betaFeedback

WorksPhysics and Astronomy

Multiscale ensemble Monte Carlo of transport and gas sensing in monolayer MoS2_2

Abstract

Two-dimensional semiconductors such as monolayer MoS2 combine a useful band gap with a thickness of three atomic planes, which places the entire carrier density within a few angstroms of the surface and makes them attractive both as transistor channels and as chemical sensors. We present a self-consistent ensemble Monte Carlo (EMC) framework for monolayer MoS2 built on the ViennaEMC solver, which couples a Boltzmann transport model to Poisson's equation with degenerate free-carrier statistics. The transport model carries wavevector-dependent acoustic, polar-optical, homopolar and intervalley phonon scattering within and between the K valleys and the spin-orbit-split Q valleys, together with remote substrate phonons, screened charged impurities and surface roughness. Every intrinsic input is computed from first principles for this work, the band structure of the K and Q valleys, the phonon energies, the sound velocities and the dielectric response, and the electron-phonon couplings from density-functional perturbation theory matrix elements at 34 explicit phonon wavevectors and from Wannier interpolation with the two-dimensional long-range treatment. The model therefore contains no fitted transport parameter and no coupling taken from another calculation, and its scattering rates are checked against an independent solution of the linearised Boltzmann equation with the same couplings. The EMC velocity-field characteristic feeds a one-dimensional channel model, verified against a direct self-consistent three-dimensional EMC simulation of the gated channel, which reproduces the measured transfer characteristics of a CVD monolayer device in air and vacuum. An ambient and adsorbate extension reproduces the measured conductivity response to oxygen partial pressure and the concentration dependence of NO2 and NH3 sensing of independent monolayer devices.

Introduction

Monolayer molybdenum disulfide (MoS2\mathrm{MoS}_2) is a direct-bandgap (∼1.8\sim1.8 eV) two-dimensional (2D) semiconductor used as a channel material for post-silicon field-effect transistors (FETs) and as a sensitive platform for gas and chemical sensing [1, 2]. Its atomic thinness gives excellent electrostatic control and exposes the entire carrier population to the surface, but it also makes transport sensitive to phonons, the supporting dielectric, charged impurities, roughness, and adsorbed species. A model of device behaviour must therefore treat these mechanisms on the same physical footing rather than absorbing them into empirical mobilities.

The ensemble Monte Carlo (EMC) method solves the semiclassical Boltzmann transport equation (BTE) stochastically by following a statistical ensemble of carriers through field-driven free flights and scattering events, self-consistently coupled to Poisson’s equation. Since its introduction for hot-electron problems [3, 4, 5] and the development of the self-scattering technique, which makes it computationally practical [6], the method has matured over five decades into a standard framework for semiconductor transport and device simulation [7, 8, 9, 10, 11, 12, 13, 14], with full-band and technology-oriented device

The K-Q valley separation

@references:VEVENT implementations [15, 16]. It is well suited to two-dimensional materials because every scattering mechanism is described by its own microscopic rate, and hot-carrier and non-equilibrium effects are captured without additional approximation. For monolayer MoS2_2, the intrinsic phonon-limited mobility and the underlying electron-phonon couplings were established from first principles by Kaasbjerg et al. [17], and the role of the substrate and of charged impurities was clarified by Ma and Jena [18]. Experimentally, intrinsic transport was probed by Baugher et al. [19] and Radisavljević and Kis [20], and the strong influence of the ambient by Ahn et al. [21] and Li et al. [22]. In parallel, first-principles methods have matured for computing carrier mobilities in 2D materials directly from electron-phonon coupling [23, 24, 25, 26], including the role of flexural (out-of-plane) phonons enabled by the broken horizontal mirror symmetry [27]. These supply both the microscopic parameters and an independent benchmark for a transport model. Here, we consider micrometre-scale, scattering-dominated channels, in which the response is carried by the mobility and the carrier density.

Because the entire carrier population lies at the surface, monolayer MoS2_2 is also a sensitive chemical transducer. This idea can be traced back to the detection of individual gas molecules on graphene [28], but the finite band gap of semiconducting transition-metal dichalcogenides removes graphene’s zero-gap limitation, yielding a much larger relative conductance response and true field-effect operation [29, 30]. Transduction proceeds through two coupled microscopic channels, charge transfer between an adsorbed molecule and the channel (a donor such as NH3_3 raises the electron density while an acceptor such as NO2_2 depletes it), and the additional Coulomb scattering from the resulting ionised adsorbates, which lowers the mobility [31, 2]. Because the entire carrier density lies within a few Å of the surface, both effects are pronounced, and room-temperature detection of NO2_2 into the ppb range and reversible responses to NH3_3, humidity and volatile organics have been demonstrated [29, 2]. The FET geometry adds a degree of control absent in passive chemiresistors, since a gate bias tunes the operating carrier density and hence the selectivity and gain of the response [2]. The same platform also extends to label-free biosensing [32]. This physics also sets the ambient baseline, as continuously adsorbed O2_2 and H2_2O dope and scatter the channel, so that vacuum annealing recovers much of the intrinsic mobility [21, 22]. Most quantitative sensing studies remain empirical. A transport-level description that links the microscopic adsorbate-channel coupling to the measured device response, and that treats sensing and intrinsic transport within the same model, is largely absent.

In this work, we combine these components into a multiscale EMC framework for monolayer MoS2_2. Multiscale refers to the sequence from the first-principles inputs through the EMC transport model to the reduced device and sensor models, each verified against the level below it. Our contributions are (i) a complete self-consistent EMC transport model with intrinsic and extrinsic scattering mechanisms specialised to the 2D limit (Sec. 2); (ii) a first-principles verification of the model’s electronic-structure and polar-phonon inputs by density-functional perturbation theory (DFPT), with eleven quantities matching the reference parametrization and experiment and the electron-phonon couplings computed from the DFPT matrix elements (Sec. 3, Sec. 5.3); (iii) a hierarchical coupling to a one-dimensional channel model, verified against a direct self-consistent EMC device simulation, that reproduces a measured device IdI_\mathrm{d}-VgV_\mathrm{g} in air and vacuum (Sec. 4.1, Sec. 5.5); and (iv) an ambient/adsorbate extension reproducing measured ambient- and gas-sensing responses against independent monolayer experiments (Sec. 4.2, Sec. 5.6). Throughout, we state which quantities are predicted and which are fitted (Sec. 5.7).

The ensemble Monte Carlo transport model

Semiclassical transport

The framework solves the steady-state BTE for the carrier distribution f(r,k,t)f(\mathbf{r},\mathbf{k},t) [33],

∂f∂t+v(k)⋅∇rf+Fℏ⋅∇kf=(∂f∂t)coll(1)\frac{\partial f}{\partial t}+\mathbf{v}(\mathbf{k})\cdot\nabla_{\mathbf{r}}f+\frac{\mathbf{F}}{\hbar}\cdot\nabla_{\mathbf{k}}f=\left(\frac{\partial f}{\partial t}\right)_{\mathrm{coll}} \tag*{(1)}

by propagating an ensemble of NN superparticles. Between collisions, each carrier undergoes a field-driven free flight governed by Newton’s law, ℏk˙=F=−qE\hbar\dot{\mathbf{k}}=\mathbf{F}=-q\mathbf{E}, so that k(t)=k(0)−qEt/ℏ\mathbf{k}(t)=\mathbf{k}(0)-q\mathbf{E}t/\hbar, with the group velocity obtained from a non-parabolic k⋅p\mathbf{k}\cdot\mathbf{p} valley dispersion

E(1+αE)=ℏ2k22m∗,v(k)=1ℏ∇kE=ℏkm∗[1+2αE](2)E(1+\alpha E)=\frac{\hbar^2 k^2}{2m^*},\qquad\mathbf{v}(\mathbf{k})=\frac{1}{\hbar}\nabla_{\mathbf{k}}E=\frac{\hbar\mathbf{k}}{m^*[1+2\alpha E]} \tag*{(2)}

where m∗m^* is the valley effective mass and α\alpha the non-parabolicity.

The stochastic free-flight duration is generated with the self-scattering (constant-Γ\Gamma) technique [6, 7, 9]. For each valley/region, a constant total rate Γ0\Gamma_0 is chosen to bound the sum of all real rates λi(E)\lambda_i(E). A free flight of duration

tf=−1Γ0ln⁡(r),r∈(0,1] uniform,(3)t_f=-\frac{1}{\Gamma_0}\ln(r),\qquad r\in(0,1]\ \text{uniform}, \tag*{(3)}

is drawn, after which a mechanism is selected by comparing a second uniform random number against the normalised cumulative rate ∑j≤iλj(E)/Γ0\sum_{j\leq i}\lambda_j(E)/\Gamma_0. The residual defines a null (self-)scattering that leaves the state unchanged. Selected mechanisms update the carrier energy and redistribute its momentum according to their differential cross-section (isotropic, polar-angle-weighted, or intervalley as appropriate). Rates are pre-tabulated per valley/region for efficiency, and the ensemble is advanced over fixed time steps Δt\Delta t during which each particle may scatter several times (Figure 1). The loop is parallelised over particles with OpenMP.

Self-consistent ensemble Monte Carlo cycle showing inputs, initialisation, Poisson solve, particle free-flight and scattering updates, charge assignment, statistical accumulation, and outputs

Figure 1. Self-consistent ensemble Monte Carlo cycle of the ViennaEMC framework [39]. Material and scattering parameters computed by DFPT (Sec. 3) are the inputs at the top. The dashed block is repeated each time step, alternating a Poisson solve with a free-flight/self-scattering update of the carrier ensemble.

The macroscopic observables follow as ensemble averages over the non-transient part of the trajectory. The drift mobility is

μ=⟨vx⟩Ex,⟨vx⟩=1NNt∑t∑n=1Nvx,n(t),(4)\mu=\frac{\langle v_x\rangle}{E_x},\qquad\langle v_x\rangle=\frac{1}{NN_t}\sum_t\sum_{n=1}^{N}v_{x,n}(t), \tag*{(4)}

with NtN_t the number of averaging steps.

Scattering mechanisms

Table 1 lists the mechanisms included. The intrinsic mechanisms are computed for this work from the DFPT matrix elements of § 2.1.4, in the deformation-potential forms introduced by Kaasbjerg et al. [17] extended to wavevector-dependent couplings. The extrinsic mechanisms follow Ma and Jena [18] and Ridley [34]. We give the 2D forms used for the dominant channels.

Valley model. The conduction band is represented by three valley types. The K valleys are isotropic non-parabolic bands, Eq. (2), with m∗=0.446 m0m^*=0.446\,m_0 and α=1.16 eV−1\alpha=1.16\ \mathrm{eV}^{-1} fitted to the computed dispersion over the thermal energy range in both in-plane directions. These values reproduce the density of states and the mean squared velocity of the trigonally warped band to 1% between 25 and 150 meV, against a band-edge curvature mass of 0.489 m00.489\,m_0. The six Q valleys are anisotropic non-parabolic bands (ml=0.646m_l=0.646, mt=1.258m0m_t=1.258m_0, α=0.3 eV−1\alpha=0.3\ \mathrm{eV}^{-1}) whose spin-orbit splitting of 71 meV separates them, for an electron of given spin, into two sets of three valleys 120° apart, one at an EQKE_{QK} above the K minimum and one at EQK+71E_{QK}+71 meV. The two sets are carried as separate valley types. A K electron scatters into the lower set through its three nearest and into the upper set through its three farthest Q valleys, or the reverse for the opposite spin. Spin-averaged rates represent this by distributing each K-Q channel equally over the two sets. The separation EQKE_{QK} is the band parameter that depends most strongly on the exchange-correlation functional and on the geometry of the layer (§ 2.1.4.2.5.1). The transport model uses the measured 118 meV [36], and § 2.1.4.2.5.2 gives the sensitivity to it.

K-valley couplings

Acoustic deformation potential (elastic, equipartition). Acoustic phonons scatter elastically with a rate proportional to the 2D density of states. The phonon energies exchanged by thermal electrons (3 to 5 meV at the wavevectors q≤2kq\leq2k) are small against kBTk_BT down to 150 K, so the elastic equipartition form applies. A monolayer deformation potential depends on the phonon wavevector, and the transport model keeps that dependence,

λac(E)=mdkBTℏ3ρvs2[1+2αE]⟨Ξ2(q)⟩θ,q=2ksin⁡(θ/2),(5)\lambda_{\mathrm{ac}}(E)=\frac{m_d k_{\mathrm{B}}T}{\hbar^3\rho v_s^2}\left[1+2\alpha E\right]\left\langle\Xi^2(q)\right\rangle_\theta,\qquad q=2k\sin(\theta/2), \tag*{(5)}

where Ξ(q)\Xi(q) is the tabulated deformation potential of the LA or TA branch, mdm_d the density-of-states mass, ρ\rho the areal mass density, vsv_s the sound velocity of the branch, and the average runs over the deflection angle θ\theta, from which the angle after scattering is also sampled. For a constant Ξ\Xi, this reduces to the usual isotropic form. The Q valleys use constant values because the direct calculation covers their couplings at two wavevectors only.

Polar-optical (Fröhlich) and homopolar coupling. The longitudinal-optical (LO) branch of the E′E^\prime mode couples through the two-dimensional Fröhlich matrix element, which is finite at q→0q\to0 and reduced at larger qq by the dielectric response of the layer itself [35]. The rate for absorption (+) and emission (−) carries the Bose occupation NLO=[exp⁡(ℏωLO/kBT)−1]−1N_{\mathrm{LO}}=[\exp(\hbar\omega_{\mathrm{LO}}/k_{\mathrm{B}}T)-1]^{-1},

λpop±(E)=mdAcell2πℏ3(NLO+12±12)∫∣gLO(q)∣2S(q,qs)q dθkcos⁡2θ±ℏωLO/E,E′=E±ℏωLO,(6)\lambda_{\mathrm{pop}}^{\pm}(E)=\frac{m_d A_{\mathrm{cell}}}{2\pi\hbar^3}\left(N_{\mathrm{LO}}+\frac{1}{2}\pm\frac{1}{2}\right)\int|g_{\mathrm{LO}}(q)|^2 S(q,q_s)\frac{q\,\mathrm{d}\theta}{k\sqrt{\cos^2\theta\pm\hbar\omega_{\mathrm{LO}}/E}},\qquad E^\prime=E\pm\hbar\omega_{\mathrm{LO}}, \tag*{(6)}

with q=∣k′−k∣q=|\mathbf{k}^\prime-\mathbf{k}|, AcellA_{\mathrm{cell}} the unit-cell area, ∣gLO(q)∣|g_{\mathrm{LO}}(q)| the tabulated first-principles matrix element of § 2.1.4, and S(q,qs)=[1+qs/q]−2S(q,q_s)=[1+q_s/q]^{-2} the static screening by the free carriers with the two-dimensional Thomas–Fermi wavevector qsq_s of the sheet density. The static form applies because the plasma energy of the sheet at the relevant wavevectors (84 to 145 meV at 1012 cm−210^{12}\ \mathrm{cm}^{-2}) exceeds the LO energy of 46 meV, so the carriers follow the phonon field, and it overestimates the screening by at most 30 % where the two energies come closest (Supplementary Note S2). The homopolar A1′A^\prime_1 mode (49 meV), whose matrix element is likewise wavevector dependent, is described by the same expression without free-carrier screening, in the K and in the Q valleys.

Intervalley scattering. Transitions between the K and K′ valleys proceed through the zero-order coupling of the E′E^\prime mode at K and through four first-order channels (the TA, LA, TO and A1′A^\prime_1-like modes at K), whose matrix elements are linear in the phonon wavevector measured from K, and the first-order intravalley TO(Γ) channel is treated in the same form. Transitions between K and Q and between Q valleys use 21 channels, each with its own phonon energy and zero-order coupling from the DFPT calculation with the electron at Q, assigned to the two spin sets by the selection rule of the valley model (Supplementary Note S2). No separate piezoelectric term is used, because in the two-dimensional treatment the long-range part of the acoustic coupling is contained in the computed Ξ(q)\Xi(q).

Extrinsic mechanisms. Charged impurities at the interface scatter through a two-dimensional Coulomb potential screened by the free carriers and by the layer (Thomas–Fermi and Rytova–Keldysh), the polar substrate contributes remote surface-optical phonon scattering with a coupling set by its dielectric discontinuity, and interface roughness is described in the Prange-Nee form, all in the forms of Ma and Jena [18] and Ridley [34] (Supplementary Note S4). The adsorbate Coulomb centres of the sensor model (§ 2.1.4.2.2) use the same screened potential.

Electron–phonon couplings. All intrinsic couplings are computed for this work (§ 2.1.4), and Supplementary Note S2 lists every value used. The acoustic deformation potential of the K valley is ΞLA=6.5 eV\Xi_{\mathrm{LA}}=6.5\ \mathrm{eV} in the long-wavelength limit and 4.8 eV at 0.30 Å−1^{-1}, whereas the TA coupling vanishes at q→0q\to0, as it must for a non-degenerate band, and reaches 1.4 eV at 0.30 Å−1^{-1}, so acoustic scattering is an LA process with a TA share of a few per cent. Obtaining this required the two-dimensional Coulomb cutoff in the electron–phonon calculation, without which a long-range piezoelectric-like term of the periodic slab overstates the TA coupling several-fold (Supplementary Note S2). The Fröhlich element is 0.35 eV at q→0q\to0 and 53 meV at 0.30 Å−1^{-1}, the homopolar element 77 meV at q→0q\to0, the zero-order K-K′ coupling 1.7×1081.7\times10^8 eV/cm, and the Q valleys have ΞL=3.0\Xi_{\mathrm{L}}=3.0 and ΞT=2.4\Xi_{\mathrm{T}}=2.4 eV. Compared with the LDA parameterization of Kaasbjerg et al. [17], the LA coupling is twice as large and the zero-order optical couplings are 0.65 to 0.8 times as large (Supplementary Note S4). No calibration factor is applied here. § 2.1.4.2.5.2 shows that the LA coupling alone accounts for the difference between the 392 cm2^2/Vs obtained with the LDA couplings and the 214 cm2^2/Vs obtained with the computed ones. Recent first-principles values span 130 to 410 cm2^2/Vs, depending on the treatment of the Q valleys and of the two-dimensional Fröhlich coupling [23, 17, 37, 26], and our value lies inside that range and inside the range of measured monolayer devices.

