You are on page 1of 30

SAND2014-20695J

1 Error and Uncertainty in Raman Thermal Conductivity Measurements


2 Thomas Beechem,1, a) Luke Yates,1, 2 and Samuel Graham2
1)
3 Sandia National Laboratories, Albuquerque, NM, USA 87123
2)
4 G.W. Woodruff School of Mechanical Engineering, Georgia Institute of Technology,
5 Atlanta, GA, 30332

Error and uncertainty in Raman thermal conductivity measurements are investigated via fi-
nite element based numerical simulation of two geometries often employed—Joule-heating
of a wire and laser-heating of a suspended wafer. Using this methodology, the accuracy
and precision of the Raman-derived thermal conductivity is shown to depend on: (1) as-
sumptions within the analytical model used in the deduction of thermal conductivity (2)
uncertainty in the quantification of heat flux and temperature and (3) the evolution of ther-
momechanical stress during testing. Apart from the influence of stress, errors of 5% cou-
pled with uncertainties of ±15% are achievable for most materials under conditions typical
of Raman thermometry experiments. Error can increase to > 20%, however, for materials
having highly temperature dependent thermal conductivities or, in some materials, when
thermomechanical stress develops concurrent with the heating. A dimensionless parame-
ter—termed the Raman stress factor—is derived to identify when stress effects will induce
large levels of error. Taken together, the results compare the utility of Raman based con-
ductivity measurements relative to more established techniques while at the same time
identifying situations where its use is most efficacious.

6 Keywords: Raman Spectroscopy, Thermometry, Thermal Conductivity, Uncertainty

a)
Electronic mail: tebeech@sandia.gov

1
7 I. INTRODUCTION

8 Predicated by micro and nanosystems whose performance and reliability is dictated by their
9 ability to dissipate heat, methods of interrogating the thermal environment are a necessity to-
10 wards further development. Furthermore, with the implementation of “new” materials such as
11 graphene and other two-dimensional (2D) solids,1 operating temperatures must not only be quan-
12 tified but so too the properties of the materials that determine these temperatures. Raman spec-
13 troscopy is capable of meeting each of these challenges due to the technique’s ability to perform
14 non-contact thermal measurements with a resolution of ∼ 1K in a manner that is unobtrusive to
15 device operation.2 The technique also provides a spatial resolution that is on par with the struc-
16 tures themselves (∼ 0.5 µm), coupled with the ability to sample in a material-specific fashion. For
17 these reasons, Raman thermometry is often utilized to examine self-heating in operating devices3–6
18 but has recently been extended to measure the thermal conductivities of materials ranging from
19 graphene7–11 and carbon nanotubes12–14 to semiconductor nanowires15 and silicon.16,17

20 The increasing use of Raman thermometry for property determination stems from advantages
21 not present in more established approaches like the 3ω technique,18 time domain thermoreflectance
22 (TDTR)19 , or microelectromechanical (MEMS) platforms.20 First, Raman thermometry directly
23 probes the material of interest. Properties can be determined without the deposition of a metallic
24 transducer layer needed in TDTR, for example, or the microfabrication necessary in 3ω or MEMS
25 approaches. This ability not only provides experimental simplification but, more importantly,
26 is imperative for the probing of 2D-solids where any contact with surrounding materials (e.g.,
27 the transducer layer) drastically influences the thermal properties.21 Second, Raman measures the
28 in-plane thermal conductivity8 of a solid whereas standard implementations of other optical tech-
29 niques can be limited to only the cross-plane direction.22 For highly anisotropic systems typical
30 of many nanostructures, the in-plane direction is particularly germane making its measurement
31 critical.

32 Despite these advantages, Raman thermal conductivity measurements remain comparatively


33 immature. As such, there is no standard implementation as different methods are used to both
34 heat the sample and measure its temperature. Depending on the sample geometry, for instance,
35 joule14 or laser-heating8 has been utilized. Similarly, both the Stokes to anti-Stokes ratio9 and
36 the peak position7–11 of the Raman signal have been employed as the thermometer. Differing
37 implementations, in turn, will affect the accuracy and uncertainty surrounding the deduction of

2
38 thermal conductivity. For example, Raman thermometry utilizing the stokes to anti-Stokes ratio
39 typically results in uncertainties of the measured temperature that are significantly larger than
40 when the peak position is employed.5 Larger uncertainties in the measured temperature translate
41 into larger uncertainty in the deduction of thermal conductivity. In the same way, uncertainties in
42 Raman thermal conductivity measurements will be different when a device is Joule-, as compared
43 to laser-, heated due the fact that electrical dissipation can be measured with greater precision than
44 the absorption of light.
45 While apparent, the implications of these differences on uncertainty and error in the resulting
46 thermal conductivity measurement have not been addressed specifically. In response, uncertainty
47 and error in Raman thermal conductivity measurements are examined through the use of a vir-
48 tual Raman experiment carried out via numerical simulation. Specifically, Raman thermal con-
49 ductivity measurements are simulated in two geometries typically employed—Joule-heating of a
50 wire and laser-heating of a suspended wafer. Using the thermal simulation, a “measured” Raman
51 temperature is quantified and then transformed into thermal conductivity via the employment of
52 analytical models used in previous experimental reports.8,14 The deduced thermal conductivity is
53 subsequently compared to the thermal conductivity input within the simulation to assess error and
54 uncertainty. Using this methodology, error and uncertainty are examined as they manifest due
55 to: assumptions within the model (Section III A), uncertainty in the measurement of temperature
56 and heat input (Section III B), and finally the evolution of thermomechanical stress (Section III C).
57 Taken together, these assessments provide a means to both compare the capability of Raman to
58 more established techniques and to identify situations where its employment is most useful.

