WorksPhysics and Astronomy
Machine learning for predicting ultralow thermal conductivity and high ZT in complex thermoelectric materials
Abstract
Efficient and precise calculations of thermal transport properties and figure of merit, alongside a deep comprehension of thermal transport mechanisms, are essential for the practical utilization of advanced thermoelectric materials. In this study, we explore the microscopic processes governing thermal transport in the distinguished crystalline material TlSbTe by integrating a unified thermal transport theory with machine learning-assisted self-consistent phonon calculations. Leveraging machine learning potentials, we expedite the analysis of phonon energy shifts, higher-order scattering mechanisms, and thermal conductivity arising from various contributing factors like population and coherence channels. Our finding unveils an exceptionally low thermal conductivity of 0.31 W m K at room temperature, a result that closely correlates with experimental observations. Notably, we observe that the off-diagonal terms of heat flux operators play a significant role in shaping the overall lattice thermal conductivity of TlSbTe, where the ultralow thermal conductivity resembles that of glass due to limited group velocities. Furthermore, we achieve a maximum value of 3.17 in the -axis orientation for p-type TlSbTe at 600 K, and an optimal value of 2.26 in the -axis and -axis direction for n-type TlSbTe at 500 K. The crystalline TlSbTe not only showcases remarkable thermal insulation but also demonstrates impressive electrical properties owing to the dual-degeneracy phenomenon within its valence band. These results not only elucidate the underlying reasons for the exceptional thermoelectric performance of TlSbTe but also suggest potential avenues for further experimental exploration.
Introduction
Thermoelectric materials have the capacity to transform heat into electrical energy without emitting pollutants [1, 2]. They lack movable components, generate no noise pollution, and are adjustable in size, rendering them indispensable in a multitude of fields, including waste heat recovery [3] and photovoltaic-photothermal cogeneration [4]. The conversion efficiency of thermoelectric materials can be determined by the figure of merit , where , , and are the Seebeck coefficient, electrical conductivity, electrical thermal conductivity and lattice thermal conductivity, respectively. In order to achieve optimal thermoelectric performance, a promising strategy involves enhancing electrical transport while simultaneously suppressing thermal transport [5]. Nevertheless, it is challenging to achieve high electrical conductivity () and low electronic thermal conductivity () simultaneously. Consequently, ultralow lattice thermal conductivity () becomes a crucial factor contributing to high thermoelectric efficiency.
Recent studies have identified several materials with relatively complex structures, including CuSbS [6], AgPbBiSe [7], CsBiAgBr [8], KCuSe [9], and CuPS [10]. These materials exhibit potential for thermoelectric applications due to their ability to achieve ultralow thermal conductivity (). Therefore, research on complex thermoelectric materials is not only of fundamental interest but also of crucial importance for the advancement of thermoelectric technology.
The Tl-Te system contains numerous complex materials with ultralow . For instance, the of AgTlTe, AgTlTe, TlCuTe, TlGeTe, TlInTe, are all lower than 0.5 W m K at room temperature [11, 12]. Moreover, the of the promising thermoelectric materials TlBiTe and TlSbTe was reported to be 0.5 W m K and 0.7 W m K, respectively, at room temperature [11]. Although experiments have shown that TlBiTe exhibits lower thermal conductivity, it also exhibits high resistivity, which significantly reduces the power factor and constrains its further enhancement as an additional component. Even with minimal TlBiTe incorporation, it can markedly diminish the thermoelectric performance of the material [13]. Compound TlSbTe represents the most closely related material to within the family. Its thermal conductivity is comparable to that of , yet its electrical conductivity is markedly superior, as evidenced by the experimental results [11]. Although there have been experimental studies on thermoelectric properties [14, 15] [11] and analysis of electronic structure [16], there is still a lack of in-depth research on its low and high thermoelectric efficiency due to its complex structure. Further comprehensive research will help us better understand and design excellent thermoelectric materials based on the complex Tl-Te system.
Despite the significant advancements in measuring ultralow in experiments, theoretical prediction and explanation of ultralow in complex compounds still remain challenging. Recent studies have explored anharmonic effects as potential sources of discrepancies of between the experiment and theory observed in highly anharmonic crystals [17, 19, 20, 21, 22] [6, 18]. These studies have revealed that lattice anharmonicity significantly influences the phonon linewidth and lattice thermal conductivity [23]. While the conventional Peierls-Boltzmann framework is adequate for describing phonon transport in systems with well-defined phonon modes, it encounters challenges in highly anharmonic systems [23, 24]. Therefore, it becomes imperative to consider the temperature-dependent frequencies evaluated by the self-consistent phonon (SCPH) approach in elucidating thermal transport in complex compounds [25, 26].
Moreover, a discrepancy exists between the experimentally measured and theoretically predicted by the conventional Peierls-Boltzmann transport equation [27]. Consequently, it becomes necessary to consider the influence of wavelike interbranch tunneling of coherence, originating from off-diagonal terms of the heat flux operator [27], to comprehensively explain this phenomenon. Overall, to accurately predict thermal transport properties in complex thermoelectric materials, the higher-order force constants, i.e., fourth-order force constants, and coherent phonons are required to explicitly evaluate the phonon energy shifts and scattering rates. Overall, to accurately predict thermal transport properties in complex thermoelectric materials, it is necessary to explicitly evaluate the phonon energy shifts and scattering rates using the higher-order force constants i.e., fourth-order force constants, and coherent phonons.
