You are on page 1of 19

One-step Measurement of Thermoelectric Material and

Transport Descriptors
D. V. Maheswar Repaka1*, A. Suwardi1*, Z. Ren2*, D. Ding3, D. Cai4,Y. Zheng1, T.
Buonassisi2,†#, and K. Hippalgaonkar1#
1
Institute of Materials Research and Engineering, Agency for Science Technology and Research,
#08-03 Innovis, 2 Fusionopolis Way, Singapore 138634
2
Singapore-MIT Alliance for Research and Technology, 1 CREATE Way, #10-01 CREATE
Tower, Singapore 138602
3
Simtech
4
IHPC

*equal contribution, †on leave from Massachusetts Institute of Technology, #correspondence to


kedarh@imre.a-star.edu.sg, buonassisi@mit.edu

Inverse design of high-performance thermoelectric materials is impeded by the lack of

high-quality experimental data. We present a method to measure a comprehensive suite of

thermoelectric materials properties within minutes, using a Bayesian inference algorithm.

Employing only two input parameters — temperature differential and external load resistance —

the algorithm extracts material (Seebeck coefficient, electrical conductivity) and transport

descriptors (energy-dependent scattering parameter, reduced Fermi potential and weighted

mobility, μ0 m3/DOS
2
).) by solving the full Boltzmann Transport equations. This approach is

validated using conventional chalcogenide thermoelectric materials (doped Bi 2Te3), and reveals

fundamental electronic transport parameters that is otherwise not possible without sophisticated

measurement techniques. To our knowledge, the described approach represents the first instance

in which a thermoelectric characterization method is invented in response to the existence of a

machine-learning method.

Keywords: Electrical transport, thermoelectrics, Bayesian inference, machine learning, material


descriptor, rapid characterization, Boltzmann transport.
To increase the return-on-investment of materials development, and address societal

problems in a timely manner, there is an urgent need to accelerate materials innovation.

Approaches inspired by the Materials Genome Initiative (1), including high-throughput

experiments (HTE) (2), machine learning (3, 4), and high-performance computing techniques,

are beginning to make major inroads into materials science (5). In particular, machine learning

fundamentally shifted the paradigm of experimental design from a traditional Edisonian

approach to a descriptive-predictive-prescriptive strategy. In the descriptive domain,

comprehensive materials databases have been developed, such as AFLOW library (6), ICSD (7),

Open Quantum Database (8), Materials Project (9) and Citrine Informatics (10), made easier to

navigate with versatile mining tools such as Matminer (11). Recently, machine learning has

dramatically reduced the computational time to calculate materials properties compared to first

principles and Density Functional Theory (DFT) calculations, without compromising accuracy,

achieving rapid descriptive-predictive learning (5). In addition, machine learning methods are

effective at handling multi-dimensional data, critical for accelerated materials development (12).

Functional materials systems such as photovoltaics (13), piezoelectrics (14) and

thermoelectrics (15-18) have been at the forefront of this revolution, whereby the methodology

for screening and designing efficient materials through machine learning algorithms has been

demonstrated. Typically, such computation-based approaches do not exceed the predictive step

within the descriptive-predictive-prescriptive framework. Here, we report a high-throughput

experimental technique to provide rapid evaluation of key thermoelectric transport parameters

using machine learning. We infer bulk thermoelectric transport properties from non-destructive

electrical measurements using Bayesian inference in combination with the analytical full

Boltzmann Transport Equations serving as our forward model. An in-house setup is used to
obtain experimental power versus load curves of a single thermoelectric leg (cuboid) subject to

different thermal differentials (∆T’s) as the discriminative parameter. From the experimental

data, we are able to infer Seebeck coefficient (S), electrical conductivity (σ), energy dependent

scattering exponent (r), reduced Fermi potential (η), as well as weighted mobility ( μw ). In

addition, we are able to deduce experimental errors such as parasitic contact resistance (RC) as

well as systematic error in temperature differential (∆Terror). Our work represents the first time

these thermoelectric descriptors can be ascertained experimentally in a one-step, high-throughput

fashion enabled by machine learning.