Self-consistent electrostatics and degenerate statistics

Charge is assigned to the mesh with a cloud-in-cell particle-mesh scheme [38], driving Poisson’s equation,

∇⋅(ε∇ϕ)=−q(p−n+ND+−NA−),(7)\nabla\cdot(\varepsilon\nabla\phi) = -q\left(p-n+N_D^+-N_A^-\right), \tag*{(7)}

solved by successive over-relaxation on the device grid in the established device-simulation manner [10]. Because accumulation-mode 2D channels reach degeneracy, the equilibrium density is obtained from Fermi-Dirac rather than Boltzmann statistics, through a two-dimensional relation between the local density and the potential that recovers the Boltzmann limit exactly for non-degenerate materials (Supplementary Note S4). Figure 1 summarises the complete self-consistent cycle.

Numerical details and statistical uncertainty

All transport results are ensemble averages, and every quoted value carries a statistical uncertainty. Bulk transport (mobilities, velocity-field characteristics and sensor transductions) is computed in a 500×500500 \times500 nm2^2 periodic box with a 1010 nm grid, about 2×1042 \times10^4 superparticles and a time step of 0.10.1 fs. The temperature-dependent mobilities average 45 000 of 60 000 steps after a 1.51.5 ps transient at a field of 0.20.2 MV/m and a sheet density of 101210^{12} cm−2^{-2}. The sensor transductions average the second half of 25 000 steps at each coverage, and the velocity-field characteristics average the last three quarters of 40 000 steps at each field. Eight independent random seeds are used for the mobilities and the transductions, four for the velocity-field characteristics and the sensitivity variants, and three for the two-valley sweep of Supplementary Note S1. The standard error is the spread over seeds divided by the square root of their number, and it agrees with the block-averaging estimate within one run (blocks of 0.2 ps, ten times the momentum relaxation time). Power-law exponents are given with the standard error of the fit. The self-consistent device simulations of § 2.1.4.2.1 use a 5×25×15 \times25 \times1 nm3^3 grid (53×21×2153 \times21 \times21 nodes), 2.4×1042.4 \times10^4 superparticles of unit weight, a time step of 0.1 fs and a Poisson solve every ten steps to 10−310^{-3} V. After a transient of 5 ps (6 ps at Vds=1V_{\mathrm{ds}} = 1 V), 3 ps are averaged, and the resulting currents carry block standard errors of 0.3 to 4 %. Table 2 collects the key results with their standard errors.

suspendedhBNSiO2_2HfO2_2Al2_2O3_3CaF2_2
μ(300 K)\mu(300\ \mathrm{K}) (cm2^2/Vs)212 (1)214 (1)200 (1)151 (2)176 (2)160 (2)
γ\gamma1.51 (0.06)1.52 (0.07)1.64 (0.09)1.92 (0.10)1.82 (0.10)1.94 (0.10)
QuantityValueStandard error
Low-field mobility of v(E)v(E), suspended (4 seeds)213 cm2^2/Vs4 cm2^2/Vs
Drift velocity at 10 MV/m, suspended (4 seeds)3.55×1043.55 \times10^4 m/s0.01×1040.01 \times10^4 m/s
Sensitivity variants (§ 2.1.4.2.5.2, 4 seeds)below 3 %
Direct device currents (§ 2.1.4.2.5.3)0.3 to 4 %
Two-valley sweep (Supplementary Note S1, 3 seeds)below 2 %

Table 2. Key results with their statistical uncertainties. The upper part gives the room-temperature mobility at 101210^{12} cm−2^{-2} (standard error over eight seeds in parentheses) and the temperature exponent of the fit over 150 to 400 K (fit standard error). The lower part gives the remaining quantities with the standard error over independent seeds or, for the device currents, over blocks within one run.

First-principles calculations

Every intrinsic parameter of the EMC transport model is computed from first principles, namely the band structure of the K and Q valleys, the phonon energies and sound velocities, the Born charges and dielectric response, and the electron-phonon couplings. We use density-functional theory and density-functional perturbation theory (DFPT) as implemented in Quantum ESPRESSO [40] with the PBE functional [41] and norm-conserving PseudoDojo pseudopotentials [42, 43] (90 Ry cutoff and a 12×12×112 \times12 \times1 mesh for the unit cell, 60 Ry and 3×3×13 \times3 \times1 for the adsorbate supercells). A projector-augmented-wave calculation [44] provides one of the geometries of the K-Q sensitivity study of § 2.1.4.2.5.1. The monolayer is treated with an in-plane-only relaxed cell and a 2D Coulomb cutoff for the electronic and dielectric response. We unfold the in-plane dielectric response of the slab to the cell-independent two-dimensional polarisability α2D=L(ε∥−1)/4π\alpha_{2\mathrm{D}} = L(\varepsilon_{\parallel} - 1)/4\pi, with LL the cell height, which sets the Rytova-Keldysh screening length r0=2πα2Dr_0 = 2\pi\alpha_{2\mathrm{D}} of the impurity model of § 2.1.1. Of the dielectric quantities, the transport model uses r0r_0 together with the substrate and environment permittivities in the screened Coulomb terms, and the Born charges and phonon energies in the polar-optical term. The slab dielectric constant itself is not used.

The transport model includes spin-orbit coupling through the Q valleys, which split by 71 meV in our calculation (70 to 75 meV in the literature [45]) and are therefore carried as two valley sets (§ 2.1.1). At K the conduction band splits by only 3 meV, far below kBTk_{\mathrm{B}}T at the temperatures studied, so both spin subbands are equally populated and the transport masses change by less than 2 %. We therefore treat the K valley as spin degenerate. For nn-type transport the large valence-band splitting is irrelevant. Electron-phonon couplings are computed in the scalar-relativistic calculation and applied to both Q sets, because spin-orbit coupling splits the Q valley into two spin-polarised subbands of the same orbital composition [45] and therefore shifts their energies without changing the matrix elements that couple them to the phonons. The K-Q separation itself, which spin-orbit coupling lowers by about 34 meV, is set from the measurement (§ 2.1.4.2.5.1).

Sound velocities are the slopes of the interpolated acoustic branches between 0.04 and 0.15 Å−1^{-1} (Supplementary Note S2). The electron-phonon matrix elements ⟨ψk+q∣∂V/∂uqνα∣ψk⟩\langle\psi_{\mathbf{k}+\mathbf{q}}|\partial V/\partial u_{\mathbf{q}\nu\alpha}|\psi_{\mathbf{k}}\rangle of the conduction band are computed with the Allen-Heine-Cardona functionality [46] of the DFPT code, in a cell of height 20.2 Å at the norm-conserving relaxed lattice constant of 3.185 Å. With the electron at K we use 21 explicit wavevectors q\mathbf{q}, covering 0.03 to 0.34 Å−1^{-1} along Γ-M, Γ-K and at 15° between them, the K point for the zero-order K-K′ transition, six wavevectors within 0.1 Å−1^{-1} of K for the first-order K-K′ couplings, and the K-Q vector. With the electron at Q we use 13

Adsorbate calculations

wavevectors, covering the intravalley directions, the three Q-Q′ transitions, Q-K′ and Q-K. This direct calculation has no two-dimensional Coulomb cutoff, so its small-wavevector acoustic couplings carry a periodic-slab artefact. The second calculation, a Wannier interpolation with the EPW code [47, 48] with the cutoff [49] and the two-dimensional long-range treatment of Poncé et al. [50], gives the acoustic couplings along Γ-M free of it, the polar-optical and homopolar ∣g(q)∣|g(q)|, and the phonon-limited mobility from the linearised Boltzmann equation. We project on the phonon eigenvectors and convert to deformation potentials following Ref. [17]. Supplementary Note S2 gives the full tables by mode, wavevector and direction, the comparison of the two calculations and the Boltzmann check. Adsorbate charge transfers and adsorption energies for § 2.1.4.2.2 come from a third set of calculations, relaxations of NO2_2, NH3_3 and O2_2 on a 4×44 \times4 supercell of the same layer with the D3 dispersion correction, the two-dimensional cutoff and spin polarisation, followed by a Bader partition of the charge density (Supplementary Note S3). Since the reference parametrization [17] is itself a DFPT result, the comparison of § 2.1.4.2.5.1 is between two first-principles calculations. § 2.1.4.2.5.2 and Supplementary Note S1 quantify how sensitive the transport results are to the level of theory, in particular through the K-Q valley separation.

Device and adsorbate extensions
Multiscale device model

The device is treated at two levels of description. The self-consistent device simulation resolves a back-gated monolayer channel in three dimensions using ViennaEMC (Figure 2). The gate is a Dirichlet plane below a 10 nm HfO2_2 oxide, the monolayer occupies one grid plane, and 10 nm of air above the sheet are included so that field lines can leave the channel. Permittivity is assigned per node (20 in the oxide, 1 in the air), and Poisson’s equation is solved with a successive over-relaxation scheme that applies the degenerate carrier statistics of § 2.1.2 on the sheet nodes only. The channel is 200 nm long and 500 nm wide, contacted through 30 nm n+n^+ regions (3×10133 \times10^{13} cm−2^{-2}) that end in ohmic reservoir contacts, and the carriers are confined to the sheet plane, on which their charge is assigned and the in-plane force is interpolated with the cloud-in-cell scheme. Together, the n+n^+ regions and their reservoirs form an ideal ohmic termination, the standard boundary of Monte Carlo device simulation, which isolates the channel electrostatics from contact injection. Schottky barriers and Fermi-level pinning of real metal contacts to MoS2_2 lie outside the simulation, whose access resistance is on the order of 100 Ω μ\Omega\,\mum, the level of the best reported semimetal contacts, for which bismuth on monolayer MoS2_2 gives 123 Ω\Omega with no Schottky barrier [51]. In the comparison with measured devices they are represented by the contact resistance of Table 3. The drain current is the drift-velocity (Shockley-Ramo) tally over the channel interior, recorded at every step after the transient, and the net contact fluxes agree with it within 3%.

Structure of the direct self-consistent EMC device simulation

Figure 2. Structure of the direct self-consistent EMC device simulation of § 2.1.4.2.1, drawn as a section along the channel with the vertical scale exaggerated 3.25-fold. The mesh shown is the simulation mesh in this plane, 53 nodes at 5 nm along the channel and 21 nodes at 1 nm across the stack. Poisson’s equation is solved on the full three-dimensional mesh spanning the gate oxide and the air above, whereas the carriers are confined to the single grid plane of the monolayer, on which their charge is assigned and the in-plane force is interpolated. The 200 nm channel is terminated by 30 nm n+n^+ regions (3×10133 \times10^{13} cm−2^{-2}) that end in the ohmic source and drain reservoirs. The channel is 500 nm wide along the direction normal to the section, meshed with a further 21 nodes at 25 nm, which completes the 5×25×15 \times25 \times1 nm3^3 cells and 53×21×2153 \times21 \times21 nodes of the simulation.

QuantityVacuumAirSource
v(E)v(E)tabulatedsameEMC (§ 2.1.4.2.5.3)
Low-field mobility of v(E)v(E)213 cm2^2/VssameEMC
vv at 10 MV/m3.6×1063.6 \times10^6 cm/ssameEMC
μint\mu_{\mathrm{int}}, intrinsic mobility213 cm2^2/Vs90 cm2^2/Vs (1 atm)EMC, ambient model of § 2.1.4.2.5.4
μ0\mu_0, channel mobility in Eq. (10)4 cm2^2/Vs1 cm2^2/Vsdevice field-effect mobility
CoxC_{\mathrm{ox}}1.21×10−41.21 \times10^{-4} F/m2^2samedevice geometry
WW, LL32 μ\mum, 1 μ\mumsamedevice geometry
VthV_{\mathrm{th}}−6-6 V46 Vdevice (the shift is compared with the charge-transfer model)
SSSS9 V/dec22 V/decdevice (interface traps, not modelled)
RcR_c160 kΩ μ\Omega\,\mum500 kΩ μ\Omega\,\mumdevice
ns(Vg)n_s(V_g), access resistanceMonte Carlo simulation (§ 2.1.4.2.5.3)

Table 3. Quantities of the device model and their sources. EMC values refer to the intrinsic mechanisms at 101310^{13} cm−2^{-2} (§ 2.1.4.2.5.3), and device values refer to the CVD monolayer FET of Ahn et al. [21] in vacuum and air.

Because such a simulation is too expensive for the ambient and sensing sweeps of § 2.1.4.2.5.4, the device response is described by a reduced model that takes its transport input from the bulk transport model and its electrostatics from the gradual-channel approximation, in the hierarchical sense of device simulation [14, 10]. The bulk EMC velocity-field characteristic v(E)v(E), computed by sweeping the driving field at the channel density (§ 2.1.4.2.5.3), is used by the reduced model as a table.

A common parametrization of such a characteristic is the Caughey-Thomas form

v(E)=μ0E[1+(μ0E/vsat)β]1/β,(8)v(E)=\frac{\mu_0 E}{\left[1+\left(\mu_0 E/v_{\mathrm{sat}}\right)^\beta\right]^{1/\beta}}, \tag*{(8)}

with a low-field mobility μ0\mu_0 and a saturation velocity vsatv_{\mathrm{sat}}. The transport model saturates through intervalley transfer (§ 2.1.4.2.5.3) with a characteristic that this form reproduces only approximately, so the reduced model integrates the tabulated v(E)v(E) directly and the closed form below is used only where the field stays in the linear range. The v(E)v(E) feeds a one-dimensional gradual-channel model. The local sheet charge follows a smooth turn-on with ideality η\eta (from the subthreshold swing SSSS), giving an effective overdrive

Vov(Vg)=ηkBTqln⁡[1+exp⁡(q(Vg−Vth)ηkBT)],η=SS(kBT/q)ln⁡10,(9)V_{\mathrm{ov}}(V_g)=\eta\frac{k_{\mathrm{B}}T}{q}\ln\left[1+\exp\left(\frac{q(V_g-V_{\mathrm{th}})}{\eta k_{\mathrm{B}}T}\right)\right],\qquad\eta=\frac{SS}{(k_{\mathrm{B}}T/q)\ln10}, \tag*{(9)}

and the current continuity Id=WCox[Vov−V(x)]v(E(x))I_d=WC_{\mathrm{ox}}[V_{\mathrm{ov}}-V(x)]v(E(x)) with dV/dx=E\mathrm{d}V/\mathrm{d}x=E is integrated along the channel, the current being the value for which the integrated length equals LL. For the Caughey-Thomas form, this yields the closed form with a linear-to-saturation transition,

Id=WLμ0CoxVovVd−12Vd21+Vd/(EcL),Vd=min⁡(Vds,Vdsat),(10)I_d=\frac{W}{L}\mu_0 C_{\mathrm{ox}}\frac{V_{\mathrm{ov}}V_d-\frac{1}{2}V_d^2}{1+V_d/(E_cL)},\qquad V_d=\min(V_{\mathrm{ds}},V_{\mathrm{dsat}}), \tag*{(10)}

where Ec=vsat/μ0E_c=v_{\mathrm{sat}}/\mu_0 is the critical field and Vdsat=VovEcL/(Vov+EcL)V_{\mathrm{dsat}}=V_{\mathrm{ov}}E_cL/(V_{\mathrm{ov}}+E_cL), and a series contact resistance 2Rc2R_c is applied self-consistently. Unlike a lumped compact model, this resolves the channel and captures the linear-to-saturation transition of the drain current. The reduced model has no fitted parameter when it is compared with the direct simulation. The parameters μ0\mu_0, vsatv_{\mathrm{sat}} and β\beta come from the bulk EMC with the same scattering mechanisms, the channel charge qns(Vg)q n_s(V_g) from the zero-bias self-consistent simulation, and the access resistance of the n+\mathrm{n}^{+} regions is the series term 2Rc2R_c. § 2.1.4.2.5.3 reports this verification. When a measured device is compared, VthV_{\mathrm{th}}, SSSS and RcR_c are device parameters, and so is the channel mobility of a contact- and defect-limited CVD device, whose reduction below the intrinsic value exceeds what charged interface impurities alone produce in the EMC (§ 2.1.4.2.5.2). Table 3 lists every quantity of the device model with its source.

Ambient and adsorbate model

A monolayer channel, whose carriers all lie within a few Å of the surface, responds to adsorbates through two coupled channels: charge transfer, which shifts the carrier density, and additional Coulomb scattering from the ionised adsorbates, which lowers the mobility. Both determine the EMC conductivity σ=qnsμ\sigma=q n_s\mu at each fractional adsorbate coverage θ\theta,

ns(θ)=n0∓ΔNθ,μ(θ)=μ[ns(θ),Nc(θ)],S(θ)=σ0σ(θ)−1,(11)n_s(\theta)=n_0\mp\Delta N\theta,\qquad\mu(\theta)=\mu[n_s(\theta),N_c(\theta)],\qquad S(\theta)=\frac{\sigma_0}{\sigma(\theta)}-1, \tag*{(11)}

where the upper (lower) sign is an acceptor (donor), ΔN\Delta N is the transferred charge at saturation, and Nc(θ)=Ncmax⁡θN_c(\theta)=N_c^{\max}\theta is the areal density of adsorbate Coulomb centres, which lower μ\mu through the screened Coulomb scattering of § 2.1.1. The transduction S(θ)S(\theta) is thus an EMC output. A

Degenerate carrier statistics in the Poisson equation

Langmuir isotherm θ(C)=KC/(1+KC)\theta(C) = KC/(1 + KC) maps coverage to gas concentration, with KK the only per-device adsorption parameter. Oxygen partial pressure is mapped analogously.

The adsorption constant has a physical basis. Langmuir kinetics with an impingement rate s0P/2πmkBTs_0P/\sqrt{2\pi mk_{\mathrm{B}}T} onto free sites and an Arrhenius desorption rate νNsiteθexp⁡(−Eb/kBT)\nu N_{\mathrm{site}}\theta\exp(-E_{\mathrm{b}}/k_{\mathrm{B}}T) give

K=s0νNsite2πmkBTexp⁡(EbkBT),(12)K = \frac{s_0}{\nu N_{\mathrm{site}}\sqrt{2\pi mk_{\mathrm{B}}T}}\exp\left(\frac{E_{\mathrm{b}}}{k_{\mathrm{B}}T}\right), \tag*{(12)}

so KK is set by the binding energy Eb=−EadsE_{\mathrm{b}}=-E_{\mathrm{ads}}, which is positive for a bound molecule and is taken from the adsorption energy EadsE_{\mathrm{ads}} of Supplementary Note S3, together with the site density NsiteN_{\mathrm{site}}, the sticking coefficient s0s_0 and the desorption attempt frequency ν\nu. With s0=1s_0=1, ν=1013 s−1\nu=10^{13}\ \mathrm{s}^{-1} and Nsite=1014 cm−2N_{\mathrm{site}}=10^{14}\ \mathrm{cm}^{-2} the prefactor is of order 10−9 Pa−110^{-9}\ \mathrm{Pa}^{-1}. A fitted KK should therefore be interpreted as an effective adsorption constant representing the unresolved distribution of adsorption sites in a device (basal plane, vacancies, and grain boundaries), so its device dependence reflects the defect density and the growth of each sample rather than the transduction model (Supplementary Note S3).

