WorksPhysics and Astronomy
Lone-Pair Stereoactivity Overrides Mass Scaling of Thermal Expansion and Heat Conduction in Non-vander Waals Layered Bismuth Oxychalcogenides
Abstract
Heavier chalcogens are generally expected to soften isostructural semiconductor lattices, raising thermal expansion and lowering thermal conductivity. Layered bismuth oxychalcogenides, Bi2O2X (X = S, Se, Te), are mixed-anion crystals whose covalent slabs and ionic sheets are bound electrostatically rather than by van der Waals forces and whose Bi3+ 6s^2 lone pairs can become stereochemically active. Despite the rapid development of Bi2O2Se for fin and gate-all-around transistors, its fundamental thermal expansion and heat transport mechanisms remain poorly understood. Here, we show that lone-pair stereoactivity, not chalcogen mass, governs both properties. Path-integral molecular dynamics shows that the volumetric thermal expansion coefficient at 300 K decreases from 8.14 x 10^-5 K^-1 in Bi2O2S to 6.72 x 10^-5 and 6.61 x 10^-5 K^-1 in Bi2O2Se and Bi2O2Te, respectively, despite a 13% larger unit cell in the telluride, whereas substituting the S mass into Bi2O2Se changes its value by only 0.6%. Lone-pair-driven Bi off-centering in orthorhombic Bi2O2S makes b-axis expansion 1.63 times a-axis expansion yet leaves in-plane thermal conduction isotropic within 3%. Phonon transport calculations including four-phonon scattering and Wigner coherence predict a thermal conductivity maximum in Bi2O2Se (1.83 and 0.95 W m^-1 K^-1 in the in-plane and cross-plane directions, respectively, at 300 K) because Bi2O2S is limited by anharmonic scattering and Bi2O2Te by slow cross-plane phonons. Thermoreflectance measurements from 300 to 500 K and in situ X-ray diffraction on Bi2O2Se single crystals agree with these predictions within 6.4%. Phonon Boltzmann transport simulations of 12-nm-wide fins, which run hotter than Fourier diffusion predicts, show that Bi2O2Se minimizes the hotspot temperature and the thermal-mismatch strain at the gate dielectric.
Introduction
Layered bismuth oxychalcogenides, (), have emerged as a semiconductor platform for post-silicon nanoelectronics and optoelectronics [1, 2, 3, 4, 5, 6]. Their Sillén-type lattice alternates covalent, fluorite-like slabs with ionic sheets Figure 1(a). Because adjacent layers carry opposite formal charges, they are bound by predominantly electrostatic rather than van der Waals (vdW) interactions, and this non-vdW layered motif combines layer-controlled epitaxy [3, 7] with air stability, small in-plane effective masses, and high carrier mobilities. , the most mature member, couples a room-temperature electron mobility of about with an epitaxial native high- oxide, -, whose atomically sharp interface suppresses short-channel effects in fin and gate-all-around field-effect transistors (FinFETs and GAAFETs) [8, 9, 10]. Vertical fins as thin as one unit cell and gate-all-around channels now deliver on-state currents from to above [11, 14] [15]. nanosheet transistors exploit small effective masses for high-mobility channels, and lattice distortions in atomically thin enable room-temperature ferroelectric switching for nonvolatile memory [4, 5, 10, 11, 12, 13, 14] [15]. Composition-tunable bandgaps further support air-stable visible to near-infrared detection in , silicon-band-edge photodetection in , and extended infrared operation in narrow-gap 16–19.

Figure 1. (a) Side and top views of , , and . Gray wireframes mark conventional unit-cell boundaries. The Bi off-centering along the axis in is indicated as Å relative to the symmetry-matched reference position. Purple, red, blue, green, and olive spheres denote Bi, O, S, Se, and Te, respectively. (b) Ground-state electron localization function (ELF) maps on the (110) planes of the three compounds, plotted on a common scale from 0 to 1. The dashed outline highlights the asymmetric lone-pair localization adjacent to Bi in . (c) Schematic thermal-strain distributions and representative thermomechanical failure scenarios in a FinFET with a channel (top) and a thermoelectric module with a n-type leg (bottom). Enlarged callouts illustrate crack initiation and interfacial failure in regions of concentrated thermoelastic strain arising from thermal-expansion mismatch. Scale bars indicate characteristic dimensions, and the color scale represents the thermal-strain magnitude.
These nanoscale architectures confine heat generation and thermal strain to dimensions comparable to or smaller than the mean free paths of heat-carrying phonons. In sub-15 nm channels, phonon–boundary scattering impedes heat removal and invalidates Fourier-based estimates of the hotspot temperature 20–22, while the mismatch between the thermal expansion of the channel and that of the gate dielectric, contacts, and substrate concentrates interfacial stress during self-heating and thermal cycling 23,24. Analogous constraints arise in thermoelectric legs, where a low lattice thermal conductivity () supports figures of merit of – near but steep gradients and contact mismatch promote microcracking and fatigue 23,24, and in silicon-integrated photodetectors, where thermally induced strain shifts the bandgap and destabilizes the photoresponse 25–30.
Predicting these failure modes requires the full anisotropic tensors of and of the coefficient of thermal expansion (CTE, ).
Both tensors are set by how bond stiffness and anharmonicity are partitioned across the layered lattice. In vdW crystals such as graphite, weak interlayer coupling allows flexural membrane modes with negative Grüneisen parameters to contract the basal plane on heating 31. Mixed-anion compounds instead combine chemically distinct anion sublattices with a built-in hierarchy of covalent and ionic bonds 32. In BiOSe, compression shortens the interlayer Bi–Se bonds more than three times faster than the intralayer Bi–O bonds [3], identifying the electrostatic interlayer contact as the compliant element of an otherwise rigid framework. Superimposed on this hierarchy is the Bi lone pair. In the revised lone-pair model, the lone pair becomes stereochemically active when antibonding cation –anion states mix with empty cation states upon symmetry lowering 33, which produces asymmetric coordination, soft low-frequency phonons, and near-minimum in lone-pair chalcogenides 34;35. Because , where is the isothermal bulk modulus, the volume, and and the heat capacity and Grüneisen parameter of mode , the same lone-pair-softened modes should also control thermal expansion, as the spontaneous-strain and negative-expansion anomalies of lone-pair ferroelectrics such as PbTiO illustrate 36.
In BiOX, lone-pair-related anharmonicity yields large Grüneisen parameters and a broad multiphonon scattering phase space, so that four-phonon (4ph) scattering and wave-like interband tunneling suppress to values comparable to those of SnSe and BiTe 37–43. The ELF maps in Figure 1(b) show that the asymmetry of the lone-pair density varies across the series. Pressure and strain studies support this lone-pair–phonon link. Hydrostatic compression suppresses Bi lone-pair activity in BiOSe and reshapes its phonon group velocities until the conduction anisotropy inverts 44, whereas -axis tensile strain in BiCuSeO, which shares the [BiO] slab, expands the lone-pair lobes, softens optical phonons, and intensifies Umklapp scattering 43. Chalcogen substitution tunes the same interaction chemically. In tetragonal BiOSe and BiOTe (), the higher-lying and more polarizable Se and Te states weaken the Bi – antibonding interaction and leave Bi centered with 44. In BiOS, the deeper S states strengthen this interaction, activating the lone pair and driving cooperative in-plane Bi off-centering into an orthorhombic ground state with 45, which hydrostatic pressure of a few gigapascals restores to 40.
This chemical tuning raises questions that existing data cannot answer. Across most chalcogenide series, heavier anions soften the lattice, so that increases (Supporting Information, Table S3) and decreases from S to Te. Whether lone-pair stereoactivity can override this mass-and-softening scaling in BiOX, and whether the basal-plane symmetry breaking of BiOS imposes the same anisotropy on expansion () and on conduction (), remain open 40. To our knowledge, no direction-resolved CTE has been reported for any BiOX compound, which precludes estimates of device mismatch strain 38;41. Thermal transport data, concentrated on BiOSe, are also inconsistent. First-principles values span 1.3–4.5 W m K in plane and 0.8–3.7 W m K cross plane 37;38;41;42;46;47, and measured values span 1.4–2.4 and 0.9–2.3 W m K, respectively 41;48–50;55, because measurements on polycrystalline pellets, supported nanosheets, or along a single axis cannot resolve the full tensor of one crystal across the 300–500 K window relevant to self-heating [8, 9, 10]. Validated anisotropic baselines for and are therefore missing, and the thermomechanical reliability envelope of BiOX devices remains undefined 56;57.
| Material | (MD) | (MD) | (PIMD) | (PIMD) | (Exp.) |
| – | |||||
| – |
Table S3. Thermal-expansion data at 300 K. (a) In-plane and cross-plane coefficients of thermal expansion of , , and obtained from classical MD and PIMD, together with the experimentally measured cross-plane CTE of . (b) Volumetric CTEs of and the literature S–Se–Te chalcogenide series used in Figure 3(b), together with the corresponding sulfide-normalized ratios and their assignment to the shaded envelope representing the prevailing trend.
Theory has not closed these gaps. For thermal expansion, the quasiharmonic approximation (QHA) retains only volume-dependent harmonic frequencies and omits the intrinsic anharmonic renormalization at fixed volume, whereas the O sublattice of the [BiO] slabs vibrates above 7 THz ( K) and therefore retains nuclear quantum effects (NQEs) up to intermediate temperatures 56–58. For heat transport, early family-wide calculations relied on static 0 K force constants within the relaxation-time approximation or the lowest-order three-phonon (3ph) Boltzmann transport equation (BTE) 38;41;42;47. Later studies added 4ph scattering for BiOSe 37;44 or wave-like coherence for BiOS at the 3ph level 40, while other studies treated individual compounds or subsets of the family with more restricted transport models 38;41;47;59. Consequently, even the sign of the difference between BiOS and BiOSe remains unsettled 40, and no study has combined temperature-renormalized force constants, quartic anharmonicity, and off-diagonal coherent transport across the series. The Wigner transport equation (WTE) provides this framework by adding, to the particle-like population conductivity, a coherence conductivity that becomes significant when phonon linewidths approach interbranch frequency spacings 60. At the device scale, BiOSe FinFETs combine ultrathin fins with high drive currents [11, 14], yet no hotspot or thermoelastic analysis has been reported for the sub-15 nm regime in which Fourier diffusion fails 20;21.
Here, we resolve the anisotropic and tensors of BiOS, BiOSe, and BiOTe from 100 to 700 K and trace both to Bi lone-pair stereoactivity. Path-integral molecular dynamics (PIMD) 61;62 driven by a machine-learned potential trained on density functional theory (DFT) data captures anharmonicity and NQEs in the lattice parameters, and a mass-substitution control separates the roles of bonding and atomic mass. Temperature-dependent effective potential (TDEP) force constants 63;64 feed iterative solutions of the BTE and WTE that include 4ph scattering and the coherence conductivity 60. In situ X-ray diffraction (XRD) and dual-axis frequency-domain thermoreflectance (FDTR) from 300 to 500 K on BiOSe single crystals benchmark the calculations, reproducing the cross-plane CTE (2.82 measured versus K calculated) and the room-temperature conductivities ( and W m K measured versus 1.83 and 0.95 calculated) within 6.4%.
The family departs from mass scaling in two ways. First, the volumetric CTE at 300 K falls from K in BiOS to 6.72 and K in BiOSe and BiOTe, although the unit-cell volume grows by 13% from S to Te. Substituting the S mass into BiOSe changes its expansion by only 0.6%, whereas lone-pair-driven Bi off-centering in BiOS creates an antipolar double-well potential and raises the heat-capacity-weighted Grüneisen parameter to 3.55, compared with 3.31 and 3.04 for the selenide and telluride. The same distortion yields but , so basal-plane symmetry breaking appears in expansion but not in conduction. Second, peaks in BiOSe rather than in the lightest member, because BiOS is limited by strong 3ph and 4ph scattering below 3 THz and by low in-plane group velocities, whereas BiOTe is limited by the lowest cross-plane group velocity in the family and exhibits the largest conduction anisotropy (). In BiOSe, 4ph scattering removes 46.8% of the cross-plane population conductivity at 300 K, and supplies 18.6% of the cross-plane total at 700 K. Finally, non-gray phonon BTE and thermoelastic simulations of 12 nm wide fins with vertically oriented layers, a gate dielectric, and a substrate show that Fourier diffusion underestimates the hotspot temperature and that BiOSe develops the lowest hotspot temperature and thermal-mismatch strain under identical heat loads.
Results and Discussion
Chalcogen-Dependent Bonding and Lone-Pair Stereoactivity
The (110)-plane ELF maps in Figure 1(b) show a pronounced one-sided localization of electron density adjacent to Bi in Bi₂O₂S, whereas the corresponding distributions remain nearly symmetric about the centered Bi sites of tetragonal Bi₂O₂Se and Bi₂O₂Te (). This asymmetry is the signature of a stereochemically active Bi³⁺ 6 lone pair and accompanies an in-plane Bi off-centering of Å along in Bi₂O₂S [Figure 1(a)], compared with in the selenide and telluride. The Bi mean-square displacements carry the same symmetry breaking [Figure 2(a)]. From 100 to 700 K, the displacement amplitudes in Bi₂O₂S exceed those of the other two compounds (), indicating a softer local potential around Bi in the sulfide. Fourfold rotational symmetry enforces in Bi₂O₂Se and Bi₂O₂Te, whereas Bi₂O₂S retains throughout the temperature range. At 300 K, Ų, compared with Ų and Ų, so the largest Bi fluctuations occur along the off-centering direction.