The performance parameter of thermoelectric materials is quantified by the dimensionless

figure of merit, zT = S2σT/κ, where S, σ, and κ represents Seebeck coefficient, electrical

conductivity, and thermal conductivity, respectively. These are first-level material descriptors —

the numerator S2σ (power factor) is governed by electronic transport properties of the material

while the denominator, κ =κl + κe, is made up of lattice and electronic thermal transport properties,

respectively. At optimal doping concentration, i.e. in the range of degenerate doping 10 19-1021

cm-3 (19), a good thermoelectric material acts as a phonon-glass electron-crystal (20), which

requires a high Seebeck coefficient and electrical conductivity as well as low thermal

conductivity. In this regime, one can approximate the thermoelectric properties in a simplistic

form through Mott’s formula for the Seebeck coefficient:

2 2

( )
2
8 π kB ¿ π
S= 2
mDOS T 3
3eh 3n

(1)


σ =neμ ; μ= ¿ (2)
mI
¿
where n is carrier concentration, μ is carrier mobility, τ is carrier relaxation time, m DOSand
¿
mI represent density of states effective mass and the inertial effective mass, respectively.

It is therefore evident that while high effective mass (m¿) and low carrier concentration

(n) can result in an enhanced Seebeck coefficient, they tend to decrease the electrical

conductivity and carrier mobility. To resolve this conflicting interdependency between S and σ,

the B-factor (quality factor) can be used:

( ) σk T
2
kB E0
B= (3)
e L

3/2 3/2
8 π e (2 me k B T ) ¿ 3/2 8 π e (2 me k B T )
where σE = 3
µ0 mDOS = 3
µ w (4)
0
3h 3h

σ E describes how well a bulk material conducts electricity for a given η. It can also be expressed
0

in term of µw , the weighted mobility. This then allows the zT to be written in terms of two

independent variables, namely the reduced Fermi potential, η = (EF-Ec/kBT), as well as the B-

factor, both of which are temperature-dependent:

2
S (η)
zT =
(k B /e)2
η
+ L(η)
B ln(1+ e )

(5)

Here, L represents the Lorenz number L = κ e/(σT) , which is related to η via S (S is the

experimental indicator of η). Empirically, it was shown that L [10-8 WΩ/K2] ≈ 1.5 + exp

( 116−μ|SV|/ K ). (21)
The B-factor was elaborated upon by Kang and Snyder(21) using the assumption of acoustic

phonon scattering, constant relaxation time, and rigid band approximation, although the original

expression was first introduced by Chasmar and Stratton (22). However, this approach is only
¿
successful when mI and η are mutually independent, which is not always the case (23). Also In

addition, the B-factor does not take into account the energy-dependent carrier scattering, r,

which reveals the dominating scattering mechanisms at a particular temperature and thus affects

zT. (21, 24). Other efforts to resolve the interdependencies between S and σ includes expressing
¿ ¿
power factor in terms of m DOS /mI (the ratio of density of states effective mass to the inertial

effective mass) which accounts for not only the difference between the band and inertial

effective masses (25), but also anisotropies in the crystal structure and band structure by

considering the complexity of the Fermi surface (26). However, it requires separate experiments

to ascertain
¿
mI ADDIN EN . CITE< EndNote >¿ Cite>¿ Author > Spitzer <¿ Author >¿ Year>1957 <¿ Year> ¿ RecNum >97<¿ Re

In addition, in order to characterize σ, the determination of carrier relaxtion time τ, is

essential. Under the relaxation time approximation, the total average scattering rates can be

expressed as τ = τ 0 E r by accounting for scattering of charges due to acoustic phonons, optical

phonons with deformation coupling, optical phonons with polar coupling, vacancies, as well as

impurities (27, 28). Typically at temperatures above the Debye temperature in bulk materials,

electron-phonon scattering is the dominant mechanism (29). High-fidelity Density Functional

Theory (DFT) calculations using deformation potentials (30) and Electron-Phonon Wannier

(EPW) methods using maximally localized Wannier functions (31) can be used to calculate

average τ directly using Fermi’s golden rule, but are computationally expensive. Lately, faster
approximations such as Electron-Phonon Averaging (EPA) (17) have been employed with some

success. Experimentally, r can only be accurately probed by time-intensive Nernst experiments,

a key bottleneck to quantifying charge transport in thermoelectric materials.