59 II. METHOD: VIRTUAL RAMAN EXPERIMENT

60 Thermal conductivity cannot be directly measured. Instead, it is temperature that is measured


61 when thermal conductivity is quantified. It is therefore necessary to translate the measured param-
62 eter, temperature, into that sought after, thermal conductivity. Models facilitate this translation by
63 describing the experiment in such a way that the temperature distribution becomes a function of the
64 thermal conductivity. Inserting different thermal conductivities into the model produces variable
65 predictions of temperature. That thermal conductivity giving the best fit between the predicted and
66 experimentally acquired temperature is deduced to be the measured value.
67 The accuracy of a thermal conductivity measurement is dependent upon more than just the ca-

3
68 pability of the experimental probe of temperature. A perfect measurement of temperature will not
69 necessarily result in a perfect measurement of thermal conductivity. Rather, only a perfect model
70 of the experiment coupled with a perfect temperature measurement will result in an absolutely
71 accurate measurement of thermal conductivity. Assessing the utility of a thermal conductivity
72 measurement thus necessitates consideration of both the temperature probe and the thermal model.
73 In any real experiment, deconvolving error arising from the model versus that of the mea-
74 surement is difficult, however. No measurement of temperature is perfectly precise or accurate.
75 Similarly, no model perfectly describes an experiment. Uncertainties in dimension, homogeneity,
76 and even the non-thermal properties (e.g., density) of the material being measured, for example,
77 dictate that the model only imperfectly images reality. Error in the resulting measurement of
78 thermal conductivity will therefore have contributions arising from both the limitations of the tem-
79 perature measurement and the model. Quantifying the contribution of each source relative to the
80 total error necessitates some level of idealization. In response, a virtual experiment is employed
81 here using three-dimensional finite element analysis simulations to assess the utility of employ-
82 ing Raman thermometry to deduce thermal conductivity. Below, details of the simulations are first
83 provided followed by the methodology used to leverage the simulations’ results into an assessment
84 of Raman based quantifications of thermal conductivity.

85 A. Joule-Heating

86 Joule-heated wires having a radius of 500 nm and a length 2L = 30 µm were simulated using
87 ANSYS-based thermal analysis software as shown in Figure 1. Similar geometries have been
88 employed in Raman based thermal conductivity measurements of carbon nanotubes and other
89 nanowires.12–14 Uniform heat generation is assumed while the ends of the wire are constrained
90 to remain at a temperature of To = 27◦ C. The surface of the wire is defined to be adiabatic in
91 line with experiments in which conduction into the air and radiation is of negligible consequence.
92 Having an aspect ratio of 60, thermal transport is one-dimensional. An analytical model of the
93 temperature distribution can therefore be written as:23
( )
q̇L2 x2
T (x) = 1 − 2 + To (1)
2k L

94 where q̇ is the volumetric heat generation, and k is a temperature independent thermal conduc-
95 tivity. With this geometry, several materials were simulated in an effort to assess the ability of

4
96 the Raman technique to deduce a range of thermal conductivities. Specifically, gallium arsenide
97 (GaAs), silicon (Si), gallium nitride (GaN), and a “graphene-like” material—termed for simplicity
98 graphene (MLG)24 —were each examined. Properties employed within the simulations are pro-
99 vided in Table I. Total heat generation, Q, was varied between each material to impose similar
100 temperature differences—∆T = (Tmax − To )—in all simulations. Mechanically, the ends of the
101 wires were fixed along the x-axis (ϵ(x) = 0) and free to expand along all other axes resulting
102 in the development of a uniaxial stress as is addressed in Section III C. To ensure accuracy of
103 the finite element simulations, meshes were refined until the temperature distribution changed by
104 less than 1%. Additionally, simulations assuming a temperature independent thermal conductivity
105 were found to match the analytical solution to within < 1%.

FIG. 1. (left) Schematic of Joule–heated wire. Colors correspond to the relative temperature distribution in
the wire. (right) Simulated temperatures (Tsim. , red dots) are utilized to calculate a measured temperature
(Tm , black square). The thermal conductivity deduced from the measured temperature (i.e., kmeas. (Tm ))
fits the simulated temperature profile well when employed with Equation 1. A much poorer fit is realized
when the actual thermal conductivity, ksim. at Tm is input into the same equation. The discrepancy is due
to the use of a temperature independent thermal conductivity in the model and can lead to errors of > 10%
in the measured thermal conductivity.

106 The virtual experiment was performed by first calculating the temperature resolved by the Ra-
107 man instrument, Tm . To facilitate this calculation, Raman photons are assumed to be collected in
108 direct proportion to the manner that incident photons are absorbed. As the Raman emitted photons
109 carry the temperature from where they were conceived, the measured temperature results from

5
110 a weighted average of all collected photons across the region probed by the laser. Weighting is,
111 therefore, dictated by the laser profile. Assuming a Gaussian beam shape having a full width at
112 half-max (FWHM) of Γ = 1 µm, Tm was calculated according to:
( )
∫L −4 ln(2)x2
Γ2
−L
Tsim. (x)e dx
Tm = ∫L (
−4 ln(2)x2
) (2)
−L
e Γ2 dx

113 where the laser has been focused in the center of the wire and Tsim. (x) is the temperature output
114 from the simulation. Equation 2 implies that variations along the radial direction are extremely
115 small in line with the assumption of 1D heat transport. Variations across the radius were less than
116 0.001K.
117 Translation of the measured temperature into a thermal conductivity takes place by replacing
118 Tsim. (x) in Equation 2 with T (x) of Equation 1 to estimate a model temperature, Tmod. .25 A least
119 squares fitting routine is then employed to deduce a measured thermal conductivity (kmeas. (Tm )),
120 by minimizing the difference between the modeled and measured temperatures. Error in the mea-
121 sured thermal conductivity, ψ, is quantified by comparison to the thermal conductivity utilized
122 within the simulation via :
kmeas. (Tm ) − ksim. (Tm )
ψ= (3)
ksim. (Tm )
123 Factors influencing the magnitude of ψ are addressed in Section III.
124 Uncertainty in the measurement of thermal conductivity is also quantified using Equation 3. For
125 example, uncertainty resulting from a temperature measurement specified to ±δT is calculated
126 by deducing thermal conductivity assuming the measured temperature is Tm ± δT rather than
127 Tm . kmeas. (Tm ± δT ) is then inserted into Equation 3 with the difference in ψ when utilizing
128 kmeas. (Tm ± δT ) versus kmeas. (Tm ) taken to be the uncertainty. Similarly, uncertainty arising
129 from a measurement of the heat input known to within ±δ q̇ is calculated by inserting q̇ ± δ q̇
130 into Equation 1 and then recalculating the thermal conductivity using this altered form of the
131 model. Uncertainty is found by inserting kmeas. (q̇ ± δ q̇) into Equation 3 and once again taking the
132 difference between the resulting ψ and that obtained with kmeas. (q̇) to be the uncertainty. The total
133 uncertainty caused by the simultaneous presence of each effect is quantified by performing each
134 of these steps simultaneously, namely by quantifying kmeas. (Tm ± δT, q̇ ± δ q̇) and inserting this
135 value into Equation 3. Analogous methods are utilized to quantify uncertainty and error for the
136 laser-heating model described in Section II B.