The charge-transfer amplitudes are ΔN=2.2×1012 cm−2\Delta N=2.2\times10^{12}\ \mathrm{cm}^{-2} for NO2_2 (acceptor) and 2.0×1013 cm−22.0\times10^{13}\ \mathrm{cm}^{-2} for NH3_3 (donor) at full coverage, with adsorbate Coulomb-centre densities Ncmax⁡=3.0×1012N_{\mathrm{c}}^{\max}=3.0\times10^{12} and 2.0×1011 cm−22.0\times10^{11}\ \mathrm{cm}^{-2} and a baseline density n0=3×1012 cm−2n_0=3\times10^{12}\ \mathrm{cm}^{-2}, all fitted. To compare them with the same surface, we relaxed NO2_2, NH3_3 and O2_2 on a 4×44\times4 cell of the monolayer at the geometry of the electron-phonon calculations, with the PBE functional, the D3 dispersion correction, the two-dimensional Coulomb cutoff and spin polarisation for the radicals, and partitioned the charge density by the Bader method (Supplementary Note S3, Tables S10 and S11). All three molecules physisorb flat on the pristine plane with adsorption energies of −0.08-0.08 to −0.18-0.18 eV and transfer 0.040.04 to 0.05 e0.05\ e to NO2_2, 0.010.01 to 0.03 e0.03\ e from NH3_3 and 0.01 e0.01\ e to O2_2. Per molecule, at one molecule per 4×44\times4 cell, the fitted amplitudes correspond to 0.030 e0.030\ e for NO2_2, two thirds of the computed transfer, with the remainder set by the carrier limit of the channel (ΔNθ≤n0\Delta N\theta\le n_0), and to 0.28 e0.28\ e for NH3_3, ten times the pristine-plane transfer, suggesting a substantial contribution from defect-related adsorption sites. The fitted adsorption constants support the same interpretation. Under the assumed kinetic prefactor, Eq. (12) makes them correspond to effective binding energies EbE_{\mathrm{b}} of about 0.50.5 to 0.60.6 eV, larger than the pristine-plane physisorption energies but below the values reported for defect-modified surfaces [52, 53]. This suggests an effective response involving both basal-plane and more strongly binding defect-related sites. The published PBE transfers without dispersion correction, 0.10 e0.10\ e for NO2_2 and 0.0090.009 to 0.069 e0.069\ e for NH3_3 [54, 2], bracket those given here.

Results and discussion
Intrinsic transport and the mobility limit

Figure 3 shows the computed temperature-dependent mobility for several substrates. Each insulator follows a power law μ∝T−γ\mu\propto T^{-\gamma} with its own exponent (fit standard errors of 0.06 to 0.10). The suspended (intrinsic) film has the shallowest slope (γ=1.51\gamma=1.51). Acoustic scattering alone gives γ=1\gamma=1 in two dimensions. The optical and intervalley channels, whose Bose factors grow with temperature, and the population of the QQ valleys, which rises from 0.04% at 150 K to 3% at 300 K and 9% at 400 K, steepen it to 1.51. Every supporting dielectric adds a thermally activated remote surface-optical phonon channel that steepens the decrease further, from hBN (γ=1.52\gamma=1.52) and SiO2_2 (γ=1.64\gamma=1.64) to the soft, polar high-κ\kappa HfO2_2 (γ=1.92\gamma=1.92). The impurity-free phonon limit of Kaasbjerg et al. [17] (γ=1.69\gamma=1.69) lies between the suspended and the supported films. Its room-temperature value of 410 cm2^2/Vs is twice ours because its LDA acoustic coupling is half of the computed one (§ 2.1.4.2.5.2). The room-temperature mobilities, 151 to 214 cm2^2/Vs with standard errors of 1 to 2 cm2^2/Vs over eight seeds, are consistent with the best measured monolayer devices [19, 20]. Against an independent Boltzmann solution with the same couplings, the transport model agrees within 1% at 150 K and 13% at 400 K (Supplementary Note S2). Figure 4 recasts this as a benchmark in which the EMC sets the intrinsic upper bound and measured devices fall below it by an amount that quantifies their extrinsic (contact/defect/substrate) limitation. The cleanest exfoliated device approaches this bound, while CVD and contact-limited devices lie one to two decades lower.

Figure 3. Temperature-dependent mobility μ(T)\mu(T) for different substrates. Each insulator has its own power-law exponent γ\gamma (μ∝T−γ\mu\propto T^{-\gamma}, given with the standard error of the fit). The suspended (intrinsic) film has the shallowest slope (γ=1.51\gamma= 1.51). Every supporting dielectric adds a thermally activated remote surface-optical phonon channel whose scattering grows with temperature and steepens the decrease, most strongly for the soft, polar high-κ\kappa HfO2_2 (γ=1.92\gamma= 1.92). Error bars are standard errors over eight seeds and are mostly smaller than the symbols. The impurity-free phonon limit of Kaasbjerg et al. [17] (γ=1.69\gamma= 1.69, dashed) is shown for reference.

Figure 4. The EMC intrinsic phonon-limited mobility sets an upper bound below the Kaasbjerg phonon limit [17], against which the measured monolayer and multilayer device mobilities [19, 55, 56, 22, 21] are benchmarked, each falling below the bound by an amount that quantifies its extrinsic (contact, defect and substrate) limitation. Monolayer and multilayer devices are distinguished by colour, and only the monolayer devices are compared quantitatively with the bound.

Substrate and dielectric dependence

Figure 5(a) shows the remote-phonon-limited mobility for each substrate at fixed density. The ordering suspended ≈\approx hBN >> SiO2_2 >> Al2_2O3_3 >> CaF2_2 >> HfO2_2 is governed by a competition between the dielectric discontinuity and the phonon stiffness. High-κ\kappa HfO2_2 has soft (∼12\sim12 meV) modes that

Figure 5. The mobility is limited both by the supporting dielectric, shown in (a) as the remote-surface-optical-phonon-limited mobility for the different substrates, and by the carrier density, shown in (b) as a non-monotonic dependence μ(ns)\mu(n_s) that is low at small sheet density where too few free carriers screen the charged impurities, rises as the added carriers progressively screen them, and falls again at high density as the stronger vertical field presses the carriers into the rough dielectric interface, so that the balance of the two opposing trends sets the peak near ns=1013 cm−2n_s = 10^{13}\ \mathrm{cm}^{-2}.

Comparison with few-layer devices

are strongly thermally occupied, whereas the crystalline fluoride CaF2_2, despite the largest dielectric step, retains three quarters of the intrinsic mobility because its single surface-optical mode is stiff (∼51\sim51 meV) and weakly populated at room temperature, consistent with proposed use of CaF2_2 as a 2D-FET insulator [57]. Figure 5(b) shows the non-monotonic density dependence, rising as free carriers screen charged impurities and falling as interface roughness takes over, with the peak near ns=1013n_s = 10^{13} cm−2^{-2}.

Validation of the inputs

Table 4 lists the computed band, phonon and dielectric parameters against the reference parametrization and experiment. Eleven quantities agree to within the expected PBE accuracy, namely the lattice constant, the K-valley and Q-valley effective masses, the band gap, the E′E' and A1′A'_1 phonon energies, the Born effective charges, both in-plane sound velocities, the two-dimensional polarisability and the Q-valley spin-orbit splitting. The inputs of the transport model therefore sit at the level of theory of the reference parametrization, while being used directly in the EMC solver. The slab dielectric constant of 5.31 in the 20.2 Å cell unfolds to α2D=6.9\alpha_{2D} = 6.9 Å and a screening length r0=43.5r_0 = 43.5 Å, against 41 to 45 Å in the literature [58, 59] and the 40 Å used in the transport model. It also fixes the strictly two-dimensional limit of the Fröhlich coupling, 0.35 eV, which agrees with the interpolated matrix element at the smallest wavevector within 5% (Supplementary Note S2).

QuantityThis workReferenceStatus
Lattice constant a0a_03.185 Å3.16 Åmatch (+0.8%)
K-valley mass mK∗m^{*}_{K} (band edge)0.489 m0m_0 (0.477 K-Γ, 0.530 K-M)0.48 m0m_0match
Q-valley masses ml/mtm_l/m_t0.646/1.258 m0m_00.64 to 0.66/1.21 to 1.31 m0m_0 [45]match
Band gap (direct, K)1.67 eV1.6 to 1.7 eVmatch
E′ phonon (LO/TO at Γ)46.2 meV∼48 meVmatch
A1′A^{\prime}_{1} phonon49.1 meV50 meVmatch
Born charges ZMo∗/ZS∗Z^{*}_{\mathrm{Mo}}/Z^{*}_{\mathrm{S}}−0.99/+0.49−1.0/+0.5match
Sound velocity vLAv_{\mathrm{LA}}6580 m/s∼6600 m/smatch
Sound velocity vTAv_{\mathrm{TA}}4020 m/s∼4200 m/smatch
2D polarisability α2D\alpha_{\mathrm{2D}} (r0r_0)6.9 Å (43.5 Å)40 to 45 Å [58, 59]match
Q-valley spin-orbit splitting71 meV70 to 75 meV [45]match
Valley gap E(Q)−E(K)E(Q)-E(K), coupling geometry265 meV118 meV (experiment [36]), 70 to 405 meV (DFT spread)geometry-sensitive
Valley gap, PAW geometry190 meV
Valley gap, aexpa_{\mathrm{exp}} with SOC113 meVtransport model uses 118 meV (§ 2.1.4.2.5.2)

Table 4. Computed band, phonon and dielectric inputs of the EMC transport model (monolayer MoS2_2, PBE, norm-conserving calculation) compared against the reference parametrization [17] and experiment. The electron-phonon couplings are listed in § 2.1.1 and Supplementary Note S2.

The separation between the K minimum and the six Q valleys at 0.51 of Γ-K is the input of Table 4 most sensitive to the level of theory and the geometry. Our PBE bands give 190 meV at the PAW-relaxed geometry and 265 meV at the norm-conserving geometry of the electron-phonon calculations, both without spin-orbit coupling. Published values range from 70 meV for the local-density approximation at a compressed lattice constant [23, 60] to 405 meV for the HSE06 functional [60], and the only direct measurement, the pressure-driven K-A crossover of Fu et al. [36], gives 118 meV. Supplementary Note S1 traces this spread to the sulphur height, the lattice constant and spin-orbit coupling, which lowers every value by 33 to 34 meV because the Q valley splits by 71 meV and the K valley by only 3 meV. At the experimental lattice constant with the relaxed sulphur height and spin-orbit coupling, the separation is 113 meV, within 5 meV of the measurement. Because the couplings are insensitive to this difference whereas the valley population is not, the transport model takes the couplings from the DFPT geometry and the separation from the measurement, 118 meV, with its own spin-orbit splitting of 71 meV, and § 2.1.4.2.5.2 gives the sensitivity to the separation.

All electron-phonon couplings, zero and first order, intravalley and intervalley, polar and non-polar, are computed from the same calculations (§ 2.1.1, Supplementary Note S2), so that no transport input is taken from another calculation. The direct and interpolated couplings agree where they overlap (acoustic couplings within 5 % for q≥0.13q \ge0.13 Å−1^{-1}, homopolar coupling within 10 %), and the transport model built from them reproduces the Boltzmann solution of the Wannier interpolation. The scattering rates agree within 5 % above 10 meV at 150, 300 and 400 K, and the phonon-limited mobility, computed with the same valley positions and without free-carrier screening, agrees within 1 % at 150 K and within 13 % at 400 K (Supplementary Note S2, Figure S6).

Figure S6. The transport model against the Boltzmann solution of the Wannier interpolation. (a) Total scattering rate of a K-valley electron versus its energy above the band edge at 150, 300 and 400 K, the state-resolved rates of the Boltzmann solution averaged in 5 meV bins (symbols) and the rates of the transport model with the same couplings (lines). (b) Phonon-limited mobility of the suspended layer, the transport model with the Q valleys at 265 meV and the polar coupling unscreened (circles, standard errors smaller than the symbols), the iterative (squares) and relaxation-time (triangles) Boltzmann solutions, and the transport model of the main text with the Q valleys at the measured 118 meV and the polar coupling screened by the free carriers (diamonds).

Sensitivity of the transport results

Table 5 quantifies how the level of theory and the couplings influence the room-temperature mobility. Varying the K-valley transport mass over the PBE-to-GW spread (0.43 to 0.53 m0m_0 around the fitted 0.446 m0m_0) changes the mobility by +6 and −27 % for the suspended layer and by +6 and −27 % on HfO2_2. The band gap affects the transport only through the non-parabolicity, so its PBE underestimate is immaterial for electron transport. Raising the K-Q separation from the measured 118 meV to the 190 meV of the PAW geometry raises the mobility by 7 and 5 %, and removing the spin-orbit splitting of the Q valleys, which doubles the density of final states at the Q edge, lowers it by 5 and 4 %. Replacing every computed coupling by the LDA couplings of Kaasbjerg et al. [17] raises the suspended mobility to 392 cm2^2/Vs, close to their impurity-free value of 410 cm2^2/Vs. Replacing only the optical couplings changes it by −4 %, and removing the first-order K-K′ channels by +8 %, so the entire difference between the two first-principles values is the LA deformation potential. Charged interface impurities on the HfO2_2 stack reduce the mobility from 188 to 173, 148, 110, 63, 30 and 13 cm2^2/Vs for densities of 1012^{12}, 3 × 1012^{12}, 1013^{13}, 3 × 1013^{13}, 1014^{14} and 3 × 1014^{14} cm−2^{-2}, respectively (standard errors 0.2 to 5 cm2^2/Vs, screening at 1013^{13} cm−2^{-2}). Even 3 × 1013^{13} cm−2^{-2}, the upper end of the sulphur-vacancy densities measured in CVD films, leaves 63 cm2^2/Vs, so the factor of about fifty between the intrinsic value and the CVD device of § 2.1.4.2.5.3 is a contact and grain-boundary limitation rather than a Coulomb-scattering effect.

VariantSuspendedHfO2_2
Model as used (m∗=0.446m0m^{*}=0.446m_0, EQK=118E_{QK}=118 meV, ΔQ=71\Delta_Q=71 meV, computed couplings)214 (1)151 (3)
m∗=0.43m0m^{*}=0.43m_0227 (1)161 (1)
m∗=0.53m0m^{*}=0.53m_0156 (2)110 (2)
EQK=190E_{QK}=190 meV (PAW geometry)229 (5)158 (2)
ΔQ=0\Delta_Q=0 (degenerate Q valleys)202 (3)145 (3)
Computed acoustic, LDA optical and first-order couplings205 (1)142 (1)
LDA couplings of Ref. [17] throughout (constant Ξ)392 (2)218 (3)
No first-order K-K′ channels232 (1)154 (4)

Table 5. Sensitivity of the room-temperature mobility (cm2^2/Vs, four seeds, standard errors in parentheses) to the band inputs, the valley model and the electron-phonon couplings, for the suspended layer and on HfO2_2 at 101210^{12} cm−2^{-2}.

The K-Q separation affects the transport model directly through the Q-valley population and the K-Q intervalley rates. Its consequences were mapped over the published range with the two-valley parameters of Pilotto et al. [61] in Supplementary Note S1 (Table S3, Figure S2). From 190 to 118 meV, the low-field mobility falls by 10 % and the drift velocity at 5 MV/m by 20 %. At 70 meV, the LDA value at a compressed lattice constant, the mobility halves. That is the factor of two between Li et al. [23] and Kaasbjerg et al. [17] which Poncé et al. [26] trace to this parameter. With the first-principles transport model, the same dependence appears in Table 5. Raising the separation to the 190 meV of the PAW geometry gains 7 %. Switching the spin-orbit splitting off, which leaves a degenerate set of six valleys at 118 meV with twice the density of final states at the Q edge, costs 5 % suspended and 4 % on HfO2_2.

EQKE_{QK} (meV)μ0\mu_0 (cm2^2/Vs)pQp_Q (0.2 MV/m)v2v_2 (10410^4 m/s)v5v_5 (10410^4 m/s)v10v_{10} (10410^4 m/s)pQp_Q (10 MV/m)
single valley286 (3)04.14 (0.11)6.06 (0.04)5.79 (0.05)0
190 (PAW geometry)291 (4)0.0033.95 (0.04)4.81 (0.04)4.40 (0.01)0.21
160284 (4)0.0083.77 (0.08)4.49 (0.05)4.23 (0.03)0.24
148283 (3)0.0133.68 (0.07)4.32 (0.03)4.16 (0.01)0.26
118 (experiment)261 (3)0.0403.26 (0.03)3.83 (0.02)3.92 (0.01)0.30
70 (LDA, 3.13 Å)140 (2)0.211.66 (0.03)2.55 (0.02)3.31 (0.01)0.41

Table S3. Low-field mobility μ0\mu_0 at 0.2 MV/m, drift velocity v2v_2, v5v_5 and v10v_{10} at 2, 5 and 10 MV/m (standard errors in parentheses) and Q-valley occupation pQp_Q at 0.2 and 10 MV/m of the two-valley (K+Q) transport model at 300 K for the suspended monolayer with intrinsic phonons only, averaged over three seeds.

Figure S2. Velocity-field characteristic of the two-valley model for the separations of Table S3 and for its single-valley limit, together with the transport model of the main text (K and the two spin-orbit-split Q sets, first-principles couplings, four seeds) and the measured saturation velocity of Smithe et al. [62]. Error bars are standard errors over seeds and are mostly smaller than the symbols.