In our work, we demonstrate a high-throughput one-step experiment: we employ

Bayesian inference using the full Boltzmann transport equations, which incorporates energy-

dependent scattering to obtain multiple transport descriptors that determine the thermoelectric

performance in an otherwise mathematically under-constrained problem. The primary advantage

of employing Bayesian inference lies in its ability to update the conclusion of a hypothesis based

on new observations(13). Here, we first measure the power-load curve with a fixed temperature

gradient across the sample. As new experimental observations are obtained: power across a

varying load with various temperature gradients, the uniform probability distribution for our

transport descriptors (prior) converge to high probability posterior distribution, P(θ| Pobs ) (i.e.

probability of θ-inferred values given Pobs-experimental results) effectively with only up to five

different temperature gradients for each material. We incorporate our domain expertise into

Bayesian inference first with a simple device model (details in supplementary information) and

subsequently using the full Boltzmann transport equations to generate the likelihood probability,

defined as P( Pobs ∨θ ). Thus, our experimental setup allows us to directly glean out the electronic

transport descriptors from a forward model. In addition, we are also able to visualize the

correlation between each of these transport descriptors, which is an additional advantage of using

Bayesian inference. (13)

Such system-level experiments can determine material and transport descriptors from a

simple, one-step device level experiment, which represents a shift in paradigm of thinking from

property-centric approach (Figure 1(a)) to sample-centric approach (Figure 1(b)). Typically, to


extract S, σ, r and n, one would have to use the method of four coefficients, comprising of four

tedious independent measurements (Seebeck, four-probe, Hall and Nernst) (32), which require

separate device configurations and forcing fields (magnetic, thermal, and electric) (Figure 1(a)).

Figure 1. (a) Conventional ‘property-centric’ measurement techniques (widely known as method of four-
coefficients) to obtain material thermoelectric and transport parameters. It consists of a series of independent
measurements ranging from Seebeck, four-probe electrical measurement, Hall measurement as well as Nernst
measurement to obtain various thermoelectric and transport descriptors. While measurements of S,  and Rc are
direct, the r,  and necessitate the use of full Boltzmann Transport Equations (b) ‘Sample-centric’ one-step
experiment using Bayesian inference to extract the same key thermoelectric and transport descriptors directly by
analyzing the Power vs Load curves using Bayesian Inference in a probabilistic fashion.

The one-step power-load experiments are performed on single leg (cuboid) bulk

polycrystalline thermoelectric materials (Figure 2(a)). For each material, up to four sets of

temperature differentials are used to impose a discriminative conditions for our inference as S, σ

and Rc have distinct temperature dependence (details about the measurements and experimental
setup, and forward model and comparison with multivariate regression are discussed in the

supplementary section). Using output power across load resistance data (P vs RL, as shown in

Figure 2(b)), we are able to extract probabilistic distributions for S, σ, RC and ΔTerror, which is

shown in Figure 2(c) for Bi0.5Sb1.5Te3. The contour plot in Figure 2 (c) represents the probability

distribution of the parameters values. Here, the Gaussian error from the noise model in the

Bayesian inference is higher than our experimental error (details in Supplementary section).

The inferred values for all samples are tabulated in Table 1. In addition, independent

experiments using PPMS were done to measure the Seebeck and electrical conductivities as well

as carrier concentration of the samples. Evidently, besides inferring transport parameters with

good accuracy, the Bayesian inference can also isolate extrinsic experimental contributions such

as contact resistance as well as error in temperature measurement, which can be useful in device

optimization process.

In all our single-leg samples, the contact resistance falls in the range of just below 1.0 Ω,

which is about two orders of magnitude higher than the intrinsic resistance of the samples itself

(Table 2). Likewise, the error in temperature difference measurements across the samples are

near to 1 K, which originates from thermal resistance across the contacts. This indicates that

more power can actually be extracted from the samples by lowering the contact and thermal

resistance, respectively.
Figure 2. (a) In-house power-load measurement setup for bulk materials which consists of 1. Teflon insulation, 2.
Embedded hot-side thermocouple, 3. Hot-side electrical connection, 4. Chip heater, 5. Sample, 6. Embedded cold-
side thermocouple, 7. Cold-side electrical connection, 8. Spring load, and 9. Copper block. (b) Experimental (points)
and Bayesian inferred (solid line) power-load curves at three different temperature gradients. (c) Probability
distribution showing the most likely values for S and σ, RC and ΔTerror for Bi0.5Sb1.5Te3. The first column shows the
correlation between S and ΔTerror, RC, and σ as well as probability distribution of S. The second column shows the
correlation between σ and ΔTerror and RC as well as the probability distribution of σ. The third column shows the
correlation between RC and ΔTerror as well as probability distribution of RC. The last column shows the probability
distribution of ΔTerror.
Table 1. Comparison between inferred values to experimentally verified values.