Despite the significant achievements in DFT-based thermal conductivity calculations [28], the computational expense of calculating thermal conductivity considering higher-order phonon scatterings is so high that it is often prohibitive [29, 30]. The use of machine Learning Potential (MLP) can significantly reduce the cost of calculating force constants while maintaining considerable computational accuracy [31]. Machine learning potential (MLP) is a potential that employs machine learning algorithms to describe the atomic potential energy surface and predict the energy and force based on the atomic environment.
Unlike machine learning models (ML-Models) [32], MLP can be combined with BTE methods to provide detailed phonon information, thereby enabling a more comprehensive understanding of the mechanism of thermal transport in the large system. Furthermore, it is capable of calculating complex systems under a multitude of conditions, including those involving defects, disorder, anharmonicity, and other complex systems that are challenging to calculate using DFT methods.
Shapeev et al. developed another machine learning potential function, the moment tensor potential (MTP) [33, 34], which achieves a good balance between computational accuracy and computational complexity. MTP can obtain high-precision second-order and third-order force constants. According to the previous research, MTP can also obtain reliable fourth-order force constants, which can accurately calculate the containing four-phonon scattering [35]. Consequently, based on the fourth-order anharmonicity [29], we have modified and developed corresponding interface using the MTP to obtain fourth-order force constants in an efficient manner.
In this study, we conducted a comprehensive investigation of the thermal transport and thermoelectric properties of the benchmark compound . To this end, we integrated machine-learning based anharmonic lattice dynamics with an unified theory of thermal transport and momentum relaxation time approximation. This integration included: (1) Construction of a machine-learning potential with DFT-level accuracy for the complex crystalline , (2) anharmonically renormalizing phonons using the SCPH approach, (3) evaluation of thermal transport considering both particle-like and wave-like phonon transport channels, (4) application of distinct scattering mechanisms, including acoustic deformation potential (ADP), ionized impurity (IMP) scattering, and polar optical phonon (POP) scattering, to systematically study the electrical transport properties, (5) exploration of the trend of the thermoelectric property value with carrier concentration and temperature variation.
Results and discussion
Crystal structure and machine learning potential
The optimized lattice constants of , calculated using LDA potentials, are Å, and Å. These values are consistent with the experimental values, which are Å and Å [16] [11]. As illustrated in Figure 1(b), the phonon spectrum of the LDA functional is devoid of imaginary frequencies, indicating the dynamical stability of . The primitive cell comprises 16 atoms: one Sb, six Te, and nine Tl. Tl atoms are divided into eight Tl and one Tl*, based on their disparate chemical environments and locations, as depicted in Figure 1(e).

Figure 1. (a) The workflow of the thermoelectric property calculations. (b) Phonon dispersion of Tl₉SbTe₆ calculated by Moment tensor potential (MTP) and DFT. A comparisons of (c) the energy per atom (d) and the atomic force between MTP and DFT calculations for Tl₉SbTe₆. (e) The conventional cell structure of Tl₉SbTe₆. The red and purple atoms represent Thallium (Tl) atoms in different chemical environments. The atoms represented by the blue color are Tellurium (Te) atoms, while the green ones are Antimony (Sb) atoms.
Figure 1(a) depicts the workflow of our work. In this work, we obtain the training dataset from Ab Initio Molecular Dynamics (AIMD) using the isobaric-isothermal (NPT) ensemble. We then train the MTP and the training principles are detailed in the Supplementary Information [33]. After obtaining the MTP, we test it on the validation dataset.
The MTP is employed to ascertain the energy and atomic forces of each configuration in the validation dataset, with the resulting root-mean-square error (RMSE) being presented in Figure 1(c-Figure 1d). The RMSE of energy per atom and atomic force is 0.00029 eV/atom and 0.023 eV/Å, respectively. These values indicate that the trained MTP is sufficiently accurate for calculating thermodynamic properties [33, 35]. Furthermore, as shown in Figure 1(b), the phonon spectrum calculated by MTP is in good agreement with the DFT results.
Anharmonic lattice dynamics and thermal transport properties
The impact of forth-order anharmonicity on the phonon spectra at finite temperatures is next examined, as depicted in Figure 2(a-Figure 2b), using the interatomic force constants (IFCs) obtained from MLP. It is evident that overall phonon hardening phenomenon is observed in crystalline Tl₉SbTe₆ as the temperature increases from 50 to 800 K. This observation suggests the presence of strong lattice anharmonicity in Tl₉SbTe₆ and highlights the importance of properly treating lattice anharmonicity in lattice dynamics [20, 23, 36] [18]. In particular, the low-frequency phonons with a frequency below 1.5 THz are predominantly influenced by the Tl atoms, as evidenced by the projected phonon density of states (PDOS) in Figure 2(b). The pronounced phonon hardening observed in the Tl-dominated modes can be attributed to the significant atomic displacements resulting from their loose bonding, as illustrated in Fig. S1 in the supplementary information. Furthermore, it can be observed that Tl atoms contribute to the low-lying flattened phonon bands in the frequency range of 0.7 to 1.0 THz. This suggests that strong phonon scattering is occurring, which suppresses the primary heat carriers, i.e., acoustic phonon modes [38] [37, 39]. In contrast, the optical modes with higher energies exhibit less increase in stiffness corresponding to the majority of Te and Sb vibrations.

Figure 2. (a) The temperature-dependent phonon spectrum from K to 800 K for . (b) Projected phonon density of states (PDOS) of at 0 K, 300 K, and 700 K. (c) Phonon populations’ thermal conductivity () as a function of the temperature of . (d) Intrinsic lattice thermal conductivity () as a function of the temperature of . The ratio represents . The solid lines represent the -axis direction, and the dashed lines represent the -axis direction.
With the zero-K/finite-temperature IFCs at hand, we proceed to calculate the temperature-dependent using three different levels of theory, as shown in Figure 2(c-Figure 2d).
It is observed that the lattice thermal conductivity () is anisotropic, with the approximately 10% smaller in the -axis or -axis direction than in the -axis direction. The lowest level of thermal transport theory, namely the HA+3ph model, yields the following values for in the -axis at 50, 300, and 800 K: 0.88, 0.14, and 0.053 W m K, respectively. As previously stated, anharmonic phonon renormalization is a significant factor in the prediction of finite-temperature in complex compounds. With the additional effect of phonon energy shifts, we advance the HA+3ph model to a more accurate SCPH+3ph model, which gives values of 1.02, 0.24, and 0.15 W m K at 50, 300, and 800 K, respectively, in the -axis direction. These values are 16%, 71%, and 183% larger than , indicating significant anharmonic phonon renormalization in Tl9SbTe6, with a substantial increase as temperature rises. This trend is also confirmed by Figure 2(a), which shows significant phonon frequency shifts with increasing temperature.
In addition, the quartic anharmonicity not only induces significant shifts in phonon frequencies but also results in pronounced 4ph scatterings. Consequently, additional 4ph scatterings were included to obtain values of 0.96, 0.19, and 0.093 W m K at 50, 300, and 800 K, respectively, in the -axis direction (see Figure 2(c)). These values represent decreases of 6%, 21%, and 38% compared with , indicating the importance of the 4ph scattering processes [18, 39, 40]. It is noteworthy that the discrepancy between the and is 6%, 26%, and 46% greater than that between the anharmonic phonon renormalized ones. This implies that phonon renormalization and 4ph scattering induced by quartic anharmonicity exert an inverse influence on ther-
mal transport.
Nevertheless, remains considerably lower than the experimentally measured , as illustrated in Figure 2(d), indicating that coherence contributions are not negligible. By further considering the off-diagonal terms of heat flux operators, the total significantly enhanced by incorporating both and [27]. In this context, can be expressed as,
where is the reduced Planck constant, is Boltzmann constant, is the primitive cell volume, and is the total number of sampled phonon vectors.
As illustrated in Figure 2(d), the ratio is greater than 0.5 at 300 K and exhibits a pronounced increase with temperature. Specifically, in the -axis direction, the ratio of reaches 1.60 at 800 K, which is 2.44 times larger than the 0.65 observed at 300 K. This observation suggests that is the dominant contribution to in crystalline TlSbTe at temperatures above 600 K. The dominant role of coherence contribution in thermal transport was also observed in other complex compounds such as CsAgBiBr [22], CuSbS [6], CsPbBr [41], and even simple cubic CsCl [40]. Although the heat transport phenomena in TlSbTe exhibit similarities to those in glasses, the strong phonon anharmonicity is the key factor contributing to coherent thermal conductivity. In contrast, in glasses, the off-diagonal contributions arise from structural disorder.
After incorporating , the total is approximately 0.27 W m K and 0.26 W m K at 400 K and 500 K, respectively. This is consistent with experimental measurements, where is approximately 0.34 and 0.30 W m K at 400 K and 500 K, respectively [11]. In previous experiment, the is the the difference between total thermal conductivity () and electronic thermal conductivity (), where is estimated via the Wiedemann Franz relationship . The Lorenz number was calculated by utilizing the single parabolic band and elastic carrier scattering assumption [11]. This implies that the little difference between our and the experimental predicted ones is negligible. Additionally, we observed that the dependence of on temperature approximately follows , which indicates the glass-like behaviour of crystalline TlSbTe.
To gain a deeper insight into the thermal transport in complex crystalline TlSbTe, we calculate the spectral and cumulative for both populations’ and coherences’ contributions, as illustrated in Figure 3(a) and (Figure 3b).

Figure 3. (a) Calculated spectral and cumulative using the SCPH+3,4ph model at 300 K. (b) Calculated spectral and cumulative coherence thermal conductivity at 300 K. (c) Three-dimensional visualization model of the contributions to the coherence thermal conductivity in the -axis direction at 300 K. (d) The same as (c) but in -axis direction. (e) The same as (c) but at 700 K. (f) The same as (d) but at 700 K.
Fig. 3(a) shows that the majority of in TlSbTe is carried by phonons with a frequency below 2 THz at 300 K. Phonons with frequencies below 0.8 THz at 300 K contribute 57% and 62% of in the -axis and -axis directions, respectively. In contrast, as depicted in Figure 3(b), the majority of from the wave-like tunneling channel is carried by phonons with a frequency less than 2.4 THz at 300 K. However, phonons with frequencies less than 0.8 THz contribute only 7% of , due to the good particle-like nature of these low-frequency phonons [41]. Overall, the data presented in Figure 3(a) and (Figure 3b) indicate that low-frequency phonons predominantly contribute to , while high-frequency phonons primarily contribute to [22] [18]. This observation can be attributed to the fact that phonon scattering rates increase with increasing temperature in TlSbTe, as illustrated in Figure 4(c).

Figure 4. (a) Squared phonon velocities of Tl₉SbTe₆ in the harmonic approximation and anharmonic renormalization approximation at 300 K. (b) Phonon scattering rates of 3ph and 4ph for Tl₉SbTe₆ at 300 K and 700 K. (c) Weighted phonon scattering phase space of 3ph and 4ph for Tl₉SbTe₆ at 300 K and 700 K.
Furthermore, we observe a dip in and a peak in around 0.85 THz. This can be attributed to the low-lying flattened modes, which are dominated by Tl atoms with strong phonon-phonon scattering. Due to the strong anharmonicity of these modes, they significantly contribute to the coherence conductivity at this frequency range. Figure 3(c) and (Figure 3d) illustrate the contributions to in TlSbTe, calculated using the SCPH+3,4ph model at 300 K and 700 K, respectively, and resolved in terms of pair phonon energies. From Figure 3(c-Figure 3e), it is evident that quasi-degenerate phonons dominate , with optical phonons (frequencies larger than 0.8 THz) contributing the most. Additionally, it can be observed that the contribution of the couplings between phonons with large frequency differences to in TlSbTe increases as the temperature rises from 300 K to 700 K in both the and directions. This phenomenon can be attributed to the observed increase in phonon scattering rates with temperature.
To gain further insight into microscopic mechanism of the ultralow in TlSbTe, we conducted a detailed study of several parameters closely related to , including phonon velocities, weighted phonon scattering phase space, and phonon scattering rates. These findings are presented in Figure 4(a-Figure 4c). Since is proportional to the square of the phonon group velocity (), we calculate frequency-dependent where is the group velocity at 300 K, as shown in Figure 4(a). It can be seen that the of TlSbTe is extremely low, with a maximum value of 5 km s. In all directions, the maximum value of is lower than that of PbTe, which is approximately 14 km s [42], with of 2 W m K at 300 K. Furthermore, for the ultralow BiOSeCl, with a thermal conductivity of 0.1 W m K at 300 K, exhibits the highest velocity of 8 km s [30], which is higher than the maximum velocity of TlSbTe. Consequently, the ultralow in TlSbTe is attributed primarily to the low phonon group velocities.
Furthermore, we observe the pronounced scattering rates in crystalline TlSbTe, as illustrated in Figure 4(b). In particular, Figure 4(b) clearly shows a strong peak in phonon scattering rates, originating from the low-lying flattening modes dominated by Tl atoms [see Figure 2(a)]. In general, the low-lying flattening modes contribute to strong phonon-phonon scattering rates, particularly four-phonon scattering rates, thereby suppressing thermal transport [37]. Therefore, the strong phonon scattering rates arising from the loose bonding of Tl atoms are another factor contributing to the ultralow in Tl₉SbTe₆.
To obtain a more comprehensive understanding of phonon scattering rates, we also calculated the phonon scattering phase space, as depicted in Figure 4(c). As illustrated in Figure 4(c), the weighted phase space for both three-phonon (3ph) and four-phonon (4ph) scattering processes increases as the temperature rises from 300 K to 700 K. It is evident that the phase space of 4ph processes is larger than that of 3ph processes, resulting in a significant source of 4ph scattering rates in Tl₉SbTe₆. As anticipated, the peak region around 0.75 THz of the 4ph phase space in Tl₉SbTe₆ is in alignment with the partial density of states (PDOS) of Tl atoms, as illustrated in Figure 2(b). As previously discussed, the low-lying flattening modes typically result in elevated 4ph scattering rates by enhancing the 4ph scattering channels. Consequently, this will further suppress the thermal transport in Tl₉SbTe₆.
Electronic transport properties
The calculated electronic transport coefficients including the Seebeck coefficient , electrical conductivity , electronic thermal conductivity , and the power factor () of Tl₉SbTe₆ based on the Onsager coefficients are presented shown in Figure 5. For non-polar crystals, acoustic deformation potential scattering is the main contributor to electric conductivity [43]. Furthermore, we also consider the effects of polar optical phonon scattering [44] and ionized impurity scattering [45].

Figure 5. The temperature-dependent (a) Seebeck coefficient , (b) electrical conductivity , (d) electronic thermal conductivity , and (e) the power factor () of the were calculated as a function of carrier concentration (holes and electrons) along different axes at 300 K, 500 K and 700 K, respectively. (c) and (d) represent the electronic relaxation time for distinct scattering mechanisms, including acoustic deformation potential (ADP), ionized impurity (IMP) scattering, and polar optical phonon (POP) scattering, respectively.
Figure 5 presents a depiction of these parameters as functions of carrier concentration at distinct temperatures of 300 K, 500 K, and 700 K, respectively. Notably, Figure 5(a) reveals that both p- and n-type doped Tl₉SbTe₆ manifest notably elevated Seebeck coefficients across all investigated temperatures, auguring well for the attainment of enhanced values. For instance, at a temperature of 300 K, the maximal Seebeck coefficients for p- and n-doped Tl₉SbTe₆ materialize at 528 and 393 V/K respectively. Our Tl₉SbTe₆ has a Seebeck coefficient that is of a similar order to that of SnSe, which has a favorable 530 V/K at room temperature [46].
The Mott formula [47] indicates that the Seebeck coefficient, a crucial parameter in elucidating the interplay between thermal and electrical characteristics in three-dimensional (3D) materials, is expressed as, , where “” and “” denote the carrier concentration and the effective mass of electron states around the Fermi level. Furthermore, the density of state effective mass is also a function of band effective mass. The latter is assumed to be a rigid band model, and its dispersion relationship is a standard parabola. They differ by a factor of : , in which is the valley degeneracy [48]. The valley degeneracy of the Tl₉SbTe₆ valence band minimum is two. One is at the point and the other one is between the and points, as shown in Figure 6. Therefore, a favorable will result in a potential high thermoelectric performance for the Tl₉SbTe₆ material.

Figure 6. Electron Localization Function (ELF) of TlSbTe in (a) 3D and (b) 2D plots. ELF=1 corresponds to complete localization and ELF=0 corresponds to complete delocalization of electrons. (c) Calculated Fermi surfaces for the valence band structure minimum of TlSbTe in the first Brillouin zone around the Fermi level. (d) Electronic band structure and (e) electronic density of states (DOS) of TlSbTe. (f) Normalized Lorenz number as a function of carrier concentration at temperatures of 300 K, 500 K and 700 K, respectively. The Lorenz number, denoted by , is a constant that is closely approximately by the Sommerfeld value, W K.
Moreover, it can be gleaned from the Mott formula that the Seebeck coefficient for both p- and n-type exhibits a dependence on temperature and doping concentration. As depicted in Figure 5(a), at lower carrier concentrations, the value of at 300 K reaches its highest point. However, as the carrier concentration increases beyond a certain threshold, a sequential decline in the Seebeck coefficient from 300 K to 700 K is observed, aligning with the anticipated result.
This nonlinear relationship between and doping concentration stems from the bipolar effect [49, 50]. When there are significant number of both electrons and holes contributing to charge transport (bipolar charge transport) the thermoelectric properties are greatly affected. This occurs when electrons are excited across the band gap producing minority charge carriers (e.g., holes in an -type material) in addition to majority charge carriers (e.g., the electrons in an -type material). Bipolar effects are observed in small band gap materials at high temperatures. Furthermore, it is notable that the Seebeck coefficient of exhibits significant anisotropy, regardless of whether the material is doped with - or -type ions.
Figure 5(b) and (e) show the direct correlation between electrical conductivity and electronic thermal conductivity , as well as carrier concentrations and temperatures both - and -type doping of TlSbTe. It can be observed that as temperature increases, conductivity decreases. Conversely, as concentration increases, conductivity increases. Furthermore, TlSbTe exhibits pronounced anisotropy in both electrical conductivity and electronic thermal conductivity, with distinct behaviors observed for - and -type doping scenarios. In the context of electron doping, it is evident that the values along the -axis (dashed lines) are notably lower than those along the -axis (solid lines). For example, at a temperature of 300 K and a carrier concentration of cm, the values of and along the -axis are 6640.35 S cm and 4.18 W m K, which are approximately three times the values observed along the -axis with 2254.63 S cm and 1.57 W m K. Conversely, under hole doping conditions, the values along the -axis are slightly higher than the corresponding -axis.
Figure 5(c) and (d) exhibit the electronic relaxation time () for distinct scattering mechanisms, including acoustic deformation potential (ADP), ionized impurity (IMP) scattering, and polar optical phonon (POP) scattering, respectively. It can be seen that of IMP is greater than that of ADP, and of ADP is greater than POP, indicating that POP has the strongest scattering. As the temperature increases, the strength of electron scattering increases, resulting in a corresponding decrease in . This is illustrated in Figure 5(d), where the total electronic relaxation time is observed to be on the order of cm. As the temperature increases from 300 K to 700 K, the electron relaxation time decreases from cm to cm for - and -type doping TlSbTe.
As illustrated in Figure 5(f), the power factor () was computed based on the Seebeck coefficient and electrical conductivity, under varying temperatures and carrier concentrations. Our findings reveal that TlSbTe exhibits substantial anisotropy in its power factor. Across intermediate temperature and carrier concentration regimes, the power factor for electron-doped TlSbTe along the -axis consistently surpasses that for hole doping. This indicates that, in order to enhance the thermoelectric performance of TlSbTe, electron doping is more advantageous than hole doping. In contrast, the effect of hole doping along the -axis is notably more pronounced and superior in enhancing the power factor. The variation of power factor with concentration and temperature is not monotonic as a function of electrical and electronic thermal conductivity. The power factor exhibits a characteristic trend whereby it initially rises with increasing carrier concentration and subsequently declines, with a peak concentration range between and cm for TlSbTe. Below this peak concentration, higher temperatures are associated with lower power factors. It is noteworthy that at a temperature of 500 K and a carrier concentration of cm, the power factor along the -axis attains its maximum value, reaching 8.57 mW m K.
Figure 6(a) and (b) show the 3D and 2D Electron Localization Function (ELF) of TlSbTe. ELF values of 1 and 0 represent complete localization and delocalization of electrons. The Te and Sb elements exhibit a stronger tendency to attract electrons than the Tl elements, suggesting a greater degree of delocalization for thallium atoms. The delocalized electrons show a stronger electron-phonon coupling [51, 52]. This is evident in Figure 2(b), where the thallium element exhibits a very pronounced peak in the low-frequency region (<1.0 Thz) for the phonon density of states [53, 54].
Figure 6(d) and (e) present the electronic band structure and the electronic density of state (DOS). TlSbTe emerges as an indirect bandgap semiconductor, characterized by its conduction band minimum (CBM) positioned between the and points, and its valence band maximum (VBM) located between the and points, resulting in a discernible bandgap of 0.279 eV. The contributions to the bottom of the conduction band and the top of the valence band predominantly arise from the -orbitals of Te atoms and Tl atoms respectively. Furthermore, the band structure reveals a dual-degeneracy phenomenon within the valence band at the point. Importantly, this observation holds profound significance for enhancing the thermoelectricity of the material, as the insights from band engineering studies [55] [48].
It is well established that electrical transport is not solely dependent on the movement of electrons, but also on the transfer of heat. This phenomenon is commonly described by the Wiedemann-Franz (WF) law, which states that the ratio of electrical conductivity to the conductivity of heat () is directly proportional to the absolute temperature . This relationship can be expressed as , where is a constant ratio known as the Lorenz number. The value of is believed to be close to the Sommerfeld value , which is approximately
[56, 57]. It is notable that most metals that considered to be “good” in terms of their conductivity satisfy this WF law. However, the Lorenz number of transition metals such as palladium (Pd), nickel (Ni), cobalt (Co), and platinum (Pt) exhibits larger deviations from the Sommerfeld value [58, 59]. Recently, a that is an order of magnitude lower than expected from WF law was observed in metallic VO near its insulator-metal transition. This phenomenon was explained in terms of the absence
of quasiparticles in a strongly correlated electron fluid in which heat and charge diffuse independently [56].
The calculated doping-dependent ratio is shown in Figure 6(f). It can be seen that when the doping concentration is high, especially when it exceeds cm, is near equivalent to , exhibiting metallic properties. The lower limit of the doping concentration for the Lorenz number being close to increases with temperature. For -type TlSbTe, the lower limit of the doping concentration for the Lorenz number to be close to at 300 K is around cm, while at 700 K it is cm. This suggests that in order to achieve enhanced thermoelectric performance under high temperature conditions, it may be necessary to adjust the doping concentration within an optimal range.
A higher value represents a higher ratio of electronic thermal conductivity to electrical conductivity, which is less conducive to improving the thermoelectric properties of the material. At lower concentrations, a substantial deviation of from the ideal value is observed, particularly when the concentration is below cm. Furthermore, the deviation of from is severer with higher temperature than lower temperature. This is consistent with the trend shown in Figure 7, where the lower the doping concentration, the closer the peak value is to the low temperature region. This also makes it difficult to use the Wiedemann-Franz law to predict the of the material by considering the experimental values of and electrical conductivity, as considering as will underestimate , thus overestimate . Consequently, it is necessary to calculate the Lorenz number as a function of temperature and doping concentration, which is the prerequisite for accurately estimating thermoelectric performance.

Figure 7. The calculated figure of merit () of (a) -type and (b) -type as a function of temperature along the - and - axes at different carrier concentrations, ranging from to .
Thermoelectric performance
The calculated of TlSbTe at different temperatures and doping concentrations is presented in Figure 7. As the doping concentration increases from cm to cm, the peak value of the value will vary from 300 K to 700 K and higher, with a greater bias towards higher temperatures. Concurrently, the peak value will initially increase and subsequently decrease, reaching a maximum value at a doping concentration between cm and cm. Due to the orientation-dependent properties of phonons and electrons, the value will exhibit a significant anisotropy.
In the case of -type doping, the value in the -axis direction is significantly higher than that in the -axis direction, which is the opposite in the case of -type. At the same temperature, the values of - and -type vary significantly. The highest value of the -type can reach a peak of 3.17 at 600 K, with a doping concentration of cm. The highest value of the -type also reaches a peak of 2.26 at 500 K, with a doping concentration of . It can be concluded that the is a highly promising thermoelectric material in the medium temperature zone (300 K-600 K), particularly for the -type doping.
Conclusion
In conclusion, our study examines the thermal and electrical transport properties of through the integration of machine-learning-potential-based self-consistent phonon (SCPH) theory, first-principles calculations, and a unified theory of thermal transport that incorporates both coherence and population conductivities. The Moment Tensor Potential, trained with Machine Learning Inter-atomic Potential (MLIP), achieves DFT-level accuracy in predicting energy and atomic forces, thereby accurately producing the phonon dispersions for crystalline . The force constants derived from the MLIP and SCPH theory indicate a phonon hardening phenomenon across the entire frequency range, which suggests significant lattice anharmonicity in . Subsequently, an ultralow for is predicted, with a value of 0.3 W mK at room temperature. This value is consistent with the experimentally measured and exhibits a glass-like temperature dependence.
Our research indicates that four-phonon interaction processes and off-diagonal elements of heat flux operators have a significant impact on thermal transport in . Specifically, the four-phonon interaction processes influence phonon energy shifts, which increase thermal conductivity. However, they also significantly enhance overall scattering rates, thereby decreasing thermal conductivity. Furthermore, we observe that in crystalline at temperatures above 600 K, the coherent component represents the dominant contributor to the lattice , with a ratio exceeding 1. By analyzing the electronic band structure, we find that also exhibits excellent electrical transport properties, attributed to the dual-degeneracy phenomenon within the valence band.
Our investigation has revealed that , with a maximum figure of merit () of up to 3.17, is a promising candidate for thermoelectric applications. This is due to the ultralow and favorable electrical transport properties. We have also elucidated the origins of the ultralow and relatively high electrical conductivity in . This has involved confirming its optimal charge carrier concentration and operating temperature range. Our study offers insights into potential improvements and application directions for the family of materials and paves the way to accelerate the accurate prediction of thermoelectric materials.
ACKNOWLEDGMENTS
We acknowledge the support from the National Natural Science Foundation of China (No.12104356 and No.52250191, No.12104236), the Opening Project of Shanghai Key Laboratory of Special Artificial Microstructure Materials and Technology (No.Ammt2022B-1), and the Fundamental Research Funds for the Central Universities. H. Gu acknowledges the support from Gusu Leading Talent (No.ZXL2021383). This work is sponsored by the Key Research and Development Program of the Ministry of Science and Technology (No.2023YFB4604100). We also acknowledge the support by HPC Platform, Xi’an Jiaotong University.
References
References
- [14]Quansheng Guo, Meghan Chan, Bryan A. Kuropatwa, and Holger Kleinke, “Enhanced Thermoelectric Properties of Variants of Tl₉SbTe₆ and Tl₉BiTe₆,” Chem. Mater. 25, 4097–4104 (2013).DOI
- [15]Guo, L. B., Yang, L., Wang, Y. X., Yang, J. M., Yan, Y. L., & Ren, F. Z. (2015). The electronic structure and thermoelectric properties of BiTl₉Te₆ and SbTl₉Te₆: First-principles calculations. J. Appl. Phys., 118, 235703.
- [16]Xiaoma Tao, Philippe Jund, Romain Viennois, and Jean-Claude Tedenac, “Physical properties of thallium-tellurium based thermoelectric compounds using first-principles simulations,” J. Phys. Chem. A 115, 8761–8766 (2011).DOI
- [17]Yi Xia, Koushik Pal, Jiangang He, Vidvuds Ozoliņš, and Chris Wolverton, “Particlelike Phonon Propagation Dominates Ultralow Lattice Thermal Conductivity in Crystalline Tl₃VSe₄,” Phys. Rev. Lett. 124, 065901 (2020).DOI
- [19]Terumasa Tadano and Shinji Tsuneyuki, “Quartic anharmonicity of rattlers and its effect on lattice thermal conductivity of clathrates from first principles,” Phys. Rev. Lett. 120, 105901 (2018).DOI
- [20]Jiongzhi Zheng, Dongliang Shi, Yuewang Yang, Chongjia Lin, He Huang, Ruiqiang Guo, and Baoling Huang, “Anharmonicity-induced phonon hardening and phonon transport enhancement in crystalline perovskite BaZrO₃,” Phys. Rev. B 105, 242303 (2022).DOI
- [21]Jincheng Yue, Jiongzhi Zheng, Junda Li, Siqi Guo, Wenling Ren, Han Liu, Yanhui Liu, and Tian Cui, “Ultralow Glassy Thermal Conductivity and Controllable, Promising Thermoelectric Properties in Crystalline o-CsCu₅S₃,” ACS Appl. Mater. Interfaces 16, 20597–20609 (2024).DOI
- [22]Jiongzhi Zheng, Changpeng Lin, Chongjia Lin, Geoffroy Hautier, Ruiqiang Guo, and Baoling Huang, “Unravelling ultralow thermal conductivity in perovskite Cs₂AgBiBr₆: dominant wave-like phonon tunnelling and strong anharmonicity,” npj Comput. Mater. 10, 4941–4948 (2024).DOI
- [23]Terumasa Tadano and Shinji Tsuneyuki, “Self-consistent phonon calculations of lattice dynamical properties in cubic SrTiO₃ with first-principles anharmonic force constants,” Phys. Rev. B 92, 054301 (2015).DOI
- [24]Yinchang Zhao, Shuming Zeng, Geng Li, Chao Lian, Zhenhong Dai, Sheng Meng, and Jun Ni, “Lattice thermal conductivity including phonon frequency shifts and scattering rates induced by quartic anharmonicity in cubic oxide and fluoride perovskites,” Phys. Rev. B 104, 224304 (2021).DOI
- [25]Yi Xia, Vinay I. Hegde, Koushik Pal, Xia Hua, Dale Gaines, Shane Patel, Jiangang He, Muratahan Aykol, and Chris Wolverton, “High-throughput study of lattice thermal conductivity in binary rocksalt and zinc blende compounds including higher-order anharmonicity,” Phys. Rev. X 10, 041029 (2020).DOI
- [26]Joon Sang Kang, Huan Wu, Man Li, and Yongjie Hu, “Intrinsic low thermal conductivity and phonon renormalization due to strong anharmonicity of single-crystal tin selenide,” Nano Lett. 19, 4941–4948 (2019).DOI
- [27]Michele Simoncelli, Nicola Marzari, and Francesco Mauri, “Unified theory of thermal transport in crystals and glasses,” Nat. Phys. 15, 809–813 (2019).DOI
- [28]Wu Li, Jesús Carrete, Nebil A. Katcho, and Natalio Mingo, “Shengbte: A solver of the boltzmann transport equation for phonons,” Comput. Phys. Commun. 185, 1747–1758 (2014).DOI
- [29]Zherui Han, Xiaolong Yang, Wu Li, Tianli Feng, and Xiulin Ruan, “Fourphonon: An extension module to shengbte for computing four-phonon scattering rates and thermal conductivity,” Comput. Phys. Commun. 270, 108179 (2022).DOI
- [30]Zhen Tong, Alessandro Pecchia, ChiYung Yam, Traian Dumitrică, and Thomas Frauenheim, “Glass-like Transport Dominates Ultralow Lattice Thermal Conductivity in Modular Crystalline Bi₄O₄SeCl₂,” Nano Lett. 23, 9468–9473 (2023).DOI
- [31]Bohayra Mortazavi, Evgeny V. Podryabinkin, Ivan S. Novikov, Timon Rabczuk, Xiaoying Zhuang, and Alexander V. Shapeev, “Accelerating first-principles estimation of thermal conductivity by machine-learning interatomic potentials: A MTP/ShengBTE solution,” Comput. Phys. Commun. 258, 107583 (2021).DOI
- [32]Xiao Wan, Wentao Feng, Yunpeng Wang, Haidong Wang, Xing Zhang, Chengcheng Deng, and Nuo Yang, “Materials discovery and properties prediction in thermal transport via materials informatics: A mini review,” Nano Lett. 19, 3387–3395 (2019).DOI
- [33]Ivan S Novikov, Konstantin Gubaev, Evgeny V Podryabinkin, and Alexander V Shapeev, “The MLIP package: moment tensor potentials with MPI and active learning,” Mach. Learn. Sci. Technol. 2, 025002 (2020).arxiv.org/abs/2007.08555
- [34]Alexander V. Shapeev, “Moment tensor potentials: A class of systematically improvable interatomic potentials,” Multiscale Model. Simul. 14, 1153–1173 (2016).DOI
- [35]Yulou Ouyang, Cuiqian Yu, Jia He, Pengfei Jiang, Wei-jun Ren, and Jie Chen, “Accurate description of high-order phonon anharmonicity and lattice thermal conductivity from molecular dynamics simulations with machine learning potential,” Phys. Rev. B 105, 115202 (2022).DOI
- [36]Minxuan Feng, Xiaoying Wang, Guimei Zhu, Cheng He, Jun Sun, Xiangdong Ding, Junichiro Shiomi, Yi Xia, Baowen Li, and Zhibin Gao, “The relation between the atomic mass ratio and quartic anharmonicity in alkali metal hydrides,” Mater. Today Phys. 44, 101423 (2024).DOI
- [37][37–40]
- [38]Zheng, J., Shi, D., Liu, S., Yan, Y., Lin, C., Chang, Z., Guo, R., & Huang, B. (2022). Effects of high-order anharmonicity on anomalous lattice dynamics and thermal transport in fully filled skutterudite YbFe₄Sb₁₂. Phys. Rev. Mater., 6, 093801.
- [41]He, Y., Matei, L., Jung, H. J., McCall, K. M., Chen, M., Stoumpos, C. C., Liu, Z., Peters, J. A., Chung, D. Y., Wessels, J., Wasielewski, M. R., Pravad, V. P., Burger, A., & Kanatzidis, M. G. (2018). High spectral resolution of gamma-rays at room temperature by perovskite CsPbBr₃ single crystals. Nat. Commun., 1609.
- [42]Zhiting Tian, Jivtesh Garg, Keivan Esfarjani, Takuma Shiga, Junichiro Shiomi, and Gang Chen, “Phonon conduction in PbSe, PbTe, and PbTe₁−ₓSeₓ from first-principles calculations,” Phys. Rev. B 85, 184303 (2012).DOI
- [49]Zhu, R., Gao, Z., Liang, Q., Hu, J., Wang, J.-S., Qiu, C.-W., & Shen, A. T. (2021). Observation of anisotropic magnetoresistance in layered nonmagnetic semiconducting PdSe₂. ACS Appl. Mater. Interfaces, 13, 37527–37534.
- [50]JJ Gong, AJ Shuai, L Li, ZB Yan, ZF Ren, and J-M Liu, “Investigation of the bipolar effect in the thermoelectric material CaMg₂Bi₂ using a first-principles study,” Phys. Chem. Chem. Phys. 18, 16566–16574 (2016).DOI
- [51]Zhibin Gao, Tao Zhu, Kangtai Sun, and Jian-Sheng Wang, “Highly anisotropic thermoelectric properties of two-dimensional As₂Te₃,” ACS Appl. Electron. Mater. 3, 1610–1620 (2021).DOI
- [52]Qin, F., Wang, X., Hu, L., Jia, N., Gao, Z., Aydenmir, U., Chen, J., Ding, X., & Sun, J. (2022). Switch of thermal expansions triggered by itinerant electrons in isostructural metal trifluorides. Inorg. Chem., 61, 21004–21010.
- [53]Zhibin Gao and Jian-Sheng Wang, “Thermoelectric penta-silacene with a high room-temperature figure of merit,” ACS Appl. Mater. Interfaces 12, 14298–14307 (2020).DOI
- [54]Zhibin Gao, Gang Liu, and Jie Ren, “High thermoelectric performance in two-dimensional tellurium: An ab initio study,” ACS Appl. Mater. Interfaces 10, 40702–40709 (2018).DOI
- [55]Wei Liu, Xiaojian Tan, Kang Yin, Huijun Liu, Xinfeng Tang, Jing Shi, Qingjie Zhang, and Ctirad Uher, “Convergence of Conduction Bands as a Means of Enhanc- ing Thermoelectric Performance of n-Type Mg₂Si₁−ₓSnₓ Solid Solutions,” Phys. Rev. Lett. 108, 166601 (2012).DOI
- [56]Sangwook Lee, Kedar Hippalgaonkar, Fan Yang, Jiawang Hong, Changhyun Ko, Joonki Suh, Kai Liu, Kevin Wang, Jeffrey J. Urban, Xiang Zhang, Chris Dames, Sean A. Hartnoll, Olivier Delaire, and Junqiao Wu, “Anomalously low electronic thermal conductivity in metallic vanadium dioxide,” Science 355, 371–374 (2017).DOI
- [57]Wen O. Wang, Jixun K. Ding, Yoni Schattner, Edwin W. Huang, Brian Moritz, and Thomas P. Devereaux, “The Wiedemann-Franz law in doped Mott insulators without quasiparticles,” Science 382, 1070–1073 (2023).DOI
- [58]Zhen Tong, Shouhang Li, Xiulin Ruan, and Hua Bao, “Comprehensive first-principles analysis of phonon thermal conductivity and electron-phonon coupling in different metals,” Phys. Rev. B 100, 144306 (2019).DOI
- [59]Xinyu Zhang, Shouhang Li, Ao Wang, and Hua Bao, “Pressure-dependent thermal conductivity in Al, W, and Pt: Role of electrons and phonons,” Phys. Rev. B 106, 094313 (2022).DOI