Device characteristics

The reduced model is first verified against the direct ensemble Monte Carlo simulation. Figure 6(a) and (Figure 6b) compare the two over Vg=0V_g = 0 to 2 V and Vds=0.1V_{ds} = 0.1 to 1 V. The reduced model is fed with the tabulated bulk v(E)v(E) of the HfO2_2-supported stack of the device (μ0=173 cm2/Vs\mu_0 = 173\ \mathrm{cm}^2/\mathrm{Vs} at the screening density of 1013 cm−210^{13}\ \mathrm{cm}^{-2}), the settled channel density ns(Vg)n_s(V_g) of the zero-bias runs (1.03×1013 cm−21.03 \times10^{13}\ \mathrm{cm}^{-2} per volt, Vth=−0.26V_{\mathrm{th}} = -0.26 V) and an access resistance of 530 Ω\Omega for the 500 nm wide device, taken as the median of the potential drops outside the channel interior. The two agree within 10 % at eleven of the thirteen bias points and within 25 % at all of them. At Vg=0V_g = 0 and 0.5 V, Vds=0.1V_{ds} = 0.1 V, the Monte Carlo simulation lies 24 and 13 % below the reduced model. There, the channel density (2.2×10122.2 \times10^{12} and 7.1×1012 cm−27.1 \times10^{12}\ \mathrm{cm}^{-2}) is below the screening density of the bulk v(E)v(E), so the local mobility is lower than the tabulated one (Figure 5(b)). At Vg=2V_g = 2 V, Vds=0.1V_{ds} = 0.1 V, it lies 8 % above, because the fixed access resistance overstates the drop in the n+^+ regions, which become more conductive with the gate. A density-resolved table v(E,ns)v(E,n_s) and a gate-dependent access resistance would remove most of both. Figure 6(c) and (Figure 6d) show the sheet density and electron energy along the channel, respectively, with the gradual decrease of the density towards the drain, as expected of a gated channel. The reduced model is therefore a verified reduced-order form of the full simulation rather than an empirical fit, with its residual error stated.

Figure 6. Direct self-consistent three-dimensional EMC simulation of the gated monolayer channel compared with the reduced one-dimensional channel model fed with the bulk EMC v(E)v(E) (no fitted parameter). (a) Output characteristics for Vg=0.5V_g = 0.5, 1, 1.5 and 2 V and (b) transfer characteristics at Vds=0.1V_{ds} = 0.1 and 0.5 V, where symbols are the Monte Carlo simulation with twice the block standard error, which is smaller than the symbols, and lines are the reduced model with the tabulated EMC v(E)v(E) including the access resistance of the n+n^{+} regions. (c) Sheet density and (d) electron energy along the channel at Vg=1.5V_g = 1.5 V for Vds=0V_{ds} = 0, 0.1, 0.5 and 1 V, where the shaded regions are the n+n^{+} contact regions.

Figure 7(a) shows the EMC velocity-field characteristic v(E)v(E) of the suspended layer. It rises linearly at low field (μ0=213 cm2/Vs\mu_0 = 213\ \mathrm{cm}^2/\mathrm{Vs}) and bends over above 1 MV/m into a saturation set by intervalley transfer to the heavier Q valleys, which hold 21 % of the electrons at 5 MV/m and 32 % at 10 MV/m, where v=3.55×106 cm/sv = 3.55 \times10^6\ \mathrm{cm}/\mathrm{s} (standard error 0.01), in agreement with the measured (3.4±0.4)×106 cm/s(3.4 \pm0.4) \times10^6\ \mathrm{cm}/\mathrm{s} [62]. The device comparisons below are made at drain fields of at most 0.1 MV/m, where v(E)v(E) is linear. Used as the input of the reduced model (Eqs. (8) through (10)), v(E)v(E) reproduces the transfer characteristic of Figure 7(b) against a CVD monolayer device [21] in vacuum and air. The model reproduces both branches, and the air→\rightarrowvacuum change is the EMC-predicted ambient response. At 1 atm, the ambient model of § 2.1.4.2.5.4 gives a mobility reduced by a factor of 2.4 and a carrier density reduced by 1.8, a 4.3-fold conductance change, represented in the device model by the measured mobility ratio of 4 and a threshold shift. The absolute channel mobility of this device, 4 cm2^2/Vs in vacuum, is two orders of magnitude below the intrinsic value (Figure 4) and is a device parameter (Table 3). Even 3×1013 cm−23 \times10^{13}\ \mathrm{cm}^{-2} charged interface impurities leave 63 cm2^2/Vs in the EMC (§ 2.1.4.2.5.2), so the remaining reduction is a contact and grain-boundary limitation outside the transport model, while the velocity-saturation field and the ambient mobility ratio are EMC predictions.

Figure 7. Reduced one-dimensional channel model fed by the EMC v(E)v(E). (a) EMC velocity-field characteristic of the multivalley transport model (four seeds, standard errors smaller than the symbols except at the lowest field), with the linear low-field relation and the measured saturation velocity of Smithe et al. [62]; the saturation above 2 MV/m is set by intervalley transfer into the Q valleys. (b) Transfer IdI_\mathrm{d}-VgV_\mathrm{g} versus the measured monolayer FET of Ahn et al. [21] in air and vacuum, with the device parameters of Table 3.

Ambient and gas sensing

Figure 8 shows the ambient conductivity response. The EMC σ(θ)\sigma(\theta) curve is populated with device measurements from Ahn et al. (seven pressures) [21] and Li et al. [22], alongside the mechanism decomposition into charge-transfer depletion and adsorbate scattering. Figure 9 shows the gas response against monolayer devices. The same EMC transduction σ(θ)\sigma(\theta) reproduces the NH3_3 and NO2_2 concentration dependence of the CVD monolayer FET of Liu et al. [63] and the NO2_2 response of the monolayer device of Kumar et al. [56], with only the adsorption constant KK differing between devices. The NO2_2 constants of the two devices differ by two orders of magnitude, which Eq. (12) attributes to effective adsorption energies of 0.47 and 0.60 eV of their site distributions. The NH3_3 data of Liu et al. were taken at zero back-gate voltage, where the baseline density is below the 3 × 101210^{12} cm−2^{-2} of the transduction used in Figure 9. At a baseline of 101210^{12} cm−2^{-2}, the donor response is larger at every coverage (−59 against −25% at θ=0.05\theta=0.05, −93 against −76% at θ=0.5\theta=0.5, standard errors about 1%), because the transferred charge is a larger fraction of the channel electrons. The shape of the response is the same, so a lower baseline rescales the fitted adsorption constant and leaves the concentration dependence unchanged. The acceptor branch at this baseline depletes the channel above θ≈0.4\theta\approx0.4 and is not used. Figure 10 makes the device-level comparison, where the NO2_2 acceptor response of the same monolayer device [56] is reproduced as a Langmuir-saturating threshold shift. Few-layer and multilayer devices, the majority of the published sensing data, are outside the scope tested for this monolayer model. Supplementary Note S5 compares the monolayer transduction with a 2-layer device [2] and a multilayer melamine-doped FET [55] as qualitative comparisons, with the screening-length argument for which features of the monolayer response apply to few-layer films and why the gate-tunable selectivity of thick films is an interlayer effect outside the model.

Figure 8. The ambient conductivity response of the channel is shown in (a) as the EMC σ(θ)\sigma(\theta) versus adsorbate coverage, overlaid with measured device data from Ahn et al. [21] and Li et al. [22], and is decomposed in (b) into the charge-transfer carrier-depletion and adsorbate-scattering contributions.

Figure 9. The same EMC transduction σ(θ)\sigma(\theta) reproduces the measured gas response of monolayer devices, shown in (a) as the NH₃ donor response and in (b) as the NO₂ acceptor response of the CVD monolayer FET of Liu et al. [63], together with the NO₂ response of the monolayer device of Kumar et al. [56], each fitted with a single per-device adsorption constant. The standard errors of the transduction over eight seeds (2 to 9 % for NO₂, below 5 % for NH₃) are drawn as bands under the lines and are narrower than the line width.

Figure 10. Device-level charge transfer, the NO2NO_2 acceptor threshold voltage versus concentration of the monolayer device of Kumar et al. [56], reproduced by the charge-transfer model as a Langmuir-saturating shift.

Scope and limitations

Table 6 states, per result, which quantities are EMC-predicted and which are fitted. The transport model rests on band parameters and electron-phonon couplings computed for this work, with no fitted transport parameter and no coupling taken from another calculation (§ 2.1.1). The room-temperature mobilities, the temperature exponents, the density peak and the substrate ordering are predictions. The device level uses the EMC v(E)v(E) and, for the verification against the Monte Carlo simulation, no fitted parameter. For the measured CVD device the threshold, the subthreshold swing, the contact resistance and the channel mobility are device parameters (Table 3). The fitted quantities are confined to the sensor transduction, where the per-gas charge-transfer amplitude, compared with first-principles values in Supplementary Table S11, and one adsorption constant per dataset, interpreted through Eq. (12), are fitted while the shape S(θ)S(\theta) is an EMC output.

ResultEMC-predicted (no fit)Fitted / external
μ(T)\mu(T) (Figure 3,Figure 4)γ\gamma, absolute μ\munone
Substrate, μ(ns)\mu(n_s) (Figure 5)ordering, non-monotonic peaknone (substrate dielectric constants and phonon energies from the literature)
DFPT inputs (Table 4, § 2.1.1)band, phonon and dielectric inputs, all electron-phonon couplingsK-Q separation from the measurement (§ 2.1.4.2.5.1)
Device I-V, Monte Carlo simulation (Figure 6)v(E)v(E) table, ns(Vg)n_s(V_g), access resistance, terminal currentsnone
Device I-V, measured CVD device (Figure 7(b))v(E)v(E) table, saturation field, ambient conductance ratio 4.3 (measured 4)VthV_{\mathrm{th}}, SSSS, RcR_c and channel mobility of the device (Table 3)
Ambient (Figure 8)σ(θ)\sigma(\theta), decompositionpressure→\tocoverage isotherm
Gas (Figure 9,Figure 10)transduction shape S(θ)S(\theta)ΔN\Delta N, NcN_c (Supplementary Table S11), KK per device (Eq. (12))

Table 6. Summary of predicted versus fitted quantities.

QuantityNO2_2 (acceptor)NH3_3 (donor)
ΔN\Delta N at full coverage (fitted)2.2×1012 cm−22.2 \times10^{12}\ \mathrm{cm}^{-2}2.0×1013 cm−22.0 \times10^{13}\ \mathrm{cm}^{-2}
Ncmax⁡N_c^{\max} (fitted)3.0×1012 cm−23.0 \times10^{12}\ \mathrm{cm}^{-2}2.0×1011 cm−22.0 \times10^{11}\ \mathrm{cm}^{-2}
Effective charge per molecule (one per 4×44 \times4 cell)0.030 e0.030\ e0.28 e0.28\ e
Bader charge, this work (pristine plane)0.0390.039 to 0.048 e0.048\ e0.0120.012 to 0.027 e0.027\ e
EadsE_{\mathrm{ads}}, this work (pristine plane)−0.17-0.17 to −0.18 eV-0.18\ \mathrm{eV}−0.10-0.10 to −0.17 eV-0.17\ \mathrm{eV}
Effective binding energy Eb=−EadsE_b = -E_{\mathrm{ads}} of the fitted KK0.47 and 0.60 eV0.47 eV
Bader charge, monolayer, PBE [54]0.100 e0.100\ e0.069 e0.069\ e
Bader charge, bilayer, PBE [2]0.097 e0.097\ e0.009 e0.009\ e

Table S11. Fitted charge-transfer amplitudes of the sensor model of § 2.1.4.2.2 of the main text compared with the Bader charge transfers per molecule computed for this work (PBE-D3, 4×44 \times4 cell, range over three sites, Table S10) and with published values.

The framework is developed for the monolayer, and all quantitative comparisons (the DFPT verification in Table 4, the transport and device results in Figure 3 through 7, and the sensor results in Figure 9 and Figure 10) are against monolayer data. Few-layer and multilayer devices are outside the tested scope and are compared in Supplementary Note S5 only qualitatively, with the physical reasons given there, and their absolute magnitudes are not claimed as monolayer validation.

Three approximations of the transport model remain at the present level of theory. First, the couplings are used isotropically although the direct calculation shows them to be direction dependent at finite wavevector (the LA coupling along Γ\Gamma-K is 15 to 20 % below that along Γ\Gamma-M), the Q-valley acoustic couplings are constant, and the K valley is an isotropic Kane band. The comparison with the Boltzmann solution of the Wannier interpolation, which carries the full wavevector and direction dependence and the warped band, bounds their combined effect at 5 % on the scattering rates and at 13 % on the phonon-limited mobility at 400 K (Supplementary Note S2). Second, the K-Q valley separation depends on the geometry and on spin-orbit coupling at the level of tens of meV (§ 2.1.4.2.5.1). The transport model uses the measured value, and § 2.1.4.2.5.2 gives the sensitivity to it. Third, the free-carrier screening of the polar coupling is static, and the plasma-frequency argument of § 2.1.1 bounds the error at 30 % of a channel that contributes a fifth of the room-temperature scattering. The Boltzmann solution of Supplementary Note S2 carries interpolated matrix elements throughout the Brillouin zone [47, 64] and so removes the first approximation for the phonon-limited mobility, where it bounds the difference at 8 % at room temperature. Carrying that treatment through the ensemble Monte Carlo would mean abandoning the tabulated-rate form on which the self-consistent device and sensor levels rest.

Conclusion

We have presented a self-consistent, multiscale ensemble Monte Carlo framework for monolayer MoS2_2 that spans first-principles transport inputs, device characteristics, and ambient and gas-sensing response. The Boltzmann transport model carries the intrinsic and extrinsic scattering mechanisms together with a self-consistent Poisson solution and degenerate carrier statistics. It feeds a reduced one-dimensional channel model, verified against the direct self-consistent EMC simulation of the gated channel, and an adsorbate charge-transfer extension. The same microscopic inputs are used from electron-phonon coupling to terminal current and sensor transduction.

The model reproduces several independent measurements. Temperature-dependent mobility follows a substrate-specific power law whose exponent rises from γ=1.51\gamma= 1.51 for the suspended film to γ=1.92\gamma= 1.92 on HfO₂, as the thermally activated remote surface-optical phonons steepen the decrease, and the mobilities lie below the Kaasbjerg phonon limit [17]. The room-temperature mobility of 151 to 214 cm²/Vs and the non-monotonic density dependence agree with the best measured monolayer devices, and the intrinsic value provides an upper bound for the extrinsic limitation of each measured device. The band, phonon and dielectric inputs, computed by DFPT, agree with the reference parametrization and experiment in eleven independent quantities, and the sensitivity of the results to the level of theory, in particular through the K-Q valley separation, is reported. Without a fitted parameter, the reduced device model reproduces the direct EMC simulation within 10 % at eleven of thirteen bias points, and with the device parameters of the measurement it reproduces the transfer characteristic of a CVD monolayer device in air and in vacuum. The adsorbate extension reproduces the ambient conductivity, the oxygen-pressure response and the NO₂ and NH₃ sensing of monolayer devices. The separation of predicted and fitted quantities shows that the transport model uses couplings computed from the DFPT matrix elements without a fitted transport parameter, with the remaining fitting confined to the sensor transduction, whose amplitudes are compared with the charge transfers and adsorption energies computed for the same surface.

The framework can be used for modelling 2D-material field-effect transistors and chemical sensors. Every intrinsic input of the transport model comes from first-principles calculations of the same material at one level of theory, is checked against an independent Boltzmann solution with the same couplings, and is carried without further adjustment to the device and sensor level.

Funding

This research was funded in whole, or in part, by the Austrian Science Fund (FWF) 10.55776/P35318 and 10.55776/DOC142.

Author contributions

L.F. conceived the study, developed the ViennaEMC extensions and simulation drivers, performed the transport, device, sensor and first-principles calculations, analysed the results, and wrote the manuscript.

Data availability

The simulation data supporting this study (ensemble Monte Carlo outputs with their seeds, the bulk and device input decks, the first-principles DFPT inputs, and the band-structure and two-valley sweep data of Supplementary Note S1, the electron-phonon matrix elements and Boltzmann-solution files of Supplementary Note S2 and the adsorbate calculations of Supplementary Note S3) are openly available in the TU Wien Research Data repository at 10.48436/87zq1-28246 [65]. The ViennaEMC framework [39] is available in the project repository.

Supplementary data

Supplementary Notes S1 (the K-Q valley separation, its published values, geometry and spin-orbit sensitivities, and the two-valley transport bound), S2 (the first-principles electron-phonon couplings of the K and Q valleys and the comparison of the transport model with the Boltzmann solution of the Wannier interpolation), S3 (the adsorbate calculations and the sensor-model amplitudes), S4 (the extrinsic scattering forms, the degenerate carrier statistics and the comparison with the reference parametrization) and S5 (the comparison with few-layer devices) follow the reference list.

Ethics statement

This work is a computational and theoretical study based on simulation and first-principles calculation. It involved no research on human participants, no human data or tissue, and no animals, and therefore required no ethics approval.

Conflict of interest

The author declares no competing interests.

References

Supplementary material

Multiscale ensemble Monte Carlo of transport and gas sensing in monolayer MoS2\mathrm{MoS}_2

Note S1 concerns the conduction-band K-Q valley separation of monolayer MoS2\mathrm{MoS}_2, the input of the transport model that depends most strongly on the exchange-correlation functional and on the geometry of the layer. It collects the published values, traces their spread to the lattice constant, the sulphur height and spin-orbit coupling with PBE calculations at fixed and relaxed geometries, and quantifies the transport consequences of the remaining uncertainty with a two-valley ensemble Monte Carlo transport model. Note S2 documents the first-principles electron-phonon couplings of the K and Q valleys. It gives the matrix elements of the direct DFPT calculation by phonon mode, wavevector and direction, identifies the periodic-slab artefact in its small-wavevector acoustic couplings, gives the zero- and first-order intervalley couplings and the polar and homopolar couplings from the Wannier interpolation with the two-dimensional long-range treatment, lists the values used in the transport model, and compares the scattering rates and the phonon-limited mobility of the transport model with the Boltzmann solution of the Wannier interpolation. Note S3 reports the adsorption energies and charge transfers of NO2\mathrm{NO}_2, NH3\mathrm{NH}_3 and O2\mathrm{O}_2 on the monolayer and compares them with the amplitudes of the sensor model. Note S4 gives the forms of the extrinsic scattering mechanisms, the degenerate carrier statistics of the Poisson equation and the comparison of the computed couplings with the reference parametrization. Note S5 compares the monolayer transduction with few-layer devices qualitatively.

Supplementary Note S1 The K-Q valley separation

The conduction band of monolayer MoS2\mathrm{MoS}_2 has its minimum at the K and K′ points and a second set of six minima, the Q (also called Λ\Lambda or Σ\Sigma) valleys, at about half of the Γ\Gamma-K distance. Their energy above the K minimum, EQK=E(Q)−E(K)E_{QK}=E(Q)-E(K), sets the Q-valley population at a given temperature and the onset of K→\rightarrowQ intervalley scattering. Of all inputs of the transport model (Table 4 of the main text), this separation is the one that depends most strongly on the exchange-correlation functional and on the geometry of the layer. The main text reports the conclusions of this analysis, and this note gives the full account.