6
TABLE I. Thermal and structural properties of simulated materials

Property GaAs Silicon GaN Graphene


W
( )
T −1.23 26
( T )−1.3 26 ( )
T −0.73 5
( T )−5.2 10,27
k ( mK ) 46 300 148 300 245 300 3000 300
E (Gpa) 84.626 13026 38828 100029
ν 0.311826 0.27826 0.18330 0.16529
αL (K −1 ) × 10−6 5.7026 2.6226 5.5931 -7.0032
−1
∆ωT ( cmK ) -0.01833 -0.0242 -0.0142 -0.0167
−1
∆ωσ ( cm
GPa ) -1.9234 -235 -1.45536 -3.1537
σ∗ 0.05 0.03 0.22 1.4

137 Despite the similarity in their calculation, the terms uncertainty and error are not interchange-
138 able. In Section III A, error is examined by assuming a perfectly accurate measurement of Tm and
139 heat flux each having no uncertainty. Section III B, on the other hand, focuses upon uncertainty
140 by calculating the range of values in thermal conductivity that result from the uncertainty present
141 in the measured values (i.e., heat input and temperature) that lead to its deduction. In all sub-
142 sequent plots, error is denoted by symbols whereas uncertainty—being indicative of a range—is
143 denoted by either contours (Figure 5) or bars (Figure 6). As will be shown, uncertainty can at
144 times counteract the effect of error resulting in a measurement that while more uncertain could be
145 more accurate.

146 B. Laser-heating

147 In the analysis of 2D-materials like graphene7–11 and molybdenum disulphide (MoS2 )38,39 , Ra-
148 man thermal conductivity measurements employ the laser as both the heater and thermometer.40,41
149 To examine this approach, simulations of a 30 x 30 x 1µm region (see Figure 2) heated by a
150 laser having a gaussian profile with a FWHM of 1µm were performed for each of the materials
151 listed in Table I except for GaAs. Explanation of GaAs’ exclusion is presented below. A square,
152 rather than circular geometry, was utilized to assess the impact of non-radial geometries typical of
153 suspended flakes or films like those examined previously.7,42 Even with the square geometry, the
154 resulting temperature distribution is to first order radially symmetric (see Figure 2) as the rectan-
155 gular domain is much larger than the laser diameter. While not mathematically pure, models of

7
156 the temperature distribution using cylindrical coordinates are most often employed.7 Error arising
157 from this approximation is addressed in Section III A.

FIG. 2. (left) Schematic of laser heated wafer. Colors correspond to the relative temperature distribution in
the wire. (right) Simulated temperatures (Tsim. , red dots) are utilized to calculate a measured temperature
(Tm , black square). Model temperatures are acquired using a numerical solution to the differential equation
described by Equation 6 (solid line/dashed line). Solid line corresponds to fitting across entire domain
along r-axis. Dashed line is a fit to the region r ≤ 8 µm where radial approximation is fully valid. While
differences between the two are typically less than 0.5K (see Inset) they lead to thermal conductivities that
may differ by > 10%.

158 Heat was input into the system on a volumetric basis according to:
( )
−4 ln(2)r 2

q̇(r, z) = (1 − ρ)q ′′ αe(−αz) e Γ2


(4)

159 where q̇ is the volumetric heat generation, q ′′ is a coefficient defining the maximum heat flux, ρ is
160 the reflectance at the probing laser wavelength of 532 nm, and α is the absorption coefficient at
161 this wavelength. Practically, the continuous function of Equation 4 was placed into the discretized
162 FEA model via a series of nested and stacked annuli each having a constant q̇(r, z) so as to build
163 up a heat generation profile matching the continuous form.
164 Owing to non-uniform heat-generation in both the r and z directions, the thermal profile is not
165 necessarily one-dimensional. If a material has a large absorption coefficient at the probing/heating
166 wavelength, heat is deposited preferentially near the surface leading to the possibility of a large
167 temperature gradient through the thickness. This was observed in GaAs, for example, where the
168 combination of a comparatively diminutive skin depth (1/α = 140nm) and thermal conductivity

8
169 results in an appreciable temperature difference through the thickness that is comparable to that in
170 the radial direction. Deduction of the thermal conductivity in this instance therefore necessitates
171 a 2D-model. Previous experimental reports7–11 have overwhelmingly focused on atomically thin
172 materials where such a complication does not arise and a 1D analysis is appropriate. To maintain
173 consistency between these previous reports and the present investigation, a 1D analysis is em-
174 ployed here necessitating conditions where such an assumption maintains its validity. For GaN
175 and graphene, the combination of a large skin depth and thermal conductivity results in through
176 thickness temperature gradients less than 1% of that in the radial direction. Thus, a 1D assumption
177 is fully valid. For Si, the 1D assumption is less robust—but reasonable—since gradients through
178 the depth reached only 6% of that in the radial direction.
179 The measured temperature was found by incorporating a weighted average both radially and
180 through the thickness according to:
( )
∫t∫R −4 ln(2)x2
Γ2
0 0
Tsim. (r, z)e e(−αz) dxdz
Tm = (
∫ t ∫ R −4 ln(2)x2
) (5)
0 0
e Γ2 e(−αz) dxdz
181 where t is the 1µm thickness and R = 15µm and once again Tsim. is the output of the simulation
182 considered to be the true temperature response. Here, it is assumed that the Raman system has
183 a confocal response that is much greater than 1µm and as such photons emanating from various
184 depths have equal probability of being collected. It is also assumed that the absorption coefficient,
185 α, at the Raman-shifted wavelength is equivalent to that of the probing wavelength.
186 Translation of the measured temperature into a thermal conductivity was accomplished via the
187 implementation of a thermal model describing the temperature distribution as a function of ther-
188 mal conductivity. Under the assumption of 1D radial heat conduction, the following ordinary
189 differential equation governs thermal diffusion:
( )
1 d dT (r) q̇(r)
r + =0 (6)
r dr dr k
190 where q̇(r) is the total heat absorbed through the thickness at a given radius. q̇(r) is calculated
191 by integrating Equation 4 through the thickness. At r = 0, a Neumann boundary condition is
192 assumed: dT (r)
dr r=0
| = 0. Meanwhile, for r = R, a Dirichlet condition is prescribed: T (R) = To .
193 Under these conditions, Equation 6 is solved using a MATLAB based numerical ordinary differ-
194 ential equation solver to obtain the model temperature distribution, T (r).43 Figure 2 demonstrates
195 that T (r) matches the simulated temperature profile thus demonstrating the capability of the nu-
196 merical solution and its applicability to deducing thermal conductivity. Practically, the measured