ΔTerror μw
S(µV/K) σ (x105 S/m) RC (Ω) η r μ(cm2/Vs)
(K) (cm2/Vs
Samples PPMS Bayesian PPMS Bayesian Bayesian Bayesian Bayesian Bayesian Bayesian PPMS
)
Bi0.4Sb1.6Te3 162 161 1.66 2.12 0.60 0.89 833 0.96 -0.44 260

Bi0.5Sb1.5Te3 189 187 1.02 0.77 0.61 0.87 134 0.78 -0.19 207

Bi0.6Sb1.4Te3 214 217 0.77 0.77 0.66 0.33 175 0.33 -0.11 319

Table 1. Comparison of Bayesian-inferred parameters and values obtained from independent PPMS measurements
for all samples.
σ =uw∗e ( k B T )(3 /2 ) ( 23r +1) F r+ ( 1/ 2)
(η)

( ) ( 2rD +1) F
1 D /2
e 2
σ= 2 2 ( k B T )r +( D / 2) r + ( D /2) −1
(η)
2 pi ℏ

Next, we obtain microscopic transport descriptors by using the full Boltzmann transport

equations, extending the device-level forward model (supplementary). Using Bayesian inference,

the energy dependent carrier scattering parameter r, reduced Fermi potential η, as well as

transport coefficientweighted mobility μw can be directly ascertained. ,The inferred microscopic

transport descriptors as well as the inter-correlation between these parameters are shown in

Figure 3(a). The values at the peak of the probability distribution for each parameter are

compared to the values obtained on the same sample using conventional measurement techniques

(i.e., Physical Property Measurement System), as summarized in Table 1. It is evident from

Table 1 that the reduced fermi level η decreases as Bi content in Sb 2Te3 increases. To understand

the physical correlation between the microscopic transport descriptors (r, η, and σ E ) as describe 0

above we first have studiedstudy closely the defect physics in BixSb2-xTe3 for x = 0.4, 0.5, 0.6 and

2. The inferred properties are plotted in Figure 3(bb). Bi2Te3 is a quintuple layered material with

“Te(1)-Bi-Te(2)-Bi-Te(1)” stacked structure along c-axis, where intralayer Bi-Te (2) bonds have

covalent or ionic nature and interlayer Te (1)-Te(1) bonds have weak van-der Waal’s interactions

(Figure 3d). The intrinsic defects, especially anti-sites and vacancies play a major role on the

type of the transport, carrier concentration and thermoelectric properties of these V 2VI3

compounds. Generally, these defects are thermally created and depend strongly on the material

processing conditions.(33) The concentration of anti-site defects is highly dependent on the


average difference in electronegativity and the difference in the atomic radii between cations and

anions.(34) Smaller difference in electronegativity and ionic radii results in higher level of anti-

site defects. Both p and n type transport behavior is observed in Bi 2Te3 due to the intrinsic anti-

site defects of BiTe (Bi at Te sites) and TeBi (Te at Bi sites), respectively. In general, the p and n-

type transport in Bi2Te3 is also affected by the Te vacancies (V Te), where the p type transport in

Bi2Te3 is associated to the excess vacancies of Te and n-type is associated to less vacancies of

Te.(35)

In our case, the observed p-type behavior in as-prepared Sb 2Te3 and concomitant n-type behavior

in Bi2Te3 are mainly due to the intrinsic anti-site defects of Sb Te (Sb at Te sites) in Sb2Te3 and

TeBi in Bi2Te3, respectively [ref].

In order to understand the change in doping concentration as more Bi is introduced in Sb 2Te3 in

our experiments, it is important to understand how antisite Sb Te and BiTe create more holes as

described below:

Sb2Te3 ⇌ 2(Sb’Te +1) + VTe +2 + 2(VSb -3) + 3/2 Te2(g) 0 + 2h’

Bi2Te3 ⇌ 2Bi’Te +2+ VTe + 2VBi -6 + 3/2 Te2(g) 0 + 2h’

Since the electronegativity difference between the cation and anion of Sb 2Te3 (XSb-XTe) is lower

than in the case of Bi2Te3 (XBi-XTe), the total number of anti-sites Bi2Te3 are less than those in

Sb2Te3. With increasing amount of Bi doping in the Sb sites in Sb 2Te3, i.e., Sb2-xBixTe3, there are

two effects on the anti-site distribution that will change the hole concentration correspondingly:

(1) the number of SbTe+1 anti-sites decreases, and (2) the number of Bi Te+1 increases; however, the

rate of decrease of (1) is higher than increase of (2), which results in the decrease of majority

hole carrier concentration (see attached figure in Supplementary Information) [ref]. Interestingly,
this trend is arrested at higher Bi doping levels – for Bi doping of x > 0.5, the number of BiTe+1

anti-site defects decreases with further increasing x reducing the number of holes [ref]. Through

this analysis, we understand the decrease of reduced Fermi potential (η) as Bi increases in our Bi

doped Sb2Te3 indicating the decrease of carrier concentration.


Figure 3. (a) Cross-correlation between inferred microscopic transport parameters as well as probability
distribution of the inferred values using Boltzmann transport equations in Bayesian algorithm. (b) Correlation
between η and μw vs r. The correlation shows that η decreases and μw increases as a function of increasing r. (c)
Effect of doping on the r, η and μw . (d) Sensitivity plot of the normalized effect of changing r, η and μw on power
factor.

Together, the extraction of η, r andσ E is the one-step direct experimental measurements of the
0

state-of-the-art transport descriptors of thermoelectric materials that have direct, independent

implication on thermoelectric materials screening.


d concentration) also
Besides lowering η, the increase in Bi content (decrease in carrier
C
leads to an increase of r (i.e. deviates from acoustic phonon scattering towards ionized impurity

scattering). This is in contrary to literature reports that associate the decrease in carrier

concentration (η) to reduced ionized impurity scattering(36-38) and weaker electron-electron

interaction, respectively. (39, 40). As discussed previously, as Bi content increases in Bi-doped

Sb2Te3, eventhough the rate of decrease of the number of Sb Te+1 anti-sites is higher than increase
of the number of BiTe+1 , the total number of defects in our system actually increases because of

the substitution of Bi in Sb sites. Therefore, it is not surprising to see that in this material system,

as Bi content increases (η decreases), r tends to increase.

In addition, we also observed that r increases (approaches zero, with decreasing σ. This

is expected since increasing r leads to lower value of carrier mobility (μ = μ0 Er ¿and thus lower

σ and higher S (29). Furthermore, the correlations between η and S as well as between μw

and σ are consistent with literature reports. (21, 30, 41)

Lastly, the weighted mobility ❑w trend with increasing Bi content follows the

experimental mobility trend obtained from hall measurements. first decreases and then increases

when Bi content increases from x = 0.5 to x = 0.6, which is consistent with the experimental

carrier mobility (μ) values observed from hall data.

In order to examine the relative importance of each of the transport parameters to the

power factor, it is useful to look at the percentage change in power factor in response to

independent variation in each of these transport parameters, as shown in Figure 3(c). The

sensitivity, defined as the ratio of the relative change in power factor to the relative change in the

input parameter value (10%), allows us to glen which transport descriptors affects powerfactor

the most. In order to determine this sensitivity plot, these parameters were varied by 10% from

their inferred values in table 1 to increase the overall power factor. It shows that r is the most

sensitive parameter that affects the power factor. This is consistent with the Boltzmann Transport

Equation, where r affects both S and σ. In fact, tuning the energy-dependent scattering, r, is

known to be the key ingredient to designing new thermoelectric materials. (42) Similarly, efforts

to account for r via computational techniques at different doping levels and temperatures are also
ongoing.(17) Lastly, it is important to note that the three descriptors are strongly material and

temperature dependent. For instance in SrTiO 3, η and μ0was shown to be correlated due to

polaron effects, which quantifies the interaction between charge carriers and lattices. (23)

Table 1. Comparison between inferred values to experimentally verified values.