Published values

Table S1 collects the published values of EQKE_{QK}, which span 70 to 405 meV. The only published direct measurement that we are aware of is the pressure-driven K-Q crossover of Fu et al. [36], who followed the direct K-K and the indirect Q-K photoluminescence lines of a monolayer under hydrostatic pressure, found them to shift at +49.4+49.4 and −15.3-15.3 meV/GPa and to cross at 1.9 GPa, and so obtained EQK=118E_{QK}=118 meV at ambient pressure. Angle-resolved photoemission on alkali-intercalated monolayers gives a lower bound of about 60 meV [66, 45].

Methodaa (Å)SOCEQKE_{QK} (meV)Reference
LDA, LDA geometry3.129no90Kormányos et al. [60]
LDA, ultrasoft3.13no70Li et al. [23]
LDA, PAW3.14no∼200Kaasbjerg et al. [17]
LDA bands, HSE geometry3.157yes207Kormányos et al. [45]
PBE, PBE geometry∼3.18yes246Kormányos et al. [45]
HSE06HSEno405Kormányos et al. [60]
PBE, PAW geometry3.181no190this work (Table S2)
PBE, geometry of the electron-phonon calculations3.185no265this work (Note S2)
PBE, aexpa_{\mathrm{exp}}, relaxed, SOC3.160yes113this work (Table S2)
Experiment, pressure crossover118Fu et al. [36]
Experiment, ARPES lower bound≳ 60Eknapakul et al. [66]

Table S1. Published values of the K-Q separation of monolayer MoS2_2. aa is the lattice constant of the calculation, SOC denotes spin-orbit coupling.

Three trends can be extracted from the table. First, the separation decreases with decreasing lattice constant. The local-density approximation, which contracts the lattice to 3.13 Å, gives the smallest separations, and Kormá nyos et al. [45] attribute the difference between their two calculations to the lattice constants used. Second, the separation depends on the functional. The HSE06 hybrid functional raises the Q valley by several hundred meV [60], whereas converged GW corrections change the separation much less than they change the band gap [45]. Third, spin-orbit coupling lowers the separation by about half of the Q-valley splitting, because the conduction band splits by 70 to 75 meV at Q and by only 3 meV at K [45]. Sohier et al. [37] add the pseudopotential and the doping level [67] to this list. The same spread appears in the calculated mobilities. Li et al. [23] obtain a phonon-limited room-temperature mobility of 130 cm2^2/Vs with EQK=70E_{QK}=70 meV and 320 cm2^2/Vs when the Q valleys are excluded, Kaasbjerg et al. [17] obtain 410 cm2^2/Vs in a K-only calculation with EQK≈200E_{QK}\approx200 meV, and Poncé et al. [26] identify this parameter as the main source of the discrepancy between first-principles mobilities of MoS2\mathrm{MoS}_2.

Sensitivity to geometry and spin-orbit coupling

To separate these trends, we recomputed the separation with one functional (PBE) at fixed and relaxed geometries, with and without spin-orbit coupling. The calculations use Quantum ESPRESSO [40] with the norm-conserving scalar-relativistic and fully relativistic PseudoDojo pseudopotentials [42, 43], a plane-wave cutoff of 60 Ry (240 Ry for the density), a 12×12×112 \times12 \times1 k-point mesh, the two-dimensional Coulomb cutoff and a cell height of 6.33aa. For the sulphur height zSz_{\mathrm{S}} above the Mo plane we either adopted the projector-augmented-wave (PAW) relaxation used for the parameter validation of Sec. 3 of the main text (zS=1.576z_{\mathrm{S}}=1.576 Å at a=3.181a=3.181 Å) or relaxed with the norm-conserving pseudopotentials at fixed lattice constant (forces below 10−410^{-4} Ry/bohr). The bands were computed on an explicit list of wavevectors that covers the K minimum and its neighbourhood, from which the effective masses follow, and the segment between 0.40 and 0.60 of K-Γ in steps of 0.01, on which the Q minimum was located and refined by a parabola through the three lowest points. Table S2 lists the eight cases and Figure S1 plots them.

aa (Å)zSz_{\mathrm{S}} (Å)SOCEQKE_{QK} (meV)qQq_QmK∗/m0m^*_{K}/m_0mQ,l∗/m0m^*_{Q,l}/m_0ΔK\Delta_K (meV)ΔQ\Delta_Q (meV)
3.1811.576 (PAW)no190.00.5100.48/0.500.59
3.1811.576 (PAW)yes155.90.5150.47/0.460.562.771.7
3.1811.564 (NC)no254.90.5110.46/0.470.59
3.1811.564 (NC)yes219.70.5150.45/0.440.582.975.2
3.1601.565 (PAW fr.)no180.60.5110.47/0.480.59
3.1601.565 (PAW fr.)yes147.50.5160.46/0.450.562.871.4
3.1601.571 (NC)no145.20.5110.48/0.490.62
3.1601.571 (NC)yes112.60.5150.48/0.460.612.770.6

Table S2. K-Q separation from PBE at fixed and relaxed geometries, with and without spin-orbit coupling. zSz_{\mathrm{S}} is the sulphur height above the Mo plane, and PAW and NC (norm-conserving) mark the pseudopotential family with which it was relaxed, “PAW fr.” the PAW fractional coordinate transferred rigidly to the smaller lattice constant. qQq_Q is the position of the Q minimum along Γ-K in units of the Γ-K distance, mK∗m^*_{K} the K-valley mass along K-Γ and K-M, mQ,l∗m^*_{Q,l} the longitudinal Q-valley mass, and ΔK\Delta_K, ΔQ\Delta_Q the spin-orbit splittings of the conduction band at K and Q. The direct gap at K is 1.71 (1.63) eV without (with) spin-orbit coupling at the PAW geometry and changes by less than 0.06 eV over the other geometries.

Figure S1. K-Q separation from PBE as a function of the sulphur height at the two lattice constants, without (filled circles, solid lines) and with (open squares, dashed lines) spin-orbit coupling, data of Table S2. The dotted line is the measured value of Fu et al. [36].

Four conclusions can be drawn from the simulations:

  1. At a fixed geometry, the separation is independent of the pseudopotential family. The norm-conserving bands at the PAW geometry reproduce the 190.0 meV of the PAW calculation, its K-valley mass (0.48 m0m_0) and its band gap (1.71 eV).

  1. The sulphur height is as important as the lattice constant. At fixed aa, the separation decreases by 5.7 to 5.8 eV per Å of sulphur height, that is by about 6 meV per 0.001 Å. The two pseudopotential families relax the height to values 0.012 Å apart at a=3.181a = 3.181 Å (1.576 and 1.564 Å), and this difference alone moves the separation from 190 to 255 meV. The latter value agrees with the 246 meV that Kormányos et al. [45] report for PBE, and the same effect explains the two published LDA values of 90 and 200 meV at almost the same lattice constant [60, 17].

  1. For relaxed geometries, the separation decreases by about 165 meV per percent of biaxial strain, so that the experimental lattice constant of 3.160 Å gives 145 meV. The measured pressure coefficient of −65 meV/GPa [36] has the same sign.

  1. Spin-orbit coupling lowers every value by 33 to 35 meV, one half of the Q-valley splitting of 71 to 75 meV, whereas the K valley splits by only 3 meV. The valence-band splitting at K of 146 to 148 meV agrees with the literature value of 148 meV [45].

At the experimental lattice constant with the relaxed sulphur height and spin-orbit coupling, the separation is 113 meV, within 5 meV of the measured 118 meV. Because the sulphur height of a monolayer is not known experimentally to the 0.005 Å that would fix the separation to 30 meV, this agreement places the measurement inside the range spanned by the PBE geometries without singling out one geometry. Our transport model therefore uses the measured 118 meV, and the range from 70 to 255 meV is examined for its transport consequences below.

Transport consequences

To map the transport consequences of the separation over its published range, independently of our own transport model, we used the two-valley (K+Q) parameters of Pilotto et al. [61], which is implemented in ViennaEMC alongside the transport model of this work and shares its valley model. It describes the K valleys by an isotropic non-parabolic band (m∗=0.47m0m^* = 0.47 m_0, α=0.94\alpha= 0.94 eV−1^{-1}) and the six Q valleys by an anisotropic non-parabolic band (ml=0.54m0m_l = 0.54 m_0, mt=1.14m0m_t = 1.14 m_0, α=1.16\alpha= 1.16 eV−1^{-1}) placed EQKE_{QK} above K, with intravalley acoustic and zero-order intravalley and intervalley optical deformation potentials for all K-K′, K-Q and Q-Q transitions from the same reference. The separation sets the valley energies and, through the final-state densities, every intervalley rate. We ran the suspended layer with intrinsic phonon scattering only at 300 K for EQKE_{QK} = 70, 118, 148, 160 and 190 meV and, as the single-valley limit of the same model, with the Q valleys placed at 1 eV. Each point is the average over three seeds of 60 000 steps of 0.1 fs after a 1.5 ps transient, and the drift velocity is the occupation-weighted mean over the two valley types. Table S3 and Figure S2 show the results.

At the PBE separation of 190 meV, the Q valleys hold 0.3 % of the electrons at low field and the mobility equals the single-valley limit of the same model within the statistical error of 2 %. At the experimental 118 meV the occupation is 4 %, the mobility is 10 % lower than at 190 meV, and the drift velocity at 2 and 5 MV/m is 17 and 20 % lower. At 70 meV the mobility halves, which reproduces the factor of two between Li et al. [23] and Kaasbjerg et al. [17]. If the true separation were the 190 meV of the PAW geometry rather than the measured 118 meV used here, the low-field results of the main text would change by 10 % (Table 5 of the main text gives +7 % for this case) and the drift velocity at 5 MV/m by 20 %. Only a separation as low as the LDA value at a compressed lattice constant, which is excluded by the measurement, would change the results by a factor of two. The low-field results of the main text (temperature and substrate dependence, device and sensor response, all at fields below 0.2 MV/m) therefore carry at most a 10 % uncertainty from the valley model.

At high field, the Q valleys are populated for every separation. Above a few MV/m they hold 21 % of the electrons at 10 MV/m even at 190 meV, and the drift velocity of the two-valley model lies 21 to 24 % below the single-valley limit of the same model at 190 meV and 37 % below it at 118 meV. The two-valley model saturates at 3.93.9 to 4.8×1064.8 \times10^{6} cm/s depending on the separation, consistent with the (3.4±0.4)×106(3.4 \pm0.4) \times10^{6} cm/s measured by Smithe et al. [62]. Our transport model, with the first-principles couplings and the spin-orbit-split Q valleys at 118 meV, saturates at 3.6×1063.6 \times10^{6} cm/s with a third of the electrons in the Q valleys at 10 MV/m (Figure S2), inside the range obtained with the Pilotto parameters and in agreement with the measurement. The comparison of the separations is made with the Pilotto parameters throughout, because the two parametrizations also differ in their optical couplings and non-parabolicity (the single-valley limit of those parameters saturates near 6×1066 \times10^{6} cm/s). The comparisons with the measured device in the main text are made at drain fields of at most 0.1 MV/m over a 1 μ\mum channel, in the linear part of v(E)v(E), and are not affected.

Data

The Quantum ESPRESSO inputs and outputs of Table S2, the wavevector list, the extraction script, the ensemble Monte Carlo driver (validation/mos2_kq_sensitivity.cpp in the ViennaEMC repository, with the compile-time override of the valley separation in the two-valley parameters) and the raw sweep data of Table S3 are part of the dataset accompanying the main text.

First-principles electron-phonon couplings

Our transport model uses electron-phonon couplings computed for this work. This note documents the two independent DFPT calculations, gives the matrix elements by phonon mode, wavevector and direction, states which values are used in the transport model and why, and compares the two calculations with each other and with the reference parametrization of Kaasbjerg et al. [17]. All calculations use Quantum ESPRESSO 7.3.1 [40] with the PBE functional [41], the norm-conserving scalar-relativistic PseudoDojo pseudopotentials [42, 43] (14 valence electrons for Mo, 6 for S, so that the conduction band is band 14), a plane-wave cutoff of 90 Ry (360 Ry for the density), a 12×12×112 \times12 \times1 k-point mesh and a hexagonal cell of height 6.33 aa. The in-plane lattice constant and the sulphur height were relaxed with the two-dimensional Coulomb cutoff [49] and the cell free in the plane only, which gives a=3.185a = 3.185 Å and zS=1.5625z_{\mathrm{S}} = 1.5625 Å (forces below 10−410^{-4} Ry/bohr, in-plane stress below 0.1 kbar). At this geometry, the K-Q separation without spin-orbit coupling is 265 meV, larger than at the PAW geometry of Note S1 because aa is larger and zSz_{\mathrm{S}} smaller, in line with the sensitivities established there. The couplings are insensitive to this difference, and the valley energy of the transport model is set from the measurement as described in Note S1.

Method

Direct DFPT matrix elements. The DFPT matrix elements gκα(k,q)=⟨ψk+q∣∂V/∂uκα(q)∣ψk⟩g_{\kappa\alpha}(\mathbf{k}, \mathbf{q}) = \langle\psi_{\mathbf{k}+\mathbf{q}} \vert\partial V / \partial u_{\kappa\alpha}(\mathbf{q}) \vert\psi_{\mathbf{k}} \rangle of the conduction band are written by the Allen-Heine-Cardona functionality of the phonon code [46], which stores them for a chosen initial state k\mathbf{k} and an explicit phonon wavevector q\mathbf{q} without any interpolation. For each q\mathbf{q}, the steps are a self-consistent calculation, a DFPT calculation of the dynamical matrix and of the self-consistent potential derivative at q\mathbf{q}, a non-self-consistent calculation at k\mathbf{k} and k+q\mathbf{k} + \mathbf{q}, and the evaluation of the matrix elements. The implementation requires norm-conserving pseudopotentials and does not support the two-dimensional Coulomb cutoff, so this calculation relies on the vacuum of the cell (cell height 20.2 Å) to isolate the layer. This is adequate for the short-range and zone-boundary couplings but leaves the long-range part of the acoustic and polar-optical couplings three-dimensional at wavevectors below about 2π/L2\pi/L, which is why the small-qq points are treated separately below. The mode-resolved couplings follow from the projection of gκαg_{\kappa\alpha} on the eigen-displacements eκανe_{\kappa\alpha}^{\nu} of the dynamical matrix, gν=∑καgκαeκαν/2Mκωνg_{\nu} = \sum_{\kappa\alpha} g_{\kappa\alpha} e_{\kappa\alpha}^{\nu}/\sqrt{2M_{\kappa}\omega_{\nu}}, and the deformation potentials are defined as in Ref. [17], Ξν=∣gν∣2Mcellων/∣q∣\Xi_{\nu} = \lvert g_{\nu}\rvert\sqrt{2M_{\mathrm{cell}}\omega_{\nu}}/\lvert q\rvert for the acoustic and Dν=∣gν∣2McellωνD_{\nu} = \lvert g_{\nu}\rvert\sqrt{2M_{\mathrm{cell}}\omega_{\nu}} for the optical modes, with McellM_{\mathrm{cell}} the mass of the unit cell. Three runs of this calculation were made, with the electron at K and 15 wavevectors for the intravalley and K-K′ and K-Q couplings (Table S4), with the electron at K and six wavevectors within 0.1 Å−1^{-1} of K for the first-order K-K′ couplings (Sec. S2.4), and with the electron at Q₁ = 0.515 ΓK and 13 wavevectors (Sec. S2.5). Figure S3 shows the valleys, the sampled phonon wavevectors and the phonon dispersion with the mode labels used throughout this note.

Figure S3. (a) Brillouin zone of the monolayer with the K and K′ corners and the six Q valleys (Q₁ to Q₆ at 0.5 of Γ-K, where the two colours mark the two spin-orbit sets, Q₁, Q₃, Q₅ and Q₂, Q₄, Q₆). The direct calculation places the electron at K and at Q₁. (b) The phonon wavevectors of the direct calculation, namely the eleven wavevectors of the first run with the electron at K, the four small wavevectors along Γ-K added for the artefact analysis, the six wavevectors near K for the first-order couplings, and the thirteen wavevectors of the run with the electron at Q₁. The dotted line is the Γ-M segment on which the Wannier interpolation gives the couplings at 30 points. (c) Phonon dispersion of the interpolation along Γ-M with the branch labels used in the tables, and the nine modes at K from the dynamical matrix of the direct calculation, with the four modes of the K-K′ channels labelled at the energies used in the transport model.

∣q∣|q| (Å−1^{-1})directionΞLA\Xi_{\mathrm{LA}} (eV)ΞTA\Xi_{\mathrm{TA}} (eV)DA1′D_{A'_1} (eV/Å)DE′D_{E'} (eV/Å)
0.050Γ\Gamma-K5.637.773.328.44
0.068Γ\Gamma-M8.30.393.216.58
0.080Γ\Gamma-K5.445.003.145.88
0.08015∘15^\circ from Γ\Gamma-M6.623.523.135.88
0.100Γ\Gamma-K5.284.122.995.12
0.132Γ\Gamma-K5.033.312.724.45
0.137Γ\Gamma-M6.550.712.634.40
0.13715∘15^\circ from Γ\Gamma-M6.562.322.594.41
0.137Γ\Gamma-K5.003.222.674.40
0.228Γ\Gamma-M5.790.941.673.71
0.263Γ\Gamma-K4.222.401.433.64
0.342Γ\Gamma-M5.000.960.783.07
K→\rightarrowK′ (q=q = K, 1.3151.315 Å−1^{-1})E′E' mode at 9.89.8 THz (40.640.6 meV), D=1.66D = 1.66 eV/Å; all other modes below 0.10.1 eV/Å
K→\rightarrowQ (q=q = Q-K, 0.6350.635 Å−1^{-1})LA 22.222.2 meV, 1.15; LO 42.742.7 meV, 1.44; A1′A'_1 49.349.3 meV, 1.04; TA 45.545.5 meV, 0.57; TA 16.516.5 meV, 0.22 eV/Å

Table S4. Electron-phonon couplings of the conduction band at K from the direct DFPT calculation by wavevector and direction. ΞLA\Xi_{\mathrm{LA}} and ΞTA\Xi_{\mathrm{TA}} are the acoustic deformation potentials, DA1′D_{A'_1} the coupling to the homopolar out-of-plane optical mode (11.911.9 THz, 4949 meV) and DE′D_{E'} the root-sum-square coupling to the two in-plane optical modes (11.211.2 THz, 4646 meV), which in this calculation contains the three-dimensional long-range part, so that the transport model uses the interpolated value of Table S5 instead. The 0.0340.034 Å−1^{-1} point lies inside the acoustic-sum-rule residual of the dynamical matrix and is not used. The rise of ΞLA\Xi_{\mathrm{LA}} along Γ\Gamma-M and of ΞTA\Xi_{\mathrm{TA}} along Γ\Gamma-K at small wavevector is the periodic-slab artefact discussed in the text. The last two rows are the intervalley transitions with the electron at K, listed by phonon mode with its energy.