9
197 thermal conductivity, km (Tm ), is acquired by minimizing the difference between the modeled and
198 measured temperatures when T (r) is inserted into Equation 5.

199 III. RESULTS

200 A. Error in Ideal Measurement

201 Even with a perfect measurement of temperature, error exists in the deduction of thermal con-
202 ductivity. Such error evolves owing to the inadequacy of the thermal model to perfectly describe
203 the temperature distribution that exists during the measurement. Here, two separate deficiencies of
204 the thermal models—employment of a radial geometry in laser-heating and use of a temperature
205 independent thermal conductivity—are shown to result in error in the assessment of conductivity
206 even when the experiment is ideal (i.e., no uncertainty/error in measurement of temperature or heat
207 flux).

208 1. Approximations of a Radial Temperature Distribution

209 In the case of laser-heating, a cylindrical model (see Equation 6) has been implemented to
210 describe a domain that is not cylindrical (Figure 2).7,42 Approximating a radial temperature distri-
211 bution requires that the temperature contours be equivalent for all paths emanating from the laser
212 spot. In a square geometry, this dictates that the temperature distribution running along the center-
213 line of the square—the r-axis—and that of the diagonal remain equivalent. As the distances of

214 these two paths have a factor of 2 difference even as the temperature at their respective termini
215 are equivalent (To ), similarity necessarily diminishes away from the laser spot. Thus, the validity
216 of the radial approximation persists for only a subset of the total domain. For the geometry em-
217 ployed here (2R = 30 µm), inspection revealed deviations between the diagonal and center-line
218 paths emerge at a distance of 8 µm from the laser spot. Approximation of a radial geometry is
219 therefore most accurate only when considering the sub-domain—r < 8µm.
220 To assess the implications of the radial approximation, thermal conductivity was calculated in
221 two different manners. First, the model temperature, T (r) of Equation 6, was evaluated utilizing
222 the entire r-axis of Figure 2 (i.e., the center-line) to deduce thermal conductivity. Second, conduc-
223 tivity was quantified utilizing only the sub-domain 0 < r < 8µm by considering To of the Dirichlet
224 condition of Equation 6 to be Tsim (8µm). The two methodologies result in temperature profiles

10
225 that differ by less than < 0.5K as is highlighted by the inset to Figure 2 . While small, these tem-
226 perature differences do reflect variations in the deduced thermal conductivity. When considering
227 a material having a temperature independent thermal conductivity, for example, the sub-domain
228 methodology measures the thermal conductivity to within 0.1% whereas utilization of the full do-
229 main leads to errors > 2%. In more realistic materials possessing a temperature-dependent thermal
230 conductivity, differences in the measured conductivity between two methodologies remain similar
231 being within 3% of one another. However, when combined with experimental uncertainty like
232 that considered in Section III B or stress effects like that highlighted in Section III C, differences
233 between the two methods can reach > 15%. Counterintuitively, the mathematically less precise
234 assumption of including the entire domain leads to higher accuracy as it serves to counterbal-
235 ance error introduced by the model’s assumption of a temperature independent thermal conduc-
236 tivity (see Section III A 2). In all subsequent results of the laser-heating geometry, the sub-domain
237 methodology (i.e., To = Tsim (8µm)) is employed as it separates the convolution of these effects
238 and most accurately reflects the intended radial nature of the measurement.
239 The above discussion emphasizes that the radial approximation applied to a rectangular do-
240 main—even a domain 30 times larger than the spot size—is capable of imparting error into Raman
241 thermal conductivity measurements. Mitigation of this error can take place by either incorpora-
242 tion of a truly radial geometry8 or implementing a second temperature measurement, for example
243 at 8 µm in this case, that ensures the approximation’s validity. While the 8 µm sub-domain is
244 particular to the geometry pursued here, the general necessity of temperature equivalence for all
245 paths within a radial approximation is applicable for arbitrary geometries. Namely, in laser heat-
246 ing experiments, thermal conductivity must be deduced within only that part of the geometry for
247 which the radial approximation remains valid. FEA models may therefore be necessary to assess
248 the applicability of using this approach in the study of arbitrarily shaped exfoliated flakes.

249 2. Approximation of a Temperature-Independent Thermal Conductivity

250 Employing a temperature independent thermal conductivity is an additional source of error


251 that exists even with an ideal quantification of temperature. To demonstrate, Figure 3 compares
252 silicon’s temperature dependent thermal conductivity to that deduced from the simulated Raman
253 measurements. In both geometries, the Raman deduced thermal conductivity is larger than the
254 actual value. The overestimation is a consequence of measuring near the peak of the temperature

11
255 distribution where the thermal conductivity is at its minimum. Since the model attempts to describe
256 an experiment in which the thermal conductivity varies using only a single value, an average value
257 influenced by the entire range must be utilized. Thus, the deduced–“average”–value must be
258 higher than the actual–“minimum”–conductivity. This can be visualized by placing the actual
259 thermal conductivity at Tm into the model, which necessarily results in an overestimation of the
260 temperature profile as shown by the green curve in Figure 1.