Figure 2. Composition-dependent Bi dynamics and chemical bonding. (a) Temperature-dependent diagonal components of the Bi mean-square displacement tensor in Bi₂O₂S, Bi₂O₂Se, and Bi₂O₂Te. (b) Static (0 K) frozen-phonon energy profiles along the antipolar -point mode that connects the tetragonal reference and the orthorhombic structure, evaluated at fixed tetragonal lattice parameters. Energies are referenced to the undistorted configuration and given per formula unit. (c) Bond-averaged pCOBI and IpCOBI and (d) sign-reversed −pCOHP and −IpCOHP for the Bi–O and Bi–X interactions in Bi₂O₂S (top), Bi₂O₂Se (middle), and Bi₂O₂Te (bottom). Gray, blue, red, and green denote Bi–O, Bi–S, Bi–Se, and Bi–Te interactions, respectively. Shaded areas show pCOBI and −pCOHP, and solid curves show their cumulative integrals. Vertical dashed lines mark . Integrated values at are listed in Table S1.
To identify the energetic origin of the distortion, we mapped the static (0 K) energy surface along the symmetry-adapted coordinate with frozen-phonon DFT calculations 65 [Figure 2(b)]. Here, is the mass-weighted amplitude of the primitive-cell -point mode that connects the tetragonal reference to the orthorhombic structure, and all energies were evaluated at fixed, symmetry-constrained tetragonal lattice parameters (§ 8.3). Along this coordinate, develops a symmetric double well. A fourth-order fit gives a negative harmonic coefficient, , minima at , and a central barrier of 7.21 meV per formula unit. The undistorted structure is therefore unstable at 0 K against antiparallel in-plane displacements of adjacent slabs. Because the dipoles of neighboring slabs cancel, the resulting antipolar structure remains centrosymmetric, and the positive quartic coefficient bounds the energy at large . and instead exhibit single wells centered at with positive of and , respectively, so restoring forces in the heavier chalcogenides prevent condensation of this mode.
The double-well barrier is a static, single-mode quantity. It excludes zero-point and thermal fluctuations, as well as the cell strain and intersite coupling that accompany the cooperative distortion, and therefore cannot be compared directly with to estimate a transition temperature. The finite-temperature stability of the orthorhombic phase is established instead by PIMD, in which the and lattice parameters of remain distinct from 100 to 700 K (Figure 3(e)), and by the absence of imaginary frequencies in the TDEP dispersions of the structure at 100 and 700 K (Figure 5(a)). This behavior agrees with the orthorhombic ground state reported for 45 and with its return to under hydrostatic pressure$40.

Figure 3. Anisotropic and composition-dependent thermal expansion of . (a) PIMD linear CTEs along the , , and axes of , , and at 300 K. The hatched bar denotes the experimental -axis CTE of . (b) Sulfide-normalized volumetric CTEs, , of and representative S–Se–Te chalcogenide series at 300 K. Thin gray lines denote series following the prevailing increase in from S to Se and Te, and the blue shaded region shows their envelope. and , which deviate from this trend, are shown separately. The dashed line marks the sulfide baseline. Values, sources, and trend assignments are given in Table S3. (c) Schematic of the in situ XRD measurement of a crystal supported on a silicon wafer holder. (d) Temperature-dependent XRD patterns of from 298 to 473 K showing the shifts of the (002), (004), (006), and (008) reflections. (e,f) PIMD lattice parameters versus temperature, showing (e) and and (f) for the three compounds. Error bars denote 95% confidence intervals from five independent trajectories.

Figure 5. Temperature-renormalized phonon dispersions and group velocities of . (a–c) TDEP phonon dispersions at 100 and 700 K, obtained at a fixed reference lattice, for (a) , (b) , and (c) . (d) Projected phonon densities of states resolving Bi, chalcogen, and O contributions, with dashed lines at 3 and 7 THz. (e) Total phonon group velocity versus frequency. (f–h) Direction-resolved group velocities , , and for (f) , (g) , and (h) .
Projected crystal orbital Hamilton populations (pCOHP) and crystal orbital bond indices (pCOBI) quantify how the bonding evolves across the series (Figure 2(c,Figure 2d) and Table S1). At the Fermi level, the bond-averaged ICOHP of Bi–O is 2.32, 2.30, and 2.19 eV per bond in , , and , about three times the Bi–X values of 0.81, 0.75, and 0.77 eV per bond. The Bi–O ICOBI remains nearly constant (0.391, 0.391, and 0.385), whereas the Bi–X ICOBI increases from 0.189 and 0.191 for S and Se to 0.203 for Te, indicating a modestly larger covalent contribution to Bi–Te bonding, in line with previous COHP analysis$66. Because the Bi–X ICOHP varies non-monotonically, this added covalency does not translate into monotonically stronger Bi–X bonds. The bonding analysis therefore establishes a stiff intralayer Bi–O framework and compliant interlayer Bi–X contacts in all three compounds but does not by itself single out the sulfide.
| Compound | ||||
| 2.322 | 0.391 | 0.808 | 0.189 | |
| 2.296 | 0.391 | 0.747 | 0.191 | |
| 2.186 | 0.386 | 0.769 | 0.203 |
Table S1. Bond-averaged and values for the nearest-neighbor Bi–O and Bi–X interactions in the fully relaxed structures. The values were averaged over 16 Bi–O and 16 Bi–X interactions in the 10-atom conventional cell. is reported in eV per bond, whereas is dimensionless.
What distinguishes the sulfide is lone-pair stereoreactivity. The deeper S $3p$ levels lie closer in energy to Bi than the Se or Te levels, strengthening the occupied Bi –S antibonding states. When the local symmetry is lowered, these states mix with empty Bi states, lowering the electronic energy and producing the one-sided density in Figure 1(b)$33,45. Coupling of this redistribution to antiparallel slab displacements produces the unstable -point mode and the double well in Figure 2(b). Static off-centering, direction-selective Bi fluctuations, and a shallow anharmonic potential thus form the microscopic basis for the anisotropic expansion and phonon transport analyzed below, paralleling lone-pair-driven anisotropy in strained 43 and in monolayer 67.
Anisotropic Thermal Expansion
Phase Verification and In Situ X-ray Diffraction
Room-temperature Raman spectroscopy and XRD verified the phase and orientation of the crystals used in the thermal measurements. The Raman spectrum (Figure S3) shows the characteristic mode at and the weaker mode near 68,69, consistent with the tetragonal phase.

Figure S3. Room-temperature Raman spectrum of the single crystal measured from the (001) surface. The prominent peak near 160 cm and the weaker broad feature at 350–365 cm are assigned to the and modes, respectively, consistent with tetragonal .
To benchmark the cross-plane CTE, we performed in situ XRD from 298 to 473 K on a crystal supported on a silicon wafer holder (Figure 3(c)), which suppresses background scattering from the heating stage. Only sharp (002), (004), (006), and (008) reflections appear (Figure 3(d)), confirming c-axis orientation, and the resolved Cu doublets indicate high crystalline quality. All (00) reflections shift continuously to lower on heating, and the room-temperature pattern is recovered after cooling. A linear fit of between 298 and 398 K, the window that brackets the 300 K PIMD value, gives , within 6.4% of the PIMD prediction of (§ 3.2).
Part of this difference may arise because the fit averages over 298–398 K, whereas PIMD provides the tangent coefficient at 300 K. Residual mechanical coupling between the crystal and the holder may also slightly modify the apparent expansion70;71.
Direction-Resolved Expansion from Path-Integral Molecular Dynamics
Isothermal–isobaric PIMD driven by neuroevolution potentials (NEPs) yields the equilibrium lattice parameters , , and from 100 to 700 K [Figure 3(e,Figure 3f)], from which the linear CTEs were obtained [Figure 3(a) and Table 1]. All lattice parameters increase monotonically with temperature. At 300 K, orthorhombic BiOS has , , and (), tetragonal BiOSe has and (), and tetragonal BiOTe has and (), following the chalcogen ionic radius (, , and ).
Table 1: Linear and volumetric CTEs () from PIMD and lattice thermal conductivities () at 300 K. Calculated conductivities are WTE values comprising the population term limited by 3ph and 4ph scattering and the coherence term. HNEMD values are listed for BiOS. The experimental BiOSe row gives from in situ XRD (linear fit over 298–398 K) and conductivities from dual-axis FDTR. The estimated electronic contribution is below 4% (§ 4.3), so the FDTR values represent .
| Compound | Space group | method | method | |||||||
| BiOS | 1.89 | 3.08 | 3.17 | 8.14 | PIMD | 1.04 | 1.07 | 0.50 | WTE | |
| BiOS | — | — | — | — | — | 1.19 | 1.23 | 0.61 | HNEMD | |
| BiOSe | 2.04 | 2.04 | 2.64 | 6.72 | PIMD | 1.83 | 1.83 | 0.95 | WTE | |
| BiOSe, exp. | — | — | — | XRD | FDTR | |||||
| BiOTe | 1.97 | 1.97 | 2.69 | 6.61 | PIMD | 1.75 | 1.75 | 0.51 | WTE |
Table 1.
The bonding hierarchy of Figure 2(c,Figure 2d) accounts for the general anisotropy . The stiff intralayer Bi–O framework restrains basal-plane expansion, whereas the weaker Bi–X contacts provide compliant interlayer connectivity. All linear coefficients are positive, and remains between 1.28 and 1.37. The electrostatic interlayer contact therefore behaves as a compliant but not membrane-like link, in contrast to graphite, where weakly coupled flexural modes with negative Grüneisen parameters contract the basal plane31.
The orthorhombic and tetragonal members differ markedly [Table 1]. In BiOS, the active lone pair breaks basal-plane fourfold symmetry, and the resulting Bi off-centering along [Figure 1(a,Figure 1b) and Figure 2(a)] produces and (), together with the largest cross-plane coefficient in the family (, ). The most expansive in-plane direction thus coincides with the off-centering direction and the largest Bi fluctuations (). In tetragonal BiOSe and BiOTe, centered Bi sites preserve , with in-plane coefficients of 2.04 and and cross-plane coefficients of 2.64 and ( and 1.37), respectively. The larger anisotropy of the telluride is consistent with the heavier Te lowering the frequencies of -polarized acoustic and interlayer breathing modes, which increases their thermal population at 300 K.
Inverted Volumetric Expansion and Mass-Substitution Control
At 300 K, PIMD yields , 6.72, and 6.61 for BiOS, BiOSe, and BiOTe, respectively [Figure 3(b), Figure S2, and Table S3], so the volumetric expansion decreases as the chalcogen becomes heavier and the cell larger. The decrease is concentrated between S and Se (17%), whereas the selenide and telluride differ by only 1.6%. The anomaly is therefore an enhanced expansion of the sulfide rather than a progressive trend. The chalcogenide series compiled in Figure 3(b), mainly binary and multinary semiconductors with ionic or mixed ionic–covalent bonding, predominantly show increasing from S to Se and Te, accompanied by lattice softening and weaker cohesion (Table S3). ZnX72 and AgGaX73 are exceptions, but BiOX shows a larger decrease, with and , outside the envelope of the prevailing trend.