Wannier interpolation. The second calculation uses the EPW code [47, 48] with the two-dimensional Coulomb cutoff in the self-consistent and DFPT steps and the two-dimensional dipole treatment of the long-range electron-phonon coupling of Poncé et al. [50]. The dynamical matrices and potential derivatives are computed on a 12×12×112 \times12 \times1 q-mesh with the dielectric tensor and Born charges, the electronic states on a 24×24×124 \times24 \times1 k-mesh, and the conduction and valence manifolds are represented by eleven maximally localised Wannier functions [68] of Mo dd and S pp character (disentanglement window −12.18-12.18 to 0.730.73 eV, frozen window up to 0.80.8 eV above the conduction-band minimum, total spread 18.918.9 Å2^{2}, convergence 10−1010^{-10} Å2^{2}). The interpolated bands reproduce the DFT bands along Γ\Gamma-M-K-Γ\Gamma, and the couplings of the conduction band at K are interpolated onto 30 wavevectors along Γ\Gamma-M from 0.040.04 to 1.141.14 Å−1^{-1}. It has the correct two-dimensional long-wavelength behaviour and is used to check the direct calculation and to supply the polar-optical coupling, which is not accessible to the direct calculation. The same Wannier representation serves the phonon-limited mobility calculation of § 11.3.0.4.

K-valley couplings

Table S4 lists the couplings of the K-valley conduction band from the direct calculation and Fig. S4 compares them with the Wannier interpolation. Four observations determine the values used in the transport model.

  1. The direct calculation carries a long-range artefact at small wavevector. Along Γ\Gamma-K its TA coupling is 7.87.8, 5.05.0, 4.14.1 and 3.33.3 eV at 0.050.05, 0.080.08, 0.100.10 and 0.130.13 Å−1^{-1}, that is ΞTAq=0.39\Xi_{\mathrm{TA}}q = 0.39 to 0.430.43 eV/Å−1^{-1}, a 1/q1/q law that cannot arise from a deformation potential, and along Γ\Gamma-M the same term appears in the LA channel (11.111.1 and 8.38.3 eV at 0.0340.034 and 0.0680.068 Å−1^{-1}). This is the piezoelectric-like coupling of the periodic slab without Coulomb cutoff, three-dimensional at wavevectors below about 2π/L2\pi/L and with the three-fold angular pattern of the hexagonal layer (longitudinal along Γ\Gamma-M, transverse along Γ\Gamma-K). The interpolation with the cutoff is free of this term, as its LA coupling tends smoothly to 6.56.5 eV and its TA coupling to zero at q→0q \rightarrow0, as required for a non-degenerate band, and the two-dimensional piezoelectric coupling that it does contain is small. An angle average over the direct calculation at 0.1370.137 Å−1^{-1} would give ΞTA=2.3\Xi_{\mathrm{TA}} = 2.3 eV and overstate the TA coupling about three-fold. The transport model uses only the artefact-free data.

  1. The genuine couplings are direction dependent at finite wavevector. The LA coupling along Γ\Gamma-K, free of the artefact by symmetry, is 5.65.6, 5.45.4, 5.35.3, 5.05.0 and 4.24.2 eV at 0.050.05, 0.080.08, 0.100.10, $0.13

and 0.260.26 Å−1^{-1} against 6.56.5, 6.46.4, 6.36.3, 6.26.2 and 5.85.8 eV along Γ-M from the interpolation, and the TA coupling after subtracting the 1/q1/q term in quadrature is about 1.01.0 eV at 0.130.13 and 1.81.8 eV at 0.260.26 Å−1^{-1} along Γ-K and at 15∘15^\circ, against 0.70.7 and 0.950.95 eV along Γ-M. The tables of § 11.3.0.3 are the interpolated Γ-M curves multiplied by the root-mean-square ratio over directions, 1.01.0 below 0.050.05 Å−1^{-1} (isotropic limit), 0.910.91 (LA) and 1.31.3 (TA) at 0.1370.137 and 0.900.90 and 1.51.5 at 0.300.30 Å−1^{-1}.

  1. The LA coupling is about twice the reference value of 2.82.8 eV [17], which was obtained from an LDA calculation with a different projection, and the TA coupling is below the reference 1.61.6 eV. The LA coupling is therefore the quantity in which the two first-principles calculations differ most, and the transport model uses the computed wavevector-dependent values without any adjustable parameter.

  1. The homopolar coupling decays with wavevector in both calculations, from 4.74.7 (interpolation) or 3.43.4 eV/Å (direct) at the smallest wavevectors to 2.72.7 eV/Å at 0.150.15 Å−1^{-1} and 0.80.8 eV/Å at 0.300.30 Å−1^{-1}, where the two calculations agree within 10%10\%. The transport model uses the interpolated A1′A'_1 matrix element as a table (see Sec. S2.3). The K-K′ intervalley coupling through the E′E' mode at K, 1.7×1081.7 \times10^8 eV/cm, is 0.650.65 of the reference value. The out-of-plane A2′′A''_2 and the E′′E'' modes couple weakly or not at all (below 0.40.4 and 0.10.1 eV/Å), consistent with the mirror symmetry of the layer, and the first-order K-K′ couplings, which vanish for q=K\mathbf{q} = \mathrm{K} exactly, are obtained from wavevectors near K (see Sec. S2.4).

Polar-optical coupling and the optical sum

The polar coupling to the LO branch of the E′E' mode can only be obtained from the calculation with the two-dimensional treatment. Table S5 and Figure S4 give the interpolated LO, TO and A1′A'_1 matrix elements, which the transport model uses directly, together with a constant-coupling representation that quantifies the effect of the wavevector dependence, namely the two-dimensional Fröhlich element of Kaasbjerg et al. [17], g(q)=g0erfc⁡(σq/2)g(q) = g_0 \operatorname{erfc}(\sigma q/2) with g0=C/Acell=97g_0 = C/\sqrt{A_{\mathrm{cell}}} = 97 meV (C=0.286C = 0.286 eV Å, Acell=8.65A_{\mathrm{cell}} = 8.65 Å2^2) and σ=5.41\sigma= 5.41 Å, and the constant homopolar element ∣g∣=52|g| = 52 meV that corresponds to DA1′=3.2×108D_{A'_1} = 3.2 \times10^8 eV/cm. The interpolated LO coupling is finite at q→0q \to0, as required in two dimensions, and is reduced with increasing wavevector by the dielectric screening of the layer itself. The strictly two-dimensional limit follows analytically from the same DFPT run [35]. Our in-plane Born charges (ZMo∗=−0.99Z^*_{\mathrm{Mo}} = -0.99, ZS∗=0.49Z^*_{\mathrm{S}} = 0.49), the E′E' eigenvector and the zero-point displacement of the Mo-S reduced mass give g0=epzp/(2ε0Acell)=0.35g_0 = e p_{\mathrm{zp}}/(2\varepsilon_0 A_{\mathrm{cell}}) = 0.35 eV, which the layer screens as g0/(1+r0q)g_0/(1 + r_0q) with r0=2πα2D=43.5r_0 = 2\pi\alpha_{\mathrm{2D}} = 43.5 Å from our in-plane dielectric constant. This form gives 132132 meV at 0.0380.038 Å−1^{-1} against the 138138 meV of the interpolation, which confirms the long-range treatment. At larger wavevectors, the interpolated coupling decays more slowly than the simple screening form (81 against 96 meV at 0.076 Å−1^{-1}) because the interpolation carries the full wavevector dependence of the slab response.

Figure S4. Electron-phonon couplings of the K-valley conduction band. Lines and symbols are the same DFPT matrix elements, interpolated through Wannier functions from a coarse mesh (lines) and computed directly at each wavevector (symbols). (a) Acoustic deformation potentials from the Wannier interpolation with the two-dimensional Coulomb cutoff along Γ-M (lines) and from the direct DFPT matrix elements of the periodic slab along Γ-M (black symbols) and Γ-K (orange symbols, Table S4). The 1/q1/q rise of the direct LA coupling along Γ-M and of the direct TA coupling along Γ-K is the long-range artefact of the slab without cutoff, whereas the direct LA coupling along Γ-K is free of it and shows the genuine direction dependence. The dotted lines are the reference values of Kaasbjerg et al. [17]. (b) Optical coupling matrix elements, the LO (E′E') and homopolar (A1′A'_1) couplings of the interpolation and the direct calculation, the polar and homopolar elements of the transport model, and the root-sum-square over the three intravalley optical modes for the interpolation and for the transport model (Table S5).

∣q∣|q| (Å−1^{-1})LOTOA1′A'_1rssconstant polarconstant A1′A'_1rssratio
0.03813827715886521011.57
0.076965681177552911.28
0.11480756986552831.18
0.15271944845552761.11
0.228601225663752641.03
0.304531614572452570.98

Table S5. Intravalley optical coupling matrix elements of the conduction band at K along Γ-M from the Wannier interpolation with the two-dimensional long-range treatment (LO and TO branches of the E′E' mode, 46 meV, and the A1′A'_1 mode, 49 meV), the two elements of the constant-coupling representation described in the text, and the root-sum-square (rss) of the intravalley optical couplings of both, in meV. The two E′E' branches cross at 0.15 Å−1^{-1} and are identified by their coupling.

The constant polar element lies a factor of 1.2 to 1.3 below the interpolated LO coupling in the thermal window and a factor of two below it at 0.3 Å−1^{-1}, whereas the constant homopolar element exceeds the interpolated value for all wavevectors above 0.07 Å−1^{-1}. All three intravalley optical modes lie within 3 meV of each other (46 to 49 meV) and affect the transport through the same emission threshold and the same final-state density, so their combined strength is the relevant quantity. Taken as root-sum-squares, the three interpolated couplings and the two constant elements agree within 20 % for ∣q∣≥0.1|q| \ge0.1 Å−1^{-1} and within 3 % above 0.2 Å−1^{-1}. Their difference reaches 28 % at 0.08 and 57 % at 0.04 Å−1^{-1}, where the phase space for a thermal electron is small. A constant-coupling representation therefore reproduces the total intravalley optical strength within 10 to 20 %, with a different division between the polar and the homopolar channel than the interpolation. The transport model avoids this difference by using the interpolated ∣g(q)∣|g(q)| of both modes directly, with the free-carrier screening applied to the polar channel only.

First-order K-K′ couplings

The first-order intervalley couplings are linear in the phonon wavevector measured from K, with ∣g∣∝∣q−K∣=∣δ′−δ∣|g| \propto|q - K| = |\delta' - \delta| for an electron at K + δ\delta scattered to K′ + δ′\delta'. They therefore vanish for q=Kq = K irrespective of the electron position, and we obtain them from matrix elements with the electron at K and the phonon at K + δq\delta q. Table S6 lists Deff/∣δq∣D_{\mathrm{eff}}/|\delta q| for ∣δq∣=0.03|\delta q| = 0.03, 0.06 and 0.10 Å−1^{-1} along K-Γ and perpendicular to it. The TA, TO and A1′A'_1-like couplings are linear in ∣δq∣|\delta q| and isotropic to within 10 %, so each is a well-defined first-order deformation potential in the sense of Ref. [17], whose angular average ∣δ∣2+∣δ′∣2|\delta|^2 + |\delta'|^2 is used by the transport model. For the weaker and less regular LA coupling we use the mean value. The zero-order coupling of the LO-like E′E' mode is recovered unchanged at every δq\delta q (1.65 eV/Å), which confirms the separation into zero- and first-order channels. Compared with the reference parametrization, the TA and TO first-order couplings are two to three times larger (four and six times in the rate) and a fourth channel, the A1′A'_1-like mode at K, is present.

Q-valley couplings

The second run of the direct calculation places the electron at Q1_1 and covers the intervalley couplings for longitudinal and transverse wavevectors (Table S7) and the intervalley transitions to the five other Q valleys, to K′ and back to K (Table S8). Three points are specific to the Q valley. The intravalley acoustic couplings are smaller than at K and less anisotropic (2.7 and 3.2 eV for the longitudinal mode with longitudinal and transverse wavevector, 3.4 and 0.3 eV for the transverse mode), so that the transport model uses ΞL=3.0\Xi_{\mathrm{L}} = 3.0 and ΞT=2.4\Xi_{\mathrm{T}} = 2.4 eV, the averages at 0.120.12 Å−1^{-1}. The flexural (ZA) mode, which does not couple at K because of the mirror symmetry of the layer, couples to the Q valley for transverse wavevectors (2.6 eV equivalent at 0.120.12 Å−1^{-1}, 0.8 eV at 0.240.24 Å−1^{-1}). This channel is not part of the transport model and would add at most a few per cent to the acoustic rate of the Q valleys. The homopolar coupling at Q is larger than at K, 5.4×1085.4 \times10^{8} eV/cm below 0.120.12 Å−1^{-1}, and decays with wavevector as at K. As a consistency check, the Q→\toK matrix elements at q=K−Q1\mathbf{q} = \mathrm{K}-\mathrm{Q}_{1} reproduce the K→\toQ elements of the first run at q=Q−K\mathbf{q} = \mathrm{Q}-\mathrm{K} (Table S4) within 1% for all five phonon channels, as time-reversal symmetry requires.

∣q∣|q| (Å−1^{-1})directionΞL\Xi_{\mathrm{L}} (eV)ΞT\Xi_{\mathrm{T}} (eV)ΞZA\Xi_{\mathrm{ZA}} (eV)DA1′D_{A'_1} (eV/Å)
0.059longitudinal2.636.4305.53
0.118longitudinal2.703.3705.36
0.118transverse3.220.292.605.51
0.237longitudinal2.951.7804.25
0.237transverse1.240.590.844.30
0.395longitudinal2.650.5301.43
0.395transverse2.501.580.061.72
0.592longitudinal2.050.2800.69

Table S7. Intravalley couplings of the Q-valley conduction band (electron at Q1=0.515 ΓKQ_{1} = 0.515\,\Gamma\mathrm{K}) from the direct calculation for wavevectors along the valley axis (longitudinal, towards Γ\Gamma) and perpendicular to it (transverse). ΞL\Xi_{\mathrm{L}} and ΞT\Xi_{\mathrm{T}} refer to the longitudinal and transverse acoustic modes, ΞZA\Xi_{\mathrm{ZA}} to the flexural mode in the same units, DA1′D_{A'_1} to the homopolar mode.

transition∣q∣|q| (Å−1^{-1})channels (meV, 10810^{8} eV/cm)used for
Q1→Q2,Q6Q_{1} \to Q_{2}, Q_{6} (60°)0.677LA 23.0, 3.87; A1′A'_1 49.2, 2.07; TO 45.4, 1.41; TA 17.4, 0.78; LO 42.4, 0.74Q→Q nearest
Q1→Q3,Q5Q_{1} \to Q_{3}, Q_{5} (120°)1.173LO 42.8, 3.15; A1′A'_1 49.1, 2.97; LA 28.5, 1.93Q→Q next-nearest
Q1→Q4Q_{1} \to Q_{4} (180°)1.35528.5, 5.09; 22.5, 2.92; 47.4, 2.43; 40.6, 1.44Q→Q opposite
Q1→K′Q_{1} \to K'1.992A1′A'_1 49.1, 4.95; LA 28.5, 4.64; LO 43.9, 2.86; 19.0, 1.56Q→K′, K→Q′ (far)
Q1→KQ_{1} \to K0.638LA 22.2, 1.15; LO 42.7, 1.44; A1′A'_1 49.3, 1.03; TA 45.5, 0.58; TA 16.5, 0.22Q→K, K→Q (near)

Table S8. Intervalley couplings with the electron at Q1Q_{1} from the direct calculation, by phonon channel (mode energy in meV, zero-order deformation potential in 10810^{8} eV/cm). Modes with couplings below 0.5×1080.5 \times10^{8} eV/cm are omitted. The Q→\toK row is the reverse of the K→\toQ transition of Table S4. The K→\toQ′ transitions to the three far Q valleys use the Q→\toK′ elements by time reversal.

Values used in the transport model

For the K valley, the transport model uses the isotropic Kane band with m∗=0.446 m0m^{*} = 0.446\,m_{0} and α=1.16\alpha= 1.16 eV−1^{-1}, fitted to the interpolated dispersion between 25 and 150 meV in both directions. Acoustic scattering follows the tables ΞLA(q)\Xi_{\mathrm{LA}}(q) and ΞTA(q)\Xi_{\mathrm{TA}}(q) of § 11.2 (6.5, 5.65 and 4.8 eV and 0, 0.95 and 1.45 eV at 0, 0.137 and 0.30 Å−1^{-1}), and polar and homopolar scattering follow the A1′A'_1 tables of Table S5, joined below 0.04 Å−1^{-1} to the analytic limit of Section S2.3. Intervalley scattering uses the zero-order K-K′ coupling of 1.7×1081.7 \times10^{8} eV/cm at 40.6 meV and the first-order couplings of Table S6, with an intravalley TO(Γ\Gamma) coupling of 3.6 eV from the interpolated TO branch. Mass density, sound velocities and phonon energies are those of the computed lattice (3.03 ×10−6\times10^{-6} kg/m2^{2}, 6580 and 4020 m/s, 49.2, 46.3 and 40.6 meV). For the Q valleys it uses the band parameters of Note S1 (ml=0.646m_{\mathrm{l}} = 0.646, mt=1.258m_{\mathrm{t}} = 1.258, α=0.3\alpha= 0.3 eV−1^{-1}), the measured separation EQK=118E_{\mathrm{QK}} = 118 meV, the computed spin-orbit splitting of 71 meV (Figure S5), the intravalley values ΞL=3.0\Xi_{\mathrm{L}} = 3.0 and ΞT=2.4\Xi_{\mathrm{T}} = 2.4 eV with the A1′A'_1 table of Table S7, and the intervalley channels of Table S8 together with the K→\toQ channels of Table S4. Each channel is a separate zero-order mechanism with its own phonon energy, and the sub-valley assignment follows from the spin selection (§ 2.1.1 of the main text). There is no piezoelectric term. The free-carrier screening of the polar coupling and the extrinsic mechanisms (remote substrate phonons, charged impurities, roughness) are those of the main text and act on all three valley types. No number in this list is taken from another calculation.