FIG. 3. Comparison of silicon’s temperature dependent thermal conductivity to that deduced by the Raman
measurement in the (a) Joule–heated and (b) laser-heated geometries under conditions of: an ideal tempera-
ture measurement and stress biasing (|σ| > 0). Differences between the measured thermal conductivity and
actual values arise due to the assumption of a temperature independent thermal conductivity in the analytical
model and biasing of the temperature measurement that occurs due to stress.

261 The amount of error stemming from the use of a temperature independent thermal conductivity
262 is dependent upon three separate factors: (1) the magnitude of the temperature excursion (2) tem-
263 perature variance across the geometry and (3) the aggressiveness by which a material’s thermal
264 conductivity changes with temperature. Each of these factors determines the degree to which ther-
265 mal conductivity varies across the geometry. Larger variations induce greater error. Increases in
266 ∆T , for example, result in a broader range of thermal conductivities and hence greater error. For
267 a given ∆T , meanwhile, the Joule–heated geometry has greater error compared to the laser heated
268 geometry since its temperature variance is greater (see Figure 3 and Figure 4). This effect is exac-
269 erbated for materials having highly temperature dependent thermal conductivities. Error increases
270 proportionally to the exponent, γ, used in the power law description of thermal conductivity’s
( T )−γ
271 dependence upon temperature—k(T ) = A 300 (see Figure 4 and Table I).

12
FIG. 4. Error in Raman based conductivity measurements at ∆T = 23K stemming from the use of analyti-
cal models having a temperature independent conductivity. As thermal conductivity becomes more sensitive
to temperature, error increases. Lines are guides to the eye.

272 The magnitude of error stemming from the deficiencies of the models is comparatively small.
273 At ∆T = 23K for instance, the error is less than 10% for each of the materials except graphene.
274 This level of error is comparable to the uncertainty reported for more established techniques like
275 TDTR and 3ω.44,45 It is important to recognize, however, that whereas the 10% associated with
276 TDTR and 3ω is an uncertainty, the 10% calculated in this section for the Raman measurement is
277 an error. It is therefore a best-case scenario. In reality, uncertainty implicit with any real measure-
278 ment of temperature or heat flux will ride atop this error as well. Use of a temperature independent
279 conductivity is not an absolute requirement, however. Numerical methods (e.g., FEA21,46 ) can eas-
280 ily incorporate temperature dependent thermal conductivities into the model and remove the need
281 of this assumption.

282 B. Experimental Uncertainty

283 Deductions of thermal conductivity emanate from measurements of temperature, Tm , and heat
284 dissipation, Q. Measurements of either parameter implicitly contain uncertainty. Traditional
285 implementations of Raman thermometry, for example, are capable of specifying temperature to
286 within only ± 1K.2,6 Similarly, the amount of heat dissipated can be quantified with only a finite
287 level of certainty. Uncertainty can stem from dissipation at the contacts rather than within the
288 nanowire in Joule-heating experiments, for example, or due to ill-quantified reflection, rather than

13
289 absorption, during laser-heating. Being deduced from these parameters, uncertainty in the mea-
290 surements of temperature and heat dissipation will therefore induce uncertainty in the deduction
291 of thermal conductivity. To then assess how uncertainty cascades, changes in thermal conductiv-
292 ity were calculated when the measured temperature was varied by ±3K or when the total heat
293 dissipation fluctuated by ±7.5%.
294 As shown in Figure 5 and Figure 6, fuzziness in the measurement of temperature can signifi-
295 cantly impact the measurement of thermal conductivity. First, the contours of Figure 5 indicate that
296 conductivity measurements necessitate temperature excursions in excess of 20K. Smaller temper-
297 ature excursions lead to substantial uncertainty in thermal conductivity that can reach in excess of
298 40%. Second, thermal conductivity’s sensitivity to Tm is geometry dependent. For a given uncer-
299 tainty in Tm , the Joule-heating geometry shows a lower uncertainty than that of the laser-heating
300 for all materials (see Figure 6). The higher sensitivity is a consequence of the larger temperature
301 gradients near the measurement of Tm for the laser- versus Joule-heating geometries (compare
302 Figure 2b and Figure 1b). Finally, if temperature measurements are biased to higher values, er-
303 rors associated with the use of a temperature independent thermal conductivity are compensated
304 leading to a more accurate measurement in total (see lower contours of Figure 5 associated with
305 Tm + 2K). Upon examining the influence of stress in Section III C, such lower temperature biasing
306 will be of consequence.
307 For ∆T > 20 K, uncertainty in the thermal conductivity measurement scales directly with the
308 magnitude of uncertainty in the measured parameters that lead to its deduction. Figure 6 indicates
309 that ± ∼ 10% uncertainty in the measurement of Tm (i.e., δTm = ±2K for ∆T = 23K) leads
310 to a measurement of thermal conductivity having an uncertainty of ∼ ±10%. The relationship is
311 independent of either the thermal conductivity of the analyte or the source of the uncertainty as,
312 in a similar fashion, a ±5% uncertainty in Q results in an uncertainty of ∼ ±5% in conductivity
313 for each of the examined materials (see Figure 5(c,d)). Considering uncertainty of each source
314 simultaneously, meanwhile, leads to a total uncertainty in conductivity that is approximately the
315 sum of the individual sources. For example, ±10% uncertainty in Tm coupled with ∼ ±5% in Q
316 leads to a thermal conductivity measurement that has an uncertainty of ∼ ±15% irrespective of
317 material or geometry.
318 The direct scaling of uncertainty indicates that under reasonable expectations of experimental
319 capability Raman thermal conductivity measurements provide uncertainty bounds that are compa-
320 rable to established techniques. To illustrate quantitatively, consider a laser-heated measurement

14
FIG. 5. Uncertainty contours in thermal conductivity measurements of silicon resulting from uncertainties
in the measured temperature of (a,b) ±3K and heat input (c,d) ± 7.5% for the (a,c) Joule–heating and (b,d)
laser-heating geometries. The black line provides the error of a conductivity measurement in which the
measurement of temperature and heat input is ideal. Lower contours correspond to (a,b) δTm > 0K and
(c,d) to δQ < 0.

