse amplitude was found to be about three times the optimum steady-state
current. A pulse cooler was integrated into a small commercial thermoelectric three-stage cooler and
it provided several degrees of additional cooling for a period long enough to operate a laser sensor.
The improvement due to pulse cooling is about the equivalent of two additional stages in a
multistage thermoelectric cooler. © 2002 American Institute of Physics.
关DOI: 10.1063/1.1489713兴

INTRODUCTION confirmation.6 –10 Such a cooler could be useful for a device

A Peltier cooler uses the thermoelectric effect to trans-
port heat with the application of an electric current. Under
steady-state conditions, the physics and engineering of such
cooling devices have been well studied1 and are in commer-
cial use. The cooling coefficient of performance and maxi-
mum temperature drop depends on the properties of the ther-
moelectric materials used through the figure of merit, Z
⫽ ␣ 2 / ␳ ␬ , where ␣ is the Seebeck coefficient, ␳ is the elec-
trical resistivity, and ␬ is the thermal conductivity. Briefly,
this is because the heat removed due to Peltier cooling, Q
⫽ ␣ TI, which is proportional to the applied current I, is
counteracted by Joule heating proportional to I 2 ␳ and the
reverse conductive heat flow proportional to ␬. Since Joule
heat is proportional to a higher power of I than the Peltier
cooling term, there will be a current, I max , which produces
the largest difference in temperature ⌬T max .
Peltier cooling occurs at the junction at the cold end of
the thermoelectric elements. Joule heating, however, occurs
uniformly throughout the thermoelectric elements. Thus
when current is applied, the cooling at the cold junction oc-
curs before the Joule heat reaches the cold end. In this way,
an applied current pulse I⬎I max can be used to temporarily
produce a difference in temperature greater than ⌬T max .
This transient behavior of a thermoelectric cooler has been
studied theoretically2–7 and there is some experimental
such as a mid-IR laser gas sensor, or any other semiconduc-
tor device11 that needs to be cold for a few milliseconds.
In this work, we attempt a thorough experimental inves-
tigation of a practical pulsed Peltier cooler to determine the
minimum set of essential parameters and their relationships.
A physical basis is used, when available, to explain these
relationships. Such relationships can then be used to design a
pulsed cooler.
In order to define terms, a summary of the theoretical
problem and results is given. The theoretical analysis of the
pulse cooling problem can be approximated into a one-
dimensional problem by assuming the n-type and p-type
thermoelectric elements have exactly the same properties ex-
cept for the opposite sign of the Seebeck coefficient. The
differential equation is
⳵ 2T I 2␳ ⳵T
2⫹ 2 ⫽ . 共1兲
⳵x A ␬ a⳵t
Here a is the thermal diffusivity, and A is the cross-sectional
area. There are two boundary conditions. At x⫽0 there is
only Peltier cooling
⳵T
⳵x
冏 x⫽0
⫽
␣ IT 共 x⫽0 兲
A␬
, 共2兲

and at x⫽l, where l is the length of the pulse thermoelectric

a兲
Electronic mail: jsnyder@jpl.nasa.gov elements, we have a hot side heat sink, T(x⫽l,t)⫽T h .

0021-8979/2002/92(3)/1564/6/$19.00 1564 © 2002 American Institute of Physics

J. Appl. Phys., Vol. 92, No. 3, 1 August 2002
FIG. 1. Three-stage pulse cooler made from a commercial three-stage cooler
with the top stage removed.
The solution to the steady-state problem for materials
with Z independent of the temperature1 gives expressions for
the current for maximum cooling I max⫽␣TcA/␳l in terms of
the cold side temperature T c , the maximum steady-state dif-
ference in temperature ⌬T max⫽Th⫺Tc⫽ZT2c /2, and the
corresponding steady-state temperature profile T ss(x)⫽T c
⫹⌬T max(2x/l⫺x2/l2).
EXPERIMENT
Hot pressed n-Bi2 Te2.85Se0.15 and p-Bi0.4Sb1.6Te3 were
used to fabricate 5.8 mm tall thermoelectric elements with 1
mm2 cross-sectional areas. The cold ends were Bi–Sn sol-
dered to a 35 ␮m thick copper foil to which was soldered a
1 mil 共25 ␮m兲 diam Chromel–Constantin thermocouple for
measurement of the temperature. The hot end in the first case
was soldered to an electrically isolated heat sink where the
heat sink temperature could be adjusted from 55 to ⫺30 °C.
In the second case, the cooler was soldered to a commercial
Marlow MI3021T-01AC three-stage cooler with the top stage
removed so that the cooling stages were connected in series
as is done in a commercial cooler 共Fig. 1兲.
I max , the steady-state current that achieves the largest
difference in temperature, was determined experimentally
from a plot 共at each hot side temperature兲 of cold side tem-
perature versus current, where the temperature was allowed
to stabilize for 10 min for each point. For the single stage
pulse cooler I max⫽0.675 A. The average ZT c of the cooler
running at ambient temperature 共conditions in Fig. 3兲 is 0.51.
The single stage pulse cooler was used for all measurements
with a hot side temperature greater than ⫺30 °C. The three-
stage pulse cooler had a somewhat larger of I max⫽0.75 A
due to the larger I max⫽1.20 A of the Marlow cooler.
The current was supplied and measured by a Keithley
2430 SourceMeter, which also measured the thermocouple
Snyder et al. 1565
FIG. 2. Example of pulse cooling data with definitions of the key variables.
temperature within 0.1 °C. The temperature data were re-
corded at speeds up to 3 Hz.
A square current pulse is used to achieve the supercool-
ing as demonstrated in Fig. 2. The temperature of the cold
junction versus the time curve that is produced has the fol-
lowing general features. Starting from the maximum steady-
state cooling (T⫽T ss ,I⫽I max) a current pulse ( P
⫽I pulse /I max) is applied at time⫽0 which supercools the
Peltier junction. The minimum temperature is achieved at
some time t min . This maximum difference between the su-
percooling temperature and the steady-state temperature T ss ,
which occurs at t min , is ⌬T pulse . When the temperature re-
turns to T ss at time t ret , the current is reduced to I max . The
temperature then rises rapidly to ⌬T postpulse above T ss . Fi-
nally, the temperature will eventually return to T ss , ready for
the next pulse.
A series of experiments were conducted in which these
parameters were taken from the temperature versus time
curve for various pulse amplitudes and hot side temperatures.
The fitting functions described below consistently fit well
with single stage data for all measurements. The triple stage
cooler shows the same trends but the curve fitting is not as
good and therefore not used in many figures.
RESULTS
The characteristic times and temperatures defined in Fig.
2 can be used to characterize the pulse cooler as a function of
independent variables such as the hot side temperature T hot ,
length of pulse thermoelectric elements l, pulse factor P, and
Z. This is useful not only to test theoretical models, but is
very helpful in designing such a cooler where a user would
1566 J. Appl. Phys., Vol. 92, No. 3, 1 August 2002
FIG. 3. Maximum pulse cooling difference in temperature as a function of
the pulse current. The analytical approximation of Babin and Iordanishvilli
共Ref. 2兲 predicts maximum cooling, ⌬T P⬁ of 29 K for this cooler. The linear
model is fit to the low P data only. Much of the data is proportional to
1⫺exp(1⫺P).
want to easily calculate the amount of increased cooling, the
time the cooling lasts, the current needed, and the time be-
tween pulses.
The magnitude of the current pulse P affects the lowest
temperature achieved characterized by ⌬T pulse and is typi-
cally the focus of studies on pulse cooling. The solution to
Eqs. 共1兲 and 共2兲 cannot be solved analytically for a current
pulse in a finite thermoelectric element. The analytical ap-
proximation of Babin and Iordanishvilli2 assumes the ther-
moelectric elements can be considered infinitely long. This
approach is most appropriate for large pulses ( P⬎2) where
the times involved are shorter than the characteristic thermal
diffusion time constant. The transcendental solution, using
the steady-state properties measured, is shown in Fig. 3 and
predicts supercooling at infinite pulse ⌬T p⬁ ⬇⌬T max/2.
A linear approximation5 to the heat equation 关Eq. 共1兲兴 is
perhaps more useful in explaining the experimental results:
⳵ 2 T I 2 ␳ T⫺T ss
⫹ ⬇ , 共3兲
⳵ x 2 A 2␬ at
where the x⫽0 boundary condition is also approximated to
⳵ T/ ⳵ x 兩 x⫽0 ⬇ ␣ IT c /A ␬ . The general solution is
T 共 x,t 兲 ⫺T ss at 冑at ⫺x/ 冑at
⫽ 共 P 2 ⫺1 兲 2 ⫺ 共 P⫺1 兲 e , 共4兲
zT 2c l l
which predicts an algebraic function for ⌬T pulse ,
⌬T pulse⫽⌬T P⬁ 冉 冊
P⫺1
P⫹1
, 共5兲
with ⌬T P⬁ ⫽⌬T max/2. This function fits well 共Fig. 3兲 to the
experimental data for small pulses with a ⌬T P⬁ slightly
smaller than that expected. Both theoretical forms give the
correct order of magnitude but both deviate significantly
from the experimental shape, particularly at large pulse am-
plitudes where both models predict a further increase in
⌬T pulse above P⫽3 which is not observed experimentally.
The experimental data fit well to the empirically determined
Snyder et al.
FIG. 4. Maximum pulse cooling compared to maximum steady-state cool-
ing. The pulse cooling is nearly proportional to the steady-state cooling.
共no theoretical basis兲 formula ⌬T pulse⫽⌬T P⬁ (1⫺e 1⫺ P )
with only one free parameter. The data at the highest pulse
amplitudes ( P⬎3) have higher standard deviations and de-
viate from the exponential trend. This may due to experimen-
tal uncertainty when measuring these fast pulses, or simply
due to the thermal diffusion time from the junctions to the
thermocouple.
The excellent fit to the empirical exponential form al-
lows us to characterize ⌬T pulse for all P with one parameter
⌬T P⬁ . It has already been suggested from the above models
that ⌬T P⬁ is related to ⌬T max . This is confirmed experimen-
tally in Fig. 4. Here ⌬T P⬁ mimics ⌬T max , decreasing as the
hot side temperature is lowered. As a consequence of the
approximate relationship, ⌬T P⬁ ⬇⌬T max/4⫽ZT 2c /8, it ap-
pears the only materials parameter that affects the minimum
temperature achieved is the thermoelectric figure of merit Z.
As the pulse amplitude increases, the minimum junction
temperature becomes lower, but for a shorter time. Figure 5
shows the decrease in t min and t ret as P increases for a single
hot side temperature. The data for t min fit very well to a ( P
⫹1) ⫺2 dependence. This can be explained by the linear ap-
proximation 关Eqs. 共3兲 and 共4兲兴 which predicts
t ret l2 ␶
t min⫽ ⫽ 2⫽ , 共6兲
4 4a 共 P⫹1 兲 共 P⫹1 兲 2
where ␶ ⫽l 2 /4a.
The data for t ret are described well by the linear model
when P is large or t ret is short. The crossover occurs at about
time t⫽ ␶ ⫽14 s. For times greater than ␶, a quasisteady-state
model described below can explain the experimental data.
The corresponding curves from Babin and Iordanishvilli’s
model2 are unphysical for P⬍3 and therefore not included in
Fig. 5.
The two characteristic times t min and t ret can be used to
estimate the time below a specified temperature in the super-
cooled state, providing the shape of the temperature pulse is
roughly constant. If using the reduced variables (T
⫺T ss)/⌬T pulse and t/t ret , the result of the linear model pre-
dicts a constant shape of temperature pulse,
J. Appl. Phys., Vol. 92, No. 3, 1 August 2002
FIG. 5. Pulse dependence of characteristic times t min and t ret . The experi-
mental results for t min 共circles兲 is shown with the fit to Eq. 共6兲. The experi-
mental results for t ret 共squares兲 is shown with the prediction t ret⫽4t min from
Eq. 共6兲. The fit with t ret⫽ ␶ ln关1⫹C/4( P⫺1) 兴 using a quasisteady-state
model is substantially better for small pulse amplitudes.
T⫺T ss
⌬T pulse
⫽4
t
冉 冑冊
t ret
⫺
t
t ret
. 共7兲
The pulse shapes predicted by the analytical model of
Babin and Iordanishvilli2 are very similar to those of the
linear model, and become nearly identical for small P. The
experimental data in Fig. 6 clearly show the transition be-
tween two regimes. For large P which take short times, the
curves are similar in shape to that predicted by the linear
model. For t⬍t min the cooling is progressively slower than in
Eq. 共7兲 as P increases, which may be due to the time for the
cooling to reach the thermocouple. It is expected that in an
actual application P⬎2 will be used, in which case the pulse
cooling shape is nearly constant and given, to reasonable
approximation, by Eq. 共7兲. For longer times 共smaller P兲
greater than the thermal diffusion time 共14 s兲 there is an
exponential decay component in the curves measured. This
provides strong deviation in the pulse shape, most noticeable
in the P⫽1.05 curve.
FIG. 6. Shape of the pulse cooling curve using reduced variables.
Snyder et al. 1567
FIG. 7. Junction heating ⌬T Postpulse after a supercooling pulse. The line
drawn is fit to ⫺6 °C data.
After supercooling, the junction rapidly heats up and
reaches ⌬T postpulse above T ss . For temperature sensitive elec-
tronics, this may be a factor in cooler design. Figure 7 shows
the maximum rise in temperature after pulse operation, de-
fined as ⌬T postpulse in Fig. 2. The magnitude of ⌬T postpulse
will depend on the amount of excess Joule heating added to
the system during pulse operation. One might expect then
that ⌬T postpulse will be proportional to the excess heating rate
( P 2 ⫺1) and a heating time, an obvious choice being t ret .
Much of the ⌬T postpulse data is extremely linear, proportional
to ( P⫺1) as seen in Fig. 7, which is off by a factor of ( P
⫹1) if using the t ret of Eq. 共6兲. Obviously, if the pulse lasts
longer 共shorter兲 than t ret , ⌬T postpulse will be larger 共smaller兲.
In practice, the optimum pulse time will be between t min and
t ret .
The time required to return to the steady-state condition
can be characterized by an exponential decay time constant,
which is derived as follows. If we define ⌬T⫽T⫺T ss then
for I⫽I max , the heat equation 关Eq. 共1兲兴 reduces to
⳵ 2 ⌬T ⳵ ⌬T
⫽ . 共8兲
⳵x2 a⳵t
Thus, after the application of pulse current, the perturba-
tion in temperature will return to the steady-state temperature
profile according to Eq. 共8兲. The series solution12 to this
equation has as the fundamental exponential decay
exp(␲2 at/4l 2 ) with time constant ␶ ⫽4l 2 / ␲ 2 a. The higher
order terms have shorter time constants and smaller ampli-
tudes. The data at long times fit well to this form 共Fig. 8兲. For
the single stage cooler the measured time constant for all hot
side temperatures is 13.7⫾0.6 s which implies a thermal dif-
fusivity of 0.010 cm2/s. Independent measurements of ther-
mal diffusivity for these Bi2 Te3 materials gives 0.007 cm2/s
at room temperature.
We can apply this thermal decay result to develop a
quasisteady-state model to explain the t⬎ ␶ regime. If we
assume that a change in current will produce a evolution
from the initial state to the final state with the primary func-
tional form being exp(⫺␲2 at/4l 2 ), then we can derive a
simple equation for t ret . This should be particularly valid for
times greater than the thermal time constant of the thermo-
1568 J. Appl. Phys., Vol. 92, No. 3, 1 August 2002
FIG. 8. Postpulse thermal decay to steady state. At long times after pulsing
the temperature returns to steady state by means of an exponential.
electric element 共14 s in that case兲. The initial state relative
to T ss , ⫺⌬T i , should be proportional to the supercooled
junction temperature approximated by Eq. 共5兲, or by fitting to
experimental data. In either case, for P⬇1 this reduces to
⌬T i ⬇C( P⫺1)⌬T max/4, where C is the proportionality con-
stant. The final state junction temperature is the steady-state
value for that current which can be found experimentally
from steady-state temperature versus current curves which
can be fit to ⌬T f ⬇C 2 ⌬T max(P⫺1)2 where C 2 is a fitting
constant approximately equal to 1. t ret is then the time is
takes the exponential ⌬T f ⫺(⌬T f ⫹⌬T f )exp(⫺t/␶) to decay
to zero, which gives t ret⫽ ␶ ln关1⫹C/4( P⫺1) 兴 . The fit to this
equation 共Fig. 5兲 is remarkably good for C⫽2, even for large
P. This suggests that this analysis is valid for larger P with
⌬T i ⫽C⌬T pulse 关from Eq. 共5兲兴 where C⫽( P⫹1).
DISCUSSION
The initial experiments of thermoelectric pulse coolers
have varying results. Early measurements that use the ther-
moelectric elements themselves as temperature probes7–9 in-
dicate more supercooling than we observed. Results in which
a separate thermocouple6,10 is used to measure the tempera-
ture are consistent with our results, but with less detail.
Despite the many dependent and independent variables,
the characteristic temperatures and times of a pulse cooler
can be modeled well by a few simple empirical relationships.
In many cases these relationships have a theoretical basis.
Such models can be simply used to judge the appropriateness
of a pulse cooler for a specified application.
The two theoretical models discussed above predict a
maximum pulse cooling ⌬T P⬁ ⫽⌬T max/2, about twice that
observed here. Both models neglect the interconnect and any
time required for junction cooling to reach the thermocouple.
To test the effect of the interconnect, a similar device was
fabricated with no copper interconnect by soldering n- and
p-type thermoelectric elements together directly with a ther-
mocouple in the solder junction. This configuration was also
used in Refs. 6 – 8. The results were essentially the same
Snyder et al.
共also found in Ref. 6兲 with ⌬T pulse 90% of that observed in
Fig. 3.
Two regimes are clearly distinguishable in the experi-
mental data. For times shorter than the thermal relaxation
time the data follow simple transient models. For longer
times thermal decay exponential forms are more appropriate.
Numerical simulations can be used to model both regimes
and can qualitatively reproduce the data reported.13
The series connection of the pulse cooler with other
stages to form a multistage pulse cooler worked well. The
performance was consistent with the results obtained for the
single stage pulse cooler despite that fact that the cooling
capability of the small cooler is much less than that of the
heat sink used for the single stage pulse cooler. For optimum
performance, the A/l ratio of the pulse cooling stage should
be optimized so that I max is the same for all stages. The heat
removal capacity of the pulse cooler is expected to be similar
to that of a standard thermoelectric cooler with similar cross-
sectional areas.
The original three stage cooler has ⌬T max⫽98.9 K 共with
an additional 1 K from pulsing兲 from a hot side temperature
of 17 °C where I max⫽1.2 A. At the same hot side tempera-
ture, the pulse cooler ⌬T max⫽92.9 K was less, primarily due
to the mismatch in I max mentioned above. At optimal pulse a
difference in cooling of ⌬T max⫹⌬Tpulse⫽102.0 K was ob-
served. For a pulsed cooler with matched I max a difference in
temperature of 108 K would be expected. This is a 9% im-
provement compared to that of the unpulsed cooler, and
about the equivalent of a five-stage cooler.
CONCLUSIONS
The characteristics of a thermoelectric pulse cooler were
analyzed by determining and modeling the characteristic
times and temperatures during the pulsing cycle for various
pulse amplitudes and steady-state temperatures. The amount
of supercooling ⌬T pulse was found to be a simple function of
the pulse amplitude steady-state temperature. At large pulse
amplitudes the amount of pulse supercooling is about 1/4
that of steady-state cooling for a wide range of temperatures.
This suggests that the thermoelectric figure of merit, Z, is the
key materials parameter for pulse cooling. The practical op-
timum pulse amplitude is found to be about three times the
optimum steady-state current.
The time to reach minimum temperature and the time
below a specific temperature can be predicted with a simple
equation, derived from simplifications to the heat equation.
Both times decrease as the current pulse is increased. After
pulse cooling, the sample will overheat by an amount that
can be predicted by a simple empirical relationship. The re-
turn to steady-state temperature follows simple exponential
decay. All the characteristic times are proportional to the
square of the thermoelectric element length, thus providing a
simple framework in which to optimize a pulse cooler’s ge-
ometry for a specific application.
A pulse cooler was integrated into a small commercial
thermoelectric three-stage cooler and it provided several de-
grees of additional cooling for a period long enough to op-
erate a laser sensor. The improvement due to pulse cooling is
J. Appl. Phys., Vol. 92, No. 3, 1 August 2002 Snyder et al. 1569

1
about the equivalent of two additional stages in a multistage R. R. Heikes and R. W. Ure, Thermoelectricity: Science and Engineering
cooler. 共Interscience, New York, 1961兲.
2
V. P. Babin and E. K. Iordanishvilli, Sov. Phys. Tech. Phys. 14, 293
共1969兲.
ACKNOWLEDGMENTS 3
J. E. Parrott, Solid-State Electron. 1, 135 共1960兲.
This work was carried out by the Jet Propulsion Labora- 4
G. E. Hoyos, K. R. Rao, and D. Jerger, Energy Convers. 17, 23 共1977兲.
5
tory, California Institute of Technology, under a contract with A. Miner, A. Majumdar, and U. Ghoshal, Appl. Phys. Lett. 75, 1176
共1999兲.
the National Aeronautics and Space Administration 共NASA兲. 6
L. S. Stil’bans and N. A. Fedorovich, Sov. Phys. Tech. Phys. 3, 460
Funding support came from the DARPA HERETIC program. 共1958兲.
The authors would like to thank Hylan Lyon of Marlow In- 7
K. Landecker and A. W. Findlay, Solid-State Electron. 3, 239 共1961兲.
dustries for providing the MI3021T-01AC cooler. Compari- 8
M. Idnurm and K. Landecker, J. Appl. Phys. 34, 1806 共1963兲.
sons with numerical simulations were done by Ronggui 9
G. E. Hoyos, K. R. Rao, and D. Jerger, Energy Convers. 17, 45 共1977兲.
Yang, A. Ravi Kumar and Gang Chen of the University of
10
R. L. Field and H. A. Blum, Energy Convers. 19, 159 共1979兲.
11
T. Yamamoto, Proc. IEEE 56, 230 共1968兲.
California at Los Angeles/Massachusetts Institute of Tech- 12
E. C. Zachmanoglou and D. W. Thoe, Introduction to Partial Differential
nology. One of the authors 共G.J.S.兲 would also like to thank Equations with Applications 共Dover, New York, 1986兲, p. 340.
Satwindar Sadhal of the University of Southern California 13
R. Yang, G. Chen, G. J. Snyder, and J.-P. Fleurial, Mater. Res. Soc. Symp.
for discussions on thermal transport theory. Proc. 共in press兲.