Mode at Kℏω\hbar\omega (meV)K-Γ (0.03 / 0.06 / 0.10 Å−1^{-1})perpendiculartransport modelreference
TA22.512.3 / 12.1 / 11.612.3 / 11.9 / 11.212.05.9
LA28.51.6 / 1.3 / 0.92.1 / 2.4 / 2.82.03.9
LO (E′E')40.6zero order, 1.65 eV/Åsame1.7×1081.7 \times10^8 eV/cm2.6×1082.6 \times10^8 eV/cm
TO (E′E')45.74.7 / 4.9 / 5.24.6 / 4.6 / 4.74.81.9
A1′A'_1-like47.43.7 / 3.4 / 2.93.9 / 3.8 / 3.73.5not included

Table S6. First-order K-K′ intervalley couplings Ξ1=Deff/∣δq∣\Xi_1 = D_{\mathrm{eff}}/|\delta q| (eV) of the conduction band from the direct calculation with the electron at K and the phonon at K + δq\delta q, for three magnitudes of δq\delta q along K-Γ and perpendicular to it; the last columns give the value used in the transport model and the reference value [17]. The flexural, E′′E'' and A2′′A''_2 modes give zero.

Figure S5. Figure S5. (a) Lowest conduction bands of PBE along Γ-M-K-Γ at the experimental lattice constant with the relaxed sulphur height (the geometry of Note S1 closest to experiment), with and without spin-orbit coupling. The K valley splits by 3 meV and the Q valley by 72 meV, and the lower Q subband lies 113 meV above K against the measured 118 meV. The transport model uses the measured separation with this splitting. (b) The Q valley along K-Γ from the same calculations, the scalar-relativistic band and the two spin-orbit subbands that the transport model represents as the sets Qlow and Qhigh. The isotropic Kane band of the K valley used in the transport model (m∗=0.446m0m^* = 0.446m_0, α=1.16 eV−1\alpha= 1.16\ \mathrm{eV}^{-1}) is fitted to the Wannier-interpolated band of the electron-phonon calculations, and is not changed by the 3 meV splitting at K.

Phonon-limited mobility from the Boltzmann equation

The same Wannier representation is used to solve the linearised Boltzmann equation for the phonon-limited electron mobility of the suspended monolayer between 150 and 400 K (EPW [64, 48]), which checks the ensemble Monte Carlo transport model by a method that shares the first-principles couplings but none of its approximations. The Boltzmann solution carries the full wavevector and direction dependence of every matrix element, the trigonally warped band and the exact angular structure of the scattering, whereas the transport model uses the isotropic tables of § 11.3.0.3, the Kane band and analytic deflection-angle distributions. Electrons and phonons were interpolated on 200×200200 \times200 meshes at a carrier density of 1012 cm−210^{12}\ \mathrm{cm}^{-2}, at which the electron gas is non-degenerate at every temperature considered, so that the mobility summed over all states coincides with the mobility over the states above the Fermi level that the code reports by default. Both the relaxation-time (SERTA) and the iterative (IBTE) solutions were obtained. The interpolation uses the DFPT geometry without spin-orbit coupling, so the Q valleys lie 265 meV above K and are empty at these temperatures. For a comparison under identical conditions, the transport model was run with the Q valleys at the same 265 meV and with the free-carrier screening of the polar coupling switched off (eight seeds, 0.2 MV/m), because the Boltzmann calculation is for the undoped layer.

Figure S6(a) compares the total scattering rate of a K-valley electron as a function of its energy. Above 10 meV the two agree within 5% at all three temperatures, below and above the optical emission threshold at 46 to 49 meV, where the rate rises by a factor of 2.7 at 150 K through the onset of A′1_1 and LO emission. The only larger deviations are single 5 meV bins at 15 to 20 meV, where the Boltzmann rates scatter, and at the threshold, where the onset of the transport model is sharper than that of the interpolated rates (14 and 12%). In the first 10 meV above the band edge, the rate of the transport model exceeds the Boltzmann rate, by 60 to 80% in the first 2.5 meV and by 10 to 20% between 2.5 and 10 meV. The elastic equipartition form of the acoustic channels neglects the phonon energy, whereas in the Boltzmann solution an electron within a few meV of the band edge cannot emit an acoustic phonon of 1 to 4 meV. These states carry 18% of the energy-weighted mobility sum at 150 K and 6% at 300 K, so the excess lowers the mobility of the transport model by about 3% at 150 K and 1% at 300 K.

Table S9 and Figure S6(b) compare the mobilities. Under the same conditions, the transport model agrees with the iterative Boltzmann solution within 1% at 150 K and 4% at 200 K, and exceeds it by 7, 8, 12 and 13% at 250, 300, 350 and 400 K. The difference grows with temperature although the total rates agree at every energy, so it does not come from the couplings but from the two approximations of the transport model that are not made in the Boltzmann solution. The deflection angle after each event is sampled from the isotropic average of ∣g(q)∣2|g(\mathbf{q})|^2 over the direction of q\mathbf{q}, and the K valley is an isotropic Kane band, whereas the trigonal warping of the real band grows above 150 meV, where an increasing fraction of the electrons resides at 350 and 400 K. Both approximations affect the momentum relaxation rather than the total rate, and the difference between the relaxation-time and the iterative Boltzmann solutions (6 % at 150 K, 1 % at 400 K) shows the size of such angular effects in this material. We therefore take 13 % at 400 K, and 8 % at room temperature, as the bound of these approximations on the phonon-limited mobility.

TT (K)Boltzmann, SERTABoltzmann, IBTEEMC, same conditionsEMC, main text
150566534530 (2)558 (2)
200359347362 (2)389 (3)
250251247265 (2)289 (2)
300188187203 (2)212 (1)
350147148167 (2)164 (2)
400119121137 (1)127 (3)
γ\gamma1.591.521.381.51

Table S9. Phonon-limited mobility of the suspended monolayer (cm2^2/Vs) from the Boltzmann solution of the Wannier interpolation (SERTA and IBTE, 200 × 200 meshes, 1012^{12} cm−2^{-2}, Q valleys at 265 meV, no free-carrier screening) and from the transport model under the same conditions (eight seeds, standard errors in parentheses), and the transport model of the main text (Q valleys at the measured 118 meV with the spin-orbit splitting, polar coupling screened at 1012^{12} cm−2^{-2}). The last row gives the exponent of a power-law fit over 150 to 400 K.

Our transport model differs from this run in two physical inputs. The Q valleys at the measured 118 meV, with the spin-orbit splitting of Note S1, add intervalley scattering and lower the mobility by 6 % at 300 K and 21 % at 400 K (exponent 1.38 to 1.60). Static free-carrier screening of the polar coupling at 1012^{12} cm−2^{-2} raises it by 5 % at 150 K, 11 % at 300 K and 17 % at 400 K (exponent 1.60 to 1.51). We apply this screening because the plasma energy of the sheet at the relevant wavevectors (84 to 145 meV for qq between 0.05 and 0.15 Å−1^{-1} at 1012^{12} cm−2^{-2}) exceeds the LO phonon energy of 46 meV, so that the free carriers follow the polar phonon field. The static form overestimates the screening by at most 30 % where the two energies are closest, which bounds its effect on the mobility used to obtain the results presented in the main text at about 4 % at 300 K.

Data

The inputs and outputs of both calculations (the explicit wavevector lists, the DFPT and matrix-element outputs of the 34 wavevectors, the Wannier interpolation output and the extraction scripts), the parameter header of the transport model (examples/singleLayerMoS2/parameterTwoValley.hpp in the ViennaEMC repository, with the reference parametrization of Ref. [17] in parameterKaasbjerg.hpp) and the coupling comparison of Figure S4 are part of the dataset accompanying the main text.

Adsorbate calculations

The sensor model of Sec. 4.2 of the main text takes the charge transferred per adsorbed molecule as the amplitude of the transduction and the binding energy through the adsorption constant. We computed both quantities for NO2_2, NH3_3 and O2_2 on the monolayer, with the same code, functional, pseudopotentials and lattice constant as the electron-phonon calculations of Note S2. Each molecule was placed 2.9 to 3.0 Å above the sulphur plane of a 4×44 \times4 supercell (48 atoms, one molecule per 7.2×1013 cm−27.2 \times10^{13}\ \mathrm{cm}^{-2}, molecule images 12.7 Å apart), on three sites, above a Mo atom, above an S atom and above the hollow site. All atoms were then relaxed to forces below 2×10−32 \times10^{-3} Ry/bohr, with the Grimme-D3 dispersion correction, the two-dimensional Coulomb cutoff, a 3×3×13 \times3 \times1 mesh, a 60 Ry cutoff and spin polarisation with the doublet (NO2_2) and triplet (O2_2) magnetisations. NO2_2 was started with its oxygen atoms towards the surface. Two further starting orientations, nitrogen towards the surface and the molecule tilted with one oxygen down, test whether the flat orientation is the minimum. The adsorption energy is

Eads=E(layer+molecule)−E(layer)−E(molecule)E_{\mathrm{ads}} = E(\mathrm{layer}+\mathrm{molecule}) - E(\mathrm{layer}) - E(\mathrm{molecule})

with all three terms in the same cell and settings, so that it is negative for a bound molecule. The Langmuir kinetics of Sec. 4.2 of the main text use the binding energy Eb=−EadsE_{\mathrm{b}} = -E_{\mathrm{ads}}, which is positive for a bound molecule. Charge transfer is taken as the Bader charge [69] of the molecule in the relaxed complex minus its valence charge. The Löwdin charges of the projected wavefunctions were also computed and are listed for reference, but they assign 0.1 to 0.2 ee more to every molecule than the Bader partition because of the spilling of the projection and are not used. A 5×55 \times5 cell (75 atoms, 2×2×12 \times2 \times1 mesh) repeats the NH3_3 case above Mo and the NO2_2 case above S.

All three molecules physisorb (Table S10). They sit 2.9 to 3.3 Å above the sulphur plane, their adsorption energies differ between the sites by less than 5 meV for NO2_2 and by at most 70 meV for NH3_3 and O2_2, and the transferred charge is small, 0.04 to 0.05 ee to NO2_2, 0.01 to 0.03 ee from NH3_3 and about 0.01 ee to O2_2. For NO2_2, the flat orientation with the oxygen atoms towards the surface is the minimum, since the tilted start relaxes back to it, and the orientation with nitrogen towards the surface is a local minimum 25 to 45 meV higher with a smaller transfer of 0.02 to 0.03 ee. The 5×55 \times5 cell changes the NH3_3 and NO2_2 values by less than 2 meV and 0.001 ee, so the 4×44 \times4 results are converged with respect to the lateral separation of the molecules. The NO2_2 transfer is half the published PBE value obtained without dispersion correction and with a different site, and the NH3_3 transfer lies at the low end of the published range. Section S3.1 compares these values with the amplitudes of the sensor model. The inputs and outputs of these calculations are part of the dataset.

MoleculesiteEadsE_{\mathrm{ads}} (eV)hh (Å)Δq\Delta q Bader (e)Δq\Delta q Löwdin (e)
NO2_2Mo−0.1733.03+0.039−0.209
NO2_2S−0.1752.99+0.048−0.200
NO2_2S, 5×55 \times5−0.1772.98+0.048−0.200
NO2_2S, started tilted−0.1742.94+0.046−0.202
NO2_2S, N towards surface−0.1303.16+0.031−0.214
NO2_2hollow−0.1763.01+0.042−0.207
NO2_2hollow, N towards surface−0.1502.98+0.022−0.224
NH3_3Mo−0.1632.91−0.026−0.173
NH3_3Mo, 5×55 \times5−0.1682.91−0.027−0.174
NH3_3S−0.101−0.012−0.169
NH3_3hollow−0.1682.91−0.027−0.175
O2_2Mo−0.0953.21+0.014−0.103
O2_2S−0.079+0.006−0.107
O2_2hollow−0.0933.30+0.011−0.105

Table S10. Adsorption energies and charge transfers of NO2_2, NH3_3 and O2_2 on the pristine monolayer (PBE-D3, 4×44 \times4 cell unless stated, and Δq>0\Delta q > 0 means electrons transferred to the molecule). hh is the height of the lowest atom of the molecule above the sulphur plane after relaxation. Published PBE values without dispersion correction are NO2_2 0.100 ee and NH3_3 0.069 ee on the monolayer [54], 0.097 and 0.009 ee on the bilayer [2].

Comparison with the sensor-model amplitudes

Table S11 compares the fitted amplitudes of the sensor model with the computed transfers and adsorption energies. Expressed per molecule with the site density of one molecule per 4×44 \times4 cell (7.2×1013 cm−27.2 \times10^{13}\ \mathrm{cm}^{-2}), the fitted amplitudes correspond to 0.030 ee for NO2_2 and 0.28 ee for NH3_3. For NO2_2, the fitted value is two thirds of the computed transfer on the pristine plane, and the remaining difference is the carrier limit of a device. The channel holds 3×1012 cm−23 \times10^{12}\ \mathrm{cm}^{-2} electrons, and once the acceptor level is filled to the point where the Fermi level drops below it, further adsorption transfers no charge. The fit obeys this bound (ΔNθ≤n0\Delta N_\theta\le n_0), and the gate-voltage dependence of the transfer computed by Late et al. [2] shows the same limit. For NH3_3, the fitted value exceeds the pristine-surface transfer ten-fold, which suggests a substantial contribution from defect-related adsorption sites, where DFT gives much larger transfers, consistent with the OFF-state operating point of the NH3_3 measurements (§ 2.1.4.2.5.4 of the main text).

The binding energies lead to the same conclusion. Under the kinetic prefactor assumed there, Eq. (12) of the main text makes the fitted adsorption constants correspond to effective binding energies EbE_b of 0.47 and 0.60 eV for NO2_2 on the devices of Kumar et al. [56] and Liu et al. [63], respectively, and 0.47 eV for NH3_3 on the device of Liu et al. [63]. These lie between the physisorption binding energies of the pristine basal plane (0.17 to 0.18 eV here, 0.28 and 0.25 eV for NO2_2 and NH3_3 in Ref. [54]) and the binding energies reported for defect-modified surfaces, 0.81 eV for NH3_3 on nitrogen-incorporated MoS2_2 carrying a sulphur vacancy [53] and 3.18 eV for NO2_2 at a sulphur vacancy [52]. The enhancement at a bare sulphur vacancy is not universal in the literature. Regis et al. [70] keep NO2_2 and NH3_3 physisorbed at 0.05 to 0.30 eV whether or not a vacancy is nearby, and find only NO chemisorbed in the vacancy, so the sites responsible for the stronger binding inferred here are not identified by the fit. KK should therefore be interpreted as an effective adsorption constant representing the unresolved distribution of adsorption sites, namely the basal plane, vacancies and grain boundaries, whose weights differ from sample to sample. Its device dependence therefore reflects the defect density, the growth method and the ambient of each device rather than a deficiency of the transduction model, and the same Arrhenius factor governs the temperature dependence of the recovery times reported for MoS2_2 sensors.

Supplementary Note S4 Extrinsic scattering forms, degenerate carrier statistics and reference comparison

Extrinsic mechanisms

The extrinsic mechanisms of the transport model (Table 1 of the main text) follow Ma and Jena [18] and Ridley [34]. Charged impurities at the interface scatter through a Thomas-Fermi- and Rytova-Keldysh-screened 2D Coulomb potential,

MechanismTypeOrigin
Acoustic deformation potential, Ξ(q)\Xi(q)intrinsic, elasticthis work
Zero-/first-order intervalley K-K′'intrinsic, inelasticthis work
Intervalley K-Q, Q-Q′' (21 channels)intrinsic, inelasticthis work
Polar-optical (2D Fröhlich, LO), ∣g(q)∣\lvert g(q)\rvertintrinsic, inelasticthis work; form of [35]
Homopolar A1′A'_1, ∣g(q)∣\lvert g(q)\rvertintrinsic, inelasticthis work
Remote surface-optical phononextrinsic (substrate)[18]
Screened charged impurity (2D)extrinsic[18, 34]
Interface roughness (Prange-Nee)extrinsic[18]
Adsorbate Coulomb / charge transferextrinsic (ambient)this work

Table 1. Scattering mechanisms in the monolayer MoS2_2 EMC transport model.

V(q)=Ze22ε0εavge−qdqs+q+r0q2,q=2ksin⁡(θ/2),(13)V(q) = \frac{Z e^2}{2 \varepsilon_0 \varepsilon_{\mathrm{avg}}} \frac{e^{-q d}}{q_s + q + r_0 q^2}, \qquad q = 2 k \sin(\theta/2), \tag*{(13)}

where qsq_s is the free-carrier screening wavevector, r0r_0 the Rytova-Keldysh length and dd the impurity setback. The polar substrate contributes remote surface-optical (SO) phonon scattering, whose coupling scales with the dielectric discontinuity 12[(εsub+εenv)−1−(εsub+εenv)−1]\frac{1}{2}[(\varepsilon_{\mathrm{sub}}+\varepsilon_{\mathrm{env}})^{-1}-(\varepsilon_{\mathrm{sub}}+\varepsilon_{\mathrm{env}})^{-1}] split between the substrate SO modes [18]. Interface-roughness scattering is included in the Prange-Nee form. The adsorbate Coulomb centres of the sensor model are described by the same potential with the areal density Nc(θ)N_c(\theta) of Eq. (11) of the main text. The Rytova-Keldysh length of the transport model is 40 Å, within 10% of the value 2πα2D=43.52\pi\alpha_{2D}=43.5 Å that follows from the computed two-dimensional polarisability (§ 2.1.4.2.5.1 of the main text).

Degenerate carrier statistics in the Poisson equation

Because accumulation-mode two-dimensional channels reach degeneracy, the equilibrium density in Poisson’s equation (Eq. (7) of the main text) follows Fermi-Dirac rather than Boltzmann statistics.

The local density and its derivative are

n(φ)=1βln⁡(1+βeφ),∂n∂φ=eφ1+βeφ,φ=qϕ−Ec+EFkBT,(14)n(\varphi)=\frac{1}{\beta}\ln(1+\beta e^{\varphi}),\qquad\frac{\partial n}{\partial\varphi}=\frac{e^{\varphi}}{1+\beta e^{\varphi}},\qquad\varphi=\frac{q\phi-E_{\mathrm{c}}+E_{\mathrm{F}}}{k_{\mathrm{B}}T}, \tag*{(14)}

with the degeneracy parameter β\beta set by the two-dimensional band-edge density, and β→0\beta\to0 recovers the Boltzmann limit exactly, so non-degenerate materials are unaffected. The derivative is used in the local linearisation of the nonlinear Poisson update, and in the direct device simulation of § 2.1.4.2.1 of the main text these statistics are applied on the sheet nodes only.