321 of silicon having a 23K temperature excursion. Assuming temperature is measured to within ±2K
322 and that the heat input is known only to within ±5%, the simulated Raman thermal conductivity
323 measurement results in a conductivity value of km = 138 ± 21 W/mK compared to an actual
324 value at this temperature of 132 W/mK. Bounds of ±15% are comparable to those achieved with
325 TDTR and 3ω.45,47 While comparable in uncertainty, the level of error associated with the Ra-
326 man measurement may be much larger than other techniques depending on the material, however,
327 owing to the larger temperature excursion necessary to realize a 10% error in the measurement
328 of temperature. For example, when measuring a material with a highly temperature dependent
329 thermal conductivity like graphene, a ∆T = 20K may result in an uncertainty of only ±10% but

15
FIG. 6. Material and geometry specific error resulting from uncertainties in the (solid) temperature mea-
surement of ±2K and (speckled) Q of ±5% at ∆T = 23K.

330 due to the inadequacy of the model result in error in excess of 20%. Therefore, it is necessary to
331 consider both error and uncertainty together when designing the parameters of a Raman thermal
332 conductivity experiment.

333 C. Thermomechanical Induced Error

334 The most precise—and thus most oft utilized—Raman thermometry technique fits the peak
335 position (i.e., the energy) of a Raman mode to deduce temperature.2,5 While precise, the peak po-
336 sition is not a direct measurement of temperature. Rather, it is a probe of the zone center optical
337 phonons, which themselves have an energy that is dependent upon the interatomic potential. In-
338 teratomic potential, in turn, is sensitive to changes of the strain state in the material. With heating,
339 materials expand or contract resulting in some finite strain. This strain alters the interatomic poten-
340 tial thereby inducing a shift in the peak position that can be implemented as a probe of temperature.
341 Being a temperature measurement deduced from strain, significant errors evolve in peak-position
342 based thermometry when thermomechanical stress develops concurrent to the heating.4 The im-
343 plications of these errors on Raman thermal conductivity measurements are investigated in the
344 following.
345 For each geometry, thermal expansion was constrained along one axis thereby inducing the
346 development of uniaxial stress, σ. Uniaxial stress was converted to an equivalent Raman shift
347 via ωσ = σ∆ωσ where ∆ωσ is a material specific calibration of the Raman shift with stress, which

16
348 is related to the Grüneisen parameter. Values for this parameter are given in Table I. The shift due
349 to temperature was calculated in a similar fashion: ωT = ∆ωT Tm where ωT is a material’s peak
350 position calibration with temperature. In line with previous reports,4,5,36 the total Raman shift is
351 the superposition of each effect: ω̄ = ωσ + ωT . The total shift was then converted into a stress
352 biased measurement of temperature, Tmσ , by the relation, Tmσ = ω̄/∆ωT . Thermal conductivity was
353 calculated by inserting Tmσ in place of Tm thus providing an assessment of the thermomechanical
354 induced error in the measurement.
355 Most materials expand with an increase in temperature. If this expansion is constrained, com-
356 pressive stress results. Compressive stress causes an up-shift in the Raman peak position. This
357 is in contrast to the down-shift in the peak position that occurs with an increase in temperature.
358 Taken together, peak position based thermometry measurements will therefore under predict the
359 actual temperature when made in the presence of compressive stress. Lower temperatures result
360 in larger, more positive, errors in thermal conductivity (see −2K bars in Figure 6). To illustrate,
361 Figure 3 displays this stress-induced over prediction of thermal conductivity for silicon.
362 Upon examination of Figure 3 and Figure 7, three aspects of stress biasing on Raman thermal
363 conductivity measurements become apparent. First, the amount of error induced by the stress does
364 not necessarily scale directly with the magnitude of stress. At the smallest temperature excursions
365 when stress is minimized relative to that at larger ∆T , thermomechanical induced error is at its
366 largest (Compare blue/green curves in Figure 3 at ∆T = 10 and 50K). Rather, it is the percent-
367 error in the measured temperature induced by the stress biasing that dictates the amount of error
368 in the measurement of thermal conductivity. Second, stress induced error is dependent upon the
369 nature of the experiment. At a given ∆T for example, Joule-heating results in less stress than the
370 laser-heating geometry and is therefore less biased by its effect (see Figure 7). Finally, the amount
371 of error induced by the stress is highly material dependent. At ∆T = 23K for example, a stress-
372 biased Raman thermal conductivity measurement results is only 10% error in silicon but over 40%
373 error in graphene. Stress, depending on the material, thus has the potential to make Raman thermal
374 conductivity measurements ineffectual. Its presence must, therefore, be considered.
375 As shown in Figure 7, a material’s sensitivity to stress in a peak-position based Raman ther-
376 mometry experiment scales directly to a dimensionless parameter termed here as the Raman stress
377 factor, σ ∗ , given by:
αL E∆ωσ

σ = (7)
∆ωT
378 where αL is the coefficient of thermal expansion (CTE) and E is the elastic modulus. The nu-

17
FIG. 7. Absolute value of error induced by stress as a function of the non-dimensional Raman stress factor
parameter, σ ∗ .

379 merator in Equation 7 can be viewed as the shift due to stress arising from thermal expansion.
380 Dividing by ∆ωT , in turn, should be interpreted as then transforming this shift into a dimension-
381 less “temperature”. Larger values therefore correspond to larger “temperatures” and hence bigger
382 biasing. This will be true in any peak-based Raman thermometry measurement whether it be for
383 assessments of device self-heating or thermal conductivity.
384 Relative to GaN, graphene seems have a smaller stress induced error for the laser-heating ge-
385 ometry despite having a larger value of σ ∗ . This is a consequence of MLG’s negative coefficient
386 of thermal expansion.48,49 With a negative thermal expansion, the stress will be tensile resulting in
387 an over-prediction of temperature. As mentioned in Section III B, an over-prediction of temper-
388 ature initially compensates for errors arising from the use of a temperature independent thermal
389 conductivity in the thermal model. Thus, the stress-induced error in graphene actually induces a
390 swing from a positive 16% error in the ideal case to a −60% error upon including the stress bias.
391 Thus, the stress induced error is, in reality, > 70% and larger than that observed in GaN in line
392 with the scaling of stress and σ ∗ presented above.

393 IV. SUMMARY

394 Raman thermal conductivity measurements offer advantages not present in more traditional
395 approaches—minimal sample preparation, small spatial resolution, material specificity and most
396 importantly a keyhole to examine in-plane transport. With these advantages, the technique’s imple-

18
397 mentation has recently accelerated as a tool particularly suited for the examination of 2D-materials.
398 Despite its accelerating implementation, the magnitude of error and uncertainty in the measure-
399 ment has not been examined in detail thus making it difficult to compare the technique’s capability
400 to more established paradigms. In response, error and uncertainty in Raman thermal conductivity
401 measurements have been examined as they emerge due to: inadequacies in the thermal models
402 utilized to deduce conductivity, uncertainty in the measured parameters—temperature and heat
403 input—and finally biasing that occurs due to thermomechanical stress.
404 Uncertainty in Raman thermal conductivity measurements scales directly with the amount of
405 uncertainty in the measurements of temperature and heat input. Other techniques of measuring
406 thermal conductivity typically have uncertainties on the order of ±10% necessitating that the Ra-
407 man technique achieve comparable performance. Having the capability of measuring temperature
408 to within only ±1K with Raman thus requires temperature excursions in excess of 10K (i.e., 10%).
409 Large temperature excursions, however, dampen the applicability of assuming a temperature-
410 independent thermal conductivity in the model used to translate the measured temperature into
411 a value of thermal conductivity. Error from this assumption is therefore present in all Raman ther-
412 mal conductivity measurements at a level proportional to the degree to which thermal conductivity
413 changes with temperature. For most materials, this error in on the order of only 5 − 10%. Raman
414 thermal conductivity measurements are therefore comparable in performance to more traditional
415 techniques—error of 5 − 10% with an uncertainty of ±10%.—for many materials. However, the
416 error can balloon to > 40% thus rendering the technique’s applicability obsolete when either ther-
417 momechanical stress substantially biases the measurement of temperature or when a material has
418 a highly temperature-dependent thermal conductivity. Sensitivity to either error source is material
419 dependent and can be identified prior to the measurement if general properties are known using
420 the dimensionless Raman stress factor derived herein or the expected temperature dependence of
421 the thermal conductivity. While generally useful, Raman thermal conductivity measurements have
422 therefore a limited range of applicability that must be considered on a material-by-material basis
423 before employment.

424 V. ACKNOWLEDGEMENTS

425 Critical review of this work by Justin Serrano and Colin Landon of Sandia National Labora-
426 tories is greatly appreciated. This work was supported by the LDRD program at Sandia National

19
427 Laboratories (SNL). Sandia National Laboratories is a multiprogram laboratory managed and op-
428 erated by Sandia Corporation, a wholly owned subsidiary of Lockheed Martin Corporation, for the
429 US DOE National Nuclear Security Administration under Contract No. DE- AC04-94AL85000.

430 REFERENCES

1
431 A. K. Geim and I. V. Grigorieva, Nature 499, 419 (2013).
2
432 T. Beechem and J. Serrano, Spectroscopy 26, 36 (2011).
3
433 M. Kuball, S. Rajasingam, A. Sarua, M. J. Uren, T. Martin, B. T. Hughes, K. P. Hilton, and R. S.
434 Balmer, Applied Physics Letters 82, 124 (2002).
4
435 T. Beechem, S. Graham, S. P. Kearney, L. M. Phinney, and J. R. Serrano, Review of Scientific
436 Instruments 78, 061301 (2007).
5
437 T. Beechem, A. Christensen, S. Graham, and D. Green, Journal of Applied Physics 103, 124501
438 (2008).
6
439 C. B. Saltonstall, J. Serrano, P. M. Norris, P. E. Hopkins, and T. E. Beechem, Review of Scientific
440 Instruments 84, 064903 (2013).
7
441 A. A. Balandin, S. Ghosh, W. Bao, I. Calizo, D. Teweldebrhan, F. Miao, and C. N. Lau, Nano
442 Letters 8, 902 (2008), doi: 10.1021/nl0731872.
8
443 W. Cai, A. L. Moore, Y. Zhu, X. Li, S. Chen, L. Shi, and R. S. Ruoff, Nano Letters (2010), doi:
444 10.1021/nl9041966.
9
445 C. Faugeras, B. Faugeras, M. Orlita, M. Potemski, R. R. Nair, and A. K. Geim, ACS Nano 4,
446 1889 (2010), doi: 10.1021/nn9016229.
10
447 J.-U. Lee, D. Yoon, H. Kim, S. W. Lee, and H. Cheong, Physical Review B 83, 081419 (2011).
11
448 S. Chen, A. L. Moore, W. Cai, J. W. Suk, J. An, C. Mishra, C. Amos, C. W. Magnuson, J. Kang,
449 and L. Shi, ACS Nano 5, 321 (2010).
12
450 I.-K. Hsu, M. T. Pettes, A. Bushmaker, M. Aykol, L. Shi, and S. B. Cronin, Nano Letters 9, 590
451 (2009).
13
452 Q. Li, C. Liu, X. Wang, and S. Fan, Nanotechnology 20, 145702 (2009).
14
453 Y. Yue, G. Eres, X. Wang, and L. Guo, Applied Physics A: Materials Science and Processing
454 97, 19 (2009).
15
455 M. Soini, I. Zardo, E. Uccelli, S. Funk, G. Koblmller, A. F. i Morral, and G. Abstreiter, Applied
456 Physics Letters 97, 263107 (2010).