Figure S2. Temperature-dependent volumetric coefficients of thermal expansion, , of (a) BiOS, (b) BiOSe, and (c) BiOTe obtained from PIMD, classical NPT-MD, and the quasi-harmonic approximation (QHA). (d) Magnitude of the NQE-induced correction, defined as , for the three compounds.
To separate atomic mass from bonding, we repeated the PIMD simulations of BiOSe with the Se mass (78.97 amu) replaced by the S mass (32.06 amu) while retaining the BiOSe potential. In classical statistical mechanics, equilibrium lattice parameters depend only on the potential energy surface, so atomic masses influence thermal expansion solely through nuclear quantum effects. The mass substitution changes by only 0.6% at 300 K relative to BiOSe (Table S3), whereas BiOS exceeds BiOSe by 21%. The enhanced expansion of BiOS therefore resides in its potential energy surface, shaped by lone-pair stereoactivity and the resulting orthorhombic distortion, and not in the lighter chalcogen. Section 2.4.2 relates this result to the heat-capacity-weighted Grüneisen parameters.
Classical molecular dynamics (MD) with the same potentials yields linear and volumetric CTEs 1.2–4.7% larger than PIMD while preserving the anisotropy and compositional trends (Table S3). This modest reduction reflects the quantum suppression of the heat capacity of the light O atoms in the [BiO] slabs, whose vibrations lie above 7 THz [Figure 5(d)], and it bounds the total quantum correction to , consistent with the small mass-substitution effect. By contrast, the QHA overestimates by about 17% for all three compounds (9.50, 7.89, and 7.70 ). Because the QHA retains only volume-dependent harmonic frequencies, it omits the explicit anharmonic renormalization at fixed volume that PIMD and MD include, which substantially reduces the expansion in this family.
The cross-plane CTEs of BiOX (2.64–3.17 ) exceed that of layered BiTe ()80 and are an order of magnitude larger than that of silicon (2.6 )74. Combined with the in-plane inequivalence of BiOS, this anisotropy concentrates stress at dielectric, contact, and substrate interfaces when BiOX is integrated into transistor channels or thermoelectric legs [Figure 1(c)], which we quantify in Section 2.5.
Lattice Thermal Conductivity
Directional and Compositional Trends
Using TDEP force constants with iterative solutions of the WTE, which combine the population conductivity limited by 3ph and 4ph scattering with the coherence conductivity , we computed and from 100 to 700 K [Figure 4(c,Figure 4d)]. All three compounds have ultralow , and in each the cross-plane direction, which expands most, conducts least. The compositional dependence is non-monotonic in both directions, with and [Figure 4(a) and Table 1]. This ordering departs from Leibfried–Schlömann–Slack (LSS) scaling 81,82, in which heavier atoms and larger cells lower the Debye temperature and sound velocities and thereby , and it differs from the monotonic decrease of . Cubic boron pnictides also violate mass scaling 83–85, but there the anomaly arises from mass-ratio-induced restrictions of the scattering phase space. In , the low in-plane conductivity of the sulfide instead reflects reduced in-plane group velocities combined with enhanced low-frequency anharmonic scattering associated with the active lone pair (§ 5).

Figure 4. Anisotropic thermal conductivity and the Bi₂O₂Se FDTR benchmark. (a) Thermal conductivities along the , , and axes of Bi₂O₂S, Bi₂O₂Se, and Bi₂O₂Te at 300 K. Solid bars denote WTE results, cross-hatched bars denote HNEMD results for Bi₂O₂S, and diagonally hatched bars denote FDTR results for Bi₂O₂Se. Error bars indicate the statistical uncertainties of HNEMD and the fitting uncertainties of FDTR. For tetragonal Bi₂O₂Se, the - and -axis components are equivalent and correspond to . (b) Representative FDTR phase lag of Bi₂O₂Se at 300 K versus modulation frequency. Symbols denote averaged data, and the solid line denotes the simultaneous fit yielding and W m⁻¹ K⁻¹. The inset shows the pump–probe configuration and the Pt/Bi₂O₂Se thermal model. (c) In-plane and (d) cross-plane thermal conductivities versus temperature, with Bi₂O₂S resolved along and in (c). Lines with symbols denote WTE results, dark-blue symbols with error bars denote HNEMD results for Bi₂O₂S at 300 and 500 K, and open symbols with error bars denote FDTR data for Bi₂O₂Se from 300 to 500 K. (e,f) Comparison of the present (e) in-plane and (f) cross-plane conductivities of Bi₂O₂Se with literature values 37;38;41;42;46–48;52–54 and 37;38;42;47;48;50;51;55, respectively.
Orthorhombic has the lowest in-plane conductivities in the family, and W m K at 300 K, despite S being the lightest chalcogen, and a cross-plane conductivity of 0.50 W m K ( and ). It combines the lowest mean in-plane group velocity in the family (0.466 km s) with the largest acoustic Grüneisen parameters and the strongest 3ph and 4ph scattering below 3 THz (§ 5.2 and § 5.3), so reduced velocities and short lifetimes act together.
Although in BiOS, its in-plane conductivities differ by only 3% (). Expansion and conduction weight the lattice dynamics differently. Expansion along reflects the strain sensitivity of the soft Bi modes polarized along the off-centering direction, whereas conduction integrates over the entire spectrum. The off-centering raises the mean -axis group velocity by 12% relative to the axis (0.494 versus 0.438 km s), and the 3ph-only population conductivity is correspondingly larger along . However, 4ph scattering removes 57.8% of the population conductivity along but only 29.5% along at 300 K (Table S4), because the soft Bi modes that carry -polarized heat current have the largest 4ph phase space (§ 5.3). The velocity advantage and the additional quartic resistance nearly cancel, so the basal-plane symmetry breaking appears in expansion but not in conduction.
| (K) | |||||||||
| 100 | 3.1024 | 2.4957 | 0.1750 | 4.4254 | 2.7847 | 0.1183 | 3.6537 | 1.1721 | 0.1359 |
| 200 | 1.7897 | 1.2705 | 0.2057 | 2.2681 | 1.3096 | 0.1667 | 1.6522 | 0.5329 | 0.1793 |
| 300 | 1.1821 | 0.8339 | 0.2097 | 2.1431 | 0.9043 | 0.1735 | 0.9467 | 0.3328 | 0.1741 |
| 400 | 1.0010 | 0.5788 | 0.2158 | 1.1892 | 0.5689 | 0.1944 | 0.6136 | 0.1977 | 0.1911 |
| 500 | 0.4550 | 0.3282 | 0.1931 | 0.5023 | 0.3260 | 0.1846 | 0.3413 | 0.1198 | 0.1670 |
| 600 | 0.2998 | 0.2168 | 0.1684 | 0.3407 | 0.2252 | 0.1641 | 0.1403 | 0.0617 | 0.1633 |
| 700 | 0.2515 | 0.1565 | 0.1682 | 0.3157 | 0.1793 | 0.1547 | 0.0942 | 0.0433 | 0.1578 |
Table S4.
Homogeneous nonequilibrium molecular dynamics (HNEMD) simulations with the same NEP, which include anharmonicity to all orders but treat the dynamics classically, provide an independent check for BiOS at 300 and 500 K [Figure 4(a,Figure 4c,Figure 4d)] and Table 1. At 300 K, HNEMD yields , , and W m K, 14–22% above the WTE values. Both methods reproduce nearly equal - and -axis conductivities, a cross-plane conductivity about half of the in-plane values, and a decrease of all three components from 300 to 500 K. The systematic offset is consistent with the classical mode occupations sampled by HNEMD, compared with the quantum statistics of the WTE. Even with this offset, the HNEMD in-plane values of BiOS remain well below the WTE value of BiOSe (1.83 W m K), so the ordering does not depend on the perturbative treatment of anharmonicity.
Tetragonal BiOSe has the highest conductivities in the family along both directions ( and W m K at 300 K, ). Its centered Bi sites preserve fourfold symmetry and confine the 4ph phase space relative to the sulfide, and its mean group velocities are the highest in the family ( and km s). In tetragonal BiOTe, remains high (1.75 W m K), close to the selenide, whereas drops to 0.51 W m K, matching the sulfide and giving the largest anisotropy in the family (). Weak scattering and an intermediate in-plane velocity (0.531 km s) sustain in-plane transport, whereas the heavy Te atoms (127.6 amu) and the compliant interlayer contact reduce to 0.249 km s, the lowest in the family, creating a cross-plane velocity bottleneck.
These conductivities (0.18–1.83 W m K over 300–700 K) place BiOX among ultralow-conductivity solids such as SnSe and BiTe 37–43. Their anisotropy (1.93–3.43) is moderate compared with that of vdW crystals such as graphite (~300) 86, MoS (~30) 87, WSe (~10) 88, -MoTe (7.6) 89, -MoTe (7.2), and -MoTe (5.6) 90,91, reflecting the electrostatic interlayer bonding. It is nonetheless large enough to shape heat removal in nanoscale devices, depending on how the layers are oriented relative to the heat sinks (§ 6).
Four-Phonon Scattering and Wigner Coherence
Four-phonon scattering and wave-like interband coherence act selectively along the two transport directions (Table S4 and S5). At 300 K, 4ph scattering reduces of BiOSe from 2.21 to 1.69 W m K in plane (23.6%) and from 1.64 to 0.87 W m K cross plane (46.8%), increasing the population anisotropy from 1.34 to 1.93. This selectivity arises because the cross-plane heat current is concentrated in low-frequency acoustic and optical modes below 3 THz, where large Grüneisen parameters and a large multiphonon phase space make higher-order scattering strongest (§ 5.3).
From 100 to 700 K, the in-plane 4ph reduction in BiOSe increases from 13.4% to 31.4% as multiphonon channels become thermally accessible, whereas the cross-plane reduction remains larger throughout, decreasing from 49.4% at 100 K to 46.8% at 300 K and 33.1% at 700 K (Table S5). Fourphonon scattering therefore enhances the population anisotropy over the entire temperature range, although its relative cross-plane effect weakens at high temperature as the spectral distribution of the heat current shifts.
The coherence conductivity provides an additional off-diagonal channel 60. The in-plane of BiOSe remains near 0.14 W m K between 100 and 700 K, whereas the cross-plane value increases by more than a factor of 2.5, from 0.038 at 100 K to 0.074 at 300 K and 0.096 W m K at 700 K. Thermal broadening of the optical linewidths increasingly bridges the small frequency spacings between adjacent cross-plane optical branches, enhancing their interband overlap.
Because of BiOSe decays approximately as under combined Umklapp and 4ph scattering (Figure S6), the coherence fraction grows with temperature, from 2.5% at 100 K to 7.9% at 300 K and 15.1% at 700 K in plane, and from 1.3% to 7.8% and 18.6% cross plane (Table S5). The coherence channel thus partially offsets the loss of population transport, and the total conductivity follows in plane and cross plane, consistent with the near- dependence measured here between 300 and 500 K (§ 4.3) and reported previously for BiOSe 48;52. The apparent near- scaling therefore results from two opposing contributions rather than from a single Umklapp-limited channel. Because increases with linewidth in this regime, strategies that lower by adding scattering, such as alloying or point-defect engineering, will be partly offset by the coherence channel at thermoelectric operating temperatures.

Figure S6. Lattice thermal conductivity of BiOS, BiOSe, and BiOTe replotted as a function of . The top row shows the in-plane components and the bottom row shows the cross-plane components. Dashed lines are linear fits to the calculated data. The near-linear dependence indicates that can be approximately described by an inverse relation with temperature over the investigated temperature range.
Across the series at 300 K, the cross-plane 4ph reduction decreases from 64.8% in BiOS to 46.8% in BiOSe and 27.4% in BiOTe, and the cross-plane decreases from 0.174 to 0.074 and 0.048 W m K. The sulfide thus combines the strongest higher-order resistance with the largest coherence channel, both reflecting the broad linewidths produced by its lone-pair-driven anharmonicity. At 700 K, supplies 78.5% of the cross-plane conductivity of BiOS because the population term has nearly vanished. In this regime, the linewidths of many low-frequency modes approach the Ioffe–Regel limit and the perturbative description of the double-well dynamics becomes less reliable, so this fraction should be regarded as semi-quantitative. By contrast, the similarly low of BiOTe at 300 K originates mainly from its low cross-plane group velocity, with only moderate 4ph suppression. BiOS is therefore predominantly scattering-limited, BiOTe is velocity-limited across the layers, and BiOSe combines relatively high velocities with moderate higher-order scattering.
Dual-Axis Thermoreflectance Benchmark
We measured and of cleaved BiOSe single crystals (thickness > 200 μm) from 300 to 500 K by dual-axis FDTR with a 61.5 nm Pt transducer (Figure 4(b), inset). Sweeping the pump modulation frequency from 0.5 to 10 MHz reduces the thermal penetration depth, , where is the thermal diffusivity, from about 0.5 to 0.1 μm cross plane and from about 0.7 to 0.2 μm in plane. Because these depths remain below the 1.94 μm beam radius, the response at high frequencies is dominated by cross-plane conduction and the Pt/BiOSe interface, whereas radial spreading contributes increasingly at low frequencies. The low thermal conductivity of Pt, relative to the more common Au, limits lateral spreading within the transducer that would otherwise mask the in-plane response of the sample. The resulting sensitivities to and have distinct frequency dependences (Figure S4), allowing both to be fitted simultaneously to a single phase spectrum, a procedure validated on anisotropic reference crystals.

Figure S4. Frequency-dependent sensitivity of the FDTR phase response of BiOSe to the in-plane thermal conductivity , cross-plane thermal conductivity , and thermal boundary conductance (TBC). The distinct frequency dependences of and enable their simultaneous determination from the same phase dataset.
At 300 K, the fit yields and W m K (Figure 4(b)), within 5.2% and 3.3% of the WTE values of 1.83 and 0.95 W m K, and the measured anisotropy ratio of 1.89 matches the calculated ratio of 1.93. FDTR measures the total conductivity. Using the reported electrical conductivity of nominally undoped bulk and a semiconductor Lorenz number, the Wiedemann–Franz law gives an electronic contribution of approximately 0.04 W m K, below 4% of the measured in-plane value between 300 and 500 K (Section 4.2.4). The measured values therefore represent . The slight theoretical overestimate lies within the experimental uncertainty and is consistent with point-defect scattering by Se vacancies or Bi antisites 93;94, which the calculations omit. Between 300 and 500 K, both measured components decrease approximately as and follow the calculated temperature dependence within the experimental uncertainty [Figure 4(c,Figure 4d)].
Previous reports show substantial axis-dependent scatter in the room-temperature of [Figure 4(e,Figure 4f)]. First-principles predictions span about 1.3–4.5 W m K in plane and 0.8–3.7 W m K cross plane 37;38;41;42;46;47, reflecting differences in force constants, scattering orders, BTE solution schemes, and the treatment of coherence. Measured values vary by about a factor of two, from 1.4 to 2.4 W m K in plane and from 0.9 to 2.3 W m K cross plane, across sample forms and measurement configurations 48;50–55. Previous single-crystal and nanostructure studies resolved only one axis 37;95–97 or were limited to sub-ambient temperatures 52, which prevented cross-checks between axes and coverage of the device-relevant window above 400 K. The present measurements determine both components on the same crystal from 300 to 500 K and agree with the WTE in magnitude, anisotropy, and temperature dependence. This agreement suggests that the spread of earlier predictions stems largely from the treatment of force-constant renormalization, 4ph scattering, and coherence, whereas the experimental spread reflects sample form and single-axis methods.
Anharmonic Lattice Dynamics and Transport Mechanisms
Temperature-Renormalized Phonon Spectra and Group Velocities
The TDEP dispersions at 100 and 700 K [Figure 5(a–Figure 5c)] contain no imaginary frequencies, confirming the finite-temperature dynamical stability of all three structures, including the phase of Bi₂O₂S. Because the force constants were sampled on a fixed reference lattice (§ 8.2), the renormalization shown here is the explicit anharmonic contribution at constant volume, which hardens most modes from 100 to 700 K, consistent with a dominant positive quartic contribution. Thermal expansion adds a quasiharmonic softening that these force constants exclude. The Raman-active modes of Bi₂O₂Se are reported to soften on heating at ambient pressure 54, indicating that the volume contribution outweighs the explicit shift for those modes. Because the lattice of Bi₂O₂Se expands by about 3% in volume between 300 and 700 K, the calculated conductivities at the highest temperatures should be regarded as upper estimates.
The projected phonon densities of states (PDOS) [Figure 5(d)] share a two-region structure across the family. Below 6 THz, heavy Bi motion dominates up to about 2.5 THz and chalcogen vibrations dominate between 2 and 5 THz, whereas the manifold above 7 THz consists of O vibrations within the slabs. From to , the maximum phonon frequency decreases from 16.5 to 12.8 THz, and the gap between the two regions widens from about 1.5 to 2.7 THz. Unlike in c-BAs and GaN83;98, this gap isolates only the O manifold and does not separate the acoustic branches from low-lying optical modes. Below 3 THz, acoustic and optical branches interpenetrate, satisfying the energy- and momentum-conservation conditions for intense multiphonon scattering. For , this low-frequency overlap is consistent with a recent combined DFT, calorimetry, and Raman study that assigned optical modes at 25.7 and 66.5 cm (0.77 and 1.99 THz) to shear-type interlayer vibrations and identified strong coupling of the lowest mode to the LA branches along –X and –A59.
Frequency-resolved group velocities [Figure 5(e–Figure 5h)] separate the dispersive acoustic and low-lying optical branches below about 3 THz from the flat, slow O manifold. The mean directional velocities summarize these spectra. For , and km s (in-plane mean 0.466 km s), and the in-plane means of and are 0.615 and 0.531 km s. This in-plane ordering, , matches that of . Because scales with the squared directional velocity, the low in-plane velocity of reinforces its strong scattering, whereas the weak scattering of partly compensates for its lower velocity relative to . The cross-plane means follow (0.522 km s) (0.298 km s) (0.249 km s). The nearly equal of and therefore arises from different limitations, namely strong anharmonic scattering in the sulfide and slow cross-plane propagation in the telluride.
Mode-Resolved Grüneisen Parameters
The volumetric mode Grüneisen parameters at 300 K (§ 8.3.3) vary most strongly below 3 THz [Figure 6(a)], where the lowest acoustic and low-lying optical states overlap, whereas the O-dominated manifold above about 6.7 THz shows a narrower distribution of moderate values. The low-frequency response is strongest in Bi₂O₂S and weakens toward Bi₂O₂Se and Bi₂O₂Te, consistent with the asymmetric lone-pair density and Bi off-centering of the sulfide [Figure 1(a,Figure 1b) and Figure 2]. The active lone pair increases the volume sensitivity of the soft Bi-dominated vibrations, whereas the symmetric Bi environments of the selenide and telluride suppress it.

Figure 6. Mode-resolved anharmonicity of Bi₂O₂X at 300 K. (a) Volumetric mode Grüneisen parameters versus frequency for Bi₂O₂S (top), Bi₂O₂Se (middle), and Bi₂O₂Te (bottom). (b) Heat-capacity-weighted atomic contributions to the Grüneisen parameters, resolved into acoustic (), low-frequency optical (), and high-frequency optical () groups.
The heat-capacity-weighted atomic decomposition [Figure 6(b)] locates the origin of these differences. The -group average decreases from 9.66 in Bi₂O₂S to 7.66 in Bi₂O₂Se and 3.97 in Bi₂O₂Te, dominated by Bi, whose partial contribution decreases from 7.52 to 6.77 and 3.56, while the chalcogen contributions are 1.97, 0.79, and 0.33 and the O contributions remain below 0.17. The group varies weakly and non-monotonically (3.28, 3.19, and 3.42), with the Bi contribution decreasing slightly from 2.05 to 1.89 and 1.83 and the chalcogen contribution increasing from 1.12 to 1.26 and 1.57. The group is dominated by O, whose contribution increases slightly from 1.96 to 2.08 and 2.16, with Bi and chalcogen contributions below 0.06. The compositional variation of the low-frequency anharmonicity therefore resides primarily on the Bi sublattice, whereas the high-frequency response belongs to the O sublattice.
These results provide a microscopic interpretation of the PIMD expansion. In the Grüneisen relation , where in practice the sum runs over irreducible points with symmetry weights , the relevant quantity is the heat-capacity-weighted average . At 300 K, decreases from 3.55 in Bi₂O₂S to 3.31 in Bi₂O₂Se and 3.04 in Bi₂O₂Te, consistent with . In the classical limit, depends only on the force constants, because the mass contribution to the product of phonon frequencies at each is independent of volume. The heat-capacity weighting introduces the only mass dependence, consistent with the small effect of mass substitution in PIMD (§ 3.3). The and groups together account for about 80.0%, 77.7%, and 74.5% of in the three compounds, identifying the low-frequency sector as the principal source of expansion, although the contribution remains appreciable at 300 K. The nearly equal of and despite their different suggests partial compensation by a smaller bulk modulus in the telluride.
The Grüneisen parameters also gauge the anharmonicity relevant to heat transport, although they do not determine the scattering rates on their own. Within LSS-type models, stronger anharmonicity lowers approximately as 81;82. The large -group Grüneisen parameter of , together with its lowest in-plane velocity, accounts for its lowest in-plane conductivity. combines a moderate -group response, the smallest average, and the highest velocities along both directions, yielding the largest conductivity. has the weakest -group anharmonicity but a slightly larger response and lower velocities than , particularly across the layers, so its conductivity remains below that of the selenide. The following section tests these interpretations against mode-resolved scattering rates and phase spaces.
Spectral Transport, Scattering Rates, and Coherence
The spectral accumulation of [Fig. 7(a–c)] shows how these modes build up the conductivity. In all three compounds, the cumulative conductivity rises steeply below 4 THz, plateaus across the phonon gap between about 5 and 7 THz, and rises again in the O-dominated manifold above 7 THz. The low-frequency rise coincides with the window below 3 THz where acoustic and low-lying optical branches overlap [Fig. 5] and mode Grüneisen parameters peak [Fig. 6(a)]. The branch-resolved contributions in the insets are defined in the 10-atom cell used for all calculations. This cell is primitive for but conventional for and , in which zone folding assigns part of the primitive acoustic dispersion to optical branch indices. Branch fractions are therefore interpreted only within each compound, and comparisons across compounds rely on the frequency-resolved accumulation, which does not depend on the choice of cell. In , the three acoustic branches carry only 17% of (TA, 6%; TA′, 6%; LA, 5%), indicating strongly suppressed acoustic transport.
The weighted phase spaces for three-phonon () and four-phonon () processes [Fig. 7(d–f)] identify the origin of this suppression. Both phase spaces expand with temperature while preserving the two-region structure imposed by the phonon gap. The largest compositional contrast occurs below 3 THz, where of reaches , exceeding the values of () and () by one and two orders of magnitude, respectively. This enlarged phase space accompanies the soft Bi-centered vibrations identified in Figure 6(b), the asymmetric lone-pair environment in Figure 1(b), and the double-well potential in Figure 2(b), and it is partly enabled by the broken basal-plane symmetry. Consistently, the mode-resolved rates at 300 K [Figure 7(g–Figure 7i)] show that below 3 THz has the highest 3ph and 4ph rates, the lowest, and an intermediate response. Although 4ph rates are smaller than 3ph rates for most high-frequency modes, 4ph processes act most strongly on the long-lived modes below 3 THz and therefore supply the additional resistance that suppresses cross-plane population transport (Table S4).

Figure 7. Spectral heat transport and anharmonic scattering in . (a–c) Spectral thermal conductivity (shaded areas, left axis) and its cumulative integral (solid lines, right axis) along the principal axes of (a) , (b) , and (c) . Insets show branch-resolved contributions of the transverse acoustic (TA, TA′), longitudinal acoustic (LA), and optical (OP) branches, defined in the 10-atom cell, which is primitive for and conventional for and . (d–f) Weighted three-phonon (upper panels) and four-phonon (lower panels) scattering phase spaces at 100, 300, and 700 K for (d) , (e) , and (f) . (g–i) Mode-resolved 3ph and 4ph scattering rates at 300 K for (g) , (h) , and (i) , compared with the Wigner limit (black dashed line) and the Ioffe–Regel limit (green dash-dotted curve).
Comparing the scattering rates with the Wigner and Ioffe–Regel limits distinguishes the two transport channels [Figure 7(g–Figure 7i)]. Below about 5 THz, scattering rates remain well below the Wigner limit (), so transport is dominated by the population channel of propagating wavepackets. In the dense optical manifold above 7 THz, 3ph rates approach or exceed the interbranch spacings () while remaining below the Ioffe–Regel limit (), the regime in which the off-diagonal coherence term becomes significant 60. The O-dominated manifold of the slabs is therefore the main source of . At 300 K, however, accounts for less than 8% of the conductivity of , so most of the optical contribution at room temperature is carried by the population channel. The coherence share grows as the linewidths broaden with temperature (§ 4.2).
Together, these results trace both family trends to the Bi lone pair. In , lone-pair activation drives static Bi off-centering and a double-well potential, which raise the heat-capacity-weighted Grüneisen parameter and the volumetric expansion and open a large low-frequency multiphonon phase space. Combined with the lowest in-plane velocity, this makes the lightest compound the poorest in-plane conductor. In , centered Bi sites and weak scattering sustain a high , whereas the lowest cross-plane velocity limits . combines moderate higher-order scattering with the highest velocities along both directions and conducts best. In all three compounds, the dense optical manifold of the slabs supplies a coherence channel whose share grows with temperature.
Device-Scale Self-Heating and Thermoelastic Response
To translate these material-level differences to device operation, we simulated steady-state phonon transport and the resulting thermoelastic response of a FinFET channel with a fin width of 12 nm, a height of 60 nm, and a length of 50 nm [Figure 8(a)]. Following fabricated FinFETs, in which the layers stand vertically;[11, 14] the axis was aligned with the fin width, so that the channel and fin-height directions both lie in the plane. The fin is covered on its top and sidewalls by a gate dielectric and rests on a substrate, and the source and drain contacts act as 300 K thermalizing boundaries (§ 8.4). Uniform volumetric heat-generation rates of and W m were applied in the drain-side upper region of the fin, corresponding to total powers of approximately 3.94 and 16.34 W per fin, or about 0.030 and 0.124 mW m normalized by the effective gate width nm. and represent the normal and elevated loads, respectively. All results in Figure 8 were obtained with modal phonon properties evaluated at 300 K, including isotope, 3ph, and 4ph scattering.

Figure 8. Device-scale phonon transport and thermoelastic response of BiOX FinFET channels. (a) Device geometry, finite-element mesh, and boundary conditions for the 12 nm wide, 60 nm high, and 50 nm long fin. The [BiO] layers stand vertically, with the crystallographic axis along the fin width. The fin rests on a substrate. The gate dielectric covers its top and sidewalls and is included explicitly only in the thermoelastic calculations. Thermalizing boundaries at 300 K are applied at the source and drain faces and at the bottom of the substrate. (b) Peak hotspot temperatures, , from the phonon BTE and maximum thermal-mismatch strains, , of BiOS, BiOSe, and BiOTe under the normal load () and the elevated load (), calculated with phonon properties evaluated at 300 K. (c) Steady-state BTE temperature field and (d) corresponding mismatch-strain field of BiOSe under the normal load. (e) Fourier temperature field and (f) corresponding mismatch-strain field under the same loading, using the same thermal model geometry as (c) and the same thermoelastic model geometry as (d).
Composition-Dependent Self-Heating
Under both loads, the peak hotspot temperatures follow [Figure 8(b)]. At the elevated load, reaches 659.43 K in , 516.14 K in , and 567.09 K in , so reduces the temperature rise above 300 K by 39.9% relative to and by 19.1% relative to BiOTe. With vertically oriented layers, heat leaves the hotspot toward the contacts along the channel and toward the substrate along the fin height, both within the plane, so the ordering follows . BiOS, whose in-plane conductivity is about 40% lower than that of BiOSe, accumulates the most heat. BiOTe nearly matches the in-plane conductivity of BiOSe, whereas its low limits lateral heat redistribution across the 12 nm fin width. The temperature field of BiOSe under the normal load [Figure 8(c)] peaks in the drain-side upper fin and decays toward the contacts and the substrate, and the corresponding heat-flux distribution is shown in Figure S7.

Figure S7. Steady-state heat-flux-magnitude distributions in the fin-on-substrate structure under (a) normal-load and (b) elevated-load volumetric heat-generation rates of and , respectively, obtained from the GiftBTE simulations.
Thermal-Mismatch Strain
The temperature fields were transferred to a thermoelastic model that additionally included the gate dielectric and used the temperature-dependent PIMD CTE tensors (§ 8.4). We report the thermal-mismatch strain , defined as the elastic part of the total strain, which vanishes for a uniformly heated, unconstrained solid. It arises from the incompatibility of nonuniform expansion within the fin and from the expansion mismatch between the channel and the dielectric and substrate. is largest in the heated upper fin, where the temperature rise and the mismatch with the dielectric coincide [Figure 8(d)]. Under the elevated load, its maximum, , reaches 4.14%, 1.82%, and 2.48% in BiOS, BiOSe, and BiOTe, respectively, following the same ordering as .
The vertical-layer orientation determines which expansion coefficients control this strain. At the sidewall interfaces, whose normal is the axis, the tangential mismatch is set by the in-plane coefficients along the channel and the fin height, whereas acts across the 12 nm width and along the top interface. BiOS combines the largest temperature rise with along the fin height, 1.63 times along the channel, producing the strongest and most anisotropic sidewall mismatch. BiOSe and BiOTe have nearly equal in-plane coefficients (2.04 and 1.97 ), so the lower strain of BiOSe originates mainly from its smaller temperature rise rather than from a smaller expansion coefficient.
Non-Fourier Transport and Model Sensitivity
To isolate non-Fourier effects, the BiOSe device under the normal load was also simulated with anisotropic Fourier conduction [Figure 8(e,Figure 8f)]. The Fourier model used the bulk conductivity tensor computed from the same modal phonon properties, together with identical thermal geometry, mesh, material properties, heat source, and boundary conditions, so that both models share the same bulk limit (§ 8.4). Both temperature fields were subsequently applied to the same thermoelastic model including the gate dielectric. The Fourier model predicts and , compared with 346.2 K and 0.66% from the BTE-based calculation. Two sub-continuum effects account for the higher BTE temperature. First, phonons with mean free paths comparable to or longer than the 12 nm fin width scatter diffusely at the channel–dielectric sidewalls, which lowers the effective in-plane conductivity of the fin along both heat-removal paths below its bulk value. Second, near the thermalizing source and drain boundaries, the phonon distribution departs from local equilibrium, producing temperature jumps that act as additional ballistic resistances absent from the Fourier description. Fourier-based estimates therefore underestimate both the self-heating and the resulting mismatch strain in sub-15 nm BiOX fins 20–22.
The BTE solver treats particle-like transport within the relaxation-time approximation and uses a fixed set of modal properties that is not updated with the local temperature, which limits its quantitative accuracy at large temperature rises 21;99;100. To assess this approximation, the elevated-load calculations were repeated with property sets evaluated at 500 and 700 K (Table S7). Higher-temperature property sets raise for all three compounds, so the absolute elevated-load temperatures depend on the property-evaluation temperature. The ordering , however, is preserved for all three sets. The BTE also omits the coherence channel. Because the coherence contribution is largest in BiOS (§ 4.2), this omission affects the sulfide most. Even the full WTE in-plane conductivity of BiOS, which includes coherence, remains about 42% below that of BiOSe, so the compositional ordering does not depend on this approximation.
Table S7: Peak hotspot temperatures (K) of BiOX fins under the elevated load ( W m), calculated using fixed phonon-property sets evaluated at , 500, and 700 K.
| Material | K | K | K |
| BiOS | 659.43 | 755.65 | 826.91 |
| BiOSe | 516.14 | 616.73 | 700.64 |
| BiOTe | 567.09 | 631.35 | 751.23 |
Table S7.
Overall, the device simulations translate the material-level trends into a consistent thermomechanical response. With vertically oriented layers, both heat removal and sidewall mismatch are governed mainly by in-plane properties, whereas the low cross-plane conductivity, which would dominate in horizontally stacked channels, plays a secondary role. Under identical geometry and loads, the high and isotropic in-plane expansion of BiOSe therefore yield the lowest hotspot temperature and mismatch strain among the three compounds. These thermoelastic calculations compare temperature-induced deformation across compositions and are not intended to predict fracture, fatigue, or device lifetime.
Conclusions
We have combined path-integral molecular dynamics, Wigner transport calculations based on temperature-dependent force constants with three- and four-phonon scattering, in situ X-ray diffraction, and dual-axis frequency-domain thermoreflectance to resolve the anisotropic thermal expansion and lattice thermal conductivity of BiOX () and to trace both to the stereochemical activity of the Bi lone pair. In BiOS, the strong Bi 6s–S 3p antibonding interaction activates the lone pair, drives in-plane Bi off-centering along , and produces an antipolar double-well potential, whereas centered Bi sites preserve tetragonal symmetry in BiOSe and BiOTe.
This lone-pair contrast overrides the mass and softening scaling followed by most chalcogenide series. The volumetric thermal expansion coefficient at 300 K decreases from in BiOS to and in BiOSe and BiOTe, although the unit cell expands by 13% from S to Te. Substituting the S mass into BiOSe changes its expansion by only 0.6%, so the enhanced expansion of the sulfide originates from its potential energy surface, consistent with its larger heat-capacity-weighted Grüneisen parameter (3.55, compared with 3.31 and 3.04). Cross-plane expansion exceeds in-plane expansion throughout the family (–). In BiOS, the -axis coefficient is 1.63 times the -axis value, whereas the in-plane conductivities differ by only 3% because the higher -axis group velocity is offset by stronger four-phonon scattering. The lattice thermal conductivity peaks in BiOSe (1.83 and 0.95 W m K in plane and cross plane at 300 K). BiOS is limited by strong anharmonic scattering below 3 THz and low in-plane group velocities, whereas BiOTe is limited by the lowest cross-plane group velocity in the family and shows the largest conduction anisotropy (3.43). In BiOSe, four-phonon scattering removes 46.8% of the cross-plane population conductivity at 300 K, and the coherence channel supplies 18.6% of the cross-plane conductivity at 700 K, partly offsetting the decay of population transport and producing the observed near- temperature dependence. X-ray diffraction () and thermoreflectance from 300 to 500 K ( and W m K at 300 K) on BiOSe single crystals agree with these predictions within 6.4%.
At the device level, phonon Boltzmann transport simulations of 12 nm wide FinFET channels with vertically oriented layers, a gate dielectric, and a substrate predict higher hotspot temperatures than Fourier diffusion (346.2 K versus 341.37 K for BiOSe under the normal load), showing that sub-continuum transport must be included in self-heating estimates at these dimensions. Because heat removal and sidewall mismatch in this orientation are governed mainly by in-plane properties, BiOSe develops the lowest hotspot temperature and thermal-mismatch strain under identical loads, followed by BiOTe and BiOS. These results identify lone-pair stereoactivity as a chemical control parameter for both thermal expansion and heat conduction in non-van der Waals layered semiconductors. They also provide validated anisotropic property sets for the thermal and thermomechanical design of BiOX transistors, photodetectors, and thermoelectric devices.
Methods
Computational Methods
First-Principles Calculations
Density functional theory (DFT) calculations were performed with the Vienna ab initio simulation package (VASP)101 using the Perdew–Burke–Ernzerhof (PBE) exchange-correlation functional102. Structural optimizations used a plane-wave cutoff of 520 eV and convergence thresholds of eV for the total energy and eV Å for the atomic forces. Reference configurations for potential training were generated by ab initio molecular dynamics (AIMD) between 50 and 750 K in supercells with -point sampling. To improve the transferability of the potentials, uniform triaxial strains from to and uniaxial strains from to were applied to the equilibrium structures. The energies, forces, and virials of the selected configurations were then recalculated with the same cutoff and a -centered -point mesh.
Neuroevolution Potentials and Temperature-Dependent Force Constants
Composition-specific NEP models 103;104 were trained for BiOS, BiOSe, and BiOTe with the GPUMD package 105. These potentials provide energies, forces, and virials with near-DFT accuracy at a small fraction of the computational cost [2] [42, 50, 55, 107]. Each NEP expresses the site energy of atom as the output of a single-hidden-layer feed-forward neural network whose input is a rotation- and translation-invariant descriptor vector constructed from Chebyshev-based radial and angular components:
where and are the numbers of hidden-layer neurons and descriptor components, and the weights and biases were optimized with the separable natural evolution strategy. Atomic forces and virials follow from derivatives of the total energy105.
For each compound, the training database was assembled from 119 AIMD trajectories, comprising temperature-ramped trajectories from 50 to 750 K and isothermal trajectories from 100 to 700 K in 100 K steps. The final dataset of each compound contained 884 configurations (224 uniaxially strained and 660 triaxially strained) and was randomly divided at the configuration level into training (90%) and test (10%) sets. Each NEP used radial and angular cutoffs of 7 and 5 Å, , , 50 hidden neurons, a batch size of 1000, and generations. The root-mean-square errors on the training and test sets are and eV atom for energies, and eV Å for forces, and and eV atom for virials, respectively (Figure S1).

Figure S1. Training and validation performance of the NEP model for Bi₂O₂X (X = S, Se, and Te). Panels (a–d), (e–h), and (i–l) present the results for Bi₂O₂S, Bi₂O₂Se, and Bi₂O₂Te, respectively. Panels (a, e, i) show the evolution of the total loss and the root mean square errors (RMSEs) of energy, force, and virial for the training and testing datasets with training generation. Panels (b, f, j), (c, g, k), and (d, h, l) compare the NEP-predicted energies, virials, and forces with the corresponding DFT reference values, respectively. The close agreement between the NEP predictions and DFT data for both datasets demonstrates the accuracy of the trained models.
Temperature-dependent second-, third-, and fourth-order interatomic force constants (IFCs) were extracted within the TDEP framework63;64 using hiPhive 108 from configurations sampled by classical NEP-driven molecular dynamics (MD). At each temperature, a supercell of the 10-atom cell (320 atoms), fixed at the DFT-relaxed lattice parameters, was equilibrated for 10 ps in the canonical () ensemble with a 1 fs time step, followed by a 50 ps production run from which 100 configurations were sampled at 0.5 ps intervals. The IFCs of all three orders were fitted jointly within a single cluster space with cutoffs of 7.5, 6.5, and 4.5 Å, respectively, by standardized ordinary least squares without regularization, subject to space-group and permutation symmetries and the acoustic sum rules. Because the same reference lattice was used at all temperatures, the TDEP force constants contain the explicit anharmonic renormalization at constant volume but not the effect of thermal expansion. PIMD was used only for equilibrium structural and thermodynamic properties (§ 8.3.4) and not for the transport force constants.
Electron Localization, Bi Displacements, Frozen-Phonon Energetics, and Bonding Analysis
The electron localization function (ELF) was calculated from the converged Kohn–Sham wavefunctions of the fully relaxed ground-state structures with the DFT settings described above. Two-dimensional slices on the (110) plane are plotted on a common scale from 0 to 1 for all three compounds. The in-plane Bi off-centering was defined as the magnitude of the -plane projection of the displacement between the Bi position in the relaxed structure and the corresponding site of the symmetry-matched reference. The Bi mean-square displacement tensor was calculated from the second-order IFCs with Phonopy65 from 100 to 700 K in 50 K steps. Its diagonal elements, with , 2, and 3 corresponding to the , , and directions, were averaged over the four Bi atoms of the 10-atom cell.
The energy landscape of the -forming distortion was evaluated with static frozen-phonon calculations at 0 K. Symmetry-constrained reference structures were first optimized. The distortion coordinate was identified by projecting the displacement between the and structures of onto the phonon eigenvectors of the reference, and the primitive-cell -point mode with the largest projection was selected. The corresponding modes of and were identified from eigenvector-subspace overlaps. This doubly degenerate -point mode folds to in the 10-atom conventional cell, in which its eigenvector was normalized as . The mass-weighted mode coordinate was defined through
where , , and are the displacement, eigenvector component, and mass of atom along direction . Total energies of the displaced structures were computed without ionic or cell relaxation, at the fixed lattice parameters of the reference, using the PBE functional, the Bi_d projector augmented-wave potential, a cutoff of 550 eV, and a -centered -point mesh. The energies were referenced to , divided by two to obtain values per formula unit, symmetrized between and , and fitted for Å amu to
where is the harmonic curvature at the high-symmetry configuration and is the leading anharmonic coefficient along the selected coordinate. For , the minima lie at , and the barrier height is . These quantities describe the static energy surface along a single mode and exclude zero-point and thermal fluctuations, cell relaxation, and coupling to other modes.
Chemical bonding was analyzed with LOBSTER109–113 from static PBE wavefunctions of the relaxed ground-state structures, using the pbeVaspFit2015 basis set with Bi 5, 6, and 6, O 2 and 2, and chalcogen sn orbitals. Projected crystal orbital Hamilton populations (pCOHP) and projected crystal orbital bond indices (pCOBI)114 were calculated from to 10 eV relative to , and their cumulative energy integrals are denoted IpCOHP() and IpCOBI(). For each compound, the results were averaged over the 16 nearest-neighbor Bi–O and the 16 Bi–X contacts of the 10-atom cell. pCOHP and IpCOHP are plotted with reversed sign, so that positive values denote bonding contributions, whereas pCOBI and IpCOBI retain their original sign convention. The values and are reported as the bond-averaged and ICOBI (Table S1).
Lattice Thermal Conductivity
The lattice thermal conductivity was evaluated with the WTE60, which generalizes the linearized BTE by including the off-diagonal elements of the heat-flux operator and yields . The population term, which is the particle-like limit of the WTE, was obtained from the linearized BTE as
where is the unit-cell volume, is the number of points, and , , and are the Bose–Einstein occupation, frequency, and group velocity of mode with wavevector and branch index . mean free displacement contains the correction , which couples mode to all other modes through the 3ph and 4ph collision operator and was updated self-consistently. Setting recovers the relaxation-time approximation (RTA)116. The coherence term, which describes wave-like interband transport between quasi-degenerate branches with overlapping linewidths, was evaluated as
where is the total scattering rate of mode , and is the velocity-operator element between branches and at the same wavevector60, with the polarization vector and the dynamical matrix117.
Population conductivities were obtained from full iterative solutions of the linearized BTE with ShengBTE and FourPhonon116;118;119, using the TDEP force constants at each temperature. The 3ph-only conductivity, , was calculated on an mesh, and the conductivity including 3ph and 4ph scattering, , was calculated on a mesh with the complete iterative collision operator. Isotope scattering was included for natural isotopic abundances. Long-range dipole–dipole interactions were included in the dynamical matrix through the high-frequency dielectric tensor and Born effective charges calculated with density functional perturbation theory65. In the transport calculations, group velocities were obtained from the -derivative of the dynamical matrix. The frequency-resolved velocities in Figure 5(e–Figure 5h) were evaluated from the same TDEP dispersions with VASPKIT115.
The mean directional velocity was defined as , averaged over all branches with -point multiplicities , and the in-plane mean of BiOS was taken as . The coherence term was evaluated on a -centered mesh with an implementation based on Phonopy and Phono3py, using linewidths that include isotope, 3ph, and 4ph scattering computed on the same mesh. The total conductivity was obtained as .
Homogeneous Nonequilibrium Molecular Dynamics
HNEMD simulations of BiOS were performed with GPUMD105;106 using the same NEP. The 320-atom structural model was replicated , yielding a simulation cell containing 16,000 atoms. The system was equilibrated at the target temperatures (300 and 500 K) and zero pressure in the isothermal–isobaric ensemble for 100 ps with a 1 fs time step. The ensemble was then switched to NVT, and a driving force with Å (m), chosen within the linear-response regime, was applied separately along the , , and axes. The thermal conductivity was obtained from the running average of the induced heat current over the 1.5 ns production trajectory. Reported values and uncertainties are the means and standard errors of five independent simulations for each direction. Because HNEMD samples classical dynamics, it includes anharmonicity to all orders but uses classical mode occupations.
Mode Grüneisen Parameters and Atomic Projections
Mode-resolved volumetric Grüneisen parameters were calculated from the 300 K TDEP force constants on the mesh used in the FourPhonon calculations. For mode ,
which was evaluated from the third-order IFCs with the implementation in FourPhonon. The complex phonon eigenvectors were exported with a modified ShengBTE build using the same second-order IFCs, crystal structures, and mesh. The participation of atomic species in mode was calculated as
where is the eigenvector of the mass-weighted dynamical matrix normalized as , and the contributions of the four Bi, two chalcogen, and four O atoms of the 10-atom cell were summed for each species. The heat-capacity-weighted partial Grüneisen parameter of species within mode group was evaluated as
where is the symmetry multiplicity of the irreducible point and
The partial contributions satisfy , and the overall average was obtained by extending the sums in eq 8 to all modes. The frequency-ordered branches were divided into the lowest three branches (), branches 4–18 (), and the O-dominated branches 19–30 (). Because the groups are defined by frequency ordering in the same 10-atom cell, each contains the same fraction of vibrational states in all three compounds (10%, 50%, and 40%, respectively). In the tetragonal compounds, for which the 10-atom cell is the conventional cell, the group therefore comprises predominantly the lower-frequency half of the primitive-cell acoustic states, and the folded acoustic states at higher frequency fall in . All folded branches were included with their associated -point weights.
Thermal Expansion and Heat Capacity
Thermal expansion was calculated by PIMD driven by the NEPs, with classical isothermal–isobaric MD and the QHA used for comparison61;62. The GPUMD PIMD integrator applies Langevin thermostats to the centroid and internal ring-polymer modes and a Berendsen barostat for pressure control, whereas the classical simulations used Berendsen temperature and pressure coupling. The PIMD simulations used 48 beads, which exceeds at 100 K by a factor of six, and a supercell containing 10,240 atoms. Separate simulations were performed at zero external pressure from 100 to 700 K in 50 K steps, each comprising steps of 1 fs, of which the first 70 ps were discarded for equilibration and the remaining 30 ps were used for structural averaging. The classical simulations used the same supercell and time step with steps per temperature, of which 700 ps were used for equilibration and 300 ps for averaging. Lattice parameters and cell volumes were sampled every 0.1 ps.
To separate the roles of atomic mass and bonding, additional PIMD simulations of BiOSe were performed with the Se mass (78.97 amu) replaced by the S mass (32.06 amu), while the BiOSe NEP, supercell, and simulation protocol, including five independent trajectories, were retained. Because the potential energy surface is unchanged, any difference relative to BiOSe arises solely from the mass dependence of nuclear quantum effects.
The linear and volumetric CTEs were calculated as
where is the lattice parameter along axis and is the cell volume. The derivatives were evaluated by central finite differences with a 50 K spacing from 150 to 650 K and by second-order forward and backward differences at 100 and 700 K, respectively, yielding tangent coefficients at each temperature. Five independent simulations were performed for each condition, and the reported CTEs and uncertainties are the means and 95% confidence intervals of the five estimates.
The constant-pressure heat capacity of used in the FDTR analysis was obtained from the PIMD simulations by differentiating a cubic polynomial fit to the temperature-dependent mean enthalpy and was converted to a volumetric heat capacity with the PIMD density. Because PIMD includes nuclear quantum effects, this heat capacity captures the suppressed contribution of the high-frequency O modes and agrees with reported measurements (Figure S5).

Figure S5. Temperature-dependent molar heat capacities at constant pressure, , of , , and obtained from the temperature derivatives of cubic-polynomial fits to the mean enthalpies from NEP-driven GPUMD simulations. Open stars denote the experimental values for reported by Zich et al. [2].
For the QHA, the temperature-dependent equilibrium volume was obtained by minimizing the Helmholtz free energy over 11 volumes spanning to of the reference volume. The harmonic force constants were constructed from DFT forces, and the vibrational free energies were evaluated with Phonopy 65;120;121. Because the QHA was restricted to isotropic volume changes, it yields only , which was compared with the PIMD and classical MD values to assess the role of explicit anharmonicity.
Device-Scale Phonon Transport and Thermoelastic Simulations
Steady-state phonon transport in a three-dimensional FinFET channel was simulated with the non-gray phonon BTE solver GiftBTE20;21, which accounts for ballistic transport and phonon–boundary scattering. GiftBTE solves the BTE within the RTA using the discrete ordinates method. The mode-resolved frequencies, group velocities, heat capacities, and relaxation times of each compound were taken from the ShengBTE and FourPhonon calculations described above, with scattering rates including isotope, 3ph, and 4ph processes. The angular space was discretized into 16 directions and the phonon properties into 15 mean-free-path bands. The BTE describes particle-like transport and therefore omits the coherence contribution. Both the normal- and elevated-load simulations used property sets evaluated at 300 K, and additional elevated-load simulations with property sets evaluated at 500 and 700 K were used to assess the sensitivity to the property-evaluation temperature (Table S7).
The device geometry and coordinate convention were adapted from previous FinFET electrothermal models and the GiftBTE FinFET benchmark21;122. The , , and directions correspond to the channel, fin-width, and fin-height directions, respectively. The fin had a width nm, a height nm, and a length nm. Following the vertical-layer orientation of fabricated FinFETs[11, 14], the crystallographic , , and axes were aligned with , , and , respectively, so that the layers lie parallel to the sidewalls. For tetragonal and , the and axes are equivalent. In the thermoelastic model, a 5.5 nm thick gate dielectric covered the top and both sidewalls within the gate region, and the fin rested on a 30 nm thick substrate. The source and drain were represented by boundary conditions at the channel end faces.
A semi-elliptical cylindrical heat-generation region spanning the fin width was placed beneath the top surface near the drain. Its cross-section lay in the – plane, with a channel-direction semi-axis of 10 nm and a depth semi-axis of 15 nm, and its center was located on the top surface at , with measured from the channel end adjacent to the hotspot. Uniform volumetric heat-generation rates of and , corresponding to total powers of approximately 3.94 and 16.34 per fin based on the discretized heat-source volume, were applied for the normal- and elevated-load cases, respectively. The generated heat was distributed among phonon modes in proportion to their modal heat capacities21;123.
For the thermal calculations, the fin and supporting base were treated using the same phonon-property set, with the bottom of the base maintained at 300 K. Thermalizing boundaries at 300 K were also applied to the channel faces adjoining the source and drain contacts21;122. The fin–dielectric interfaces were represented by adiabatic, diffusely reflecting phonon boundaries, following the GiftBTE FinFET benchmark21. The lateral faces of the base were treated as symmetry planes between neighboring fins 21;123–125. For the Fourier benchmark, the fin and supporting base were assigned the bulk conductivity tensor computed within the RTA from the same modal properties, and the thermal geometry, mesh, material properties, heat source, and boundary conditions were kept identical, so that the BTE and Fourier models share the same bulk limit. The same unstructured mesh was used for the BTE and Fourier calculations. Thermal and thermoelastic meshes contained 16,930–125,289 and 22,370–157,457 nodes, respectively. Relative to the finest-mesh results, the maximum deviations were 0.41% in and 4.41% in (Table S6).
Table S6: Mesh-convergence results for the sequential GiftBTE–COMSOL simulations of BiOSe under the normal load ( W m) and elevated load ( W m). and denote the peak hotspot temperature and maximum thermal-mismatch strain, respectively.
| Normal load, | Elevated load, | ||||
| (K) | (%) | (K) | (%) | ||
| 16930 | 22370 | 346.2 | 0.66 | 514.0 | 1.82 |
| 32630 | 42466 | 346.3 | 0.65 | 515.1 | 1.83 |
| 65031 | 82951 | 346.7 | 0.68 | 514.7 | 1.85 |
| 125289 | 157457 | 346.8 | 0.68 | 516.1 | 1.85 |
Table S6.
A sequential, one-way thermal–mechanical approach was adopted. The calculated temperature fields were transferred to the COMSOL Multiphysics thermoelastic mesh, which additionally included the gate dielectric as a bonded elastic domain to account for elastic constraint and thermal-expansion mismatch. The dielectric temperature was prescribed from the adjacent fin surface and assumed uniform through its thickness. This approximation neglects heat transport through the gate stack and deformation-induced changes in thermal transport.
Stationary, small-strain linear-elastic calculations were performed using these temperature fields as thermal loads. The channel was described by the anisotropic elastic stiffness tensor of each compound, taken from the first-principles calculations of Hu et al. 66, whereas the dielectric and substrate were treated as isotropic solids with the Young’s moduli, Poisson’s ratios, and CTEs listed in Table S8. Displacement continuity was enforced at all interfaces. The bottom face of the substrate was constrained in the normal direction, its lateral faces were assigned symmetry (roller) conditions, and all remaining surfaces were traction-free. The temperature-dependent PIMD coefficients , , and were supplied as tangent CTEs, so that the thermal strain along each crystallographic axis relative to the stress-free reference temperature of 300 K was
Table S8: Material parameters adopted for the HfO gate dielectric and LaAlO substrate in the isotropic thermoelastic model. The listed values were treated as temperature-independent model parameters.
Table S8.
The thermal-mismatch strain was defined as the elastic part of the total strain, , which vanishes for a uniformly heated, unconstrained solid. The reported is the spatial maximum within the channel of the largest-magnitude principal value of .
Experimental Methods
Materials
Millimeter-scale single crystals were obtained from SixCarbon Technology and used for Raman spectroscopy, temperature-dependent XRD, and FDTR measurements.
Raman Spectroscopy
Room-temperature Raman spectra were collected from the (001) surface of a crystal before the thermal measurements with a Renishaw inVia micro-Raman system using 514 nm excitation at 2.5 mW. Spectra were acquired from 100 to 500 cm through a 50× objective with three accumulations of 60 s each.
Temperature-Dependent X-ray Diffraction
Temperature-dependent XRD was performed on a crystal with lateral dimensions of about 3 mm × 3 mm and a thickness above 50 m, supported on a silicon wafer holder to suppress parasitic scattering from the heating stage. Measurements were carried out on a Malvern Panalytical Empyrean diffractometer equipped with an Anton Paar HTK 1200N high-temperature chamber, using Cu radiation ( Å, 45 kV, 40 mA). Patterns were collected in ambient air from 298 to 473 K in 25 K steps over – with a step size of . The room-temperature pattern was consistent with the tetragonal structure (No. 139), and a pattern recorded after cooling to room temperature confirmed that heating caused no structural degradation. At each temperature, the lattice parameter was calculated from the positions of the reflections and their Miller indices using Bragg’s law. The cross-plane CTE was obtained from a linear fit of between 298 and 398 K as , which corresponds to the calculated averaged over the same interval, and its uncertainty was derived from the uncertainty of the fitted slope.
Frequency-Domain Thermoreflectance
The anisotropic thermal conductivity of BiOSe was measured by FDTR following our previous work126;127. A 445 nm pump laser, intensity-modulated from 0.5 to 10 MHz, and a continuous-wave 532 nm probe laser were focused onto the sample through a 20× objective. The spot radius, determined by beam-offset scans, was about 1.94 m128. Pump and probe powers of 1.5 and 2 mW kept the steady-state temperature rise below about 15 K. For measurements from 300 to 500 K, the sample was mounted on a temperature-controlled heating stage, and data were acquired after the temperature had stabilized at each set point.
A 61.5 nm Pt transducer was sputtered onto the freshly cleaved (001) surface of a bulk crystal with a thickness above 200 m. Pt was chosen instead of the more common Au because its lower thermal conductivity ( W m K at 300 K) limits lateral spreading within the transducer and increases the sensitivity to . The Pt thickness was measured with an Alpha-Step D300 profilometer (KLA-Tencor) on a Si reference wafer coated in the same deposition run. was calibrated at each measurement temperature by FDTR on a reference sample in which the same Pt film was deposited on a 2 m SiO layer, and the calibrated values were fixed in the analysis of BiOSe. The pump–probe phase lag was recorded with a balanced detector and demodulated with a lock-in amplifier (HF2LI, Zurich Instruments). At each temperature, measurements were performed at three locations on the sample surface, each scanned three times.
The phase data were analyzed with a transfer-matrix solution of anisotropic heat diffusion in a multilayer127. Because the sensitivities to and have distinct frequency dependences over 0.5–10 MHz (Fig. S4), , , and the Pt/BiOSe thermal boundary conductance were fitted simultaneously to the complete phase spectrum. The volumetric heat capacity of BiOSe at each temperature was taken from PIMD (Section 4.1.7). All fixed parameters are listed in Table S2. The uncertainties of the fitted conductivities were evaluated by error propagation, assuming a 3% uncertainty in each fixed parameter.
Because FDTR measures the total thermal conductivity, the electronic contribution was estimated from the Wiedemann–Franz law, . We used the in-plane electrical conductivity reported for nominally undoped bulk BiOSe single crystals, S m at 300 K48, and approximated its temperature dependence above room temperature as owing to phonon-limited carrier scattering. A Lorenz number of W K, representative of a nondegenerate semiconductor dominated by acoustic-phonon scattering, was used49. The resulting estimate, W m K over 300–500 K, corresponds to less than 4% of the measured in-plane conductivity, and the cross-plane contribution is smaller still because of the anisotropic electrical conductivity. The electronic contribution was therefore neglected, and the measured values were compared directly with the calculated .
Acknowledgements
This work was supported by the National Key R&D Program of China (Grant No. 2024YFB4405700), the Research Grants Council of Hong Kong (Project No. 16214626), and the National Natural Science Foundation of China (Grant No. 52676057). It was also funded by the Guangdong Basic and Applied Basic Research Foundation (Grant Nos. 2025A1515012898 and 2026A1515011290), the Opening Project of the State Key Laboratory of Advanced Glass Materials (Grant No. 2025KF27), and the Frontier Technology Research for Joint Institutes with Industry Scheme sponsored by the Center on Smart Sensors and Environmental Technologies at HKUST. Furthermore, the authors gratefully acknowledge the HKUST Fok Ying Tung Research Institute and the National Supercomputing Center in Guangzhou (Nansha Sub-center) for providing the high-performance computational resources used in this study.
Conflicts of Interest
The authors declare no competing financial interest.
Data Availability
The data supporting this study are available from the corresponding author on reasonable request.
Machine-learning potential validation
Bond-resolved chemical-bonding metrics
QHA benchmark for volumetric thermal expansion
Raman characterization of
FDTR parameters and sensitivity analysis
| Layer | (J cm K) | (W m K) | (W m K) | Thickness (nm) |
| Pt | 2.85 [1] | 36.9 | 36.9 | 61.5 |
| G1 | – | fitted | – | – |
| BiOSe | 1.96 | fitted | fitted | infinite |
Table S2. Material properties in the FDTR fitting model.
Detailed comparison of calculated CTEs
Thermal-expansion data at 300 K. (a) In-plane and cross-plane coefficients of thermal expansion of , , and obtained from classical MD and PIMD, together with the experimentally measured cross-plane CTE of . (b) Volumetric CTEs of and the literature S–Se–Te chalcogenide series used in Figure 3(b), together with the corresponding sulfide-normalized ratios and their assignment to the shaded envelope representing the prevailing trend.
(a) Direction-resolved CTEs of .
CTEs are given in at 300 K. Superscripts and denote the crystallographic and directions of orthorhombic , respectively. Replacing the Se atomic mass with the S atomic mass while retaining the potential changes its PIMD volumetric CTE by only 0.6% at 300 K.
| Material series | Included in shaded envelope | Source | |||||
| 8.14 | 6.72 | 6.61 | 0.826 | 0.812 | No—this work | This work | |
| 1.512 | 1.733 | 1.993 | 1.146 | 1.318 | Yes | [3] | |
| 5.07 | 6.03 | 6.12 | 1.189 | 1.207 | Yes | [4] | |
| 5.81 | 6.04 | 6.36 | 1.040 | 1.094 | Yes | [4] | |
| 4.87 | 5.04 | 5.10 | 1.035 | 1.047 | Yes | [5] | |
| 3.84 | 5.00 | 4.74 | 1.302 | 1.234 | Yes | [6] | |
| 6.09 | 6.70 | 6.99 | 1.100 | 1.148 | Yes | [7] | |
| YX, B1 phase | 3.70 | 4.10 | 4.30 | 1.108 | 1.162 | Yes | [8] |
| 4.67 | 4.81 | 5.18 | 1.030 | 1.109 | Yes | [9] | |
| 2.625 | 1.965 | 2.010 | 0.749 | 0.766 | No | [10] | |
| 4.81 | 4.72 | 4.79 | 0.981 | 0.996 | No | [11] |
Table 2. (b) Source data for the sulfide-normalized volumetric-CTE comparison in Figure 3(b).
Volumetric CTEs are given in at 300 K and ambient pressure. The normalized ratios were calculated from the values before rounding. The shaded envelope in Figure 3(b) comprises the series for which both and exceed unity. , ZnX, and are therefore not included in the envelope and are shown separately. Values digitized from the corresponding published curves at 300 K and 0 GPa. For cubic materials reported in terms of linear CTE, .
Higher-order phonon scattering and temperature scaling
TABLE S4. Temperature-dependent Peierls population conductivities calculated with three-phonon scattering only () and with combined three- and four-phonon scattering (), together with the Wigner coherence contribution (), for (a) BiOS, (b) BiOSe, and (c) BiOTe. All values are given in W m K.
(a) BiOS.
(b) BiOSe.
| (K) | ||||||
| 100 | 6.3641 | 5.5118 | 0.1399 | 5.9833 | 2.9518 | 0.0378 |
| 200 | 3.2882 | 2.7134 | 0.1401 | 2.6062 | 1.3733 | 0.0568 |
| 300 | 2.2091 | 1.6887 | 0.1439 | 1.6430 | 0.8741 | 0.0736 |
| 400 | 1.7702 | 1.3152 | 0.1385 | 1.2614 | 0.7159 | 0.0795 |
| 500 | 1.4371 | 1.0029 | 0.1442 | 0.9108 | 0.4979 | 0.0930 |
| 600 | 1.2965 | 0.8841 | 0.1412 | 0.7666 | 0.4630 | 0.0900 |
| 700 | 1.1322 | 0.7764 | 0.1384 | 0.6258 | 0.4188 | 0.0957 |
Table S4b.
(c) BiOTe.
| (K) | ||||||
| 100 | 5.9847 | 3.8578 | 0.0974 | 1.6608 | 1.0593 | 0.0260 |
| 200 | 3.1327 | 2.3818 | 0.0981 | 0.9063 | 0.6190 | 0.0340 |
| 300 | 2.1726 | 1.6426 | 0.1117 | 0.6355 | 0.4614 | 0.0475 |
| 400 | 1.6752 | 1.2597 | 0.1132 | 0.5177 | 0.3484 | 0.0507 |
| 500 | 1.4799 | 1.0909 | 0.1030 | 0.4154 | 0.2932 | 0.0500 |
| 600 | 1.3125 | 0.9689 | 0.1022 | 0.3311 | 0.2322 | 0.0530 |
| 700 | 1.2051 | 0.8013 | 0.1038 | 0.2974 | 0.1971 | 0.0525 |
Table S4c.
| (K) | ||||||
| 100 | 19.6 | 37.1 | 67.9 | 6.6 | 4.1 | 10.4 |
| 200 | 29.0 | 42.3 | 67.7 | 13.9 | 11.3 | 25.2 |
| 300 | 29.5 | 57.8 | 64.8 | 20.1 | 16.1 | 34.3 |
| 400 | 42.1 | 52.2 | 67.8 | 27.2 | 25.5 | 49.2 |
| 500 | 27.9 | 35.1 | 64.9 | 37.0 | 36.1 | 58.2 |
| 600 | 27.7 | 33.9 | 56.0 | 43.7 | 42.2 | 72.6 |
| 700 | 37.8 | 43.2 | 54.0 | 51.8 | 46.3 | 78.5 |
Table S5. Temperature dependence of the four-phonon-induced reduction () and Wigner coherence fraction () of BiOS, BiOSe, and BiOTe. The four-phonon-induced reduction and coherence fraction, both expressed as percentages, are defined as and . Values were calculated from the unrounded conductivities reported in Table S4. (a) BiOS.
| (K) | ||||
| 100 | 13.4 | 49.4 | 2.5 | 1.3 |
| 200 | 17.5 | 47.3 | 4.9 | 4.0 |
| 300 | 23.6 | 46.8 | 7.9 | 7.8 |
| 400 | 25.7 | 43.2 | 9.5 | 10.0 |
| 500 | 30.2 | 49.1 | 12.6 | 15.7 |
| 600 | 31.8 | 39.6 | 13.8 | 16.3 |
| 700 | 31.4 | 33.1 | 15.1 | 18.6 |
Table S5. (b) BiOSe.
| (K) | ||||
| 100 | 35.5 | 36.2 | 2.5 | 2.4 |
| 200 | 24.0 | 31.7 | 4.0 | 5.2 |
| 300 | 24.4 | 27.4 | 6.4 | 9.3 |
| 400 | 24.8 | 32.7 | 8.2 | 12.7 |
| 500 | 26.3 | 29.4 | 8.6 | 14.6 |
| 600 | 26.2 | 29.9 | 9.5 | 18.6 |
| 700 | 33.5 | 33.7 | 11.5 | 21.0 |
Table S5. (c) BiOTe.
Device-scale simulations
The thermal meshes comprise the fin and substrate, whereas the thermoelastic meshes additionally include the gate dielectric. Their mesh-node counts are denoted by and , respectively. The thermal and thermoelastic meshes are linked by spatial mapping of the calculated temperature fields.
The HfO Young’s modulus and Poisson’s ratio follow the ALD-film analysis of Gaskins et al. [12]; the reported in-plane CTE was adopted as an effective isotropic value. Its nominal density was selected within the reported 9.7–9.8 g cm range. For LaAlO, and were calculated from the Voigt–Reuss–Hill bulk and shear moduli, GPa and GPa [13], using the standard isotropic relations. The nominal density is consistent with the reported room-temperature value of 6.52 g cm. The LaAlO CTE was evaluated at 300 K as using the experimental volume fit of Chakoumakos et al. [14].
References
- [1]T. Yagi, K. Ohta, K. Kobayashi, N. Taketoshi, K. Hirose, and T. Baba, “Thermal diffusivity measurement in a diamond anvil cell using a light pulse thermoreflectance technique,” Meas. Sci. Technol. 22 (2010) 024011. https://doi.org/10.1088/0957-0233/22/2/024011.
- [2]J. Zich, A. Sojka, P. Levinský, M. Mišek, K.-H. Ahn, J. Navrátil, J. Hejtmánek, K. Knížek, V. Holý, D. Nuzhnyy, F. Borodavka, S. Kamba, and Č. Drásar, “Phonon properties and unconventional heat transfer in a quasi-two-dimensional Bi₂O₂Se crystal,” Phys. Rev. Materials 9 (2025) 054603. https://doi.org/10.1103/PhysRevMaterials.9.054603.
- [3]H. K. Hieu, N. T. Hong, P. T. M. Hanh, and B. K. Benazzouz, “Temperature-dependent structural and elastic properties of CdS, CdSe, and CdTe compounds studied by statistical moment method,” Rom. J. Phys. 69 (2024) 603. https://doi.org/10.59277/RomJPhys.2024.69.603.
- [4]M. Boudjendlia, Etude de premier principe des propriétés physiques des alliages ternaires SrₓCa₁₋ₓS, SrₓCa₁₋ₓSe et SrₓCa₁₋ₓTe par la méthode FP-LAPW, Doctoral Thesis, Université Badji Mokhtar, Annaba, 2018. https://biblio.univ-annaba.dz/wp-content/uploads/2022/09/These-Boudjendlia-Moussa.pdf.
- [5]S. Tomar et al., “Ab initio studies of structural, electronic, optical, elastic and thermal properties of copper thallium dichalcogenides (CuTlX₂: X = S, Se, Te),” Chalcogenide Lett. 16 (2019) 1–19. https://chalcogen.ro/1_TomarS.pdf.
- [6]P. Goel et al., “Phonons and anisotropic thermal expansion behavior of NiX (X = S, Se, Te),” J. Appl. Phys. 125 (2019). https://doi.org/10.1063/1.5094756.
- [7]B. Bekki et al., “Theoretical study of structural, elastic and thermodynamic properties of Cu₂MgSnX₄ (X = S, Se and Te) quaternary compounds,” Comput. Condens. Matter 18 (2019) e00339. https://doi.org/10.1016/j.cocom.2018.e00339.
- [8]T. Seddik et al., “External temperature and pressure effects on thermodynamic properties and mechanical stability of yttrium chalcogenides YX (X = S, Se and Te),” Physica B 428 (2013) 78–88. https://doi.org/10.1016/j.physb.2013.07.014.
- [9]S. Sharma, A. S. Verma, R. Bhandari, and V. K. Jindal, “Ab initio studies of structural, elastic and thermal properties of copper indium dichalcogenides (CuInX₂: X = S, Se, Te),” Comput. Mater. Sci. 86 (2014) 108–117. https://doi.org/10.1016/j.commatsci.2014.01.021.
- [10]Z. Zhu et al., “A high-throughput framework for lattice dynamics,” npj Comput. Mater. 10 (2024) 258. https://doi.org/10.1038/s41524-024-01437-w.
- [11]S. Sharma, A. S. Verma, and V. K. Jindal, “Ab initio studies of structural, electronic, optical, elastic and thermal properties of silver gallium dichalcogenides (AgGaX₂: X = S, Se, Te),” Mater. Res. Bull. 53 (2014) 218–233. https://doi.org/10.1016/j.materresbull.2014.02.021.
- [12]J. T. Gaskins, P. E. Hopkins, D. R. Merrill, S. R. Bauers, E. Hadland, D. C. Johnson, D. Koh, J. H. Yum, S. Banerjee, B. J. Nordell, M. M. Paquette, A. N. Caruso, W. A. Lanford, P. Henry, L. Ross, H. Li, L. Li, M. French, A. M. Rudolph, and S. W. King, “Review—Investigation and review of the thermal, mechanical, electrical, optical, and structural properties of atomic layer deposited high-k dielectrics: Beryllium oxide, aluminum oxide, hafnium oxide, and aluminum nitride,” ECS J. Solid State Sci. Technol. 6 (2017) N189–N208. https://doi.org/10.1149/2.0091710jss.
- [13]M. A. Carpenter, S. V. Sinogeikin, J. D. Bass, D. L. Lakshtanov, and S. D. Jacobsen, “Elastic relaxations associated with the Pm3̄m–R3̄c transition in LaAlO₃: I. Single crystal elastic moduli at room temperature,” J. Phys.: Condens. Matter 22 (2010) 035403. https://doi.org/10.1088/0953-8984/22/3/035403.
- [14]B. C. Chakoumakos, D. G. Schlom, M. Urbanik, and J. Luine, “Thermal expansion of LaAlO₃ and (La,Sr)(Al,Ta)O₃, substrate materials for superconducting thin-film device applications,” J. Appl. Phys. 83 (1998) 1979–1982. https://doi.org/10.1063/1.366925.