Uncertainty analysis of thermoreflectance measurements

Jia Yang, Elbara Ziade, and Aaron J. Schmidt

Citation: Review of Scientific Instruments 87, 014901 (2016); doi: 10.1063/1.4939671

View online: http://dx.doi.org/10.1063/1.4939671
View Table of Contents: http://aip.scitation.org/toc/rsi/87/1
Published by the American Institute of Physics

Articles you may be interested in

Analysis of heat flow in layered structures for time-domain thermoreflectance
Review of Scientific Instruments 75, (2004); 10.1063/1.1819431

Modeling optical absorption for thermoreflectance measurements

Review of Scientific Instruments 119, 095107095107 (2016); 10.1063/1.4943176

Thermal conductivity measurements of non-metals via combined time- and frequency-domain

thermoreflectance without a metal film transducer
Review of Scientific Instruments 87, 094902094902 (2016); 10.1063/1.4962711

Understanding and eliminating artifact signals from diffusely scattered pump beam in measurements of rough
samples by time-domain thermoreflectance (TDTR)
Review of Scientific Instruments 87, 064901064901 (2016); 10.1063/1.4952579

Thermal property microscopy with frequency domain thermoreflectance

Review of Scientific Instruments 84, 104904104904 (2013); 10.1063/1.4824143

Simultaneous measurement of thermal conductivity and heat capacity of bulk and thin film materials using
frequency-dependent transient thermoreflectance method
Review of Scientific Instruments 84, 034902034902 (2013); 10.1063/1.4797479
 REVIEW OF SCIENTIFIC INSTRUMENTS 87, 014901 (2016)

Uncertainty analysis of thermoreflectance measurements

Jia Yang, Elbara Ziade, and Aaron J. Schmidt
Department of Mechanical Engineering, Boston University, Boston, Massachusetts 02215, USA
(Received 18 August 2015; accepted 26 December 2015; published online 12 January 2016)
We derive a generally applicable formula to calculate the precision of multi-parameter measurements
that apply least squares algorithms. This formula, which accounts for experimental noise and uncer-
tainty in the controlled model parameters, is then used to analyze the uncertainty of thermal property
measurements with pump-probe thermoreflectance techniques. We compare the uncertainty of time
domain thermoreflectance and frequency domain thermoreflectance (FDTR) when measuring bulk
materials and thin films, considering simultaneous measurements of various combinations of thermal
properties, including thermal conductivity, heat capacity, and thermal boundary conductance. We vali-
date the uncertainty analysis using Monte Carlo simulations on data from FDTR measurements of an
80 nm gold film on fused silica. C 2016 AIP Publishing LLC. [http://dx.doi.org/10.1063/1.4939671]

I. INTRODUCTION II. BACKGROUND ON TDTR AND FDTR

Optical pump-probe techniques based on thermore- We review the TDTR and FDTR techniques here to
flectance, such as time domain thermoreflectance (TDTR)1–3 provide context for the uncertainty analysis. Both techniques
and frequency domain thermoreflectance (FDTR),4 have been use the same measurement geometry, shown in Figure 1(a).
established as powerful and flexible tools for studying the A pump laser beam is focused on the sample surface to
thermal properties of materials.5 These properties include create a Gaussian-shaped heat source, while a probe laser
the thermal conductivity and heat capacity of solids,1,4,6,7 beam is focused at the same spot and measures the change
liquids,8,9 thin films,10–12 and the thermal conductance of in reflectivity. For small temperature changes, the change in
interfaces.13–16 In addition, TDTR and FDTR have been reflectivity is proportional to the change in surface temperature
developed into microscopy techniques capable of map- via the coefficient of thermoreflectance.19 The sample is
ping spatial variations in thermal properties with sub-micron usually coated with a thin transducer layer such as a metal
resolution.7,17,18 that strongly absorbs light at the pump wavelength and has a
Analysis of TDTR and FDTR experiments requires a large coefficient of thermoreflectance at the probe wavelength.
comparison of the experimental data and a model of heat We typically use gold films with our FDTR system, which
transfer in the system under study. The properties being has a 532 nm probe laser, and aluminum films with our
measured are treated as free parameters in the model and are TDTR system, which has an 800 nm probe laser. More details
adjusted by an optimization technique such as least squares on our implementations of these techniques can be found
minimization to match the model to the data. The confidence elsewhere.18,21
intervals associated with measured values are a critical aspect In FDTR, the pump beam is modulated at multiple
of any measurement technique, particularly techniques being frequencies, and the phase lag between the pump beam and
used to validate physical models and obtain values that are probe beam is measured with a lock-in amplifier. In TDTR, the
inputs in device simulations. Despite the widespread use of pump and probe beams originate from a pulsed laser source
TDTR and FDTR for these purposes, however, a thorough and are separated by a variable optical delay. The pump pulses
analysis of their uncertainty has not been done. heat the sample and the probe pulses monitor the change in
In this paper, we derive a generally applicable formula to surface reflectivity as a function of the delay time. The train
calculate the precision of multi-parameter measurements that of pulses that comprise the pump beam is given an additional
use least squares algorithms. This formula, which accounts periodic modulation at one or more frequencies to enable
for both experimental noise and uncertainty in the controlled lock-in detection. Either the amplitude, the phase, or the ratio
model parameters, is then used to calculate and compare the of the in-phase to out-of-phase components of the lock-in
uncertainty of TDTR and FDTR. We consider the cases of a amplifier signal can be used as the observable parameter.
bulk material and a thin film on a substrate and examine fitting In this paper, we use the phase signal for consistency with
various combinations of two, three, and four unknown thermal FDTR. In both TDTR and FDTR, the unknown properties
properties, including the thermal conductivity, heat capacity, are treated as free parameters in a heat transfer model. These
and multiple thermal boundary conductances. We validate the are adjusted to minimize the difference between the predicted
uncertainty analysis using Monte Carlo simulations on data and measured phase at each frequency or delay time; this
from FDTR measurements of an 80 nm gold film on fused can be accomplished using a technique such as least squares
silica. minimization.