Comparison with the reference parametrization

The LDA parametrization of Kaasbjerg et al. [17] gives ΞLA=2.8\Xi_{\mathrm{LA}}=2.8 and ΞTA=1.6\Xi_{\mathrm{TA}}=1.6 eV, first-order couplings of 5.9, 3.9 and 1.9 eV, homopolar and K-K′' couplings of 4.1×1084.1 \times10^{8} and 2.6×1082.6 \times10^{8} eV/cm, and a Fröhlich constant of 97 meV. Against these, the computed LA coupling is twice as large and the TA coupling smaller, the first-order TA and TO channels are two to three times larger, the zero-order optical couplings are 0.65 to 0.8 times as large, and the polar coupling is larger and wavevector dependent. The first-order K-K′' couplings of the transport model are 12.0 eV (TA, 22.5 meV), 2.0 eV (LA, 28.5 meV), 4.8 eV (TO, 45.7 meV) and 3.5 eV (A1′A'_1-like, 47.4 meV), and the intravalley TO(Γ\Gamma) coupling 3.6 eV (Table S6). Table 5 of the main text shows that the LA deformation potential alone accounts for the difference between the mobility obtained with the reference parametrization and with the computed couplings.

Comparison with few-layer devices

The framework of the main text is a monolayer model, and its quantitative claims are monolayer claims. Few-layer and multilayer devices are nevertheless the majority of the published sensing data, so Figure S7 compares the monolayer transduction with a 2-layer NH3_3 and NO2_2 device [2] and a multilayer melamine-doped FET [55], as a qualitative comparison rather than validation. The direction of the response is expected to be the same in a few-layer film, because adsorbates act only on the top surface, and charge transfer and adsorbate Coulomb scattering act on the layers within the screening length of MoS2_2, which is a few layers [71, 72]. For a bilayer, the entire film lies within that length, so the monolayer σ(θ)\sigma(\theta) gives a qualitative reference, with KK and ΔN\Delta N absorbing the layer dependence of adsorption and screening. The first-principles NO2_2 transfer differs by 3% between the monolayer and the bilayer (0.100 and 0.097 ee [54, 2]), whereas the two NH3_3 values differ by an order of magnitude (Table S11), which is why the NH3_3 comparison is only qualitative. In thick films, the back gate acts on the bottom layers and the adsorbate on the top layers, the interlayer resistance network decouples the two, and the gate-tunable selectivity reported for few-layer devices [2] is an interlayer effect outside the model. An extension to few-layer films requires a layer-resolved model with interlayer coupling and screening and lies outside the scope of this work.

Figure S7. Qualitative comparison with few-layer devices, not validation. (a) NH3_3 and (b) NO2_2 response of a 2-layer device [2] against the monolayer transduction with a fitted adsorption constant. (c) Transfer characteristics of a multilayer FET before and after melamine doping [55], reproduced by the reduced device model as a donor shift of ΔVth≈−21\Delta V_{\mathrm{th}} \approx-21 V.

References

  1. [1]Radisavljevic B, Radenovic A, Brivio J, Giacometti V and Kis A 2011 Nature Nanotechnology 6 147–150 https://doi.org/10.1038/nnano.2010.279
  2. [2]Late D J, Huang Y K, Liu B, Acharya J, Shirodkar S N, Luo J, Yan A, Charles D, Waghmare U V, Dravid V P and Rao C N R 2013 ACS Nano 7 4879–4891 https://doi.org/10.1021/nn400026u
  3. [3]Kurosawa T 1966 Monte Carlo calculation of hot electron problems Proceedings of the International Conference on the Physics of Semiconductors (Journal of the Physical Society of Japan, Supplement 21) (Kyoto, Japan) pp 424–426 URL https://www.jps.or.jp/books/jpsjs/21/jpsj.21s.424.pdf
  4. [4]Fawcett W, Boardman A D and Swain S 1970 Journal of Physics and Chemistry of Solids 31 1963–1990 https://doi.org/10.1016/0022-3697(70)90001-6
  5. [5]Kurosawa T and Maeda H 1971 Journal of the Physical Society of Japan 31 668–678 https://doi.org/10.1143/jpsj.31.668
  6. [6]Rees H D 1969 Journal of Physics and Chemistry of Solids 30 643–655 https://doi.org/10.1016/0022-3697(69)90018-3
  7. [7]Jacoboni C and Reggiani L 1983 Reviews of Modern Physics 55 645–705 https://doi.org/10.1103/RevModPhys.55.645
  8. [8]Price P J 1979 Monte Carlo calculation of electron transport in solids Lasers, Junctions, Transport ed Willardson R K and Beer A C (New York: Elsevier) chap 4, pp 249–308 https://doi.org/10.1016/S0080-8784(08)60267-7
  9. [9]Jacoboni C and Lugli P 1989 The Monte Carlo Method for Semiconductor Device Simulation (Wien: Springer-Verlag) https://doi.org/10.1007/978-3-7091-6963-6
  10. [10]Selberherr S 1984 Analysis and Simulation of Semiconductor Devices (Wien: Springer-Verlag) https://doi.org/10.1007/978-3-7091-8752-4
  11. [11]Hess K (ed) 1991 Monte Carlo Device Simulation: Full Band and Beyond (Boston: Kluwer Academic) https://doi.org/10.1007/978-1-4615-4026-7
  12. [12]Moglestue C 1993 Monte Carlo Simulation of Semiconductor Devices (London: Chapman & Hall) https://doi.org/10.1007/978-94-015-8133-2
  13. [13]Tomizawa K 1993 Numerical Simulation of Submicron Semiconductor Devices (Boston: Artech House) URL https://api.semanticscholar.org/CorpusID:116364667
  14. [14]Jungemann C and Meinerzhagen B 2003 Hierarchical Device Simulation: The Monte-Carlo Perspective (Wien: Springer-Verlag) https://doi.org/10.1007/978-3-7091-6086-2
  15. [15]Fischetti M V and Laux S E 1988 Physical Review B 38 9721–9745 https://doi.org/10.1103/PhysRevB.38.9721
  16. [16]Fischetti M V 1991 IEEE Transactions on Electron Devices 38 634–649 https://doi.org/10.1109/16.75176
  17. [17]Kaasbjerg K, Thygesen K S and Jacobsen K W 2012 Physical Review B 85 115317 https://doi.org/10.1103/PhysRevB.85.115317
  18. [18]Ma N and Jena D 2014 Physical Review X 4 011043 https://doi.org/10.1103/PhysRevX.4.011043
  19. [19]Baugher B W H, Churchill H O H, Yang Y and Jarillo-Herrero P 2013 Nano Letters 13 4212–4216 https://doi.org/10.1021/nl401916s
  20. [20]Radisavljevic B and Kis A 2013 Nature Materials 12 815–820 https://doi.org/10.1038/nmat3687
  21. [21]Ahn J H, Parkin W M, Naylor C H, Johnson A T C and Drndić M 2017 Scientific Reports 7 4075 https://doi.org/10.1038/s41598-017-04350-z
  22. [22]Li J, Peng S, Jin Z, Tian H, Wang T and Peng X 2023 Crystals 13 1501 https://doi.org/10.3390/cryst13101501
  23. [23]Li X, Mullen J T, Jin Z, Borysenko K M, Buongiorno Nardelli M and Kim K W 2013 Physical Review B 87 115418 https://doi.org/10.1103/PhysRevB.87.115418
  24. [24]Restrepo O D, Krymowski K E, Goldberger J and Windl W 2014 New Journal of Physics 16 105009 https://doi.org/10.1088/1367-2630/16/10/105009
  25. [25]Gunst T, Markussen T, Palsgaard M L N, Stokbro K and Brandbyge M 2017 Physical Review B 96 161404 https://doi.org/10.1103/PhysRevB.96.161404
  26. [26]Poncé S, Li W, Reichardt S and Giustino F 2020 Reports on Progress in Physics 83 036501 https://doi.org/10.1088/1361-6633/ab6a43
  27. [27]Fischetti M V and Vandenberghe W G 2016 Physical Review B 93 155413 https://doi.org/10.1103/PhysRevB.93.155413
  28. [28]Schedin F, Geim A K, Morozov S V, Hill E W, Blake P, Katsnelson M I and Novoselov K S 2007 Nature Materials 6 652–655 https://doi.org/10.1038/nmat1967
  29. [29]Perkins F K, Friedman A L, Cobas E, Campbell P M, Jernigan G G and Jonker B T 2013 Nano Letters 13 668–673 https://doi.org/10.1021/nl3043079
  30. [30]Donarelli M and Ottaviano L 2018 Sensors 18 3638 https://doi.org/10.3390/s18113638
  31. [31]Cho B, Hahm M G, Choi M, Yoon J, Kim A R, Lee Y J, Park S G, Kwon J D, Kim C S, Song M, Jeong Y, Nam K S, Lee S, Yoo T J, Kang C G, Lee B H, Ko H C, Ajayan P M and Kim D H 2015 Scientific Reports 5 8052 https://doi.org/10.1038/srep08052
  32. [32]Sarkar D, Liu W, Xie X, Anselmo A C, Mitragotri S and Banerjee K 2014 ACS Nano 8 3992–4003 https://doi.org/10.1021/nn5009148
  33. [33]Lundstrom M 2000 Fundamentals of Carrier Transport 2nd ed (Cambridge, UK: Cambridge University Press) https://doi.org/10.1017/CBO9780511618611
  34. [34]Ridley B K 1982 Journal of Physics C: Solid State Physics 15 5899–5917 https://doi.org/10.1088/0022-3719/15/28/021
  35. [35]Sohier T, Calandra M and Mauri F 2016 Physical Review B 94 085415 https://doi.org/10.1103/PhysRevB.94.085415
  36. [36]Fu L, Wan Y, Tang N, Ding Y m, Gao J, Yu J, Guan H, Zhang K, Wang W, Zhang C, Shi J j, Wu X, Shi S F, Ge W, Dai L and Shen B 2017 Science Advances 3 e1700162 https://doi.org/10.1126/sciadv.1700162
  37. [37]Sohier T, Campi D, Marzari N and Gibertini M 2018 Physical Review Materials 2 114010 https://doi.org/10.1103/physrevmaterials.2.114010
  38. [38]Hockney R W and Eastwood J W 1988 Computer Simulation Using Particles 1st ed (Boca Raton: CRC Press) https://doi.org/10.1201/9780367806934
  39. [39]Filipovic L et al. 2026 ViennaEMC: Ensemble Monte Carlo framework for semiconductors and metals institute for Microelectronics, TU Wien. URL https://github.com/ViennaTools/ViennaEMC
  40. [40]Giannozzi P, Andreussi O, Brumme T, Bunau O, Buongiorno Nardelli M, Calandra M, Car R, Cavazzoni C, Ceresoli D, Cococcioni M, Colonna N, Carnimeo I, Dal Corso A, de Gironcoli S, Delugas P, DiStasio Jr R A, Ferretti A, Floris A, Fratesi G, Fugallo G, Gebauer R, Gerstmann U, Giustino F, Gorni T, Jia J, Kawamura M, Ko H Y, Kokaji A, Kückübenli E, Lazzeri M, Marsili M, Marzari N, Mauri F, Nguyen N L, Nguyen H V, Otero-de-la Roza A, Paulatto L, Poncé S, Rocca D, Sabatini R, Santra B, Schlიფ M, Seitsonen A P, Smogunov A, Timrov I, Thomhauser T, Umari P, Vast N, Wu X and Baroni S 2017 Journal of Physics: Condensed Matter 29 465901 https://doi.org/10.1088/1361-648X/aa8f79
  41. [41]Perdew J P, Burke K and Ernzerhof M 1996 Physical Review Letters 77 3865–3868 https://doi.org/10.1103/PhysRevLett.77.3865
  42. [42]Hamann D R 2013 Physical Review B 88 085117 https://doi.org/10.1103/PhysRevB.88.085117
  43. [43]van Setten M J, Giantomassi M, Bousquet E, Verstraete M J, Hamann D R, Gonze X and Rignanese G M 2018 Computer Physics Communications 226 39–54 https://doi.org/10.1016/j.cpc.2018.01.012
  44. [44]Dal Corso A 2014 Computational Materials Science 95 337–350 https://doi.org/10.1016/j.commatsci.2014.07.043
  45. [45]Kormányos A, Burkard G, Gmitra M, Fabian J, Zólyomi V, Drummond N D and Fal’ko V 2015 2D Materials 2 022001 https://doi.org/10.1088/2053-1583/2/2/022001
  46. [46]Lihm J M and Park C H 2020 Physical Review B 101 121102 https://doi.org/10.1103/PhysRevB.101.121102
  47. [47]Giustino F 2017 Reviews of Modern Physics 89 015003 https://doi.org/10.1103/RevModPhys.89.015003
  48. [48]Lee H, Poncé S, Bushick K, Hajinazar S, Lafuente-Bartolome J, Leveillee J, Lian C, Lihm J M, Macheda F, Mori H, Paudyal H, Sio W H, Tiwari S, Zacharias M, Zhang X, Marzari N and Giustino F 2023 npj Computational Materials 9 156 https://doi.org/10.1038/s41524-023-01107-3
  49. [49]Sohier T, Calandra M and Mauri F 2017 Physical Review B 96 075448 https://doi.org/10.1103/PhysRevB.96.075448
  50. [50]Poncé S, Royo M, Stengel M, Marzari N and Gibertini M 2023 Physical Review B 107 155424 https://doi.org/10.1103/PhysRevB.107.155424
  51. [51]Shen P C, Su C, Lin Y, Chou A S, Cheng C C, Park J H, Chiu M H, Lu A Y, Tang H L, Tavakoli M M, Pitner G, Ji X, Cai Z, Mao N, Wang J, Tung V, Li J, Bokor J, Zettl A, Wu C I, Palacios T, Li L J and Kong J 2021 Nature 593 211–217 https://doi.org/10.1038/s41586-021-03472-9
  52. [52]Kushwaha A, Bharti N R, Sharma A, Kedia S K, Gupta G and Goel N 2024 ACS Sensors 9 5966–5975 https://doi.org/10.1021/acssensors.4c01812
  53. [53]Kashyap A, Chakraborty B, Hazarika T, Chouhan S, Kakati B and Kalita H 2025 Materials Advances 6 3828–3840 https://doi.org/10.1039/D5MA00232J
  54. [54]Yue Q, Shao Z, Chang S and Li J 2013 Nanoscale Research Letters 8 425 https://doi.org/10.1186/1556-276X-8-425
  55. [55]Al Mamun M S, Takaoka T, Wang Z and Komeda T 2025 ChemistrySelect 10 e202403982 https://doi.org/10.1002/slct.202403982
  56. [56]Kumar S, Sharma A, Sao A K, Singh J P, Chowdhuri A and Tomar M 2025 Chemical Physics Impact 10 100847 https://doi.org/10.1016/j.chphi.2025.100847
  57. [57]Illarionov Y Y, Banshchikov A G, Polyushkin D K, Wachter S, Knobloch T, Thesberg M, Mennel L, Paur M, Stöger-Pollach M, Steiger-Thirsfeld A, Vexler M I, Waltl M, Sokolov N S, Mueller T and Grasser T 2019 Nature Electronics 2 230–235 https://doi.org/10.1038/s41928-019-0256-8
  58. [58]Berkelbach T C, Hybertsen M S and Reichman D R 2013 Physical Review B 88 045318 https://doi.org/10.1103/physrevb.88.045318
  59. [59]Laturia A, Van de Put M L and Vandenberghe W G 2018 npj 2D Materials and Applications 2 6 https://doi.org/10.1038/s41699-018-0050-x
  60. [60]Kormányos A, Zólyomi V, Drummond N D, Rakyta P, Burkard G and Fal’ko V I 2013 Physical Review B 88 045416 https://doi.org/10.1103/PhysRevB.88.045416
  61. [61]Pilotto A, Khakbaz P, Palestri P and Esseni D 2022 Solid-State Electronics 192 108295 https://doi.org/10.1016/j.sse.2022.108295
  62. [62]Smithe K H H, English C D, Suryavanshi S V and Pop E 2018 Nano Letters 18 4516–4522 https://doi.org/10.1021/acs.nanolett.8b01692
  63. [63]Liu B, Chen L, Liu G, Abbas A N, Fathi M and Zhou C 2014 ACS Nano 8 5304–5314 https://doi.org/10.1021/nn5015215
  64. [64]Poncé S, Margine E R, Verdi C and Giustino F 2016 Computer Physics Communications 209 116–133 https://doi.org/10.1016/j.cpc.2016.07.028
  65. [65]Filipovic L 2026 Ensemble Monte Carlo and DFPT simulation data for charge transport and gas sensing in monolayer MoS₂ [Dataset] URL https://doi.org/10.48436/87zq1-28246 https://doi.org/10.48436/87zq1-28246
  66. [66]Eknapakul T, King P D C, Asakawa M, Buaphet P, He R H, Mo S K, Takagi H, Shen K M, Baumberger F, Sasagawa T, Jungthawan S and Meevasana W 2014 Nano Letters 14 1312–1316 https://doi.org/10.1021/nl4042824
  67. [67]Brumme T, Calandra M and Mauri F 2015 Physical Review B 91 155436 https://doi.org/10.1103/physrevb.91.155436
  68. [68]Marzari N, Mostofi A A, Yates J R, Souza I and Vanderbilt D 2012 Reviews of Modern Physics 84 1419–1475 https://doi.org/10.1103/RevModPhys.84.1419
  69. [69]Henkelman G, Arnaldsson A and Jónsson H 2006 Computational Materials Science 36 354–360 https://doi.org/10.1016/j.commatsci.2005.04.010
  70. [70]Regis N M, Da Silva J L F and Lima M P 2024 Materials Today Communications 38 107710 https://doi.org/10.1016/j.mtcomm.2023.107710
  71. [71]Das S and Appenzeller J 2013 physica status solidi (RRL) - Rapid Research Letters 7 268–273 https://doi.org/10.1002/pssr.201307015
  72. [72]Castellanos-Gomez A, Cappelluti E, Roldán R, Agraït N, Guinea F and Rubio-Bollinger G 2013 Advanced Materials 25 899–903 https://doi.org/10.1002/adma.201203731

Paper details

vlnc:W_xvjv7gvx
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents