Origin of anomalous temperature dependence of Nernst effect in narrow-gap semiconductors

Ryota Masuki,1, ∗ Takuya Nomoto,1 and Ryotaro Arita1, 2 1Department of Applied Physics, The University of Tokyo,7-3-1 Hongo, Bunkyo-ku, Tokyo 113-8656 2RIKEN Center for Emergent Matter Science, 2-1 Hirosawa, Wako, Saitama 351-0198, Japan (Dated: November 5, 2020) Based on the Boltzmann transport theory, we study the origin of the anomalous temperature de- pendence of the Nernst coefficient (ν) due to the phonon-drag mechanism. For narrow-gap semicon- ductors, we find that there are two characteristic temperatures at which a noticeable peak structure appears in ν. Contrarily, the Seebeck coefficient (S) always has only one peak. While the breakdown of the Sondheimer cancellation due to the momentum-dependence of the electron relaxation time is essential for the peak in ν at low T , the contribution of the valence band to the phonon-drag current is essential for the peak at higher T . By considering this mechanism, we successfully reproduce ν and S of FeSb2 for which a gigantic phonon-drag effect is observed experimentally.

Introduction. The thermoelectric effect has been ex- In this Letter, we investigate another mechanism of tensively studied for various materials due to its versatile anomalous T dependence of the Nernst effect in narrow- potential applications such as power generation, energy gap semiconductors. We first derive an electron-phonon conversion, and temperature sensing [1–3]. While the coupled quantum Boltzmann equation based on the longitudinal Seebeck effect has been exploited in many Keldysh formalism and calculate the Seebeck (S) and thermoelectric devices, those using the transverse Nernst Nernst coefficient (ν) [26–28]. In the approximation effect are of great interest since they have many advan- with a constant electron relaxation time (the so-called tages: One can design flexible structures covering a heat constant-τ approximation), ν vanishes at low T due to source with a scalable generation of a large thermopower the Sondheimer cancellation. However, if we take ac- and high energy conversion efficiency [4–9]. count of the momentum dependence of τ, a noticeable peak structure appears. On top of that, if the bandgap The thermoelectric effect is usually governed by the ε is sufficiently small, another peak originating from the diffusion of electrons and becomes monotonically weak g valence band appears at higher T . This result suggests below the room temperature [10]. However, in some sit- that when τ has momentum dependence in a narrow-gap uations, there can be another significant contribution at semiconductor, ν(T ) will have a characteristic double- low temperature (T ), which is called the phonon-drag peak structure. On the other hand, we find that S(T ) effect. In the presence of strong electron-phonon interac- always has one featureless peak, even if we go beyond tion, a momentum transfer from nonequilibrium phonons the constant-τ approximation. to electrons occurs, and the thermopower can be dramat- ically enhanced [11]. The phonon-drag effect is partic- We then examine whether this mechanism plays a cru- ularly pronounced in some semiconductors with a long cial role in the gigantic phonon-drag effect in FeSb2. phonon lifetime, and is regarded as a promising mech- FeSb2 is a correlated narrow-gap semiconductor with a anism to make high-performance thermoelectrics below large effective mass [3, 24, 29] and exhibits a remark- the room temperature [12]. ably large Seebeck and Nernst effect at cryogenic tem- peratures [24, 30–33]. The maximum value of |S| and The quantitative description of phonon drag is a fasci- |ν| reaches 45 mV/K at 10K [31] and 3.2 mV/(KT) at nating problem to explore, having a long history [13–19]. 7K [24], respectively. While electron-correlation in FeSb Regarding the Seebeck effect, intensive studies have been 2 is considerably strong [23, 24, 29, 34–37], it has been performed for several semiconductors in a wide range of recently shown that phonon drag is responsible for the T and carrier concentrations [12, 20–22]. On the other colossal thermopower in this compound [22, 25, 29, 33]. hand, it is not fully understood in which materials and

arXiv:2010.11001v2 [cond-mat.mtrl-sci] 4 Nov 2020 in which conditions the phonon-drag contribution par- For S(T ), we show that the present Boltzmann ap- ticularly enhances the Nernst coefficient. One represen- proach gives a result consistent with that obtained by tative example is FeSb2, for which a huge phonon-drag an elaborate microscopic calculation based on linear re- effect is observed. Interestingly, it has been shown that sponse theory [22, 38]. S(T ) has a single peak, regard- its Nernst effect exhibits a characteristic T dependence less of whether or not the momentum dependence of τ with multiple peaks, which are absent in the Seebeck co- is taken into account. On the other hand, we find that efficient [23, 24]. Recently, it has been proposed that this ν(T ) has a double peak structure when we go beyond distinct difference in the longitudinal and transverse ther- the constant-τ approximation. The result shows a good moelectric effect can be understood by the phonon-drag agreement with the experiment [24]. Our result indi- coupling to multiple in-gap states [25]. cates the importance of the momentum dependence of 2

τ in narrow-gap semiconductors, which will be a useful The transverse component of the linear response coeffi- guiding principle to control the phonon-drag effect and cient L12,yx is calculated as design efficient thermoelectric devices. Z 3 Method. Starting from the Keldysh formalism [27], we gs d k Nernst L12,yx = 3 (−evky)[δfqp(k)]ph-drag, derive the Boltzmann transport equation that includes −(∇xT )/T (2π) the effect of impurity state up to linear order of impurity where the linear response tensor L12 is defined by the concentration. The numerical calculation shows that the relation j = σE + L12(−∇T )/T . Then, the Nernst coef- effect of impurity state on S and ν is negligible within our ficient is written as approximation. Hence, we start from the conventional 1  σ L  Boltzmann transport equation that disregards the effect ν = S − yx + 12,yx , (1) of the impurity state to simplify the discussion here. De- B σxx L12,xx tails of the treatment that considers the effect of impurity using the linear response coefficients in the weak field state is explained in the supplemental materials [39]. limit. L ,L can be calculated from the above dis- The phonon-drag effect on the thermoelectric effect is 12,yx 12,xx cussion and σxx, σyx can be calculated within the frame- calculated from the Boltzmann transport equation in a work of the conventional Boltzmann transport equa- method similar to that by Cantrell and Butcher [40]. In tion [41]. the following, we show a result for a simple case where For typical semiconductors, we consider one conduc- one electron band and one phonon mode are involved. We tion band, one valence band, and one longitudinal acous- neglect the effect of the nonequilibrium electron distribu- tic (LA) mode. We assume that the conduction and tion on the phonon distribution function. The change of the valence band have isotropic quadratic dispersions, the electron distribution function due to the phonon-drag and the LA phonon has an isotropic linear dispersion effect is 2 2 ~Ed with phonon velocity cL. We use |gq| = q for the Z 3 0 3 2ρcL Seebeck d k d q τph ωq ~ electron-phonon coupling constant, where Ed is the defor- [δfqp(k)]ph-drag = −τel,k 3 3 2 (2π) (2π) kBT mation potential and ρ is the mass density. It is possible q q ×(∇qωq) · (∇T )(Pkk0 − Pk0k), to analytically carry out all the angular integration un- where der these assumptions and rewrite L12,xx and L12,yx by single integrals as q 2π 2 eq eq 0 eq Pkk0 = 2 |gq| (1 − fqp(k))fqp(k )Nph(q) Z ∞ ~ cond e~ 3 L12,xx = − 2 dkk τel,c,kFc,ph(k), (2) 1 3 0 3π mc 0 ×δ(ωq − (εk − εk0 ))(2π) δ(k − k − q) ~ e2 Z ∞ Lcond = −B ~ dkk3τ 2 F (k), (3) is the transition amplitude of scattering from momen- 12,yx z 3π2m2 el,c,k c,ph 0 c 0 tum k to k mediated by phonon with momentum q. fqp for the conduction band contribution, where mc and and Nph are the distribution function of electrons at the τel,c,k are the effective mass and the electron relaxation quasiparticle peak and of the phonons respectively. τel,k is the relaxation time of electron excitation of momen- time for the conduction band, respectively. The subscript c stands for the conduction band. Fc,ph(k) is defined as tum k. τph is the phonon lifetime. εk and ωq repre- sent the dispersions of the electron band and the phonon Seebeck [δfqp,c(k = (k, 0, 0))]ph-drag mode respectively. gq is the electron-phonon coupling Fc,ph(k) = . (4) Seebeck τel,c,k(−∇xT )/T constant. Using this expression for [δfqp(k)]ph-drag, the phonon-drag contribution to the Seebeck coefficient can where δfqp,c(k = (k, 0, 0)) is the change of electron dis- be calculated as tribution function in the conduction band evaluated at Z 3 Seebeck momentum (k , k , k ) = (k, 0, 0). The valence band con- gs d k [δfqp(k)] x y z S = − (−ev ) ph-drag , val val ph-drag σ (2π)3 kx (∇ T ) tribution to L12 tensor (L12,xx, L12,yx) can be written in x a similar expression [39]. The electric conductivity tensor where σ is the electric conductivity and gs(= 2) is the can be calculated as spin degeneracy. Note that hereafter the temperature Z ∞ cond e~ 3 gradient is assumed to be in the x direction. In order σ = − dkk τ F (k), (5) xx 3π2m el,c,k c,E to calculate the Nernst coefficient, we consider the case c 0 2 Z ∞ where external magnetic field is applied in the z direc- cond e ~ 3 2 σyx = −Bz 2 2 dkk τel,c,kFc,E(k), (6) tion and retain the terms that are first order both in 3π mc 0 ∇T and in B. The change of distribution function that where Fc,E(k) is the electric field correspondent of contributes to the phonon-drag effect on the Nernst co- Fc,ph(k), which is defined as efficient is  eq  1 [δfqp,c((k, 0, 0))]E ∂fqp,c Nernst Seebeck Fc,E(k) = = evc,k . (7) [δfqp(k)]ph-drag = τel,k e(vk × B) · ∇k[δfqp(k)]ph-drag. E τ ∂ε ~ x c,k c,k 3

[δfqp,c]E is the change of the distribution function in- duced by the external electric field. The valence band 0 0 val (a) ε = 1.0 eV (b) ε = 0.03 eV term of the electric conductivity tensor σ can be cal- -0.2 g -0.2 g culated in an analogous way [39]. -0.4 -0.4 The chemical potential is determined from the con-

[mV/K] -0.6 -0.6 dition of charge neutrality, where the bandgap εg and S -0.8 -0.8 the donor binding energy εb play a crucial role. Note that the electron concentration in the impurity state is 0.03 1 0.2 nd× β(ε −ε −µ) , which is apparently different from e c,k=0 b /2+1 0.02 0.1 the Fermi-Dirac statistics because spin up and spin down 0.01 [mV/(KT)]

ν electron cannot occupy the same impurity site at the - 0 0 same time due to the strong Coulomb repulsion [41]. 0 10 20 30 40 50 60 0 10 20 30 40 50 60 We consider the impurity scattering and the phonon T [K] T [K] scattering process to determine 1/τel,c,k = 1/τc,imp,k + 1/τc,ph,k. For the impurity scattering, we adopt the Brooks-Herring model [42], in which a Yukawa-type FIG. 1. Phonon-drag contribution to the Seebeck coefficient S and the Nernst coefficient ν for (a) a moderate-gap semi- screened Coulomb potential is employed to represent the conductor and (b) a narrow-gap semiconductor. We set the impurity potential. bandgap εg and the donor binding energy εb as (a) εg = 1.0 eV and ε = 0.05 eV and (b) ε = 30 meV and ε = 5 meV. Z 2 b g b 1 4πεbnd xdx For other parameters, see the text. = 3 2 2 2 . τc,imp,k ~k 0 (x + qD/(2k ))

2 2 nde where nd is the donor concentration and q = . D kB T The dielectric constant  is calculated from ε and m as than εg, we can neglect the contribution of the valence q b c 13.6 eV me band and Eq. (1) is simplified as  = 0 × . The electron scattering rate by the εb mc phonon can be calculated from the Boltzmann transport cond val cond val ! S σyx + σyx L12,yx + L12,yx equation as ν = − cond val + cond val (8) Bz σxx + σxx L12,xx + L12,xx −1 1  ∂f eq  Z d3k0 d3q cond cond ! qp,c q q S σyx L12,yx = −kBT 3 3 (Pc,kk0 + Pc,k0k). ' − + . (9) τc,ph,k ∂εk (2π) (2π) cond cond Bz σxx L12,xx q Here, Pc,kk0 is the phonon mediated transition amplitude It should be noted that we can show that in the conduction band. We assume that the hole relax- cond cond ation time for the valence band is equal to that of the σ /Bz L /Bz eτ yx = 12,yx = el,c conduction band. cond cond σxx L12,xx mc Results and Discussion. Let us start with model cal- culations for the following two representative cases: A in the constant-τ approximation, so that the first and sec- semiconductor with a moderate bandgap εg = 1.0 eV ond term in Eq. (9) cancel with each other (the so-called and a narrow bandgap εg = 30 meV. For εb, we use 50 Sondheimer cancellation) and eventually ν becomes neg- meV for the former and 5 meV for the latter. The other ligibly small. parameters are the same for both cases. Namely, we set However, if we consider the momentum dependence of −8 −3 nd = 1.0 × 10 A˚ , cL = 5000 m/s, Ed = 1.0 eV, and τel,c,k, this cancellation does not happen and ν(T ) can 3 ρ = 5.0 g/cm . We assume that the valence and conduc- be finite even at low T . In Fig. 2, we show 1/τel,c,k, tion band have the same effective mass (=me). We set Fc,ph(k) and Fc,E(k) as a function of k(= |k|). We see −7 −T/T∗ τph(T ) = 1.0 × 10 × 10 s, where T∗ = 20 K. The that the peak of Fc,ph(k) extends to k larger than that parameters are in the same order as the fitting curve of of Fc,E(k), for which τel,c,k is longer. In this situation, the previous calculation result [22]. the second term in Eq. (9) dominates over the first term In Fig. 1, we show S(T ) and ν(T ) for the moderate- (see Eqs. (2), (3), (5), (6)). Therefore, the momentum gap case (a) and narrow-gap case (b). For S(T ), we see dependence of τel,c,k is crucial for the formation of the that there is only one peak around T = 10 K for both peak in ν(T ) around 10 K. (a) and (b). On the other hand, while ν(T ) has only one Let us now move on to the case of higher T . As peak for (a) but two peaks for (b). Namely, on top of is discussed in the supplemental materials [39], when the peak around T = 10 K, one additional peak appears εg = 5 meV and T ∼ 30 K, the contribution of the valence around 40 K. band to the linear response coefficient (i.e., σval and Lval) The origin of these two peaks in ν(T ) can be under- becomes comparable to that of the conduction band (i.e., cond cond cond stood in terms of Eq. (1). When T is sufficiently lower σ and L ). As we can see from Table I, σyx and 4

14 2 ] -1 10 (a) impurity scatt. 10 (a) [s 0 phonon scatt. 12 10 10 total [ Ω cm] ρ -2 el, c , k 10 ]

10 -1 -2 1/τ 10 C 10 3 (b) 10-4 [m -6 | ]

H 10 6 (b) 160 -8 | R 10 120

[m/s 0 C/(N s)]

4 4 80 -10 (c) (k) [10 2 40 [mV/K] -20 c,ph S ( k ) -30

0 F c,E 0 4 calculation 0 0.02 0.04 0.06 0.08 0.1 experimental data (d) − F −1 k [Å ] 3 2 FIG. 2. Momentum-dependence of (a) electron scattering rate and (b) F (k) and F (k) (see Eqs. (4) and (7) in the [mV/(KT)] 1 c,E c,ph - ν text) for the narrow-gap semiconductor in Fig. 1(b) at 8 K. 0 0 10 20 30 40 50 60 T [K]

val cond val σyx have opposite signs. Contrarily, L12,yx and L12,yx FIG. 3. (a) Electric resistivity, (b) Hall coefficient, (c) See- have the same sign, so that the second term in the r.h.s beck coefficient, and (d) Nernst coefficient of FeSb2 compared of Eq. (8) becomes dominantly larger than the first term with the experimental result [24]. at T ∼ 30 K, and eventually, −ν(T ) becomes large. How- ever, it should be noted that ν(T ) vanishes in the limit of high T since S(T ) becomes negligibly small. Thus ν(T ) has a peak at intermediate T ∼ 30 K. This mechanism linear response theory [22]. The electric resistivity and 4 the Hall coefficient reach a plateau around T = 10 K to 20 does not work when εg is larger than 1 eV ∼ 10 K, so that the double peak structure in ν(T ) appears only in K. In this temperature range, a large fraction of the elec- narrow-gap semiconductors. trons in the impurity state is excited to the conduction band while the valence band is still almost completely occupied. Above this plateau, the carrier concentration TABLE I. Signs of the conduction band and the valence band increases due to excitation from the valence to the con- contribution to the linear response coefficients. duction band, which is consistent with the scenario for cond val cond val σxx σxx σyx /Bz σyx /Bz the anomalous T dependence of ν(T ) in narrow-gap semi- + + + − conductors. These results clarify that the origin of the cond val cond val L12,xx L12,xx L12,yx/Bz L12,yx/Bz huge phonon-drag contribution to ν of FeSb2 at low tem- − + − − peratures is the momentum dependence of τel,k and the small εg. Finally, let us consider the case of FeSb2. We show the Conclusion. We investigate the phonon-drag contri- result for S(T ) and ν(T ) in Fig. 3. In the calculation, the bution to the Seebeck and the Nernst coefficient using effective mass of the conduction and the valence band are the Boltzmann transport equation. We identify that the set to mc = mv = 5me [22, 33]. The bandgap, impurity momentum-dependent electron relaxation time is crucial binding energy are set to εg = 28 meV, εb = 6 meV, re- to the large negative Nernst coefficient at low temper- −8 spectively [24]. We set cL = 3100 m/s, nd = 1.9 × 10 ature, which breaks the Sondheimer cancellation. Fur- −3 A˚ . The donor concentration nd is determined so that thermore, we find that the Nernst coefficient has another the carrier concentration at low T be the same order as peak at higher temperature if the bandgap is sufficiently the experiment [24]. The deformation potential is set to small. Considering this mechanism, we calculate the elec- Ed = 0.85 eV, which is a typical value for semiconduc- tric resistivity, the Hall coefficient, the Seebeck and the 3 tors. The mass density is set to ρ = 8.2 g/cm . The Nernst coefficient of FeSb2, and succeed in reproducing phonon lifetime τph(T ) is determined to reproduce the the experimental result. Our results propose another pos- experimental result of the thermal conductivity [22]. sible origin of the anomalous temperature dependence of We see that the result is in good agreement with the the Nernst coefficient of FeSb2, which provides a pos- experiment [24]. Note that the result for S(T ) is also sibility to utilize the phonon-drag effect to design good consistent with the previous result based on microscopic thermoelectric materials. 5

Acknowledgements. We would like to thank J. M. Tom- [14] H. P. R. Frederikse, Thermoelectric power of germanium czak for his fruitful comments. This work was supported below room temperature, Phys. Rev. 92, 248 (1953). by a Grant-in-Aid for Scientific Research (No. 19K14654, [15] T. H. Geballe and G. W. Hull, Seebeck effect in germa- No. 19H05825, and No. 16H06345) from Ministry of Ed- nium, Phys. Rev. 94, 1134 (1954). [16] T. H. Geballe and G. W. Hull, Seebeck effect in silicon, ucation, Culture, Sports, Science and Technology, and Phys. Rev. 98, 940 (1955). CREST (JPMJCR18T3) from the Japan Science and [17] C. Herring, T. H. Geballe, and J. E. Kunzler, Phonon- Technology Agency. drag thermomagnetic effects in n-type germanium. I. gen- eral survey, Phys. Rev. 111, 36 (1958). [18] W. Dunstan and D. W. Sear, The Nernst-Ettingshausen coefficient in n-type silicon, Journal of Physics C: Solid State Physics 7, 157 (1974). ∗ [email protected] [19] Y. Gurevich and O. Mashkevich, The electron-phonon [1] X. Shi, L. Chen, and C. Uher, Recent advances drag and transport phenomena in semiconductors, in high-performance bulk thermoelectric materials, Physics Reports 181, 327 (1989). International Materials Reviews 61, 379 (2016), [20] G. D. Mahan, L. Lindsay, and D. A. Broido, https://doi.org/10.1080/09506608.2016.1183075. The Seebeck coefficient and phonon drag in sili- [2] H. Alam and S. Ramakrishna, A review on the enhance- con, Journal of Applied Physics 116, 245102 (2014), ment of figure of merit from bulk to nano-thermoelectric https://doi.org/10.1063/1.4904925. materials, Nano Energy 2, 190 (2013). [21] M. Fiorentini and N. Bonini, Thermoelectric coefficients [3] J. M. Tomczak, Thermoelectricity in correlated narrow- of n-doped silicon from first principles via the solution gap semiconductors, Journal of Physics: Condensed Mat- of the boltzmann transport equation, Phys. Rev. B 94, ter 30, 183001 (2018). 085204 (2016). [4] M. H. Norwood, Theory of Nernst generators and re- [22] H. Matsuura, H. Maebashi, M. Ogata, and H. Fukuyama, frigerators, Journal of Applied Physics 34, 594 (1963), Effect of phonon drag on Seebeck coefficient based on https://doi.org/10.1063/1.1729315. linear response theory: Application to FeSb2, Journal [5] Z. Yang, E. A. Codecido, J. Marquez, Y. Zheng, J. P. of the Physical Society of Japan 88, 074601 (2019), Heremans, and R. C. Myers, Scalable Nernst thermoelec- https://doi.org/10.7566/JPSJ.88.074601. tric power using a coiled galfenol wire, AIP Advances 7, [23] P. Sun, N. Oeschler, S. Johnsen, B. B. Iversen, and 095017 (2017), https://doi.org/10.1063/1.5003611. F. Steglich, FeSb2: Prototype of huge electron-diffusion [6] Y. Sakuraba, Potential of thermoelectric power genera- thermoelectricity, Phys. Rev. B 79, 153308 (2009). tion using anomalous Nernst effect in magnetic materials, [24] P. Sun, W. Xu, J. M. Tomczak, G. Kotliar, Scripta Materialia 111, 29 (2016), viewpoint Set No. 57: M. Søndergaard, B. B. Iversen, and F. Steglich, Highly Contemporary Innovations for Thermoelectrics Research dispersive electron relaxation and colossal thermoelec- and Development. tricity in the correlated semiconductor FeSb2, Phys. Rev. [7] M. Ikhlas, T. Tomita, T. Koretsune, M.-T. Suzuki, B 88, 245203 (2013). D. Nishio-Hamane, R. Arita, Y. Otani, and S. Nakatsuji, [25] M. Battiato, J. M. Tomczak, Z. Zhong, and K. Held, Large anomalous Nernst effect at room temperature in a Unified picture for the colossal thermopower compound chiral antiferromagnet, Nature Physics 13, 1085 (2017). FeSb2, Phys. Rev. Lett. 114, 236603 (2015). [8] A. Sakai, S. Minami, T. Koretsune, T. Chen, T. Higo, [26] L. Keldysh, Diagram technique for nonequilibrium pro- Y. Wang, T. Nomoto, M. Hirayama, S. Miwa, D. Nishio- cesses, Zh. Eksp. Teor. Fiz. 47, 1515 (1964). Hamane, F. Ishii, R. Arita, and S. Nakatsuji, Iron-based [27] R. E. Prange and L. P. Kadanoff, Transport theory for binary ferromagnets for transverse thermoelectric conver- electron-phonon interactions in metals, Phys. Rev. 134, sion, Nature 581, 53 (2020). A566 (1964). [9] A. Sakai, Y. P. Mizuta, A. A. Nugroho, R. Sihomb- [28] J. M. Luttinger, Theory of thermal transport coefficients, ing, T. Koretsune, M.-T. Suzuki, N. Takemori, R. Ishii, Phys. Rev. 135, A1505 (1964). D. Nishio-Hamane, R. Arita, P. Goswami, and S. Nakat- [29] J. M. Tomczak, K. Haule, T. Miyake, A. Georges, and suji, Giant anomalous Nernst effect and quantum-critical G. Kotliar, Thermopower of correlated semiconductors: scaling in a ferromagnetic semimetal, Nature Physics 14, Application to FeAs2 and FeSb2, Phys. Rev. B 82, 085104 1119 (2018). (2010). [10] J. He, M. G. Kanatzidis, and V. P. Dravid, High perfor- [30] P. Sun, N. Oeschler, S. Johnsen, B. B. Iversen, and mance bulk thermoelectrics via a panoscopic approach, F. Steglich, Huge thermoelectric power factor: FeSb2 ver- Materials Today 16, 166 (2013). sus FeAs2 and RuSb2, Applied Physics Express 2, 091102 [11] L. Gurevich, Thermoelectric properties of conductors, J (2009). Phys (Moscow) 9, 477 (1945). [31] A. Bentien, S. Johnsen, G. K. H. Madsen, B. B. Iversen, [12] J. Zhou, B. Liao, B. Qiu, S. Huberman, K. Esfar- and F. Steglich, Colossal Seebeck coefficient in strongly jani, M. S. Dresselhaus, and G. Chen, Ab initio op- correlated semiconductor FeSb2, EPL (Europhysics Let- timization of phonon drag effect for lower-temperature ters) 80, 17008 (2007). thermoelectric energy conversion, Proceedings of the [32] P. Sun, N. Oeschler, S. Johnsen, B. B. Iversen, and National Academy of Sciences 112, 14777 (2015), F. Steglich, Narrow band gap and enhanced thermoelec- https://www.pnas.org/content/112/48/14777.full.pdf. tricity in FeSb2, Dalton Trans. 39, 1012 (2010). [13] C. Herring, Theory of the thermoelectric power of semi- [33] H. Takahashi, R. Okazaki, S. Ishiwata, H. Taniguchi, conductors, Phys. Rev. 96, 1163 (1954). A. Okutani, M. Hagiwara, and I. Terasaki, Colossal See- beck effect enhanced by quasi-ballistic phonons dragging 6

massive electrons in FeSb2, Nature communications 7, 1 An ARPES and ab initio study, Phys. Rev. Research 2, (2016). 023190 (2020). [34] Q. Jie, R. Hu, E. Bozin, A. Llobet, I. Zaliznyak, C. Petro- [38] M. Ogata and H. Fukuyama, Range of validity of vic, and Q. Li, Electronic thermoelectric power factor Sommerfeld–Bethe relation associated with Seebeck co- and metal-insulator transition in FeSb2, Phys. Rev. B efficient and phonon drag contribution, Journal of 86, 115121 (2012). the Physical Society of Japan 88, 074703 (2019), [35] C. Petrovic, Y. Lee, T. Vogt, N. D. Lazarov, S. L. https://doi.org/10.7566/JPSJ.88.074703. Bud’ko, and P. C. Canfield, Kondo insulator description [39] Supplemental materials. of spin state transition in FeSb2, Phys. Rev. B 72, 045103 [40] D. G. Cantrell and P. N. Butcher, A calculation of the (2005). phonon-drag contribution to the thermopower of quasi- [36] H. Takahashi, R. Okazaki, Y. Yasui, and I. Terasaki, Low- 2d electrons coupled to 3d phonons. I. general theory, temperature magnetotransport of the narrow-gap semi- Journal of Physics C: Solid State Physics 20, 1985 (1987). conductor FeSb2, Phys. Rev. B 84, 205215 (2011). [41] N. W. Ashcroft and N. D. Mermin, Solid State Physics [37] A. Chikina, J.-Z. Ma, W. H. Brito, S. Choi, P. S´emon, (Holt-Saunders, 1976). A. Kutepov, Q. Du, J. Jandke, H. Liu, N. C. Plumb, [42] D. Chattopadhyay and H. J. Queisser, Electron scatter- M. Shi, C. Petrovic, M. Radovic, and G. Kotliar, Corre- ing by ionized impurities in semiconductors, Rev. Mod. lated electronic structure of colossal thermopower FeSb2: Phys. 53, 745 (1981).