S(µV/K) σ(S/m) Table 1. Comparison between inferred values to experimentally ve

Materials PPMS Bayesian PPMS Bayesian Bayesian

Bi0.4Sb1.6Te3 162 161 1.66 x 1.52 x 1.39 x 105

105 105
Bi0.5Sb1.5Te3 189 187 1.02 x 7.69 x 4.84 x 104

105 104
In summary, we have shown that one-step simple power-load experiment using Bayesian

inference with a well refined forward model in the form of full Boltzmann transport equations

with energy-dependent scattering consideration, we are able to extract key thermoelectric

transport descriptors (energy dependent scattering parameter r, reduced fermi potential η, as well

as transport coefficientweighted mobility, ❑w ❑❑ ). More importantly, it also enables us to


rapidly elucidate the inter-correlation between these descriptors. This allows us to probe into the

deepest layer of physical understanding of any thermoelectric material in a high-throughput

fashion. Going forward, our one step technique will be an indispensable complement to other

high throughput materials discovery effort, and will provide an avenue for building a database of

thermoelectric descriptors on a wide variety of compounds. We have also shown that Bayesian

inference is not only an effective diagnostic tool, but can potentially be used as prescriptive tool

that can eventually improve our state-of-the-art understanding of materials in general.


References

1. J. P. Holdren, Materials genome initiative for global competitiveness. National Science and
technology council OSTP. Washington, USA, (2011).
2. M. L. Green et al., Fulfilling the promise of the materials genome initiative with high-throughput
experimental methodologies. Applied Physics Reviews 4, 011105 (2017).
3. A. Gopnik, Making AI more human. Scientific American 316, 60-65 (2017).
4. M. I. Jordan, T. M. Mitchell, Machine learning: Trends, perspectives, and prospects. Science 349,
255-260 (2015).
5. G. Crabtree, S. Glotzer, B. McCurdy, J. Roberto, "Computational materials science and chemistry:
accelerating discovery and innovation through simulation-based engineering and science,"
(USDOE Office of Science (SC)(United States), 2010).
6. S. Curtarolo et al., AFLOWLIB. ORG: A distributed materials properties repository from high-
throughput ab initio calculations. Computational Materials Science 58, 227-235 (2012).
7. A. Belsky, M. Hellenbrandt, V. L. Karen, P. Luksch, New developments in the Inorganic Crystal
Structure Database (ICSD): accessibility in support of materials research and design. Acta
Crystallographica Section B: Structural Science 58, 364-369 (2002).
8. J. E. Saal, S. Kirklin, M. Aykol, B. Meredig, C. Wolverton, Materials design and discovery with
high-throughput density functional theory: the open quantum materials database (OQMD). Jom
65, 1501-1509 (2013).
9. A. Jain et al., Commentary: The Materials Project: A materials genome approach to accelerating
materials innovation. Apl Materials 1, 011002 (2013).
10. M. Hutchinson. (2016).
11. L. Ward et al., Matminer: An open source toolkit for materials data mining. Computational
Materials Science 152, 60-69 (2018).
12. J.-P. Correa-Baena et al., Accelerating Materials Development via Automation, Machine
Learning, and High-Performance Computing. Joule, (2018).
13. R. E. Brandt et al., Rapid photovoltaic device characterization through Bayesian parameter
estimation. Joule 1, 843-856 (2017).
14. S. Somnath et al., Ultrafast current imaging by Bayesian inversion. Nature communications 9,
513 (2018).
15. P. Gorai, V. Stevanović, E. S. Toberer, Computationally guided discovery of thermoelectric
materials. Nature Reviews Materials 2, 17053 (2017).
16. A. o. Furmanchuk et al., Prediction of seebeck coefficient for compounds without restriction to
fixed stoichiometry: A machine learning approach. Journal of computational chemistry 39, 191-
202 (2018).
17. G. Samsonidze, B. Kozinsky, Accelerated Screening of Thermoelectric Materials by First‐
Principles Computations of Electron–Phonon Scattering. Advanced Energy Materials, 1800246
(2018).
18. M. Yamawaki, M. Ohnishi, S. Ju, J. Shiomi, Multifunctional structural design of graphene
thermoelectrics by Bayesian optimization. Science Advances 4, eaar4192 (2018).
19. M. Jonson, G. Mahan, Mott's formula for the thermopower and the Wiedemann-Franz law.
Physical Review B 21, 4223 (1980).
20. G. J. Snyder, E. S. Toberer, in Materials For Sustainable Energy: A Collection of Peer-Reviewed
Research and Review Articles from Nature Publishing Group. (World Scientific, 2011), pp. 101-
110.
21. S. D. Kang, G. J. Snyder, Transport property analysis method for thermoelectric materials:
material quality factor and the effective mass model. arXiv preprint arXiv:1710.06896, (2017).
22. R. Chasmar, R. Stratton, The thermoelectric figure of merit and its relation to thermoelectric
generators. International journal of electronics 7, 52-72 (1959).
23. W. Wunderlich, H. Ohta, K. Koumoto, Enhanced effective mass in doped SrTiO3 and related
perovskites. Physica B: Condensed Matter 404, 2202-2212 (2009).
24. J. Mao et al., Manipulation of ionized impurity scattering for achieving high thermoelectric
performance in n-type Mg3Sb2-based materials. Proceedings of the National Academy of
Sciences 114, 10548-10553 (2017).
25. M. Cagnoni, D. Führen, M. Wuttig, Thermoelectric Performance of IV–VI Compounds with
Octahedral‐Like Coordination: A Chemical‐Bonding Perspective. Advanced Materials 30,
1801787 (2018).
26. Z. M. Gibbs et al., Effective mass and Fermi surface complexity factor from ab initio band
structure calculations. npj Computational Materials 3, 8 (2017).
27. Y. I. Ravich, B. Efimova, V. Tamarchenko, Scattering of current carriers and transport phenomena
in lead chalcogenides. physica status solidi (b) 43, 11-33 (1971).
28. I. I. Ravich, Semiconducting lead chalcogenides. (Springer Science & Business Media, 2013), vol.
5.
29. S. Ahmad, S. Mahanti, Energy and temperature dependence of relaxation time and Wiedemann-
Franz law on PbTe. Physical Review B 81, 165203 (2010).
30. J. Yan et al., Material descriptors for predicting thermoelectric performance. Energy &
Environmental Science 8, 983-994 (2015).
31. S. Poncé, E. R. Margine, C. Verdi, F. Giustino, EPW: Electron–phonon coupling, transport and
superconducting properties using maximally localized Wannier functions. Computer Physics
Communications 209, 116-133 (2016).
32. J. P. Heremans et al., Enhancement of thermoelectric efficiency in PbTe by distortion of the
electronic density of states. Science 321, 554-557 (2008).
33. T. Zhu, L. Hu, X. Zhao, J. He, New Insights into Intrinsic Point Defects in V2VI3 Thermoelectric
Materials. Advanced Science 3, 1600004 (2016).
34. J.-M. Zhang et al., Stability, electronic, and magnetic properties of the magnetically doped
topological insulators Bi 2 Se 3, Bi 2 Te 3, and Sb 2 Te 3. Physical Review B 88, 235131 (2013).
35. C. B. Satterthwaite, R. W. Ure, Electrical and Thermal Properties of ${\mathrm{Bi}}_{2}$ ${\
mathrm{Te}}_{3}$. Physical Review 108, 1164-1170 (1957).
36. D. Long, J. Myers, Ionized-Impurity Scattering Mobility of Electrons in Silicon. Physical Review
115, 1107-1118 (1959).
37. D. Chattopadhyay, H. J. Queisser, Electron scattering by ionized impurities in semiconductors.
Reviews of Modern Physics 53, 745-768 (1981).
38. H. Brooks, in Advances in Electronics and Electron Physics, L. Marton, Ed. (Academic Press,
1955), vol. 7, pp. 85-182.
39. P. P. Debye, E. M. Conwell, Electrical Properties of $N$-Type Germanium. Physical Review 93,
693-706 (1954).
40. R. Hartman, Temperature Dependence of the Low-Field Galvanomagnetic Coefficients of
Bismuth. Physical Review 181, 1070-1086 (1969).
41. A. Zevalkink et al., A practical field guide to thermoelectrics: Fundamentals, synthesis, and
characterization. Applied Physics Reviews 5, 021303 (2018).
42. J. Mao et al., Manipulation of ionized impurity scattering for achieving high thermoelectric
performance in n-type Mg<sub>3</sub>Sb<sub>2</sub>-based materials. Proceedings of the
National Academy of Sciences, 201711725 (2017).

Acknowledgements

We would like to acknowledge Dr. Pawan Kumar for his contribution in scientific discussion
about the thermoelectric forward model and its experimental verification. We would also like to
acknowledge the contribution of Amol Paranjape and Marcella Marissa in the data collection
process.

You might also like