0034-6748/2016/87(1)/014901/11/$30.00 87, 014901-1 © 2016 AIP Publishing LLC

014901-2 Yang, Ziade, and Schmidt Rev. Sci. Instrum. 87, 014901 (2016)

the layer thickness, d, and the thermal boundary conductance

(TBC) to the next layer, G. A diffusion model is only valid
when the energy carriers are in local thermal equilibrium, an
assumption which can break down on short length and time
scales due to factors such as ballistic phonon transport in
crystals and weak electron-phonon coupling in the transducer
layer.22–26 In these situations, a different model, such as the
Boltzmann transport equation, is needed to accurately describe
nonequilibrium heat flow.
The uncertainty of the measured thermal properties,
which we take in this paper as the ±1/e confidence intervals
around the best fit mean values, is affected by three factors:
experimental noise in the phase data; uncertainty in the
controlled parameters in the model, such as the focused laser
spot radii, the lock-in amplifier’s reference phase, and all
other thermal properties of the sample; and the particular
combination of unknowns being determined. In Sec. III, we
derive an analytical formula for the uncertainty that accounts
for all of these factors. The best fit values may also be biased
by systematic errors from factors such as experimental error
or incorrect assumptions about the physics in the heat transfer
model. We do not consider the effect of any systematic errors.

III. UNCERTAINTY ANALYSIS

The effect of experimental noise on the uncertainty of
multi-parameter estimation using the least squares method
has been studied in the framework of regression analysis.27,28
A Taylor expansion approach leads to a matrix relationship
between the uncertainty in the experimental data and the
uncertainty in the unknown parameters. In this section, we first
briefly review this approach in the context of TDTR and FDTR
measurements, and then extend it to account for uncertainty
in the controlled parameters in the heat transfer model. This
method can be applied to any measurement where a known
model is fit to one or more observable parameters using a least
squares algorithm.
FIG. 1. (a) Sample configuration for FDTR and TDTR. (b) Typical phase In the following analysis, we use FDTR as an example,
data measured by FDTR and the best model fit for an 80 nm gold film on a
fused silica substrate. (c) Typical TDTR phase data and the best model fit for where phase data points are measured at multiple modulation
a 70 nm aluminum film on a fused silica substrate. frequencies. For a TDTR measurement, the procedure is
the same, except that the phase data points are measured
at multiple delay times. Suppose that there are p unknown
FDTR experiments typically access the frequency parameters, x u , u = 1, 2, . . . , p, being estimated by fitting a
response of the sample in the kHz to MHz range, while TDTR model to measured phase data. The model calculates the
experiments can probe frequencies up to the GHz range.20 As phase lag of the reflected probe signal, φ, as a nonlinear
we discuss later, both methods depend on accurate knowledge function of the unknown parameters and the pump modulation
of the heat capacity per unit area of the transducer, while frequency, ω,
FDTR is more sensitive to the laser spot radii and the in-plane
thermal conductivity of the transducer. Sample data from a φ = f (ω, XU ), (1)
typical FDTR measurement of fused silica coated with 80 nm
where XU is a p × 1 vector of unknown parameters x 1, x 2, . . . ,
of gold are shown in Fig. 1(b), along with the best fit of the
x p . The least squares estimate of the unknown parameter
heat transfer model, while sample phase data and the best
model fit from a TDTR measurement of fused silica coated vector XU , denoted by X U , minimizes the sum of squared
with 70 nm of aluminum are shown in Fig. 1(c). errors R
In this paper, we use a two-dimensional diffusion model M

for anisotropic heat conduction in layered media.3,4 As shown R= [φi − f (ωi , XU )]2, (2)
in Fig. 1(a), each layer of the sample is modeled with five i=1

physical parameters: the volumetric heat capacity, C, the where φi , i = 1, 2, . . . , M, are the M measured phase data
cross-plane and in-plane thermal conductivities, κ ⊥ and κ ∥ , points at frequencies ωi .
014901-3 Yang, Ziade, and Schmidt Rev. Sci. Instrum. 87, 014901 (2016)

When each f (ωi , XU ) is differentiable with respect to XU , The least squares estimate of XU , still denoted as X
U ,
U will satisfy
X now satisfies

∂ R(XU ) ∂ R(XU , XC )
|  = 0, u = 1, 2, . . . , p. | XU = 0,u = 1, 2, . . . , p. (8)
∂ x u XU
(3) ∂ xu
From Eq. (7) and Eq. (8), we have
Taking a first-order Taylor expansion of f (ωi , XU ) around the
U , which we denote as X ⋆ , leads to the result27 M
mean of X  ∂ f (ωi , XU , XC )
U ([φi − f (ωi , XU , XC )] )| XU
i=1
∂ xu
U = (JU⋆ ′ JU⋆)−1 JU⋆ ′(Φ − F(XU⋆ )) + XU⋆ ,
X (4)
= 0,u = 1, 2, . . . , p. (9)
where Φ is an M × 1 vector of measured phase data φi , F(XU⋆ ) We denote the mean values of X U and XC as XU⋆
and XC⋆,
is an M × 1 vector of the calculated phases f (ωi , XU⋆ ), and JU⋆ respectively. The corresponding elements are denoted as x ⋆u ,
is the Jacobian matrix u = 1, 2, . . . , p, and x ⋆c , c = p + 1, p + 2, . . . , N, respectively.
∂ f (ω1, XU ) ∂ f (ω1, XU ) When XU and XC are in a small neighborhood around XU⋆
*. ∂ x1
|X⋆
U
···
∂xp
|X⋆
U +/ and XC⋆, f (ωi , XU , XC ) can be approximated by a first-order
.. .. .. Taylor expansion around both XU⋆ and XC⋆
JU⋆ = ...
. //
. . . // . (5)
..
∂ f (ω M , XU ) ∂ f (ω M , XU )
// f (ωi , XU , XC ) ≈ f (ωi , XU⋆ , XC⋆)
|X ⋆ ··· |X⋆
∂ x1 U ∂xp U- p
∂ f (ωi , XU , XC )
+ | X ⋆ , X ⋆ (x u − x ⋆u )
,
∂ xu U C
The variances of the unknown model parameters in X U are u=1
then given by the diagonal elements of the variance-covariance N
∂ f (ωi , XU , XC )
matrix of Eq. (4). + | X ⋆ , X ⋆ (x c − x ⋆c ),
∂ x c U C
The procedure described above relates the uncertainty in c=p+1
the least squares estimate of multiple unknown parameters, i = 1, 2, . . . , M. (10)
U , to the uncertainty of the measured data, Φ. We now
X
extend this approach to also account for uncertainty in the Substituting f (ωi , XU , XC ) from Eq. (10) into Eq. (9), we have
controlled model parameters. Suppose that there are a total M

of N parameters in the model. We split the N parameters ([φi − f (ωi , XU⋆ , XC⋆)
into two groups: p unknown parameters x u , where u = 1, i=1
2, . . . , p, and N − p controlled parameters x c , where c p
∂ f (ωi , XU , XC )
= p + 1, p + 2, . . . , N. We use column vectors XU , with − | X ⋆ , X ⋆ ( x̂ u − x ⋆u )
∂ x u U C
elements x 1, x 2, . . . , x p , and XC , with elements x p+1, x p+2, . . . , u=1
N
x N , to denote the unknown parameters and controlled  ∂ f (ωi , XU , XC )
parameters, respectively. The calculated phase lag φ can then − | X ⋆ , X ⋆ (x c − x ⋆c )]
c=p+1
∂ x c U C
be written as
∂ f (ωi , XU , XC )
φ = f (ω, XU , XC ) (6) × | X ⋆ , X ⋆ ) = 0, (11)
∂ xu U C

and the sum of squared errors is given by for u = 1, 2, . . . , p, where x̂ u are the elements of the least
squares estimate vector X U . We use matrix operations to
M
 simplify Eq. (11). For i = 1, 2, . . . , M, we denote Φ as a
R= [φi − f (ωi , XU , XC )]2, (7) vector with elements φi , and F(XU⋆ , XC⋆) as a vector with
elements f (ωi , XU⋆ , XC⋆). We use Jacobian matrices to represent
i=1

where φi , i = 1, 2, . . . , M are still the M measured phase data the first order derivatives of f (ωi , XU , XC ) in Eq. (11), so
points at frequencies ωi . that

∂ f (ω1, XU , XC ) ∂ f (ω1, XU , XC )
|X⋆ , X⋆ ··· |X⋆ , X⋆
*. ∂ x1 U C ∂xp U C +/
.. .. ..
JU⋆
. //
= ... . . . //
∂ f (ω M , XU , XC ) ∂ f (ω M , XU , XC )
.. //
|X⋆ , X⋆ ··· |X⋆ , X⋆
, ∂ x1 U C ∂xp U C-
(12)
∂ f (ω1, XU , XC ) ∂ f (ω1, XU , XC )
|X⋆ , X⋆ ··· |X⋆ , X⋆ +
*. ∂ x p+1 U C ∂xN U C /
.. .. ..
and JC⋆
.. //
= .. . . . // .
∂ f (ω M , XU , XC ) ∂ f (ω M , XU , XC )
.. //
|X⋆ , X⋆ ··· |X⋆ , X⋆
, ∂ x p+1 U C ∂xN U C-
014901-4 Yang, Ziade, and Schmidt Rev. Sci. Instrum. 87, 014901 (2016)

Rewriting Eq. (11) in terms of Φ, F(XU⋆ , XC⋆) and Eq. (12), where cov[x u i , x u j ] is the covariance of x u i and x u j . If we
we then have let Pij be element (i, j) of Var[ X U ], the standard deviation
√
U − XU⋆ ) − J ⋆(XC − X ⋆)) = 0.
JU⋆ (Φ − F(XU⋆ , XC⋆) − JU⋆( X
′ of the ith element of XU will be Pii, which we take to

C C
be the uncertainty of that parameter. For the case of a single
(13)
unknown fitting parameter and no experimental noise, Eq. (15)
When JU⋆ ′ JU⋆
is non-singular, Eq. (13) can be rearranged to is equivalent to the uncertainty estimate for TDTR given in
express the least squares estimate X
U in terms of the measured Ref. 7.
phase data Φ, the controlled parameter vector XC , and the
Jacobian matrices,
U = (JU⋆ ′ JU⋆)−1 JU⋆ ′(Φ − F(XU⋆ , X ⋆) − J ⋆(XC − X ⋆)) + XU⋆ . IV. UNCERTAINTY OF TDTR AND FDTR
X C C C
(14) A. Fitting multiple parameters

Compared to Eq. (4), Eq. (14) additionally accounts for Typically, multiple unknown parameters are determined
uncertainty in the controlled model parameters. Assuming that simultaneously from a FDTR or TDTR measurement. The
the experimental noise for each data point in Φ is normally correlation between these unknowns in the thermal model
distributed with zero mean, and that the elements of XC are will influence whether a least squares algorithm will be able
normally distributed around their mean values, the elements to find a unique global minimum in the parameter space
of X
U will be asymptotically normally distributed around their and will strongly affect the confidence intervals calculated
mean values.29 by Eq. (15). We can quantify this effect using the correlation
We calculate the variance-covariance matrix of X U from coefficient, which describes the strength of linear dependence
Eq. (14) to obtain the variances of the individual elements of between parameters.30 For parameters x i and x j , each with
U , which we then use to determine the standard deviations
X a finite variance of σi2 and σ 2j , respectively, the correlation
for each of the unknown parameters. Given that Φ and XC are coefficient is equal to cov[x i , x j ]/(σi σ j ),30 where cov[x i ,x j ]
independent vectors, the variance-covariance matrix of X U is is the covariance of x i and x j . Then for a least squares estimate
given by U , the correlation coefficient between the ith and jth element,
X
U ] = (JU⋆ JU⋆)−1 JU⋆ (Var[Φ]
′ ′ r ij, is given by
Var[ X
Pij
+ JC⋆Var[XC ]JC⋆ )JU⋆(JU⋆ JU⋆)−1.
′ ′
(15) r ij =  , (18)
Pii Pjj
If we assume that the elements of Φ and XC are all independent
variables, we can write Var[Φ] and Var[XC ] as where P is (JU⋆ ′ JU⋆)−1σ 2, a simplified form of Eq. (15), where
we assume constant variance σ 2 in phase data and zero
σφ2 1 0 0 ···
variance in controlled parameters. We replace Var[Φ] with
σφ2 2
*. +
0 0 · · · // a constant variance, because for the condition of non-constant
Var[Φ] = ..
.
σφ2 3
/
.. 0 0 · · · // variance, the frequencies or delay times with more noise
.. .. .. ..
/ will contribute more to the relationship between parameters,
, . . . .- resulting in a biased correlation coefficient. For a pair of
and parameters, a correlation coefficient with an absolute value
of unity indicates that those two parameters have the same
σc21 0 0 ···
σc22
*. +
0 0 · · · //
Var[XC ] = ..
.
0 0 σc23
/
· · · // (16)
..
.. .. .. ..
/
, . . . .-
where σφ2 k , k = 1, 2, 3..., are the variances of the measured
phase data points φk at modulation frequencies ωk , and
σc2l , l = 1, 2, 3..., are the variances of the l controlled parame-
ters. Thus, using the experimentally determined variances in
the phase data and the variances for the controlled parameters,
we can determine the variances of the unknown parameters
from the diagonal elements of Eq. (15),
U ]
Var[ X
σ 2x u cov[x u1, x u2] cov[x u1, x u3] ···
*. 1 +
.. cov[x u2, x u1] σ 2x u cov[x u2, x u3] · · · //
= .. cov[x , x ] cov[x ,2x ] σ 2x u · · · /// ,
/
u3 u1 u3 u2
3
.. .. .. .. /
..
.- FIG. 2. Calculated phase sensitivities to thermal conductivity, κ, volumetric
. . .
, heat capacity, C, and thermal boundary conductance, G, for a FDTR mea-
(17) surement of fused silica coated with 80 nm of gold.
014901-5 Yang, Ziade, and Schmidt Rev. Sci. Instrum. 87, 014901 (2016)

TABLE I. Material properties for sensitivity calculation.

C κ d G
Material (106 J m−3 K−1) (W m−1 K−1) (nm) (MW m−2 K−1)

Au 2.48 ± 0.07 200 ± 10 80 ± 1.5 40

SiO2 1.63 1.38 5 × 105

effect on the observable phase variable and therefore cannot

be separately determined with least squares fitting.
Sensitivity analysis is a useful tool to visualize the
correlation between model parameters. Here, we take the
phase sensitivity to a parameter x to be ∂φ/∂ ln x, where φ
is the phase calculated from the thermal model. For example,
in Fig. 2, we consider a FDTR measurement of fused silica
coated with 80 nm of gold, assuming pump and probe 1/e2
beam spot radii of 2.8 µm and 2.3 µm, respectively. The
sensitivity is plotted for κ of the fused silica, C of the
fused silica, and the TBC between gold and fused silica, G.
The material properties used for the calculation are listed in
Table I. At modulation frequencies greater than ∼1 MHz,
the sensitivity to κ and C are closely correlated, because at
high frequencies heat transfer approaches the one-dimensional
(1D) limit and the surface temperature is determined by FIG. 3. (a) Measurement configuration for a bulk substrate. (b) Measured
√ phase noise as a function of frequency for a typical FDTR measurement with
the thermal effusivity, Cκ. At lower frequencies, the three lock-in amplifier bandwidth of 1.1 Hz.
sensitivities separate and all three properties can be uniquely
determined with high precision from one measurement that
spans the full frequency range shown in Fig. 2. We note that the B. Comparison of FDTR and TDTR for bulk materials
elements of the Jacobian matrices in Eq. (12) are equivalent to
We now use Eq. (15) to calculate the uncertainty of TDTR
the phase sensitivities divided by the parameter values, so that
and FDTR measurements of a bulk material with variable
the measurement uncertainty given by Eq. (15) incorporates
thermal conductivity. The sample configuration is shown in
the effect of correlation between the unknown parameters.
Fig. 3(a). We assume an 80 nm gold transducer, 1/e2 spot radii
In Table II, we show the correlation coefficients and
of 6.9 ± 0.1 µm for the pump and probe beams, respectively,
measurement uncertainties of the three parameters in Fig.
and a TBC value for the interface of 100 MW m−2 K−1.
2, calculated with two different frequency ranges using Eq.
The values of the remaining parameters and their standard
(15). Here we define one normalized standard deviation
deviations are listed in Table III. We incorporate experimental
as the uncertainty in a parameter. In our calculation, we
noise in the calculation by using the noise profile shown
used forty logarithmically spaced frequency points and have
in Fig. 3(b) for Var(Φ). This plot was made by computing
assumed a constant phase noise of 0.3◦ for all frequencies,
the standard deviations of 200 phase measurements at each
the uncertainties in the properties of the gold film shown in
frequency for an actual FDTR measurement of the sample
Table I, and 0.05 µm uncertainty in both pump and probe spot
shown in Fig. 3(a), using a lock-in amplifier bandwidth
radii. For a frequency range of 10 kHz–30 MHz, the absolute
of 1.1 Hz. For the TDTR calculation, we use the same
values of the correlation coefficients are all well below 1 and
number of data points and assumed a constant phase noise
are associated with low uncertainty. On the other hand, for a
of 0.1◦ over a time delay range of 100 ps–6 ns, although
frequency range of 1 MHz–30 MHz, C and κ of the fused silica
we note that the noise in TDTR data generally increases
are strongly correlated because heat transfer is mainly in the
at low frequencies because of increased 1/ f noise and also
1D diffusion regime, leading to significantly more uncertainty
increases at longer delay times due to a lower signal-to-noise
in these two parameters.
ratio.21

TABLE II. Correlation coefficients and calculated uncertainties for a three-

TABLE III. Material properties used for uncertainty calculations.
parameter fit.
C κ d
Correlation coefficient
Uncertainty (%) Material (106 J m−3 K−1) (W m−1 K−1) Anisotropy (nm)
Frequency r r r
Au 2.48 ± 0.07 150 ± 7.5 1 80 ± 1.5
range (Hz) (G, C SiO2) (G, κ SiO2) (C SiO2, κ SiO2) G C SiO2 κ SiO2
Silicon 1.65 ± 0.05 143 ± 7 1 5 × 105
104 − 3 × 107 −0.26 0.6 −0.77 17.7 5.4 13.7 Thin film 2 0.1–1000 1 ± 0.1 500 ± 5; 100 ± 2
106 − 3 × 107 −0.82 0.84 −0.99 8.9 56 66.2 Bulk 2 0.1–1000 1 ± 0.1 5 × 105
014901-6 Yang, Ziade, and Schmidt Rev. Sci. Instrum. 87, 014901 (2016)

FIG. 4. Calculated uncertainties (normalized standard deviations) for two and three parameter fits of a bulk substrate, plotted as a function of substrate thermal
conductivity, κ s . (a) Uncertainty of G and κ s for FDTR. (b) Uncertainty of G and κ s for TDTR. The solid and dashed lines are computed for 1 MHz and
10 MHz modulation frequencies, respectively. (c) Uncertainty of G, C s , and κ s for FDTR. (d) Uncertainty of G, C s , and κ s for TDTR.

Figure 4(a) shows the uncertainty for a two-parameter frequencies. Both techniques show similar uncertainties
FDTR measurement of κ s and G, assuming a known for κ s , but TDTR with a 10 MHz modulation frequency
volumetric heat capacity of Cs = 2 × 106 J m−3 K−1 with has a significantly lower uncertainty when measuring the
an assumed uncertainty of 3%, while in Fig. 4(b) we TBC between a transducer and a low thermal conductivity
show equivalent results for TDTR at two modulation material.

FIG. 5. Uncertainty in substrate thermal conductivity, κ s , determined by FDTR and TDTR with two-parameter fits of κ s and G, plotted as a function of (a) the
uncertainty of the transducer thickness, and (b) the uncertainty of the pump spot 1/e 2 radius, for pump and probe 1/e 2 spot radii of 6.9 µm.
014901-7 Yang, Ziade, and Schmidt Rev. Sci. Instrum. 87, 014901 (2016)

In many cases, the heat capacity of the sample is well

known from literature values and a two-parameter fit is
sufficient. Next, we consider the case where the heat capacity
is not known and calculate the uncertainty for three-parameter
fits of κ s , Cs , and G. The results for FDTR and TDTR are
shown in Figs. 4(c) and 4(d), respectively. We see that FDTR
has lower uncertainty when measuring both κ s and Cs for
the widest range of materials, although TDTR with a low
modulation frequency can also extract all three parameters
if the substrate conductivity is greater than ∼10 W/mK. FIG. 6. Sample configuration used for calculating the uncertainty of thin film
measurements.
TDTR with a high modulation frequency again has the
lowest uncertainty when measuring the TBC to a low thermal of the transducer thickness is important for both TDTR and
conductivity material. FDTR, while uncertainty in the laser spot size has a more
We now consider two of the main sources of uncertainty significant effect on FDTR.
for TDTR and FDTR measurements of bulk materials: the heat We typically use atomic force microscopy (AFM) to
capacity per unit area of the transducer (C · d), and the laser perform a step height measurement of the transducer thickness
spot geometry. In Fig. 5, we plot the uncertainty in substrate with a precision of ∼2 nm. For TDTR measurements, the
thermal conductivity, κ s , resulting from two-parameter fits of transducer thickness can also be determined using picosecond
κ s and G for the sample configuration shown in Fig. 3(a). acoustics.31 To determine the pump and probe spot radii,
In Fig. 5(a), we plot the uncertainty as a function of the we use a closed-loop piezoelectric stage to perform a two-
standard deviation of the gold transducer thickness, d, holding dimensional knife-edge measurement with a typical repeat-
the transducer heat capacity, C, constant, while in Fig. 5(b) we ability of ±0.01 µm.18 The sample surface should be within
plot uncertainty as a function of the standard deviation of the the depth of focus of the objective lens during measurement
pump spot radius. Figure 5 shows that accurate determination to avoid defocusing of the laser spots. The depth of focus,

FIG. 7. Calculated uncertainties for 100 nm and 500 nm thin films on a silicon substrate, plotted as a function of the thin film thermal conductivity κ t .
(a) Uncertainty of G1, G2, and κ t for FDTR of a 500 nm thick film. (b) Uncertainty of G1, G2, and κ t for TDTR of a 500 nm thick film. (c) Uncertainty of G1,
G2, and κ t for FDTR of a 100 nm thick film. (d) Uncertainty of G 1, G 2, and κ t for TDTR of a 100 nm thick film.
014901-8 Yang, Ziade, and Schmidt Rev. Sci. Instrum. 87, 014901 (2016)

FIG. 8. Calculated uncertainties for four-parameter FDTR measurements, plotted as a function of the thin film thermal conductivity, κ t . (a) Uncertainty of
κ t , C t , G 1, and G 2 for a 500 nm film. (b) Uncertainty of κ t , κ s , G 1, and G2 for a 500 nm film. (c) Uncertainty of κ t , C t , G1, and G2 for a 100 nm film.
(d) Uncertainty of κ t , κ s , G1, and G 2 for a 100 nm film.

∆z, around the minimum spot size can be calculated with and a 100 nm thick film. The uncertainties of these three
Gaussian beam optics,32 parameters are plotted as a function of κ t in Fig. 7. For these
 calculations, a value of 100 MW m−2 K−1 was used for G1
( )2
πwo2 w and G2, and the rest of the material properties are given in
∆z = ± − 1, (19) Table III. We used a frequency range of 104–3 × 107 Hz for the
λ w0
FDTR calculations and a modulation frequency of 10 MHz
where λ is the optical wavelength, w is the 1/e2 Gaussian spot for the TDTR calculations. We see that the measurements
radius, and w0 is the minimum focused radius. For example, are insensitive to G2 when the thermal conductivity of the
if w0 = 5 µm and λ = 800 nm, a tolerance of w = 5 ± 0.1 µm thin film is low. In addition, there is more uncertainty in
requires that ∆z = ±19.6 µm. For w0 = 1 µm and λ = 800 nm, the measurement of the 100 nm film than the 500 nm film
a tolerance of w = 1 ± 0.05 µm requires that ∆z = ±1.3 µm. because the thermal conductivity of the thin film and the
We use a CCD camera to focus on the knife edge during the TBCs become more closely correlated as the film becomes
spot size measurement. For sample measurements, the height thinner. Generally, TDTR has lower uncertainty in cross-
of the sample is adjusted until the CCD image is in sharp focus, plane transport properties for thinner films because of the
ensuring that the sample surface is at the location where the higher sensitivity to G1. For thin film measurements with
spot profiles were measured. either TDTR or FDTR, the uncertainty can be reduced by first
measuring G1 using a much thicker film and then performing
a two-parameter fit of k t and G2 for the thin film.
C. Comparison of FDTR and TDTR for thin films
Finally, we consider using FDTR to simultaneously
The measurement of a thin film on a substrate introduces measure four unknown parameters for thin film samples.
more potential unknowns in the heat transfer model. A typical In Fig. 8(a), we consider the case of a 500 nm thick film
sample configuration is shown in Fig. 6, where we have chosen with variable thermal conductivity on a silicon substrate,
a silicon substrate for the calculations. We first consider fitting where both the thermal conductivity, κ t , and heat capacity
three unknown parameters: the thin film thermal conductivity, of the thin film, Ct , are unknown, in addition to the two
κ t , and the two TBCs, G1 and G2, for a 500 nm thick film, unknown TBCs, G1 and G2. In Fig. 8(b), we consider the same
014901-9 Yang, Ziade, and Schmidt Rev. Sci. Instrum. 87, 014901 (2016)

FIG. 10. (a) Sample configuration of a gold film on a fused silica substrate.
(b) Measured phase noise with different lock-in amplifier bandwidths.

of fitted parameters were then fit with normal distributions to

FIG. 9. Flow chart of the Monte Carlo simulation used for uncertainty
analysis.
extract the mean values and standard deviations, which can
be directly compared to the diagonal elements of Eq. (15).
We implemented the Monte Carlo simulation in Matlab
500 nm film but now fit the thin film thermal conductivity,
using the lsqnonlin function, which employs a trust-region-
the substrate thermal conductivity, and the two TBCs. An
reflective least squares algorithm. This algorithm is a gradient
example of this type of four-parameter measurement can be
based method that requires initial guesses for the unknown
found in Ref. 33. It is also possible to characterize thin films
parameters. To ensure that the best fit values correspond to
with anisotropic thermal conductivity by treating the in-plane
a unique global minimum, our code randomly chooses the
and cross-plane thermal conductivities as separate unknown
initial guesses from a uniform distribution that spans 0.1 to 10
parameters, provided that some of the other unknowns, such
times our best initial guess values.
as the heat capacity or substrate thermal conductivity, are
We use FDTR measurements of a fused silica sample
determined by separate measurements.34 In Figs. 8(c) and
to compare the Monte Carlo simulation with Eq. (15). The
8(d), we repeat the uncertainty calculations for the case of a
sample configuration is shown in Fig. 10(a). The fused
100 nm film of variable thermal conductivity. In these cases,
silica was coated with gold by electron-beam evaporation
we see that it is only possible to fit a subset of the four unknown
at a deposition rate of 1.5 Å/s and chamber pressure of
parameters, and only for a limited range of thin film thermal
3 × 10−6 Torr. The thickness of the gold film was determined
conductivity values.
to be 81 ± 1.5 nm by AFM, and the pump and probe spot
radii used for the measurement were 2.8 ± 0.03 µm and
2.3 ± 0.02 µm, respectively, determined by a knife-edge
V. VALIDATION USING A MONTE CARLO SIMULATION
measurement.
A Monte Carlo simulation can be used as an alternative We consider two scenarios using the same set of
means to obtain the distribution of values estimated by least experimental data. In scenario 1, we fit κ Au, CAu, and GAu-SiO2,
squares algorithms. We used this approach to validate our assuming the literature values for the properties of the fused
analytical solution. Figure 9 shows the flow chart of our silica substrate,35 while in scenario 2, we fit κ SiO2, CSiO2, and
simulation. Each controlled parameter in the heat transfer GAu-SiO2, assuming that κ Au and CAu were determined from a
model is assumed to have a normal distribution about separate measurement using a reference substrate. For each
its mean value with a standard deviation estimated from scenario, we consider separately the impact of experimental
separate measurements. We obtained the distributions of fitted noise and uncertainty in the thickness of the gold film.
parameters by repeatedly performing a least squares fit on real The impact of noise was studied by varying the bandwidth
FDTR data while randomly selecting values of the controlled of the lock-in amplifier from 1.1 Hz to 81 Hz and performing
parameters from their distributions. The resulting histograms 200 FDTR measurements on the sample at each bandwidth
014901-10 Yang, Ziade, and Schmidt Rev. Sci. Instrum. 87, 014901 (2016)

setting. The standard deviation of the measured phase at each using only the data files obtained with the lock-in amplifier
frequency is plotted in Fig. 10. We represent the total amount bandwidth of 4 Hz. The results for scenario 1 and 2 are shown
of noise for each bandwidth setting as in Figs. 11(b) and 11(d), respectively. In all cases, the Monte

 Carlo results are in good agreement with the analytical results.
The mean best fit values obtained for κ Au, CAu, and
 N

φnoise = (σφ i )2/N, (20) GAu-SiO2 are listed in Table IV. We note that the volumetric
i=1
heat capacity of the gold film is ∼9% lower than the bulk
where σφ i is the standard deviation of the 200 phase value.35 Assuming a constant specific heat, this suggests that
measurements at the ith frequency and N is the total number the evaporated gold film is ∼9% less dense than bulk gold. The
of frequency points. The 200 data files at each noise level properties of gold thin films are sensitive to the deposition
were run through the Monte Carlo simulation to obtain conditions. Rutherford backscattering and AFM thickness
distributions in the fitted parameters, holding the uncertainty measurements by Malinský et al. showed that gold films
in the controlled parameters constant. The distribution of sputtered on silicon and glass at a deposition rate of 5 Å/s can
values for each unknown parameter was then fit with a be as much as 50% less dense than bulk gold.36 Based on our
normal distribution to obtain the standard deviation, which own FDTR measurements and scanning electron micrographs
we again use to represent the uncertainty in that parameter. of many gold films deposited on reference substrates with no
The results for scenario 1 and 2 are shown in Figs. 11(a) adhesion layer, we observe that the volumetric heat capacity
and 11(c), respectively. The impact of uncertainty in the gold and grain size of the gold films are strongly affected by the
film thickness was studied by performing the Monte Carlo deposition rate and the deposition method. For films on the
simulation for five values of uncertainty in film thickness, order of 100 nm thick, we observe that CAu can vary from

FIG. 11. Comparisons of Eq. (15) and a Monte Carlo simulation. For scenario 1, the uncertainty in κ, C, and G for the gold film is plotted as a function of
(a) phase noise and (b) the standard deviation of the gold film thickness. For scenario 2, the uncertainty in κ, C, and G for fused silica is plotted as a function of
(c) phase noise and (d) the standard deviation of the gold film thickness. In all cases, the Monte Carlo results, plotted in open symbols, are in good agreement
with the analytical results.
014901-11 Yang, Ziade, and Schmidt Rev. Sci. Instrum. 87, 014901 (2016)

TABLE IV. Material properties for the Au/SiO2 sample. 7C. Wei, X. Zheng, D. G. Cahill, and J.-C. Zhao, Rev. Sci. Instrum. 84, 071301

(2013).
ρc p κ d G 8O. Wilson, X. Hu, D. Cahill, and P. Braun, Phys. Rev. B 66, 224301
Material (106 J m−3 K−1) (W m−1 K−1) (nm) (MW m−2 K−1) (2002).
9A. J. Schmidt, M. Chiesa, X. Chen, and G. Chen, Rev. Sci. Instrum. 79,

Au 2.26 219 81 ± 1.5 51 064902 (2008).

10R. Costescu, M. Wall, and D. Cahill, Phys. Rev. B 67, 054302 (2003).
SiO2 1.63 1.38 5 × 105
11A. J. Schmidt, R. Cheaito, and M. Chiesa, J. Appl. Phys. 107, 024908

(2010).
12M. N. Luckyanova, J. Garg, K. Esfarjani, A. Jandl, M. T. Bulsara, A. J.
∼80% of the bulk value at a deposition rate of 4 Å/s to as much Schmidt, A. J. Minnich, S. Chen, M. S. Dresselhaus, Z. Ren, E. A. Fitzgerald,
as 97% of the bulk value at a deposition rate of 0.5 Å/s. We and G. Chen, Science 338, 936 (2012).
13H.-K. Lyeo and D. Cahill, Phys. Rev. B 73, 144301 (2006).
also find that κ Au can vary considerably due to the dependence 14P. E. Hopkins, M. Baraket, E. V. Barnat, T. E. Beechem, S. P. Kearney, J. C.
of grain size on the deposition method and rate; the values Duda, J. T. Robinson, and S. G. Walton, Nano Lett. 12, 590 (2012).
of κ Au we measured with FDTR are consistent with values 15P. E. Hopkins, L. M. Phinney, J. R. Serrano, and T. E. Beechem, Phys. Rev.

determined from electrical conductivity measurements via the B 82, 085307 (2010).
16J. Yang, E. Ziade, C. Maragliano, R. Crowder, X. Wang, M. Stefancich,
Wiedemann-Franz Law.11 For these reasons, it is important
M. Chiesa, A. K. Swan, and A. J. Schmidt, J. Appl. Phys. 116, 023515
to accurately characterize κ and C of the transducer using (2014).
a reference substrate that is placed close to the samples of 17S. Huxtable, D. G. Cahill, V. Fauconnier, J. O. White, and J.-C. Zhao, Nat.

interest in the deposition chamber. Mater. 3, 298 (2004).

18J. Yang, C. Maragliano, and A. J. Schmidt, Rev. Sci. Instrum. 84, 104904
In conclusion, we have presented an uncertainty analysis
(2013).
for the TDTR and FDTR techniques that accounts for 19M. G. Burzo, P. L. Komarov, and P. E. Raad, IEEE Trans. Compon. Packag.

both experimental noise and uncertainty in the controlled Technol. 28, 39 (2005).
20K. C. Collins, A. A. Maznev, J. Cuffe, K. A. Nelson, and G. Chen, Rev. Sci.
model parameters. We derived a generally applicable formula
Instrum. 85, 124903 (2014).
for multi-parameter measurements that use least-squares 21A. J. Schmidt, X. Chen, and G. Chen, Rev. Sci. Instrum. 79, 114902 (2008).
algorithms and applied it to TDTR and FDTR measurements 22A. J. Minnich, J. A. Johnson, A. J. Schmidt, K. Esfarjani, M. S. Dresselhaus,

of bulk materials and thin films, considering two, three, and K. A. Nelson, and G. Chen, Phys. Rev. Lett. 107, 95901 (2011).
23K. T. Regner, D. P. Sellan, Z. Su, C. H. Amon, A. J. McGaughey, and J. A.
four-parameter measurements of various thermal properties.
Malen, Nat. Commun. 4, 1640 (2013).
We found that both methods are suitable for a wide range of 24A. Majumdar and P. Reddy, Appl. Phys. Lett. 84, 4768 (2004).
situations and that TDTR has lower uncertainty when measur- 25R. B. Wilson, J. P. Feser, G. T. Hohensee, and D. G. Cahill, Phys. Rev. B 88,

ing the thermal boundary conductance to materials with low 144305 (2013).
26R. B. Wilson and D. G. Cahill, Nat. Commun. 5, 5075 (2014).
thermal conductivity, while FDTR can generally determine 27G. Seber and C. Wild, Nonlinear Regression, Wiley Series in Probability
more properties from a single measurement. The uncertainty and Statistics (John Wiley & Sons, 2003), pp. 21–23.
analysis was validated with a Monte Carlo simulation. We 28G. Seber and A. Lee, Linear Regression Analysis (John Wiley & Sons, Inc.,

believe this work will be useful to researchers applying the Hoboken, NJ, 2003), pp. 35–42.
29R. I. Jennrich, Ann. Math. Stat. 40, 633 (1969).
TDTR and FDTR techniques and to researchers using other 30R. E. Walpole and R. H. Myers, Probability and Statistics for Engineers and
metrology techniques that use least-squares algorithms. Scientists (Prentice-Hall, Inc., Englewood Cliffs, NJ, 1993), pp. 403–405.
31X. Zheng, D. G. Cahill, P. Krasnochtchekov, R. S. Averback, and J. C. Zhao,
1C. A. Paddock and G. L. Eesley, J. Appl. Phys. 60, 285 (1986). Acta Mater. 55, 5177 (2007).
2W. S. Capinski and H. J. Maris, Rev. Sci. Instrum. 67, 2720 (1996). 32E. Hecht, Optics, 4th ed. (Addison Wesley, 2002).
3D. G. Cahill, Rev. Sci. Instrum. 75, 5119 (2004). 33E. Ziade, J. Yang, G. Brummer, D. Nothern, T. Moustakas, and A. J. Schmidt,
4A. J. Schmidt, R. Cheaito, and M. Chiesa, Rev. Sci. Instrum. 80, 094901 Appl. Phys. Lett. 107, 091605 (2015).
(2009). 34V. V. Medvedev, J. Yang, A. J. Schmidt, A. E. Yakshin, R. W. E. van de
5D. G. Cahill, P. V. Braun, G. Chen, D. R. Clarke, S. Fan, K. E. Goodson, Kruijs, E. Zoethout, and F. Bijkerk, J. Appl. Phys. 118, 085101 (2015).
P. Keblinski, W. P. King, G. D. Mahan, A. Majumdar, H. J. Maris, S. R. 35D. R. Lide, CRC Handbook of Chemistry and Physics (Taylor & Francis,

Phillpot, E. Pop, and L. Shi, Appl. Phys. Rev. 1, 011305 (2014). Boca Raton, FL, 2007).
6J. Liu, J. Zhu, M. Tian, X. Gu, A. J. Schmidt, and R. Yang, Rev. Sci. Instrum. 36P. Malinský, P. Slepička, V. Hnatowicz, and V. Švorčík, Nanoscale Res. Lett.

84, 034902 (2013). 7, 241 (2012).