20
16
457 S. Perichon, V. Lysenko, B. Remaki, D. Barbier, and B. Champagnon, Journal of Applied
458 Physics 86, 4700 (1999).
17
459 S. Perichon, V. Lysenko, P. Roussel, B. Remaki, B. Champagnon, D. Barbier, and P. Pinard,
460 Sensors and Actuators A: Physical 85, 335 (2000).
18
461 D. G. Cahill, Review of Scientific Instruments 61, 802 (1990).
19
462 D. G. Cahill, Review of Scientific Instruments 75, 5119 (2004).
20
463 J. Seol, I. Jo, A. Moore, L. Lindsay, Z. Aitken, M. Pettes, X. Li, Z. Yao, R. Huang, and
464 D. Broido, Science 328, 213 (2010).
21
465 W. Jang, Z. Chen, W. Bao, C. N. Lau, and C. Dames, Nano Letters 10, 3909 (2010).
22
466 Here, we assume that optical measurements are performed in a backscattering arrangement
467 where light impinges perpendicularly to a sample surface. Thermal transport parallel to the
468 sample surface is considered in-plane. Cross-plane transport is perpendicular to the surface or,
469 equally, parallel to the wavevector of the incoming light.
23
470 F. Incropera, D. De Witt, T. Bergman, A. Lavine, B. DeWitt, D. Dewitt, L. Van Hove, C. DeWitt-
471 Morette, and P. Cartier, Fundamentals of heat and mass transfer, 6th ed. (Wiley, 2007).
24
472 It is to be stressed that graphene is not being simulated. Rather, a fictitious material having the
473 thermal and structural properties of suspended monolayer graphene but with finite thickness is
474 employed. This distinction allows simulations to be implemented with an identical geometry for
475 all materials thereby facilitating a 1 to 1 comparison of the Raman technique’s capability apart
476 from geometric considerations.
∫L −4 ln(2)x2
Γ2
25 −L T (x)e dx
477 Mathematically, the model temperature is written as: Tmod. = ∫L −4 ln(2)x2
where T (x) is
e Γ2 dx
−L
478 found using Equation 1.
26
479 CINDAS, “Thermophysical properties of matter database,” (2014).
27
480 Interpolated results of Lee et al.10 utilized in simulations. Power law presented here only for
481 comparison to other materials.
28
482 M. Yamaguchi, T. Yagi, T. Sota, T. Deguchi, K. Shimada, and S. Nakamura, Journal of Applied
483 Physics 85, 8502 (1999), compilation and indexing terms, Copyright 2007 Elsevier Inc. All rights
484 reserved 04057901415 0021-8979.
29
485 C. Lee, X. Wei, J. W. Kysar, and J. Hone, Science 321, 385 (2008).
30
486 M. A. Moram, Z. H. Barber, and C. J. Humphreys, Journal of Applied Physics 102, 023505
487 (2007).

21
31
488 W. Qian, M. Skowronski, and G. S. Rohrer, “Structural defects and their relationship to nucle-
489 ation of gan thin films,” (1996).
32
490 W. Bao, F. Miao, Z. Chen, H. Zhang, W. Jang, C. Dames, and C. N. Lau, Nat Nano 4, 562
491 (2009).
33
492 A. Sarua, A. Bullen, M. Haynes, and M. A. K. M. Kuball, Electron Devices, IEEE Transactions
493 on 54, 1838 (2007).
34
494 P. Puech, G. Landa, R. Carles, and C. Fontaine, Journal of applied physics 82, 4493 (1997).
35
495 I. De Wolf, Semiconductor Science and Technology , 139 (1996).
36
496 T. Beechem, A. Christensen, D. Green, and S. Graham, Journal of Applied Physics 106, 114509
497 (2009).
37
498 T. Mohiuddin, A. Lombardo, R. Nair, A. Bonetti, G. Savini, R. Jalil, N. Bonini, D. Basko,
499 C. Galiotis, and N. Marzari, Physical Review B 79, 205433 (2009).
38
500 S. Sahoo, A. P. Gaur, M. Ahmadi, M. J.-F. Guinel, and R. S. Katiyar, The Journal of Physical
501 Chemistry C 117, 9042 (2013).
39
502 R. Yan, J. R. Simpson, S. Bertolazzi, J. Brivio, M. Watson, X. Wu, A. Kis, T. Luo, A. R.
503 Hight Walker, and H. G. Xing, ACS nano 8, 986 (2014).
40
504 J. S. Reparaz, E. Chavez-Angel, M. R. Wagner, B. Graczykowski, J. Gomis-Bresco, F. Alzina,
505 and C. M. Sotomayor Torres, Review of Scientific Instruments 85, (2014).
41
506 Recently, a two-laser Raman approach has been reported in which the probing and heating laser
507 are separated.40 For simplicity, we focus here only on the more oft implemented single laser
508 approach.
42
509 Z. Luo, H. Liu, B. T. Spann, Y. Feng, P. Ye, Y. P. Chen, and X. Xu, Nanoscale and Microscale
510 Thermophysical Engineering 18, 183 (2014).
43
511 The ODE was solved using the built-in MATLAB function bvp4c. MATLAB Release 2013b,
512 The MathWorks, Inc., Natick, Massachusetts, United States.
44
513 H.-K. Lyeo and D. G. Cahill, Physical Review B (Condensed Matter and Materials Physics) 73,
514 144301 (2006).
45
515 Y. K. Koh, S. L. Singer, W. Kim, J. M. Zide, H. Lu, D. G. Cahill, A. Majumdar, and A. C.
516 Gossard, Journal of Applied Physics 105, 054303 (2009).
46
517 T. Westover, R. Jones, J. Y. Huang, G. Wang, E. Lai, and A. A. Talin, Nano Letters 9, 257
518 (2009).

22
47
519 T. S. English, L. M. Phinney, P. E. Hopkins, and J. R. Serrano, Journal of Heat Transfer 135,
520 091103 (2013).
48
521 S. Linas, Y. Magnin, B. Poinsot, O. Boisron, G. Forster, Z. Han, D. Kalita, V. Bouchiat, V. Mar-
522 tinez, and R. Fulcrand, arXiv preprint arXiv:1411.7840 (2014).
49
523 It should be emphasized that MLG is a graphene–like fictitious material. For example, the CTE
524 of graphene remains under debate.48 As such, while the conclusions reached in this section are
525 consistent with the properties of Table I, they may not be necessarily representative of graphene
526 under all circumstances.

23

You might also like